Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
5.6 MB
Raw Blame

Changelog

All notable changes to GBrain will be documented in this file.

Historical entries use generic identifiers for the retired hosted provider and its removed commands. Those identifiers and any sample command containing them are non-executable archival labels, not supported configuration. Dates, measurements and contributor credits are retained; no result has been reassigned to another provider. Original identifiers and attribution are available in the pre-removal Git revision 6040075c6cb95be5881cc2e1b76ef7d71f4e5d29 (retained on 2026-09-23).

[0.60.737.0] - 2026-11-17

Conversation extraction keeps wishes and preferences again: the extractor no longer asks the model for a slot on every fact.

Fixed

  • Since 0.60.735.0, gbrain extract-conversation-facts asked the model for a slot (where a person lives, who employs them or their role) on every fact. That request made it save fewer facts, with wishes, preferences and ideas down about 15%. The main extraction call now sends the prompt and response shape it used before 0.60.735.0, so no new fact carries a slot, and a --force re-extraction adds none: skip the paid run the 0.60.735.0 notice suggested for them. A fact stored with a slot under 0.60.735.0 keeps it, and the facts block still works out its validity window, until the part of its page it came from is extracted again (an edit there, or --force, which re-extracts every page); the new rows carry no slot and read as they did before 0.60.735.0. Fact text is kept either way.

[0.60.736.0] - 2026-11-17

Contributors: the pre-merge check now accepts changes to the macOS helper's files under apple/helper/, which build and test only in their own macOS workflow.

Changed

  • .gstack/pregate.json lists apple/helper/** as having no local test run, the same way scripts/bench/** is listed. The helper's Swift tests run on every change to apple/** in the Apple helper workflow. Nothing changes for gbrain users. (#9627)

A private fact restated as world and held for the owner's accept stays hidden in more places from callers who cannot read the private fact: think leaves it out, the reply to the held write reads like an ordinary save, and a restatement written onto another page about the same entity is held too.

Fixed

  • think now leaves a held fence row out of the pages it reads and the evidence it delivers for callers who cannot read the private fact, as get_page and search already do. The owner's callers and the local CLI read the same as before.
  • The reply to a held write (remember, put_page and the other page writes) named the owner's proposal in contested.proposal_ref, which told a caller who guessed a claim that a more private copy of it exists. For a caller who cannot read that private fact, the reply, its outcome and write_request copies, get_write_request and list_write_requests now leave the reference out, so the reply reads like a saved write. The owner's callers, the local CLI and callers who can read the private fact still get it, and a contested reference for any other reason (a write that tried to replace a more trusted fact) is unchanged.
  • A page write whose ## Facts row restated a more trusted private fact about the page's own entity, kept off that page (a remembered fact with no page row, or a row on a holding page), was stored as world. It is now stored private with a proposal, like a restatement on the same page, and gbrain trust confirm tp<N> makes it world. Like the same-page hold, this applies to the owner's own local writes and sync too, when the private copy is more trusted than the writer.
  • A credential with a private_facts: granted_sources grant reads a held row again in get_page, search and think, as it did before the held-row page reads, when it can read every private copy the row restates. A row restating a copy the grant can't read (on a member or restricted page, or on no page) stays held from it. Known: such a credential's put_page still restores a held row it can see, so its edit or delete of that row doesn't stick and an edit can leave two rows.
  • A remote get_page or get_versions of a page stored in segments reported a size that counted held rows. It now reports the size of what the caller can read.
  • Still not held: a restatement about another entity, a reworded claim, captured prose and extracted facts. A writer that cannot read the private fact still can't read its own held row back (page reads and recall leave it out) until the owner accepts it, which a saved row would not do.

A private fact restated as world and held for the operator's accept no longer reaches remote callers through page, chunk, search and edit reads: get_page, get_versions, fetch, search, query, search_exact, get_chunks, the entity card and the edit_page diff leave the held row out for callers who cannot read the private fact.

Fixed

  • A remote put_page that restated a private fence row as world (#9196), or a remote remember that restated a private fact (GST-2875), stored the new fact private until the operator accepts the widening, but the page body kept the restatement in a world cell. Every remote caller who cannot read the private fact (another member, an unbound token) read it through get_page (whole, include_content, lines, grep, section, outline, as_of, known_as_of and windows around a chunk), get_versions with include_body (known_as_of bodies and diffs too), fetch, keyword and hybrid search, query, search_exact (snippets, and the match itself, on pages stored in segments too), get_chunks (on any source) and the entity card. A recorded-time read at the slug a page had before a rename leaves it out too, and search_exact still finds a page's public hit when an earlier hit in the same text is a held row. Those reads now leave out a world fence cell whose fact is stored narrower than world, including a row a chunk boundary cut in half and chunks written before the upgrade, so no rechunk is needed. The owner's callers, the local CLI and a member reading their own member page read exactly what they read before. If the held-row lookup fails, a page read fails rather than return the page, and search drops every fence row.
  • Such a caller's put_page keeps the held row it could not read. Writing back a page read without it (a get_page, edit, put_page round trip, a whole new body, or a body with the lines around it deleted) restores the row exactly as stored, so the operator's accept still widens it. A row the writer sends at the held row's number cannot replace or reword it. A different claim moves to a new row number, and the same claim keeps every stored cell except a narrower visibility the writer chose. A member session under GBRAIN_ENFORCED_MEMBER reads such a page the same way, so its put_page keeps the held row and its edit_page diff leaves it out. The owner's own writes and the local CLI's are unchanged.
  • The edit_page diff showed a remote caller the held row as unchanged context when an edit landed near it, on inline and segmented pages alike. The diff now leaves it out for callers who cannot read the private fact, as get_page does.
  • An edit elsewhere in a held page (a hand edit synced from the repository, or a local put_page) could re-derive the held row as world without the accept when the private copy was less trusted than user_confirmed. A held row now stays private whatever the writer's tier until the operator accepts it. Accept still makes it world, and reject and undo still work.
  • A remote writer could skip the GST-2875 hold by first saving the same claim as a world fact about another entity. The hold looked for an existing world copy anywhere in the source, found that one, and stored the restatement on the confirmed private fact's entity as world, so remote recall returned it. The world copy now has to be on the same entity.
  • Not covered: the page file in the brain repository still carries the world cell, so anyone who reads the repository directly, or a Markdown export, reads it. Restatements that were never held (capture, extract_facts, a restatement written onto another page or entity, a reworded claim, and a remember whose private copy is no more trusted than the writer) stay world as before.

Link context no longer shows remote callers private or held fact rows: get_backlinks, get_links, traverse_graph (walks and hops), the relationship-chain evidence on search, query and recall results, the entity card, context_pack and delta's belief changes leave that text out for callers who may read only world facts, and context_pack's always-loaded core block leaves held rows out for them too.

Fixed

  • A link's stored context is a short excerpt of the page around the reference. When the reference sat in or next to a page's ## Facts table, the excerpt carried rows of that table, including rows saved private and world rows held until the operator accepts a widening. Remote callers who may read only world facts (another member, an unbound token) read that text through get_backlinks, get_links, traverse_graph walks and hops, the chain evidence on search, query and recall results, the entity card's edges and context_pack's cards. An attendance edge, stored on the attendee's page but cut from the meeting page, carried the meeting's rows the same way. For those callers, a context cut from a page is now returned only when it reads, word for word, inside that page as it is now and as the caller may see it; otherwise it comes back empty. That also covers a page whose table was removed by an edit that left its links unchanged. The page bodies for one read are loaded together, for at most 200 pages; contexts cut from pages beyond that come back empty for those callers. A context written by add_link or derived from frontmatter is unchanged unless it holds a facts or takes table marker.
  • unify-types' preserve_notes stored the first 240 characters of a converted page as the new link's context, facts and takes tables included. It now leaves those tables out, so links written from now on store a shorter excerpt for every reader, the owner included. Contexts it stored before are emptied for callers who may read only world facts when they hold a table marker.
  • delta with beliefs: true returned each changed link's stored context in its before and after values. Those callers now get the link change without the context.
  • With memory.core.enabled on, context_pack's core memory block showed those callers a core page's held rows. It now leaves them out, as get_page does.
  • The owner's remote callers and the local CLI read stored links exactly as before. Stored links are not rewritten and no migration runs.

A connection gbrain throws away now frees its slot in the connection pool even when the database host has stopped answering, and the background job worker keeps running while PostgreSQL shuts down or after a TCP server has stopped.

Fixed

  • When gbrain discards a Postgres connection (a cancelled bounded read, an aborted transaction, a statement that would not cancel) and the server or pooler at the other end has gone silent, the pool slot comes back within about 2 seconds. Before, the slot stayed taken until the operating system gave up on the socket, so a process with a small pool could stop being able to connect. Follow-up to #8155.
  • gbrain jobs work no longer exits when a claim meets a PostgreSQL server that is shutting down and refusing new connections (SQLSTATE 57P03), or one that has stopped listening on TCP (connect ECONNREFUSED). The worker reconnects and claims again on its next poll, as it already did for a vanished Unix socket. Follow-up to #9540, reported and first fixed by @javieraldape in #9624.

A pull request landed as one squashed commit now reuses its passed full test run on master again, even when master was merged into it with no file change just before landing.

Fixed

  • scripts/ci-reuse.ts: a one-parent push to master looked for a passed full run only on the head GitHub names for the landed pull request. Landing first merges master into the branch, and when that merge changes no file the new head has the tested tree but only a pull-request run, so every landing ran Test and E2E in full again (about 18 minutes instead of about 4). The push's tree now decides, as in the remote gate. The newest full run on any head with that tree wins, and those heads are the pull request head, its first parents with the same tree and the heads of recent dispatch and push runs. A newer failed or in-progress full run on the tree still runs in full, and what counts as a complete full run is unchanged. A hand revert or direct push whose tree matches a passed full run now reuses it too. docs/ci-event-parity.md, "Master-push reuse", says so.

When a scoped graph walk or typed hop chain ends early, the next call it suggests now keeps the caller's source scope.

Fixed

  • traverse_graph returns a graph_walk_truncated notice when a walk stops at its cap, with a ready-to-run retry one level shallower. That retry dropped the caller's source_id and all_sources, so an agent that followed it walked its default source instead of the one it asked for. The retry now carries both, and still leaves them out when the call didn't set them (#8498).
  • The same holds for a hops chain that finds no edges, stops at an empty hop or hits a chain cap: its suggested traverse_graph call now keeps source_id and all_sources (#8498).

Tests

  • test/degraded-stamp-paths.serial.test.ts checks the retry for a walk scoped to one source, an all-sources walk and an unscoped walk.
  • test/traverse-graph-hops.test.ts checks the no_edges and empty_hop suggestions for the same three cases.

Long-running gbrain processes recover from a lost database connection: a bounded read cancelled while the database stops answering frees its connection at once, and the background job worker keeps running through a PostgreSQL shutdown instead of exiting.

Fixed

  • A bounded read (a statement with a time limit) whose caller cancels it after it has a connection now sends the database a cancel request, discards that connection and returns at once. Before, the call waited until the database answered, and the connection stayed reserved the whole time, so a process that hit a stalled socket or pooler again and again could run out of connections. (#8155)
  • gbrain jobs work no longer exits when PostgreSQL shuts down and removes its local Unix socket (connect ENOENT <dir>/.s.PGSQL.<port>). The worker reconnects and claims again on its next poll, the same as for other lost connections. A first connect to a socket path that doesn't exist still fails at once. (#9540)

The Rust engine setup script is safer on shared and already-used machines: each run logs to its own private folder, it ignores PGHOST, PGHOSTADDR and PGSERVICE, and it no longer resets an existing gbrain database role.

Fixed

  • scripts/dev/rust-engine-bootstrap.sh now writes each run's step logs to a new private folder, and refuses a GBRAIN_DEV_LOG_DIR that another user can write to.
  • The script unsets PGHOST, PGHOSTADDR and PGSERVICE for itself, so its database steps always reach the local Postgres 16 cluster.
  • The script creates the gbrain role only when it is missing. An existing role keeps its password and flags; if it cannot be used, the step fails with a message that names the setting to change.

Contributors working on the Rust engine can now set up an Ubuntu machine with one command: scripts/dev/rust-engine-bootstrap.sh installs the pinned Bun, Postgres 16 with pgvector, Rust and the napi-rs CLI, and prints a pass/fail line per step.

Added

  • scripts/dev/rust-engine-bootstrap.sh: a non-interactive, rerunnable dev setup for Ubuntu. It installs the Bun version from package.json, starts Postgres 16 with pgvector and a test database, installs the Rust stable toolchain and the napi-rs CLI, runs bun install, and writes ~/.gbrain-rust-dev.env. It does not change gbrain itself.
  • docs/research/rust-engine/dev-environment.md: how to run the script, what each step does, how to run the engine parity tests, measured timings and known problems.

Install guides now say Bun's installer needs unzip, show how to get it, and give a tested route for machines where packages can't be installed.

Documentation

  • README.md, INSTALL_FOR_AGENTS.md, docs/INSTALL.md, AGENTS.md and the company-brain tutorial say Bun's installer stops with error: unzip is required to install bun when unzip is missing, and give the check (command -v unzip) and the apt, dnf and apk install commands.
  • INSTALL_FOR_AGENTS.md#no-unzip is a tested route for Linux machines with no unzip and no package manager: it puts Bun's own build in ~/.bun/bin using only curl and tar, and checks that gbrain runs from a fresh login shell.
  • docs/guides/troubleshooting.md has a row for the unzip is required error.

gbrain git commit-push --untrack now takes every file under a declared, ignored database-only folder, .json and other non-Markdown files included.

Fixed

  • In a managed checkout, gbrain git commit-push --untrack -- <folder> refused untrack_not_db_only for a folder declared under storage.db_only and ignored by .gitignore when it held .json (or other non-Markdown) files, because each file was matched by its page slug. A file is now matched by its path under the declared folder. A Markdown page still needs its slug under the folder; a code-extension file is checked against the brain and refuses when it is a live page, when the source's strategy imports code (sync imports it as a page), when the source's last finished sync has not seen it at its current version (run gbrain sync --source <id>, then rerun), or when the brain cannot be read. Symlinks out of the folder, folders reached through a symlink and sibling folders that only share the name refuse. Each refusal's fix now names what is missing for that file, and --dry-run --untrack lists exactly the files the real run removes.
  • The managed sync note about pages kept from the working tree no longer tells a page under a declared storage.db_only folder to be committed.
  • For a source rooted in a folder of its repository (brain/, with brain/gbrain.yml), --untrack read the repository's top-level gbrain.yml: every file refused untrack_not_db_only, and a folder only a stray top-level gbrain.yml declared was admitted, so the next sync deleted its pages. In a managed checkout every untrack now reads the source's folder from the brain, read-only for at most 10 seconds, and refuses every file when the brain cannot be read. On PGLite, stop a running gbrain serve or sync before an untrack. The refusal names the gbrain.yml it read.

A large capture's receipt now reports the same date as the page slug it created, and five tests that failed now and then pass every run.

Fixed

  • gbrain capture --file with a file over the page size limit took the slug date and the receipt's captured_at from two separate clock reads, so a capture that ran across UTC midnight could print a receipt dated one day for a page slug dated the next. Both now come from one timestamp. (#7980)
  • The forget test for twin inline and segmented pages froze no clock, so a run that crossed UTC midnight stamped the two pages with different dates and failed. Both halves now run under one frozen time. (#8674)
  • The sync dry-run test compared request rows read in no fixed order. It now orders them by id and compares whole rows, so it also catches a dry run that edits a request. (#8416)
  • The private-queue recovery test gave its orphan job a 1 ms lease and relied on it running out before reconcile looked. The test now expires the lease explicitly. (#8739)
  • The reranker failure test freed a random port and assumed nothing would take it again. It now points the dead endpoint at port 1, where nothing listens. (#9616)

The query facts block now shows a later retraction right after the fact it calls off.

Added

  • When the facts block shows a fact about a person or company, a later fact about the same entity that retracts it or calls it off ("retracted", "scratch that", "I was wrong", "cancelled", "called off", "fell through", "never happened", "just a rumour" and similar) follows it, with its own date header, inside the block's usual share of the token budget and row count. The later fact pairs only when it is dated after the fact, its cue is not negated ("not cancelled", "wasn't called off" and "isn't a rumour" never pair), and it carries the same slot (dimension and value) or, leaving out the entity's name and the cue words, every word it has is in the fact or it shares at least two words with it. It is read under the caller's read policy, so a retraction a remote caller or another member cannot read never appears, and a forgotten or withdrawn fact never appears. No model call; a question with no such later fact gets exactly the block it got before.

The brain owner's own remote agents now get the owner's private facts on the entity and context_pack cards, in search's facts about an entity, in the list of facts saved without an entity and in forget's similar-facts check.

Fixed

  • The entity card (recent facts and active_fact_count) and the context_pack cards built from it, the entity facts and saved facts that search and query attach, extraction_pending's unattributed list and forget's similar_active decided fact visibility from the raw remote flag, so a credential bound to the brain owner got world facts only. They now use the same owner rule as recall, think and find_trajectory, and the owner's reads use the owner-aware page rule, so a fact from a private page the owner reads stays visible.
  • Nothing widens for anyone else: another member, a labelled-only member, an unbound token, a call with no credential and the owner speaking for an audience of several read what they read before, and a private-facts grant works where it already did. Member access still applies to the owner, and withdrawn facts stay out of every read.

The brain owner's own remote agents now get the owner's private facts from trajectories, ontology reads and write attribution, as they already did from recall, query and think.

Fixed

  • find_trajectory, ontology_get, ontology_conflicts and volunteer_chronicle decided fact visibility from the raw remote flag, so a credential bound to the brain owner got world facts only and a trajectory built from private facts came back empty. They now use the same owner rule as recall and think. get_write_attribution finds a private fact for the owner's admin credential the same way.
  • The owner also gets facts taken from a private page it can read on the ontology reads and in write attribution, as find_trajectory and recall already allowed.
  • A newer observation from a diary page no longer hides the older value behind it on ontology_get and volunteer_chronicle, and no longer makes a false conflict on ontology_conflicts. Diary observations are left out before each value resolves, for every remote caller, and still never reach one.
  • Nothing widens for anyone else: another member, a labelled-only member, an unbound token, a call with no credential and the owner speaking for an audience of several read no private fact they did not read before. Page rules, member access and source grants still apply to the owner, and withdrawn facts stay out of every read.

Files in a folder whose name starts with two dots, such as ..drafts or ..., are no longer refused or reported as outside the brain by sync, git commit-push, doctor, frontmatter, skills and the other checks listed below.

Fixed

  • Many path checks asked "does this path start with two dots?" to catch a path that goes up and out of a folder. That question also matched a normal folder named ..drafts, ..x or ..., so a file inside the brain was refused, skipped or called "outside". All of these checks now look at the first folder name exactly, through one shared helper. Remote file uploads into such a folder are still refused; a separate fix covers that check.
  • gbrain git commit-push -- ..drafts/a.md commits the file instead of failing with "outside the checkout", and a gbrain.sourceRoot of ..drafts now scopes the commit to that folder instead of being dropped (which scoped it to the whole checkout).
  • Sync markers no longer report a file that was never committed as committed when the source sits in a .. folder of its repository.
  • Skill packs and the skills catalog (get_skill, list_skills) read a SKILL.md that sits in, or is linked to, an in-pack ..shared folder.
  • gbrain frontmatter expects the slug ..drafts/note for ..drafts/note.md, as for any other folder, instead of reporting a false slug mismatch; its backups of two note.md files in different .. folders no longer overwrite each other.
  • Doctor's git-writer and foreign-writer checks, brain-repo hardening (including a core.hooksPath of ..hooks), the first-turn brain rule, extract-benchmark, extract-atoms transcript labels, skillpack update bases, sync applicability, fence-write and reconcile path states, and the LongMemEval call plan treat such folders as inside.
  • Real escapes stay refused with the same messages as before: ../x, an absolute path elsewhere, a sibling folder that shares the root's name prefix, a symlinked parent that points outside, and another Windows drive.

Added

  • bun run verify runs check:relative-dotdot, which refuses a new .startsWith('..') containment check in src/ and points at relativeEscapesRoot in src/core/path-confine.ts.

On Postgres, file records now report their size as a plain number, and closing the engine a second time no longer runs its shutdown steps again. PGLite already worked this way.

Fixed

  • On Postgres, getFile and listFilesForPage returned size_bytes as a BigInt. PGLite returned a number, and the declared type is a number. Both engines now return a number. The output of the gbrain files CLI and the MCP file operations does not change, because those paths already converted the value.
  • On Postgres, a second disconnect() of an engine that was already closed ran the registered shutdown hooks again. A second disconnect() now does nothing, as on PGLite.

[0.60.735.0] - 2026-11-17

Conversation archive pages written while a brain was personal are now kept to the owner as soon as a command makes the brain shared, instead of staying readable by every member until the next archive run.

Fixed

  • gbrain access member add and gbrain access member enable (when they make two or more enabled members), gbrain config set brain.audience shared and gbrain access owner set now apply the owner-only rules to the conversation archive before they return. member add and member enable do it before the member is added, enabled or bound to a token, and change nothing if it fails; a member add refused for its own input (a bad kind, email or group) changes nothing either. Each command prints one line on stderr per archive source; --json output is unchanged. (#9563)
  • Still waiting, and named by gbrain transcripts archive status and gbrain doctor: a brain that turns shared from client usage with no command run (gbrain transcripts archive run applies it), a shared brain with no owner (gbrain access owner set <member>), and access rules kept in git (add the conversations/ owner-only rule to .gbrain/access.yml). Pages an archive run already in progress writes into new folders stay readable by members until the next run; with maintenance off, that run is manual.

Forwarding an email with only an FYI note no longer opens a "waiting on them" loop.

Fixed

  • A forward whose own note asks nothing ("FYI, see below.") over someone else's question opened an unanswered_outbound loop, against the rule in docs/guides/open-loops.md that an FYI or forward without a question opens nothing. Since #8941 a forward keeps its whole body, and the loop detector counted a question mark anywhere in that body. It now reads only the note above the forward separator (Gmail's "Forwarded message" line or Apple Mail's "Begin forwarded message:"). A forward whose note does ask still opens a loop, and forwards still keep their full content.
  • The same reading applies to mail sent to you, so a "Thanks!" reply to someone's FYI forward counts as the reply again instead of leaving a reply-owed loop open.
  • A loop that already opened this way is not closed by the upgrade. It closes when the other person replies, or with gbrain loops drop <id>.

Keyword search finds the people behind a reworded relationship question again.

Fixed

  • On a brain with no embedding provider, a question that names a company in words its people's pages don't use ("Who is on the payroll at Acme?") returned the company page and little else. Since v0.60.338.0 (#7135) the title step's any-word retry read page titles only, so it no longer saw the people whose dated timeline names the company. For a local caller the retry now reads titles and timelines again while at most 5,000 pages match the question's words, and titles only past that, so a date or a very common word in a question still can't pull every dated timeline into ranking. Remote MCP callers are unchanged.
  • On the public one-hop relationship set (870 keyless rows), recall at 5 for reworded questions went from 0.381 to 0.573 with relational retrieval off and from 0.512 to 0.661 with it on, on both PGLite and Postgres.
  • On a 1,000,000-page Postgres brain the retry adds one bounded statement, 0.03 to 0.14 s in the cases measured.

gbrain repair open-facts now keeps a fact private once its end date has passed, and opens facts tables that sit below the timeline.

Fixed

  • A private fact whose valid_until date has passed stays private, and the preview counts it under a new reason, lapsed. Before, the repair could open a fact that was no longer true.
  • A fact in a ## Facts table below the timeline sentinel now opens in place. Before, the apply looked only above the timeline and reported every such fact fact_left_private.
  • The preview warns that a fact saved private on purpose before the cut looks the same as a legacy one, and that --ids-file limits the run to the ids you list.

Seven small fixes from outside reports: Claude Code transcript ingest, date-only search windows, the unit runner's failure log, doctor's Hermes hook check, get_page grep on tables, skill-catalog rebuilds and Windows sync paths.

Fixed

  • gbrain transcripts ingest skips the state files Claude Code 2.1 keeps inside each session directory, instead of failing each with unknown format and holding --since last back (#9365, reported and first fixed by @Masashi-Ono0611).
  • A date-only since in search, query and image query starts at UTC midnight, so a one-day window covers that UTC day whatever the database's timezone (#9533, reported and first fixed by @DanCanadian).
  • The unit runner's .context/test-failures.log keeps each failing test's own error and diff instead of the next test's (#9315, reported and first fixed by @Masashi-Ono0611).
  • gbrain doctor recognizes a gbrain turn-rule hook added by hand to an existing Hermes hooks: section, so harness_wiring no longer warns about it (#8846, reported and first fixed by @Masashi-Ono0611).
  • get_page grep returns the matching table row instead of the whole table (#9104, reported and first fixed by @Masashi-Ono0611).
  • gbrain setup keeps a skill catalog's --routing table when it rebuilds a stale catalog (#9332, reported and first fixed by @Masashi-Ono0611).
  • Sync on Windows accepts a page whose stored path uses the other path separator, instead of refusing it with page_identity_changed (#9118, reported and first fixed by @yuriysan).

Five small fixes from outside reports: connector sync on a brain with an active writer, Gmail bodies that are HTML labelled as plain text, repair captured-facts for Codex sessions, doctor --only on checks that need no database, and regen:all --check for the CLI flag registry.

Fixed

  • gbrain connectors sync <provider> run in the foreground now imports chat history on a brain whose writes go through the writer, the way gbrain transcripts ingest already does, instead of being refused. A --background or scheduled sync, and the connector_sync operation, still pass no write context. (#9124, reported and first fixed by @Masashi-Ono0611)
  • A Gmail message part labelled text/plain that is really an HTML document is now stripped to readable text instead of being stored with its raw tags and CSS. (#8937, reported and first fixed by @Masashi-Ono0611)
  • gbrain repair captured-facts now recognises facts captured from Codex sessions, in both the live and the archived Codex session folders, instead of counting them as unclassifiable. (#9554, reported and first fixed by @Masashi-Ono0611)
  • gbrain doctor --only <check> on a check that needs no database now reports "No finding" when the check has nothing to say, instead of a warning that the database checks stopped early. A connection that really failed still reports the stop. (#9527, reported and first fixed by @Masashi-Ono0611)
  • bun run regen:all --check now fails when the committed CLI flag registry is out of date, by comparing it with a fresh render. (#9019, reported and first fixed by @sx4im)

A token or OAuth client can now be allowed to read private facts in the sources its grant names, without changing any fact (gbrain auth rescope --private-facts granted-sources).

Added

  • gbrain auth rescope --token <name> --private-facts granted-sources (or --client <id>) lets that one credential's remote reads see facts saved with visibility private, only in the sources its grant names explicitly; --private-facts none turns it off on the next request. Every existing and new token and client starts at none (Discussion #8194).
  • It never opens a fact held on, provenanced from or about a private, readable_by, folder-ruled, quarantined or member page; context_pack, delta, writes and audience or attested reads stay world-only. Only the local CLI on the brain host can set it.
  • gbrain auth list (and --json), gbrain auth clients and the rescope before/after output show the new axis.

A new Postgres test now fails if the local title search stops using the index for rare words or the cheap early stop for common ones.

Added

  • A test against the real Postgres planner checks the candidates step of the local title search. A word on a few pages must read matches from the full-text index, and a word on nearly every page must walk the primary key and stop after the page cap. The test changes no behavior.

Commands no longer print a "wedged" migration notice asking for a forced retry once a newer migration has completed since; doctor and apply-migrations name each source still in conflict and its next step instead.

Fixed

  • Since v0.60.725.0 every ordinary command (remember, recall, config get and the rest) printed Upgrade migration(s) 0.53.0 failed several runs in a row and are wedged with an ask-the-user --force-retry fix on a brain whose upgrade ledger held three old partial runs of a migration, even though a newer migration had completed since and gbrain apply-migrations called it history. The notice now follows apply-migrations, which treats that migration as history, so ordinary commands print nothing about it. A stuck migration with no newer migration past it still gets the notice.
  • The shared-skills migration (v0.53.0) reports a source conflict where an operator looks, and never advises a forced retry: gbrain doctor (connected) warns with each source in conflict, its code and its next step, and goes quiet once none is left; gbrain apply-migrations --dry-run, --yes and post-upgrade name the same sources. They read the migration's own read-only preview, which takes under a second on a brain with five sources of 20,000 files each.
  • A source conflict now carries a code. shared_skills_unmanifested (gbrain errors shared_skills_unmanifested) means a source's skills/ folder holds skills with no skillpack.json declaring them: declare the ones to share in a skillpack.json at the source's root, or keep the source out of shared-skills adoption with gbrain sources shared-skills <id> off. Other conflicts carry the code of the error that stopped the inventory.
  • A named run of a wedged v0.53.0 (gbrain apply-migrations --migration 0.53.0 --yes) still refuses, now with the read-only preview as its fix instead of a forced retry.

Tests

  • A source with no local path (a database-only source such as the conversation archive's) is pinned never to answer a path lookup, on PGLite and Postgres.

gbrain doctor now names any purge or withdrawal guard on Postgres that still lacks its ledger check, and a migration run no longer hides a failed guard re-create.

Added

  • New doctor check guard_ledger_checks. On Postgres it names each purge or withdrawal guard whose body is missing the check that refuses a write role row-level security keeps from the ledger, so that role could still store purged content again or save a withdrawn fact active. It names the role that owns each guard. When the role doctor connects as may re-create them, the fix is gbrain apply-migrations --yes. Otherwise it prints the SQL the owning role runs. On PGLite, which has no row-level security, it reports not applicable (GST-2885, follow-up to #7632 and #9517).

Fixed

  • Every migration run re-creates the guards its role owns. That step used to swallow every error. Now only a guard another role owns stays silent (SQLSTATE 42501). Any other failure is reported on stderr with a line pointing at guard_ledger_checks. A concurrent update of the same function is retried (GST-2885).

Documentation

  • docs/ENGINES.md ("Purge and withdrawal guards"): a bulk fact insert whose row matches a withdrawal needs an UPDATE policy on facts for the write role, beside its SELECT policy (GST-2885).

The nightly heavy upgrade matrix no longer dies with exit 141 on a large doctor report, and the IPC takeover test no longer fails on slow macOS runners (#6049).

Tests and CI

  • tests/heavy/pg_upgrade_matrix.sh read doctor's top-level status with grep -o | head -1 under pipefail. Doctor's one-line JSON report now holds more than one pipe buffer of "status" matches, so head closed the pipe early, grep took SIGPIPE and set -e ended the script with exit 141 and no message, which turned the nightly Heavy Tests red (#6049). It now reads with sed -n, which consumes the whole stream, and a report with no status fails that shape by name. test/scripts/pg-upgrade-matrix-status.test.ts fails before the fix and passes after it.
  • The "simultaneous processes elect one provider" IPC test asked the elected provider through the 250 ms production budget, so a slow darwin-x64 runner read it as unavailable. The test now pauses the provider for 400 ms on every run and asks with a test-sized wait. resolveViaIpc takes an optional { timeoutMs } like its sibling clients; the default is unchanged.

New gbrain repair purged-change-reasons finds page-history change reasons that still quote a claim you purged before the #8635 fix, and redacts them after you agree.

Added

  • Before #8635 was fixed, a sync that read a page's gbrain-reason trailer just before a fact purge could still store the purged text as that page version's change reason. gbrain repair purged-change-reasons re-reads each reason the purge recorded from the source's Git checkout, counts the page versions that still hold it (for the whole brain and for each source, never printing the text), and with --apply --expect <hash> --yes sets exactly those to [purged], the value a purge writes. The same text under another file, or from a commit that gives the file two reasons, is kept unless you pass --include-ambiguous. It runs only when you name it, and never on upgrade, sync or doctor. Repo-less sources have nothing to check and the preview still succeeds. Pages, files and Git history are not changed (GST-2886, follow-up to #8635).

The research maps and the draft plan for a Rust storage engine are now in the repository, so anyone can read what gbrain does today and how the Rust work is meant to proceed.

Added

  • docs/research/rust-engine/: maps of TypeScript gbrain as it is today (the BrainEngine interface, schema, Postgres features, CLI, MCP tools, brain format, background jobs, test suites, subsystems) and of gbrain-evals as the parity oracle, with an index in docs/research/rust-engine/README.md. They are research notes pinned to one commit, not a description of current behavior.
  • docs/plans/PLAN.md: the draft, phased plan for a Rust storage engine. It is pending review and approves no build phase. Nothing changes for gbrain users.

On a read-only mirror, the held file-drift comparison no longer points at previews that find nothing, and a pass that only re-held what a retry asked for no longer counts as progress.

Fixed

  • On a read-only mirror, gbrain repair file-drift --held keeps each differing page's verdict and evidence but no longer names a --prefer preview, because those previews read every mirror page as database-only. It says the page stays held until the source is writable again (gbrain sources mirror-writable <id>).
  • A stopped sync pass whose only rewritten holds were re-checked because sync unblock or sources retry-held asked, or that brought a transient hold to its attempt cap, no longer resets the abandoned-pass count.
  • The sync-unblock runbook's pass_abandoned row, the repair guide (mirror comparison, complete: false on a held dry run) and the sources status key-file entry (by_code counts every hold) now match what gbrain does.

On PGLite, a managed gbrain sync that holds or skips many pages in a row now stops promptly on SIGTERM and keeps its timers running, instead of going silent for seconds at a time.

Fixed

  • PGLite runs Postgres inside the gbrain process and answers queries without ever handing control back to the event loop, so a managed pass that settled pages without waiting on a write (pages held for file drift, for example) could run for several seconds at a time with no timer, progress watchdog or signal handler able to run. On a test mirror whose pass held page after page, the longest stall was 8.3 seconds and a SIGTERM took 7.8 seconds to reach its handler; with this fix the longest stall was 89 milliseconds and the handler ran 21 milliseconds after the signal. Each step of a PGLite managed pass now lets the event loop take a turn first. Postgres syncs are unchanged (GST-2914).

gbrain sync stopped by SIGTERM or SIGHUP now records why its pass stopped.

Fixed

  • A managed gbrain sync that gets SIGTERM or SIGHUP in the middle of a pass saves the pass as resumable with stop reason terminated, prints its result and resume command, and exits 143 (129 for SIGHUP). gbrain sync status reads pass_terminated instead of guessing pass_interrupted, and the drain's JSON stop reads terminated instead of deadline. A pass stopped this way that committed nothing still counts toward pass_abandoned. A second signal, or a run still saving or printing its result 3 seconds after the handler starts, exits as before. On PGLite the handler can start tens of seconds after the signal, as it already did before this change, so a supervisor that kills the process sooner leaves the pass open as today; it reads pass_interrupted and the plain sync resumes it. Ctrl-C and sync jobs inside jobs work are unchanged (GST-2869).
  • On a read-only mirror, gbrain repair file-drift --held no longer names a --prefer preview for a held deleted page whose file differs either; it names gbrain sources mirror-writable <id> like the other mirror pages. The sync-unblock runbook's pass_abandoned row now says that a retry whose re-check changed the hold's code or file bytes still counts as progress once.

Questions about the past now see earlier values in the facts block, with when each was told.

Added

  • On a history cue (a history word, or a date that has passed), the query facts block also takes the named entity's ended facts (superseded, or past their valid_until), each with its own validity window. Forgotten and withdrawn facts, and facts from a deleted and purged page, never appear, and the caller's read policy applies as before.
  • When a question names one date, each fact header in the block says ; told after DATE for a fact said after it and ; learn date unknown when the saying date is unknown.

Fixed

  • The facts block shows recall's contested label: (contested by #ID) on a fact in a pending supersede proposal whose other fact the caller can read; a failed lookup reports contested_unavailable.

The facts block shows when a moved or changed value stopped being true: conversation facts carry a slot (where someone lives, who employs them, their role), and query works out each value's validity window when it reads.

Added

  • gbrain extract-conversation-facts asks the model for one slot per fact (location, employer or role) and stores it on the fact's ontology columns (dimension, value, value_hash, dim_status), on managed and unmanaged brains, with or without a git repository. Other fact extractors keep their prompt byte for byte.
  • The query facts block (and the temporal fact reserve) shows an older slot value as "valid FROM to UNTIL", ending where the next value starts, when both came from the same conversation page or the newer fact states the change in plain English. Without such a cue, both values stay open and nothing is labelled. A stored end date wins, and only facts the caller can read take part.
  • gbrain extract-conversation-facts --dry-run prints an estimate of the paid run: segments, input and output tokens, and dollars at the extraction model's price.

Changed

  • The ontology (ontology_get, ontology_propose, the chronicle conflict check, the dimension census and gbrain repair ontology-facts) ignores conversation facts, which are never closed in storage, so an extracted slot never shadows, closes or absorbs an owner's observation. A conversation fact that holds the exact key an ontology observation needs gives up its slot and keeps its text. Ontology observations report exactly as before.
  • A re-extraction or a retention purge that keeps a conversation fact as expired history clears its slot, so the slot never blocks the page's next facts.

For existing brains

  • Nothing is re-extracted automatically. Preview with gbrain extract-conversation-facts --force --dry-run, then run gbrain extract-conversation-facts --force to add slots.

Code blocks on pages made with gbrain publish get a Copy button, and gbrain mcp expose prints one-click links that add the brain to Cursor and, for a public brain, to Claude.

Added

  • Every code block on a published page, including a password-protected page once it is unlocked, shows a Copy button in its top-right corner. It copies the block's exact text, works from the keyboard, and reads "Copied" for a moment. Pages opened from a file or a plain-http host fall back to the browser's older copy command. Code blocks also get a bordered panel that scrolls sideways, so long lines no longer stretch the page. Contributed by @Pierre-Bouquet (#9453).
  • gbrain mcp expose ends its Next block with a Cursor line: Cursor's install link, which opens Cursor's MCP install prompt with the brain's URL filled in. With --funnel a Claude line comes first: Anthropic's custom connector install link, which opens claude.ai's add-connector form the same way. A tailnet-only publish gets no Claude line, because claude.ai connects from Anthropic's servers and can't reach the tailnet. --status repeats the links when everything verifies, and --json adds them as install_links (only when there is at least one). The links carry no credential, and a URL that isn't plain https gets none.
  • docs/mcp/README.md has a "One-click client links" section with ready links for hosted gbrain.io, docs/mcp/CLAUDE_DESKTOP.md puts the Claude shortcut above its manual steps, and docs/guides/remote-mcp.md shows the new line in its example output.

Fixed

  • A published page whose markdown contains a closing </script> tag, such as a code block that shows HTML, now renders that text instead of breaking the page. Before, the tag ended the page's script early, so the rest of the page was lost and any HTML after it skipped the sanitizer.
  • The password-protected page now renders unlocked content through the same sanitizing function as an open page, instead of patching its unlock script by string replacement.

Contributed by @Pierre-Bouquet

Three agent-facing docs that had hit their size caps now have room again; no cap was raised.

Documentation

  • The guard registry and self-test section moved out of docs/TESTING.md into its own page, docs/ci-guards.md. Links to its anchors, guard error messages and the guard baselines now point at the new page.
  • The key-files entries for page windows and large page writes moved from core-services-2-continued.md to a new core-services-2-page-window.md, and the entries from src/core/sql-query.ts through src/core/utils.ts moved from core-utilities-2.md to a new core-utilities-2-continued.md. Each full page links its continuation from its last line.
  • llms-full.txt links "Connect a hosted brain" (docs/guides/hosted-harness-access.md) instead of inlining it. It applies only once a brain runs on another machine, and the inlined AGENTS.md already routes those agents to it.

A .gbrain-scan-allow line can now allow one secret-scan rule in matching files, instead of skipping every rule there.

Added

  • .gbrain-scan-allow accepts path:<glob> pattern:<rule>, which allows only the named rule (for example jwt on newsletter unsubscribe links) in the files the glob matches; every other rule still blocks a push there. Path globs and fingerprints keep their meaning. A malformed rule line or an unknown rule name allows nothing and is named by the push refusal, gbrain git scan and bootstrap verify, and gbrain git scan lists each hit a rule line allowed with the line that allowed it (allowed[], allowlist_problems[] in --json).

Four failures that used to happen silently now either work or report themselves: connector pages on a managed brain get facts again, takes extract --dry-run sends nothing to the model unless you also pass --yes, turning off mention linking records each removed link in the belief ledger, and a claude-cli call that hits the chat timeout fails only that call.

Fixed

  • On a managed brain, Slack and meeting pages that a connector publishes are queued for fact extraction again. The connector used to call a queue path that a managed brain refuses, and the refusal was swallowed, so those pages never got facts. Each import's publication now queues the extraction itself and keeps --no-extract and the 50-page limit per run. Gmail, calendar and GitHub imports queue no facts on a managed brain, the same as on a classic brain. On a classic brain, a page whose extraction could not be queued is now logged to ingest_log and reported on stderr instead of being dropped. Slack pages synced before this fix get their facts from gbrain extract-conversation-facts --source-id <id> --types slack --yes (#9552, reported by @lloydsilvertwo).
  • gbrain takes extract --from-pages --dry-run without --yes lists and counts the candidate pages and calls no model, as its "preview" wording promised. It used to send every candidate page to the chat model without --yes. Its output says nothing was sent and no claims were classified, and --json carries dry_run: true with claims_extracted: null. --dry-run --yes keeps the old preview, which classifies the candidates and writes nothing (#9547, reported and first fixed by @Kkkkkuro).
  • Turning off mention linking for a source that the belief ledger covers now writes one link.retracted event for each mention link it removes, so the ledger no longer asserts links that are gone. The turn-off and the mention reconcile take a per-source lock in a fixed order, so the two cannot deadlock (#9548, reported and first fixed by @akinduroifedayo).
  • When a claude-cli call hits the chat timeout (GBRAIN_AI_CHAT_TIMEOUT_MS, 300 s by default), the caller now gets the timeout error instead of "claude-cli adapter aborted". Fact extraction records it as a provider error for that page and moves on, so one hung call no longer ends the whole conversation-facts backfill (#9553, reported by @lloydsilvertwo).

The access commands now say when a page is left readable by no member, and what to run.

Fixed

  • gbrain access member disable lists the pages naming the member that no enabled member reads remotely any more, with the gbrain repair visibility --slug <slug> --readers <member>,<owner> write that gives one back.
  • gbrain access who-can-read explains a page no enabled member reads remotely, whether because its readers name only disabled or removed members or because its readable_by and the folder rules covering it admit nobody in common.
  • gbrain access rule add (and the other rule edits) warns in its preview about pages the change leaves readable by no member, such as a page written into a person's folder before that folder's rule named them.
  • gbrain auth list shows a token bound to a disabled or removed member as refused; --json keeps status and adds member_status.
  • A thin client whose credential is bound to a disabled member now says so (access_member_disabled) instead of "credentials may have been revoked".

gbrain skillpack harvest now harvests a skill whose paired file sits in a folder that starts with the skill's name, such as skills/foo-extra/ next to skills/foo/.

Fixed

  • A paired source in a sibling folder whose name starts with the skill's name was treated as a file of the skill folder, so the harvest stopped with path_traversal and copied nothing. The harvest now copies it to the same path as any other paired source. Symlinks and files outside the repo are still refused.
  • A harvest that refuses a paired source now stops before it writes anything. Before, it could copy the skill folder first and leave it behind, so a second try stopped with slug_collision.
  • A paired source reached through a linked folder that points outside the host repo is now refused with path_traversal. Before, harvest refused only a linked file, not a linked folder.

Keeping a deletion on a read-only mirror now sticks: the held page stops asking again on every re-check.

Fixed

  • On a read-only mirror, gbrain repair file-drift --held --side deleted --state tombstone_edited --prefer database lists each page keep_deleted and, once the user agrees to the preview's hash, records the decision on its hold instead of trying to remove a file gbrain never writes. The hold stays, reported carried with reason mirror_keep_deleted and needs_human: false in gbrain sources status, gbrain sync status and doctor. A re-check of the same file bytes keeps the decision; a new upstream edit or gbrain restore asks again; a purge moves it to the purged_differs hold, where it stays while the file bytes are the same. The caller's CLI grant must cover the page, as for every other file-drift repair. Nothing is deleted, restored or written into the mirror, and writable sources behave as before.

A design plan for an optional Rust storage engine is now in the docs; no code ships and nothing changes for any brain.

Added

  • docs/designs/RUST_ENGINE.md: the plan for a third storage engine, written in Rust, behind the same engine contract as PGLite and Postgres. It covers the parity tests and gbrain-evals bake-off, a shadow mode that runs Rust beside an existing brain and generates a head-to-head report, an opt-in swap with a way back, and a public engine benchmark in gbrain-evals. It sets the bar that the Rust engine must be equal or better than both Postgres and PGLite on every measured axis, including concurrent writes, graph traversal, vector search and memory. The target is 10x better than Postgres and pgvector, and equal or better is the gate to ship. The Rust engine has no SQL dialect and serves typed engine methods only. It refuses raw SQL. It does not copy PGLite's locking. It has no whole-brain lock and no busy error, and a second process can always read and write. Each lock and queue call site becomes a typed engine operation with a stated guarantee and a concurrency test. The embedded engine comes first and is built so it can later serve many users. A full multi-user Rust server that replaces Postgres and pgvector is the end state. The new section "Path to a multi-user server" lists those later phases, and each one needs the maintainer's approval before it starts. Until that server passes the same bar, brains with writers on more than one machine stay on Postgres. Every switch it names ships off.
  • docs/designs/rust-engine/research/: twelve read-only maps of the current code that the plan cites, and docs/designs/rust-engine/review-record.md: the review record for the plan.

A new research page shows how proven databases let many users and agents write at once without whole-brain locks, and what the Rust engine should copy.

Documentation

  • New docs/research/rust-engine/multi-user-prior-art.md compares about 40 databases, storage engines, search indexes and access-control systems on concurrency, crash safety, replication, access control, backup, multi-process use and proof of correctness.
  • It recommends optimistic, serializable transactions checked at commit (the FoundationDB design) over a small proven store, with redb leading the bake-off, and deterministic simulation testing from the start.
  • It maps each Postgres lock and queue pattern gbrain uses today (job claims with SKIP LOCKED, advisory locks, row locks, TTL leases, LISTEN/NOTIFY, RLS) to a Rust engine replacement without PGLite-style whole-brain locking, and lists what may still wait.
  • It finds that the Rust rewrite of Postgres is not a fit as a dependency: it keeps the Postgres lock model, is AGPL-licensed and is not yet production-ready.

The access docs now say plainly that restricted pages are still embedded by the brain's configured embedding provider.

Documentation

  • readable_by, folder rules and visibility: private control who reads a page, not which configured providers process it. The memory-boundaries guide ("Where text goes"), the access-designations guide and the company-brain tutorial now say that every page is embedded with the configured provider, that gbrain cannot tell a local endpoint from a remote one, and that a self-hosted provider keeps restricted text on your own hardware. Access designations do not exclude a page from embedding.

[0.60.734.0] - 2026-11-17

Repairs keep a page's safety flag or quarantine, gbrain facts relink keeps a fact's origin and exact dates, and consolidate no longer ends a fact because it was said again.

Fixed

  • gbrain repair rows-behind and the other managed file repairs no longer strip the write gate's instruction_like flag or quarantine from a page. A repair now publishes through the same write gate a sync of those bytes runs, with the file's path, so an owner allow rule scoped to that path holds during a repair as it does during sync. Rows-behind and overwritten-edits comparisons ignore those database-only markers, so a flagged page is no longer reported behind its own file. Under write_gate.external_mode=reject, a repair of external instruction-like content now refuses with write_gate_rejected, as sync already holds that file, and the refusal names gbrain trust allow --source <id> --uri-prefix <path> (#9511).
  • gbrain facts relink no longer re-stamps a fact's trust tier and write origin with the relink request, and no longer rounds its valid_from and valid_until to the second. The page's facts table carries the stored instant to the millisecond, so the page projection writes the same instant back. A fact stored finer than a millisecond is left unlinked and counted claim_unfenceable instead of being rounded (#9260).
  • The upkeep consolidate step no longer ends a fact at the next copy of the same claim. A fact now ends at the next fact in its cluster that says something else, by claim text and typed value, so a restated value stays active from the date it was first stated, every copy ends together when the value changes, and a typed metric whose value changed under the same wording still ends the older value. Facts consolidated before this release keep the dates they have.

New gbrain repair open-facts lets the operator open facts saved private before the world default, after a preview, with an undo.

Added

  • gbrain repair open-facts --source <id> --created-before <ISO-8601> previews the active private facts saved before that instant, counts every fact it leaves private by reason (private, missing or quarantined pages, member pages, withdrawn, contested, expired, superseded), and --apply --expect <hash> opens exactly that window in place, keeping ids and trust tiers (Discussion #8194).
  • Facts with no page and transcript-lane facts are listed apart and open only with --class unpaged:<lane>; --ids-file narrows the set; --undo <ledger> re-privates what a run opened.
  • The command runs only from the CLI on the brain host: remote callers, member mode (GBRAIN_ENFORCED_MEMBER), repair --all, doctor and the remediation plan never run it. Nothing changes until an operator runs it.
  • A fact made private on purpose before the cutoff is opened with the legacy ones, so read the preview or narrow the run with --ids-file.
  • A fact with an end date (valid_until) in the past that was never expired or superseded is opened like a current one; a follow-up leaves it private.
  • A fact whose table sits in a page's timeline is left private; a follow-up opens it.

A remote remember that restates a confirmed private fact as world is now stored private until the brain's operator accepts it, as page writes to that fact's own page already were.

Fixed

  • A remote caller that remembered the exact claim of a private fact on the same entity, more trusted than itself, got a new world copy that every other connected client could read, so an agent that knew a private claim could publish it. The new fact is now stored private with a trust proposal that carries the widening; gbrain trust review lists it and gbrain trust confirm tp<N> makes it world. Remembers with no private copy on that entity, with a world copy already present, or with an explicit private visibility are unchanged.
  • Two gaps remain until #9574 lands: the page that carries a held fact still shows its text to remote get_page and search, and a writer who first saves the same claim about a different entity is not held.

Ten small fixes from outside reports: agent-session repair, doctor, the prompt hook, sync, transcript ingest and enrich.

Fixed

  • gbrain repair agent-session-facts no longer counts sessions that already carry a speaker sidecar toward another session's repeated opening, so an owner session whose prompt was replayed by marked workers is no longer listed as ambiguous (#9521, reported and first fixed by @Masashi-Ono0611).
  • gbrain doctor --only memory_writeback connects the database instead of reporting it unreachable (#9513, reported and first fixed by @Masashi-Ono0611).
  • The Retrieval Reflex recipe's health check and setup steps run only its own doctor check, so an unrelated failing check no longer fails the integration (#9228, reported and first fixed by @johnerik).
  • A facts-absorb job that finished by skipping (superseded, page missing, extraction off) no longer counts as proof the page was extracted, so a later rewrite that keeps the body still queues extraction (#9248, reported and first fixed by @Masashi-Ono0611).
  • Task notifications and teammate messages no longer get the brain-first nudge, in the main prompt hook and in the turn-rule hook; prompts a person types are unchanged (#9163, reported and first fixed by @Masashi-Ono0611).
  • A failed commit-ledger location lookup is retried on the next call instead of being remembered for the life of the process (#8870, reported and first fixed by @sx4im).
  • gbrain sync --all --parallel N caps how many sources run at once, as its help says, and --max-sources still wins when both are given (#8951, reported and first fixed by @Masashi-Ono0611).
  • gbrain transcripts ingest skips Claude Code's per-project sessions-index.json instead of reporting an unknown format on every run (#9115, reported and first fixed by @Masashi-Ono0611).
  • The autopilot-install doctor test names its installer target, so on macOS it can no longer load a real LaunchAgent (#9545, reported and first fixed by @Kkkkkuro).
  • gbrain enrich stops asking the provider for a prompt cache, because its system prompt is below the smallest cacheable prefix and the cache was written on every call but never read (#9541, reported and first fixed by @mattverlaque).

gbrain repair agent-session-facts finishes its preview on brains with tens of thousands of saved facts, and gbrain setup claude-code no longer fails its re-run after Claude Code rewrites its settings file.

Fixed

  • gbrain repair agent-session-facts (and doctor's agent_session_facts check) no longer times out on a large brain. The preview fingerprinted the source's facts once for every user-attributed fact saved from a session, before it knew whether that session was an agent's; it now looks for the owner's own copy only for facts whose session it already judged evidenced or ambiguous, using the existing fingerprint index, and it reads each take's contributing facts in one pass instead of one lookup per take. On a generated brain with 143,640 facts and 7,045 sessions the preview went from a statement timeout to about 1.5 seconds (#9503).
  • gbrain setup claude-code re-run no longer exits 1 with turn_hook_not_loaded (and doctor's harness_wiring no longer warns) when Claude Code has rewritten ~/.claude/settings.json without the _gbrain key on gbrain's hook entries. The per-turn hook check now recognises an entry the setup receipt owns, the same way setup itself does (#9505).
  • gbrain errors migration_host_work_owed explains the doctor warning instead of printing "Unknown error code".
  • gbrain sources git-reconcile treats a git command that ran past its time limit or was killed from outside (git_command_stopped) as a retryable failure (exit 4) instead of asking the user (exit 1). A git command stopped by its output limit or by a git crash still asks the user, since a retry stops the same way.
  • gbrain git commit-push refuses with git_unavailable when it cannot list conflicted files, instead of reading a failed listing as "no conflicts". A write-through file refused for unresolved conflicts is kept in the pending ledger, so doctor reports it and --pending retries it, while later write-throughs still commit the ledger's other files, and the advice now says to keep the changes from both sides rather than to pick one side, since one side may be a write gbrain made.
  • A git child that crashes (SIGSEGV, SIGBUS, SIGILL, SIGFPE or SIGABRT) is reported as a crash with a git fsck suggestion, not as a kill from outside gbrain, and output too large for one string is reported as git_command_stopped (output_limit) instead of an empty listing.

Tests and CI

  • The unit suite runs on 36 shards instead of 34, with refreshed shard weights, which brings the predicted time per shard from about 595 s to about 562 s.
  • A test now covers scripts/check-migration-projection.ts and its baseline, and macOS 26 validation records the tree it tested so a passed run can stand as evidence for the same tree.

gbrain frontmatter validate checks every path you give it, and doctor, gbrain serve --http, the update check and the agent-session repair each get a small fix.

Fixed

  • gbrain frontmatter validate a.md b.md validated only the last path and could report ok for a bad file named earlier. It now checks every path, walks each directory, checks a file named twice once, refuses before checking anything when a path is missing, and lists the paths as targets in --json when there is more than one.
  • Doctor no longer tells you to run gbrain sync for a gbrain-owned content folder on a brain with persistence off, which sync refuses. The check still fails, and it now names the gbrain import command that works for that folder.
  • gbrain transcripts archive status says when maintenance.auto is false, because free upkeep then doesn't archive conversations by itself (an installed autopilot still does). The line names gbrain transcripts archive run.
  • gbrain serve --http no longer treats a request refused for a missing or wrong token as activity, so unauthenticated traffic can't keep idle upkeep from running (#9484).
  • If your computer's clock moved backwards, the update check no longer goes quiet. A saved answer stamped more than 5 minutes in the future counts as stale and is replaced, so a notice that your version was withdrawn still reaches you (#8795, also reported in #9157).
  • gbrain repair agent-session-facts reads only the listed sources' facts when it looks up takes, so its cost follows the source instead of the whole brain.

Local searches on a very common word no longer spend the keyword time limit in the title step.

Fixed

  • The title step of hybrid search now ranks at most 5,000 matching pages for local callers, the ones with the lowest page ids, so the cut keeps the same pages on every call. A word that sits in many page timelines used to make it rank every matching page and run into the 8 s keyword statement limit on a large brain, a second time after the keyword step (#7082). When the cut applies, pages whose title holds the words come first and hybrid search reports keyword_candidates_incomplete. A query that matches 5,000 pages or fewer keeps its exact result, including a page found only through a timeline row that no chunk holds. Remote callers' results do not change.

A restricted Postgres write role can no longer save purged or withdrawn content again by lacking a read policy on the purge and withdrawal ledgers: the guard now refuses the write with a message naming the policy to add.

Fixed

  • The seven purge and withdrawal guard triggers (facts, takes, pages, segment derivations and the derivation sweep's scope note) read their ledgers as the writing role. A role with write policies on the guarded table but no policy on the ledger read it as empty, so a purged claim, take, page or derivation could be stored again while the purge reported done, and a withdrawn fact could be saved or set active again. Each guard now refuses such a role with writer_coordinator_required (unreadable_state), the same refusal the writer fences and derived-state maintainers use, and its hint names the table and the docs/ENGINES.md recipe. The guards keep running with the writer's own rights (no SECURITY DEFINER). Migration v318 applies it to existing brains. Owners, roles that bypass row-level security and PGLite see no change. (#7632, #7659)
  • After upgrading on Postgres, check the guards. v318 can only replace a guard function the migrating role owns. One owned by another role stays fail-open, and v318 names it in a single stderr line starting guards:. docs/ENGINES.md ("Purge and withdrawal guards") has a query that lists the seven guards and whether each now fails closed. For any that doesn't, its owner runs ALTER FUNCTION public.<name>() OWNER TO <the role gbrain migrates as>, then gbrain apply-migrations --yes (or the next gbrain start) replaces it.

A purged claim no longer comes back as a page's change reason when a sync publishes after the purge.

Fixed

  • A sync reads each page's change reason from its gbrain-reason commit trailer before it saves the page. If a fact purge committed in between, the saved page history still stored the purged text as that page's reason. The save now checks the purge's tombstone for the exact commit and file when it writes the history row, so the reason is stored empty, while the page content still imports and other files' reasons from the same commit are kept. This applies to classic and managed sync (#8635, #8591).

A large read-only mirror source whose sync passes only re-check old holds no longer parks as pass_abandoned with nothing settled, and its status and repair output now say what is actually held.

Fixed

  • A sync pass that stopped while it re-checked old holds now counts as progress when it cleared a hold's re-check trigger, so a few such passes no longer build the abandoned streak. Passes that re-check the same holds every time (slug conflicts) still end pass_abandoned, and so do passes that only re-hold a page edited again before each of them: once that hold has been held three times, gbrain sync status reports needs_human: true with that page's own decision.
  • A pass now reads the earlier Git versions of the held pages it re-checks one batch at a time, just before that batch commits, instead of for the whole round before its first commit. On a source with many such holds the first commit came only after the whole history read.
  • gbrain sync status leads with the pass's own next step when the holds are not what stops the source, and next.why for pass_interrupted and pass_abandoned names the step the stopped pass was last seen in (the level read, or the history read with how far it got).
  • On a read-only mirror, gbrain repair file-drift --held compares each held page with its file. A differing page reads prefer_file or needs_human with its evidence, a matching one release_hold, and only a page whose file is gone reads no_drift with database_only. The comparison still writes nothing.
  • A windowed held dry run reports complete: false while held pages remain past its window, and details.window.next_command names the next window.
  • gbrain sources status --json counts git_holds.by_code over every hold of the source, not only the listed ones.

Larger sources still need a per-class decision flow for their held pages; this change only makes the counts and the comparison true.

Saved facts in query results now say when they were said, collapse their copies, match on word stems, and end at the right date when replaced.

Fixed

  • The facts block, temporal fact reserve, facts arm and recall date headers print the observed date of the page each fact came from, read under the caller's page access; a failed read reports fact_observed_date_unavailable instead of a silent observed unknown.
  • Copies of one claim take one row in the facts block (dated as first stated, (stated N times, last DATE)), so restated facts stop crowding out older values; facts that differ by one word keep their rows.
  • Question ranking of saved facts matches on stems under the brain's full-text config (an exact word still ranks above a stem-only match) and ignores dates, years and auxiliary verbs, so an as-of question admits the facts it names and a fact sharing only a year with the question is not admitted.
  • Every supersession writer ends the replaced fact by one rule. It ends on the newer fact's valid_from when it falls inside the older fact's window, else at the older fact's own earlier end or now, and never before the older fact began. New writes only.

[0.60.733.0] - 2026-11-16

gbrain transcripts archive no longer has a day page refused by the write gate over its name, and a run exits 1 exactly when its errors list is not empty.

Fixed

  • Day page names always pass the write gate. A cron job name cut at 24 characters no longer ends in a hyphen, and a label's repeated hyphens fold into one and its leading or trailing hyphen drops.
  • Every day a run fails to write is listed in the run's errors (in --json) with its day and code, and a shared brain with no owner lists the archive_owner_unset hold there too. run and backfill exit 1 exactly when that list is not empty, and 0 otherwise.
  • --dry-run reports write_mode: coordinator when the real run would write through the write coordinator, even before the archive source exists.
  • Days recorded by gbrain transcripts archive adopt-legacy are still skipped after the brain turns shared and the archive moves to its owner-only source.

The conversation archive reads OpenClaw's Gateway in far fewer process starts, and gbrain transcripts archive status no longer reads any session.

Changed

  • The OpenClaw Gateway reader asks chat.history for pages of up to 1,000 messages within a 4 MiB budget, and reads up to 4 sessions at once. Each openclaw gateway call is its own process, so on a test install with 350 sessions a read went from 162 process starts and 164 s to 52 starts and 16 s. A Gateway older than 2026.10 that refuses the byte budget is asked again without it. Pages of one session stay in order, conversations keep their sessions.list order, and a failed session still fails the whole read before any day is written.
  • gbrain transcripts archive status and gbrain doctor --only transcript_archive_freshness report what the last archive run read for each reader (conversations, newest session day, time and any error) instead of running the readers. They start no openclaw process and parse no session file. A reader no run has read yet says so. Each real run records this in transcripts.archive.last_read.<source>; a dry run records nothing.

Fixed

  • Doctor warns when no archive run has read a reader successfully within transcripts.archive.stale_days, for example when maintenance passes keep stopping at their deadline before that reader, and suggests gbrain transcripts archive run.

A keyword search that timed out on a cold Postgres cache comes back within about a minute instead of ten.

Fixed

  • After hybrid search's keyword statement hits the keyword statement timeout, gbrain serve and MCP sessions still skip that exact statement (keyword_arm_failed: skipped_after_timeout), and now also re-run it once in the background 30 seconds later, then after 60, 120 and so on seconds while it keeps failing, never past search.keyword_timeout_negative_cache_seconds (default 600, unchanged). The first re-run that answers ends the skip. A re-run uses the same statement, scope and bound, starts only when the pool has no waiting query and keeps a connection free for other queries, runs one at a time per brain, throws its rows away, writes no timeout counter row and is not recorded as a search. Before, one timeout on a cold cache left every identical search keyword-less for ten minutes even after the cache had warmed (#7082).

A stuck shared-skills migration is reported as owed work everywhere, with a read-only first step, instead of being called history.

Fixed

  • A migration that stopped three or more runs in a row on a source conflict and still owes host work (v0.53.0 today) is no longer called "history, not outstanding work" because a newer migration completed. gbrain apply-migrations --list, the notice on other commands and gbrain serve startup, and gbrain doctor (a warning) now say the same thing: the work is still owed, it does not clear without a person, and the first step is gbrain apply-migrations --migration 0.53.0 --dry-run --json, which lists each source's conflict and changes nothing. They give the order that works: review each conflict, clear the wedge with gbrain apply-migrations --force-retry 0.53.0, then run the command the inspect output names for that source, or gbrain apply-migrations --yes. That inspect command no longer refuses a wedged version as "not registered", and running a wedged migration by name (for example the accept command a conflict message names) still exits 2 but says it is wedged and names those steps. Nothing runs the migration again by itself, and no exit status changes.

bun run verify now fails when a schema migration would take every page it touches out of search until gbrain projections drain rebuilds them, unless the migration is listed with the reason search keeps answering (#9192).

For contributors

  • bun run check:migration-projection, part of bun run verify, fails on a migration whose SQL, or whose handler's own source, updates a page column the page revision trigger watches (UPDATE pages ... SET or an INSERT INTO pages ... DO UPDATE SET upsert) or writes tags. The dollar-quoted body of a CREATE FUNCTION or CREATE PROCEDURE is skipped; a DO block is read.
  • A migration that has to do this goes in scripts/migration-projection-baseline.tsv with the reason search keeps answering, or why it cannot. v153 and v189 are listed. A listed migration that no longer matches fails as stale.
  • The guard cannot see a handler that changes pages through helpers or functions defined elsewhere, SQL assembled at run time, or writes made by a database function the migration calls. Review those by hand.

gbrain git commit-push works on a checkout with a very large index, no longer commits a file that still has merge conflicts, and OpenClaw skill setup no longer fails on a busy machine.

Fixed

  • gbrain git commit-push read each git command's output with Node's default 1 MiB limit, so on a large checkout (one report had about 385,000 tracked files) naming a folder killed git and the run failed with git_unavailable "git ls-files -v failed: exit -1". It now reads git's whole output, so the commit goes through. A git command stopped before it finishes (its time limit, or a signal such as the out-of-memory killer) now fails with git_command_stopped and its reason, still exit 1, instead of git_unavailable, and its partial output is never used, so a cut-off listing can no longer pick the wrong files to stage or restore. A push, fetch or remote listing stopped after the local commit still reports push_unconfirmed (exit 4), and a write-through file whose commit was stopped stays in the pending ledger for the next retry.
  • gbrain git commit-push refuses a named file that still has unresolved merge conflicts in the index (for example after a merge, rebase or stash pop that stopped on a conflict) with commit_paths_unmerged, before it stages anything. It used to commit the file, conflict markers included, as if it were resolved. Nothing is staged or committed, and the message says how to resolve the file first. Every gbrain write that commits named files goes through the same check.
  • gbrain setup openclaw checked its installed skill plugin by resolving one turn against the live per-turn deadline of 50 ms, so on a loaded or slow machine the check timed out and setup failed with skill_resolver_selftest_failed. The check now gives that one turn 10 seconds, because it tests that a turn resolves, not how fast. Live turns keep the 50 ms deadline.

After an upgrade from before v0.51, gbrain projections drain rebuilds the pages that are out of search first, and on Postgres it rebuilds 4 pages at once by default, so every page answers search much sooner.

Fixed

  • The search-index rebuild that an upgrade from before v0.51 starts used to work through pages in queue order, so pages with a facts fence could wait hours behind pages that were still answering from their old index. gbrain projections drain now rebuilds the pages that are out of search first, then seals images, then rebuilds the pages still answering from their pre-upgrade index. --dry-run lists rows in the same order. A follow-up to #9192.
  • On Postgres the drain rebuilds several pages at once, each on its own pooled connection: 4 by default, or fewer when the pool is small, and --concurrency N goes up to the pool size less 2. On a 50,000-page test brain with 30,000 fenced pages behind an 18 ms round trip, every page was searchable after about 86 minutes at the default. PGLite, gbrain serve on PGLite and the drain inside gbrain migrate --to still rebuild one page at a time.
  • A refused or dropped extra connection lowers the width by one and the drain carries on; stderr names the new width. It never fails the drain.
  • The upgrade's [153] line gives the rebuild time at the default width, names the faster --concurrency command with its time, and counts only the Markdown and code pages the drain rebuilds, not images.

The check that runs before every search stays fast while a large brain's pages wait to be rebuilt after an upgrade from before v0.51.

Fixed

  • After an upgrade from before v0.51, pages still answering from their old index were walked one by one by the readiness check that runs before every search, so the check cost 21 ms on a test brain with 20,000 such pages and 126 to 149 ms with 200,000. A new partial index (migration v317) holds only the pages that are out of search, so the check now takes about 2 ms and the index is empty once the rebuild finishes. On Postgres the index builds without blocking writes; nothing is rewritten. Follow-up to #9472 (#9192).

gbrain doctor checks for facts left behind by purged pages in about a second per million facts, instead of timing out on a large brain. The purged_page_facts check used to run its exclusion tests (a withdrawal lookup that fingerprints each claim, plus the superseded and open-loop checks) on every active fact before asking whether the fact's page was gone, so on a brain with a few million facts it ran past its 15-second budget and reported nothing. It now finds the facts whose page is gone first and runs those tests only on them. It counts exactly what it counted before, and gbrain repair purged-page-facts is unchanged.

Fixed

  • Doctor purged_page_facts: on a generated brain with 1,000,000 active facts and nothing to report, the check took 6.7 to 7.0 s on Postgres and 24 to 26 s on PGLite, and now takes about 1.0 s on both.

A release's list of changes shows each change's headline alone, even when its changelog entry continues on the headline's line.

Fixed

  • scripts/changelog-entry.sh --pr-list cuts each fragment's first bold line at its closing **. Before, a fragment whose body text followed the headline on the same line put that whole paragraph, with a stray **, into the release notes.

For contributors

  • test/release-workflow.test.ts covers a fragment with body text after its headline.

Three fixes from outside bug reports: retired atoms keep their trust tier, a provider answer that cannot be parsed is no longer retried, and doctor stops recommending a sync that refuses a source.

Fixed

  • An atom a re-extraction retires keeps the trust tier and write origin its extraction gave it, and comes back at that tier when a later extraction produces it again. It used to be stored as unknown with no origin, and a revived atom stayed unknown. Every derived retirement (chronicle events, synthesis outputs, gbrain repair stale-atoms) now stamps its pages this way. Reported and first fixed by @howardpark (#9456).
  • A model provider that answers with a success status and a body that is not one JSON document now fails with the non-retryable code provider_response_unparseable. A facts job that hits it ends on its first attempt instead of paying for the same failing call again, gbrain facts retry-dead leaves it alone, and its recorded error names the parse failure without keeping any response text. Reported by @haumanto and first fixed by @Masashi-Ono0611 (#9466).
  • gbrain doctor's sync_freshness no longer fails a source whose folder sits inside gbrain's own content directory, and no longer tells you to run gbrain sync for it, which refuses such a source with sync_not_applicable. It counts the source as a content directory and names gbrain import <dir> --source <id> instead. Reported by @railsloes and first fixed by @Masashi-Ono0611 (#9471).

The daily stable check sees merged important fixes again, and stops with a clear error if it ever loses access to pull requests.

Fixed

  • The daily stable check's GitHub token can read pull requests again. A recent permissions change had left it unable to see any pull request, so the search for merged stable:important fixes came back empty and the check reported nothing to promote.
  • Each run now logs how many labelled pull requests the search found. If the search sees no merged pull request at all, the check stops red and names the missing pull-requests: read permission instead of quietly reporting nothing.

gbrain upgrade and the stable release notes now start with a short summary of the behaviour changes since the version you had.

Added

  • Once the new version is installed, gbrain upgrade (through gbrain post-upgrade) prints the summary before its migrations run: the changes that need you first, then changed defaults, then a count of fixes and additions, and a pointer to gbrain doctor --only behavior_changes for the full text. Each change shows its first sentence. Under --json the summary goes to stderr with the other human lines, so the JSON document on stdout is unchanged.
  • When stable moves, its GitHub release notes open with the same summary for the range since the previous stable release, above the changelog entries. A summary too long for the page by itself is left out, so the notes always fit.
  • Rows written before kind: existed are counted as not classified yet. The generated behaviour-change table now carries each row's kind, so new rows land in the right group.

[0.60.732.0] - 2026-11-16

Every pull request is now checked against stable's promise: removing a command, flag, MCP tool or parameter, config key or exit code needs a deprecation record whose 14-day notice has passed.

Added

  • bun run check:public-surface, part of bun run verify. It compares this tree with the merge base against docs/contracts/public-surface.json, a list built from the code by scripts/build-public-surface.ts (refreshed by bun run regen:all). It refuses a pull request that drops a listed entry, or changes an exit code, before a deprecations.d/<id>.md record for it has reached its remove_after date. A cli-command or mcp-tool record also covers that command's flags or that tool's parameters.
  • Removals the promise does not cover go through with no record. That means anything missing from the list at the current stable release (the stable tag), and any removal in a pull request that adds a kind: breaking row with exception: security or exception: data-loss. The check fetches the tag when the clone lacks it, and in CI a fetch that fails or takes more than 20 seconds fails the check. Until a stable release carries the list, the check says so and counts everything as on stable; a revert then adds a row with exception: revert that names the pull request it reverts.
  • New deprecation records must give at least 14 days of notice. An announced remove_after date can move later but never earlier, and a merged record keeps the surface and name it announced.
  • New behavior-changes.d rows need kind: front matter (breaking, action, default, fix or add). A breaking row needs an exception or a deprecation: <id> naming a record whose date has passed.
  • A change to the CLI or agent-contract goldens needs a new behaviour-change row, or a <!-- public-surface: no-behavior-change: <reason> --> line in the changelog fragment.

For contributors

  • docs/STABILITY.md "How it is checked" says what the check enforces and what still rests on review: --json fields, the page format, wording, behaviour beyond names, the two-stable-releases rule, and the flags of the 124 CLI-only commands without a curated help record.
  • A branch converted with bun run release:unstamp gets rows without kind:; add one to each before pushing (docs/RELEASING.md).

The public-surface check no longer fails with "no merge base … even after deepening the clone" on shallow CI checkouts whose head is more than one commit past master.

Fixed

  • bun run check:public-surface deepened a depth-1 CI checkout with --filter=tree:0, which GitHub's server fails on for a shallow clone that is not partial. It now deepens with --filter=blob:none, and when a fetch fails it prints git's own error under the FAIL line.

The public-surface check no longer passes silently when it cannot read master's copy of a file it compares with.

Fixed

  • After a shallow CI checkout is deepened with --filter=blob:none, bun run check:public-surface fetches file contents at the merge base on demand. When that fetch failed, it read master's manifest as missing and skipped the removal check. It now tells a file that is missing on the merge base from one it could not read. Under CI=true it fails with git's own error, and locally it skips.
  • A file already on master whose name git quotes (one with non-ASCII characters) now counts as on master, so an old changelog.d waiver no longer covers a new golden change.

A page its owner purged stays purged after a restore, even when the purge waited behind a private derived page.

When an owner purges a page that a derived page depends on for its privacy, the purge waits until that derived page can be labelled private. Before this change, if an automatic purge removed the page first, or the next purge run failed or was deferred, the page was deleted without the purge records that tell a restore to delete it again, so restoring an older backup brought its text back. gbrain now remembers that the owner's purge is owed, by page id only, and whichever run finally removes the page writes those records, on repo brains and on brains without a repository. Restoring the page clears the reminder, so a page its owner brings back is never purged later by it.

A restore now also removes the same rows a live page purge removes. Replaying a purge from the purge mirror used to delete the page but leave its write requests, file rows, derived generation items and their effects, which still held the purged text. Replay and the live purge now share one sweep. If replay cannot finish a sweep (a store that does not exist on the brain aside), the restore stops with purge_replay_incomplete and names the purge it could not replay, instead of reporting success. A new doctor warning, purge_mirror, says when recent purges are missing from this machine's purge mirror (for example when gbrain runs in a home directory that is removed after each run), so a restore from an older backup would not replay them. The fix is to give every writer one durable GBRAIN_HOME. Thanks to @oakleaf for the replay and purge mirror findings (#9469).

Fixed

  • An owner purge that waited behind a private derived page writes its purge records whichever later run removes the page (an automatic expiry, or a run after a failed or deferred one), on every brain.
  • Restore-time purge replay removes write requests, file rows, derived generation items and their effects, the same rows the live purge removes, and fails with purge_replay_incomplete instead of reporting a partial sweep as done.

Added

  • Doctor check purge_mirror: a warning (never a failure) when purges recorded in the last 30 days have no line in this machine's purge mirror. It prints counts and the mirror's path, never a page name or text.

An upgrade from before v0.51 keeps a page in search while the index rebuilds, when its old index can't hold withdrawn, held or redacted text, and says how long the rebuild will take.

Fixed

  • Upgrading from a release before v0.51 used to take every page out of keyword and vector search until gbrain projections drain finished: about 9 hours on a 50,000-page brain behind an 18 ms database round trip. Now a Markdown or code page indexed at or after the safe-chunk fence keeps answering from its old index until its own rebuild, as long as it has no facts or takes fence, no private-key text and no quarantine marker. Pages with a facts or takes fence, pages with private-key text, quarantined pages, images and pages indexed before the safe-chunk fence still wait for the rebuild, so a brain made mostly of entity pages with a Facts fence gets little from this change. Access checks are unchanged, a page that changes after the upgrade leaves search until it is rebuilt, and no embedding calls are made. Reported by @jonahberg (#9192).
  • The upgrade's [153] line says how many pages keep answering, how many wait and why, and how long the rebuild takes at the round trip it measured to the database. gbrain doctor still counts every page waiting for the rebuild, and says results may be incomplete only while some page is out of search.

[0.60.731.0] - 2026-11-16

docs/STABILITY.md now says what stays the same on the stable channel, and how much warning you get before something changes.

Added

  • A "What stays the same on stable" section in docs/STABILITY.md. It covers command names and their documented flags, MCP tools and their parameters, documented config keys, the page and frontmatter format, exit codes and the top-level --json fields of documented commands, and upgrades from any stable release of the last six months. A removal or rename is announced in a stable release first, keeps working with a warning, and is removed no sooner than 14 days and two stable releases later. Security and data-loss fixes are the one exception. The section also lists what the promise does not cover, including very large brains during an upgrade until #9192 is fixed.
  • deprecations.d/README.md, which says how a planned removal is recorded and announced before it happens.
  • The README's Stability section points to the new promise.

Stable now waits for a green upgrade test from every recent stable release.

Changed

  • The daily stable check promotes a release only after the upgrade matrix ran green on that release's exact tag and tree. With no run yet it starts one; a red run holds stable with a warning and is not started again by the schedule.
  • The upgrade matrix's full run also upgrades from the stable releases users may still be on, which are the last 8 stable releases plus the newest one of each older month, back 6 months. A test run from before this change no longer counts.
  • While the upgrade matrix's own break issue is open, stable holds. Only a security or data-loss fix that fails on no from-version the issue doesn't name may pass it.
  • promote-stable.yml takes an optional matrix_run_url, checks it the same way and records it as matrix_run in stable.json and the stable tag. A maintainer can still promote without one, recorded as matrix_run: null.

Doctor, init, call import and gbrain serve report and record the right thing in seven small cases.

Fixed

  • gbrain doctor's grant check is ok when the only operations a grant misses are ones the owner adds by name, and its message names them with the command that adds them.
  • gbrain import --calls finds a call already imported from any folder above the one you import, and reports a call whose page is still being written as pending instead of writing a second copy; with --accept-pending it still exits 10 when that write waits on a stale owner host.
  • A folder whose .git is an empty file is no longer taken as a git checkout.
  • Release notes drop a trailing internal issue code, like (ABC-12), from each line of the change list.
  • gbrain init --force --embedding-model <model> records the new model and width in the database as well as the config file, so the two agree (#9451).
  • gbrain serve names upgrade migrations that haven't run, once at startup on stderr, and gbrain doctor reports them too, also on a brain an older release last opened whose local upgrade ledger is empty.

For contributors

  • The key-files page google-and-loops.md was at its 60,000-byte cap, so its harness and MCP entries moved unchanged to google-and-loops-more.md; the Google source entry now says what an unmanaged --dry-run returns.

Reported by @Loprz (#9451).

Stable moves within a day only for security, data-loss and stable-regression fixes; everything else waits for the weekly move.

Changed

  • An important fix, the kind that moves the stable channel early, is now only a security fix, a data-loss fix or a fix for a regression in a release that is on stable. A fix that unblocks one setup, or a regression only alpha releases have, rides the weekly move unless a maintainer promotes by hand. gbrain upgrade --help, docs/STABILITY.md, docs/RELEASING.md ("Daily stable check") and changelog.d/README.md say so.
  • scripts/build-behavior-changes.ts (in check:regen-all) refuses a changelog fragment whose important: class is anything but security, regression or data-loss, naming the three and docs/RELEASING.md#daily-stable-check, the same way it refuses a bump: other than patch. A # comment after a front-matter value is no longer read as part of it.
  • Tests: test/scripts/build-behavior-changes.test.ts covers the refused and accepted important: values.

[0.60.730.0] - 2026-11-16

A dry run of a Google or GitHub source no longer writes pages (#9449).

Fixed

  • gbrain sync --source <id> --dry-run on a Google (Gmail, Calendar, Contacts) or GitHub source of a brain without managed persistence fetched from the provider and wrote what it fetched: pages, chunks and soft-deletes. It now fetches and writes nothing, prints that the source has no preview, and returns the dry_run result, as Slack and meetings sources already did. A brain with managed persistence still refuses a connector dry run before any provider call (#9449).
  • Tests: test/sync-connector-dry-run-9449.test.ts runs a dry run of every connector kind on PGLite and Postgres and checks that nothing was fetched or written.

Reported by @kostjakla-del-ok. First fixed by @Masashi-Ono0611.

Supersedes #9450

[0.60.729.0] - 2026-11-15

Fifteen small fixes: the doctor's grant repair grants every operation it counted except the ones you approve by name, call import handles single files, look-alike names and pending writes, a folder whose own .git is not a repository is no longer taken as a checkout, storage adopt reports what it did, and doctor's type check reads through an index.

Fixed

  • gbrain doctor (grant_new_ops_available) now prints a repair that grants every operation it counted as missing from a token's or OAuth client's grant, not only put_pages, and its preview is the same command with --dry-run, which shows each changed grant line as before and after. Membership operations (join_brain, sync_brain_skills, leave_brain), which you approve by name, are left out of that repair and listed apart with their own command. The apply still needs your agreement.
  • gbrain import <file> --calls imports that one call file instead of reporting success with nothing imported; a file that is not .txt or .md is reported as an error.
  • Call import keeps two files apart when a redaction pattern makes their names look the same, so the second file is imported instead of being treated as the first. Pages imported before this release are still found on a rerun.
  • Call import reports a write that was accepted but is still being saved as pending, with the command that reads its receipt, and exits 10 instead of counting it as a failure.
  • A folder whose own .git is not a repository (an empty .git folder, or a .git file pointing at a missing directory) is no longer taken as its own checkout, and when that folder sits untracked inside a real repository, gbrain no longer writes its managed marker into that repository, which made writes elsewhere in it fail with writer_coordinator_required.
  • gbrain doctor (type_proliferation) reads the distinct source and type pairs through a new index on pages (schema migration 316) instead of scanning every page, which took over 20 seconds on a large brain.
  • gbrain storage adopt --apply on a hosted shelf now indexes the cards it committed, so the shelf's sync checkpoint no longer lags until a manual sync; when it cannot, it prints the sync command to run.
  • gbrain storage adopt --apply --json counts adopted, skipped, refused and pending records and gives each write its receipt id.
  • gbrain storage adopt reports a record cut off by its 64 KiB read as truncated, and says to move the ## Record section to the top, instead of reporting the record as missing.
  • Atom extraction skips the storage shelf, and doctor no longer names the shelf as a source with held atoms.
  • gbrain storage-find reports a card asked for by its exact title as an exact hit, not as a closest match.
  • An empty storage location path now says to write / for an item at the root of its bucket or repository.
  • Sync no longer tells you to restart a gbrain serve that already runs your gbrain version.
  • Owner sessions whose first message starts with a Claude Code slash command (such as /clear or /model) are no longer classed as launched by an agent (#9390).

Changed

  • The graduation tests bound and retry their git calls to the remote, so a hung fetch fails fast with a clear error.

[0.60.728.0] - 2026-11-15

Five small fixes: doctor sees a project-level Codex registration, a stray .git no longer claims a checkout, the host's CLI can read a write receipt another principal owns, keyless search keeps working when its fill step fails, and the conversation archive's folder-cap notice stays accurate.

Fixed

  • gbrain doctor (harness_wiring) now also reads the workspace's .codex/config.toml, so a gbrain MCP server registered there for one project counts as registered instead of being reported as unwired (#9398). Reported by @Masashi-Ono0611.
  • An enclosing .git that is not a repository (for example an empty ~/.git folder another tool left behind) no longer counts as the checkout that owns a brain folder under it, so gbrain stops writing its marker there and stops refusing writes under that folder with writer_coordinator_required (#9430). Reported by @garrietv and first fixed by @Masashi-Ono0611.
  • gbrain write-request <id> on the brain host now shows the receipt of a write that ran under another principal, such as a fact extraction submitted by an MCP client, which gbrain sources writer failures names as each row's next step. Remote callers and request ids two principals share still answer not found (#9426). Reported by @cupit01.
  • Keyless search no longer counts a fill row the model never answered as judged, leaves such rows out of the fill, and returns the plain keyword results when the any-word fill step fails instead of failing the whole search.
  • The conversation archive's archive_rescope_folder_cap notice shows only while the conversations/ rule still keeps the folder to the owner, and on a source already at the 256-rule limit it says that gbrain access rule export then import fails until another rule is removed.

Changed

  • The refresh-barrier TLA+ nightly now runs in its own job and checks the adoption-plus-foreign-commit case one write at a time; the two-write run is kept as RefreshBarrierAdoptFull.cfg and runs by hand.

New page, docs/STABILITY.md: which release the default install follows, how a release becomes stable, and how to pin a version or go back.

Added

  • docs/STABILITY.md explains that the README install command follows the stable channel, what a release has to pass before it becomes stable, when stable moves, how to stay on stable, pin an exact release or return to an earlier one, where behaviour changes are announced, and what the faster alpha channel is for. Nothing in gbrain changes; the page describes what it already does.
  • The README has a short Stability section that links to the page, and llms.txt / llms-full.txt list it.

Published release notes now open with one line per change and the pull request that made it.

Changed

  • The release workflow puts a list at the top of each GitHub release: one line per change the release folds, taken from the change's changelog entry, with its pull request number. A release that never published is listed in the next one. CHANGELOG.md is unchanged.
  • bun run verify fails when a test reads a changelog.d file from the repository, because the next release deletes that file and the test breaks.
  • changelog.d/README.md says every pull request needs a changelog entry, a tests-only one included.

gbrain setup --transcripts DIR imports your exported call transcripts into the company brain with credentials removed, me/facts and company/facts become quotable facts pages, and setup says what the conversation archive is doing (GST-2731).

Added

  • gbrain setup claude-code|codex --shape company|both --transcripts DIR (also asked at a terminal, after the docs folder) imports .txt and .md call or meeting transcripts from any recorder into the company source, one page per file at company/calls/<date>-<title>. Every credential in the text (API keys, tokens, passwords, plus your own redaction patterns) is replaced by a <REDACTED:…> marker before anything is stored, with the same redaction the conversation archive uses, and the file's path is redacted the same way. Pages are keyword-only: embedding them waits for your approval, with the conversation archive's hold. The facts backstop skips them; with a chat model key set, scheduled atom and event extraction can still read them within their budgets, and setup prints a what leaves this machine line saying which is on, or that nothing leaves. Each file is imported once. A rerun reports it unchanged, and a file changed since keeps its first page. Above 500 files setup prints the import command instead. --transcripts needs the company part on this brain, so it is refused with --team.
  • Setup prints who can read the call pages. That is everyone who can read the company source on a brain that serves one person, or the members the source's access rules allow on a brain that serves several. The pages carry no readers of their own.
  • gbrain import DIR --calls [--prefix company/calls] [--source company], the same import on its own (implies --no-embed).
  • Page type facts (base packs; gbrain-base-v2 is now 1.7.0). me/facts and company/facts are created with type: facts, and the instructions setup writes for the agent say to quote only a facts page's dated lines and to add a new fact as a new dated line.
  • Doctor check facts_undated_lines: warns about a facts page with a fact line that has no date, naming the page and the line numbers (never the text), and asks you for the dates. It changes nothing.
  • Setup's report ends with one line on the conversation archive: on, off, or offered (a brain that may serve several people is only offered it).
  • Docs: docs/guides/brain-layout.md.
  • Tests: test/import-calls.test.ts, test/doctor-facts-undated-lines.test.ts, and the transcripts cases in test/setup-brain-layout.test.ts.

[0.60.727.0] - 2026-11-15

  • v0.60.726.0 was not published, so this is the first release you can install with the conversation archive; it carries everything in the 0.60.726.0 section of the changelog, plus one test fix.
  • The conversation archive is now on by default for personal brains; turn it off with gbrain transcripts archive off (#9433)
  • New gbrain transcripts archive saves each day of your agent conversations, OpenClaw included, as one redacted, searchable page (#9432)
  • Archive pages rank a little below your own notes, and gbrain doctor warns when the archive falls behind (#9433)
  • Tests only: the archive upgrade-notice test keeps passing after a release (#9437, #9441)

The upgrade-notice test for the conversation archive keeps passing after a release folds its changelog entry (#9437).

Fixed

  • test/upgrade-transcript-archive-notice.test.ts no longer reads the changelog.d fragment that every release deletes, so the full test suite passes on a release commit again. Tests only; no behavior change.

[0.60.726.0] - 2026-11-15

  • The conversation archive is now on by default for personal brains; turn it off with gbrain transcripts archive off (#9433)
  • New gbrain transcripts archive saves each day of your agent conversations, OpenClaw included, as one redacted, searchable page (#9432)
  • Archive pages rank a little below your own notes, and gbrain doctor warns when the archive falls behind (#9433)

New, off by default: gbrain transcripts archive writes one redacted page per conversation per local day from your agent harnesses, plus a daily index (GST-2768).

Added

  • gbrain transcripts archive run | backfill --from <day> | status | adopt-legacy --dir <folder> | on | off. Off until gbrain transcripts archive on; nothing is scheduled yet. Pages go to conversations/<harness>/<day>/<label>-<hash8> and conversations/days/<day> in the conversations source, redacted before any write; reruns change nothing; the first run pins the timezone so days never re-slice. Backfilled pages are keyword-searchable at once and their embedding waits for gbrain embed --held. On a shared brain every page names its readers (the owner, plus an existing chat-<hash8> group for a verified group chat); with no owner named, days are held with archive_owner_unset. Works on repo-less brains (database is the record), personal repo brains (files committed with gbrain git commit-push, then synced) and managed brains (the write coordinator).

  • OpenClaw: live sessions are read through the Gateway (openclaw gateway call, with OpenClaw's own config or OPENCLAW_GATEWAY_URL / OPENCLAW_GATEWAY_TOKEN, or GBRAIN_OPENCLAW_GATEWAY_URL / GBRAIN_OPENCLAW_GATEWAY_TOKEN to override them), checked against OpenClaw 2026.10.9 and 2026.11.3. The archive also reads OpenClaw's retained session files (.jsonl.reset.<stamp>, zstd .jsonl.deleted.<stamp>.<hex>.zst and .jsonl.bak.<stamp>) and merges them with the Gateway's copy of the same session, so a backfill works with the Gateway down and messages from before a reset are kept; migration and Doctor copies are skipped, and gbrain transcripts ingest still reads live files only. A day page is rewritten only when it keeps every message it already held, and backfill re-redacts a finished day when the redaction rules changed since it was written. The openclaw child gets only PATH, HOME, locale and the resolved Gateway URL and token, and a working-folder .env can no longer set any OPENCLAW_* or gbrain Gateway variable. Backfill reads large session folders in a bounded way and resumes where the last run stopped.

  • New error codes: archive_gateway_unreachable, archive_gateway_auth, archive_gateway_failed, archive_owner_unset, archive_redaction_failed, archive_run_busy (one run per source at a time), archive_audience_unproven (the brain's audience could not be checked, so nothing was written).

  • When a brain becomes shared, the next archive run first keeps every earlier archive page to the owner, then records the new audience. A failed day index write is retried on the next run.

  • Docs: docs/guides/conversation-archive.md.

  • Tests: test/transcripts-archive-*.test.ts (PGLite, including a real git remote and child CLI commit-push), test/transcripts-openclaw-gateway.test.ts, test/transcripts-openclaw-archived-files.test.ts, test/transcripts-openclaw-files-only.test.ts.

Known limits

  • Redaction works on what each day's own text marks as a secret. A value that a later day labels as a credential stays visible in the earlier day's page.
  • Pages written while the brain was personal are closed on the next archive run after it becomes shared, so they stay readable to other members for up to one maintenance interval.

The conversation archive is now on by default for personal brains, runs from the maintenance pass, ranks a little below your own pages, and has a doctor check (GST-2768).

Changed

  • transcripts.archive.enabled unset now resolves on when the brain is personal (this default flip is approved by Garry's 2026-11-15 "do it" on GST-2768) (brain.audience personal, or the audience heuristic reads it as personal) and off on a shared or unknown brain, a thin client and a remote caller; an explicit on or off always wins, and a config read failure resolves off. A shared brain is only offered the archive. Existing personal brains see a one-time upgrade notice with the off command, gbrain transcripts archive off.
  • Search and query: an archive page ranks a little below a written page that scores close to it (rank prior 0.8), and one result list keeps at most two archive hits per day, unless the query asks about a conversation or a date. search.rank_priors false turns both off.
  • gbrain init and gbrain setup <harness> state the archive in their first-run settings (on, or offered on a shared brain), what leaves the machine, and how to turn it off.

Added

  • The maintenance pass runs the archive: unattended upkeep (a serve's idle tick, gbrain maintain --unattended and its schedule) and the autopilot tick archive the forward days, then one capped backfill run (transcripts.archive.backfill_days_per_run) of older days until none are left, at most once per transcripts.archive.maintenance_floor_min (default 360). It never runs when the archive resolves off, never fails the pass, and records its outcome in transcripts.archive.last_maintenance. One pass is bounded by count and by ARCHIVE_MAINTENANCE_BUDGET_MS (120 s, coordinator waits included). A busy run (archive_run_busy) or an unproven audience (archive_audience_unproven) is a recorded skip the next pass retries. On a managed brain the pass writes through the write coordinator as put_page requests under the local CLI identity (local_cli); a write accepted but not yet published counts as accepted and is not written again; called inside a coordinated write it skips with archive_inside_publication. Backfilled pages stay keyword-only until gbrain embed --held; no facts extraction or Dream synthesis runs on archive pages.
  • When a brain archived while personal becomes shared, the owner-only re-scope now adds one rule per folder that holds an archive page (replacing any rule already on those folders, a wider one included) instead of one conversations/ rule, so a group conversation archived afterwards keeps its group. A personal history longer than the per-source rule limit (256 folders) is kept to the owner by one rule on conversations/ instead (archive_rescope_folder_cap, reported by status and doctor), so group conversations archived afterwards stay owner-only too; widening that rule also opens the earlier days to its readers. When the re-scope waits (no owner: archive_rescope_owner_unset; rules kept in git: archive_rescope_git_rules), gbrain transcripts archive status and doctor say so in one line with the next step.
  • Embedding planning reads a page's archive hold marker and budget row in one statement instead of two.
  • Doctor check transcript_archive_freshness: warns when a harness's newest archived day is more than transcripts.archive.stale_days (default 2) behind its newest session day (fix: gbrain transcripts archive run, or backfill --from <day> for a wider gap), when no maintenance pass has run the archive in that long (fix: gbrain autopilot --install), and lists held days with their codes. Off is ok; a harness with no readable sessions is informational.
  • Docs: docs/guides/conversation-archive.md (default-on behaviour, what leaves the machine, the off switch, backfill and the held embedding, the doctor check, retiring a home-made pipeline).
  • Tests: test/transcripts-archive-default.test.ts, test/transcripts-archive-maintenance.test.ts, test/search-archive-rank-prior.test.ts, test/doctor-transcript-archive-freshness.test.ts, test/upgrade-transcript-archive-notice.test.ts, test/transcripts-archive-maintenance-managed.test.ts, test/transcripts-archive-run-on-demand.test.ts, test/transcripts-archive-rescope-skipped.test.ts, test/page-embedding-budget-statements.test.ts, and the archive line in test/init-first-run.test.ts.

[0.60.725.0] - 2026-11-15

Doctor's Git checks now finish on large brains, and a check that could not see everything says health unknown instead of ok.

Changed

  • db_ahead_of_git and strict_facts_outside_git read the 800 most recently changed pages of each source instead of every page, and their messages say how far back they looked. gbrain sources git-gap <source-id> still reads every page.
  • When that census runs out of its time budget or cannot run, db_ahead_of_git, strict_facts_outside_git and git_index_lag now report a warning with health unknown instead of ok.
  • Doctor runs the checks that catch lost writes before its full scans.

Fixed

  • strict_facts_outside_git no longer says every fact has a Git home when a fenced fact sits on a page older than that window. It reports health unknown, and gbrain sources git-gap <source-id> counts the whole source.

A batch of small fixes: folder attach retries, dropped webhook events, cut-off stdio replies, and clearer help and doctor text.

Fixed

  • Retrying a managed folder attach with the same --request-id replays the first attempt's default-search choice instead of failing with idempotency_conflict.
  • Two different GitHub events on one item within 30 seconds are both processed; a redelivery of the same event is still deduplicated.
  • Local serve and the thin-client bridge deliver a large stdio reply in full when stdin closes right after it.
  • Atoms stamped by repair visibility can be adopted by a later generation again, without widening their visibility (#9386).
  • A migration that keeps failing gets the migrations_pending notice with the apply-migrations --force-retry <version> step.
  • Doctor's derived_visibility check counts only active sources, so it no longer suggests a repair that finds nothing.
  • repair purged-page-facts: a preview that approves nothing prints an apply command that reports nothing to do.
  • A DNS failure on a Supabase pooler URL points at this machine's DNS and network instead of a paused project (#8531).
  • sources status --json prints each held file's fix with the --brain pin, so it runs as written (#9319).
  • A hosted install asking setup for the skill resolver is told the command that works without a local brain (#9245).

Changed

  • gbrain setup codex and --remove keep a .bak.<timestamp> copy of each config.toml or hooks.json they change and print its path.
  • gbrain setup codex prices the search-mode cost matrix with OpenAI models.
  • embed --help names the --stale --facts --source <id> form for facts saved with remember; upgrade --help no longer mentions a postinstall step; doctor's consent wording says gbrain makes no provider request on its own.
  • The enrich skill says to leave out empty sections instead of writing [No data yet] (#7061).

Keyless search finds pages that hold most of the question's words. On a brain with no embedding provider, a question used to miss any page that lacked even one of its words whenever another page held all of them: "SSO Starter plan" found the pricing and decisions pages and never the call notes that mention SSO and Starter. Now, when the first page of results comes back shorter than the limit, search asks the keyword index once more for pages that hold only some of the words and adds them after the rows it already had. Every row search returned before comes back exactly as it did: same pages, chunks, order and scores. The added rows go through the same stages as any other row (the caller's read policy, redirects, superseded-page warnings, the evidence and injection checks when they are on, the alias and facts steps of search and query, evidence delivery), inside the same request with the same decision time budget. They never repeat a page, never take the page past its limit, score below the rows above them and carry keyword_fill: true (also in lean remote rows); --explain-target says the page matched only some of the words. If adding them would change any other row, or a check that is on was skipped for them, they are left out and the result is the one search returns today; a row from a source on decide.egress.deny_sources is treated like any other row from that source, which today means it is returned and counted as judged (GST-2786); a failed or timed-out lookup is reported like any other incomplete keyword arm. Only the search and query operations (the CLI commands and MCP tools of those names) add these rows, because they are where the result is checked against the one without them; features that call search internally (think, recall, enrichment, whoknows and the rest) get the rows they got before. Later pages (an offset above zero), questions in Chinese, Japanese or Korean (their keyword match already works term by term) and brains with embeddings are unchanged. There is no new setting.

On LongMemEval S run keyless through the query operation (470 scored questions, top 10), recall_all@10 rose from 80.64% to 91.06% (+10.43 points, 95% interval +7.87 to +13.19; 49 questions better, none worse), recall_all@5 from 72.13% to 81.06% and MRR from 0.8726 to 0.8770; none of the 451 questions with a hit ranked it lower, and the rows returned without the added pages were the same in all 500 questions. The search function gives the same numbers when asked for the added rows; with the clock frozen its rows without them were identical in every field in all 500 questions, and when not asked it added nothing. The extra lookup ran on 103 of 500 questions and added a median 37.8 ms (p95 55.3 ms) to the search on PGLite; through the query operation those questions took a median 125 ms longer (p95 210 ms), because the operation's later steps run twice to prove the other rows did not change. On Postgres a cold search whose first page is short costs 65 database round trips, the same as a search whose all-words match is empty (master: 55). With the evidence or injection check on, the added rows cost one more provider request.

Itemized changes

  • src/core/search/hybrid.ts, src/core/search/hybrid/keyword-only.ts: the search function adds the rows only for a caller that passes anyWordFill, which the search and query operations do; a Chinese, Japanese or Korean question skips the extra lookup. The keyless path's row stages are one function, called once for the rows search returns today and, when that page is short, once more for the any-word candidates with the first page held as context, so duplicates, collapsed copies and redirect targets of a returned page are dropped. The added rows are scored below the page's lowest row, checked, and appended up to the limit.
  • src/core/search/decide-stage.ts: gateFillRows runs the evidence and injection checks on the added rows with what is left of the request's one time budget and adds their counts to the request's one record; if the first page's check failed open, or the added rows cannot be judged, none are added and the result says so.
  • src/core/ops/search.ts: search and query run their own steps on the rows without the added pages, then again with them, and return the second result only when every other row, the alias record, the confidence grade and the delivered text are the same. A page the alias step already returns is not added a second time.
  • src/core/search/keyword-identifier-probe.ts, src/core/pglite-engine.ts, src/core/postgres-engine.ts, src/core/types.ts: the keyword arm takes orFill, which runs only the any-word statement (the all-words rows are already in hand) and tags its rows keyword_relaxed and keyword_fill; without it the arm behaves as before. On Postgres the extra lookup remembers its own statement timeouts apart from an ordinary search's, so a timed-out extra lookup never makes the next search skip its keyword statement.
  • src/core/search/lean-rows.ts, src/core/search/explain-target.ts: lean rows keep keyword_fill; the explain trace names the added arm and says the page matched only some words.
  • Tests: test/search/keyless-any-word-fill.test.ts (PGLite and Postgres) with property tests at the search function and at the search and query operations that the rows without the added pages come back unchanged; test/search/keyless-any-word-fill-gates.serial.test.ts for the checks, the shared time budget and the reported failures; test/keyword-timeout-negative-cache-6495.test.ts for the separate timeout memory. The C5 and evidence off-path goldens gain the appended rows and their counts only (the think prompt golden is unchanged), and the export-surface golden gains the new fields.

Two sync tests are easier to trust: one no longer fails on a busy runner, and the other says why a page failed when it fails. Neither changes what it checks.

Tests

  • test/sync-pass-reader-gaps-6712.test.ts: the delegated-deadline test slept on the real clock and then checked the deadline 50 ms before it expired, so a runner that woke a timer 50 ms late failed it (seen once on CI, 5 of 5 passes locally). It now drives setTimeout and performance.now with Bun's fake timers, with the same 600 ms deadline, 400 ms window and 450, 350 and 500 ms steps. A probe that stalls the event loop for 60 ms after the first slice fails the old test and passes the new one. A probe that stops a committing slice from extending the deadline still fails the new one.
  • test/e2e/persistence-preparation-lock-wait-postgres.test.ts: when the drain doesn't end synced, the test prints every page that didn't commit. For a page that failed with storage_error, that printout carried only the generic public message, so a CI failure (one page, storage_error, while the test held its lock) couldn't be traced: it passed 44 of 44 times locally, 24 of them with four copies running at once. The printout now includes each page's recorded failure detail (the error class, the database error text, the statement and the source frame), so the next failure names its cause.

Codex starts the gbrain MCP server and hooks even when bun is not on its PATH.

Fixed

  • After gbrain setup codex, a Codex session started without bun on its PATH (the desktop app, a launcher, a terminal opened before the shell profile changed) failed with /usr/bin/env: 'bun': No such file or directory and showed no gbrain tools, with nothing telling you why. gbrain's hook launcher, which every MCP server entry and hook setup writes runs through, now records the Bun that installed gbrain, so those entries start from any PATH. The same gap closes for Claude Code, Hermes and OpenClaw entries that run through the launcher. The compiled binary needs no Bun and is unchanged. (GST-2784)
  • gbrain doctor reports a launcher that records no Bun, or one whose Bun is gone, instead of calling the wiring healthy.

To take this fix

  • gbrain upgrade (its post-upgrade step rewrites the default install's launcher), or rerun gbrain setup codex. Neither changes config.toml, hooks.json or Codex's trust entries. A launcher under another gbrain home is rewritten by rerunning the setup that wrote it; gbrain doctor names it.

A large source whose sync keeps stopping at its run limit now finishes catching up instead of being reported as abandoned.

Fixed

  • A sync pass that saved at least one page or hold before its run limit stopped it now counts as progress, so gbrain sync status and gbrain doctor no longer report pass_abandoned for a source that is still moving. A pass that saved nothing still counts toward the abandoned streak, and the open pass row shows how many entries it has settled so far.
  • On a read-only mirror, a page you edited in gbrain whose file is gone upstream is kept as a database-only page. Sync results and status report these as carried instead of listing them as holds to fix, and no command advises --prefer database on a mirror, since gbrain never writes a mirror's files.
  • A symlink whose target does not exist is skipped and counted (skipped.symlink_dangling) instead of held as an unreadable file, and an older hold for one is released. A dangling link at a path that still has a live page stays held, and so does a committed file whose working copy alone was replaced by a dangling link.
  • The commit notifier no longer starts a sync for a source with sync turned off (config.syncEnabled=false). Turning sync back on makes the notifier watch it again.

gbrain setup claude-code --shape company (or both, and the same for codex) gives a brain a company part: a company source in a folder gbrain owns, six empty company pages, your docs folder imported read only, and lines in the agent's instructions that say which pages to read. The pages are company/facts, company/decisions, company/pricing, company/customers, company/calls and company/playbooks. Each is an empty form marked status: stub, so nothing about the company is invented, and a re-run never overwrites a page you have edited. --docs DIR imports that folder's markdown once without writing into it; above 500 files setup prints the import command instead of running it. On a brain that serves one person the company pages join default search. On a brain more than one person reads they stay out of it, and setup asks you before adding them. --shape both sets up the personal and company parts together. --team creates nothing for the company part and explains where a shared company brain lives, either in a hosted workspace or on a self-hosted brain that serves several people. If the company folder already holds files, setup leaves it alone, keeps the company lines out of the agent's instructions and asks you to move the files before you run it again.

Itemized changes

  • src/commands/setup-brain-layout.ts: the company and both shapes, the company starter pages, the company source and its read-only docs import, company lines in the managed gbrain:brain-layout block (now version 2), the team route, and new options in the brain_shape question. Personal pages are written before the company source is added, so they stay in your default source. A company folder that already exists and is not an empty folder (a file included) is refused instead of adopted. The re-run command setup hands back keeps --team, and after a refused company folder it asks you first instead of re-running by itself. Both reply formats of sources add --json are read when checking whether the company source joined default search.
  • src/commands/setup.ts: --shape company|both, --docs DIR and --team for claude-code and codex.
  • Docs: docs/guides/brain-layout.md covers the company part and a team company brain.
  • Tests: test/setup-brain-layout.test.ts.

Imported conversations no longer keep three kinds of secret that the last cookie and .netrc fix still let through. gbrain transcripts ingest now also redacts a .netrc password when the entry is written over several lines (machine, login and password each on its own line), a curl cookie passed with -b on its own line of a command split with trailing backslashes (what Chrome's "Copy as cURL" produces), and a quoted cookie value inside escaped JSON (sid=\"…\"). Before this change all three were copied into the searchable page as plain text. Cookie names, curl's other arguments and ordinary prose near these shapes stay as they were.

What to do: conversations imported earlier keep what they stored. gbrain doctor asks for a fresh audit; run gbrain transcripts audit-secrets --json to list affected pages (slugs and hit counts, never values). The sources push gate and the compiled-context scan are unchanged.

Still not caught: a curl cookie written with bash's $'…' quoting (-b $'sid=…', which Chrome's "Copy as cURL (bash)" uses for a cookie holding a quote or non-ASCII text); a .netrc entry or curl command split across two messages; and a cookie header whose opening and closing escaped quotes differ in depth (sid=\\\"…\"), where that value and the pairs after it stay. In the other direction, a line that is only default or machine <word> followed by a line that is only password <word> has that word redacted even in prose.

Itemized changes

  • src/core/secret-scan-labeled.ts: a password line inside a multi-line .netrc stanza (opened by a machine or default line, kept open by login, account and password lines, blank lines and # comments) is redacted. The line must be only password <value>, so prose such as "machine learning" followed by "password managers are useful" stays.
  • src/core/secret-scan-cookie.ts: a -b on a line that continues a curl command (each earlier line ends in \) counts as curl's cookie argument; a cookie value quoted with escaped quotes (\", or \\\" one level deeper) is read. A -b after a continued line that is not curl (git checkout -b) stays.
  • src/core/secret-scan.ts: the transcript-lane scan carries the .netrc stanza and the curl continuation from line to line.
  • src/core/transcripts/secret-audit.ts: TRANSCRIPT_SECRET_AUDIT_VERSION 6 → 7, because the audit now finds these shapes on stored pages.
  • Tests: test/transcripts-secret-cookie-header.test.ts, test/transcripts-secret-corpus.test.ts.
  • Docs: docs/guides/data-ingestion.md#credential-redaction and the key-files entries.

[0.60.724.0] - 2026-11-15

  • sources_resolve_path no longer shows that a committed file exists, or its frontmatter slug, to a member who could not read the page that file would become. Before, such a file was hidden only once it had a page. The brain owner's own connections are unchanged.
  • A page the owner asked to purge, whose purge was held, now writes its purge tombstones when a later consented gbrain pages purge-deleted removes it, so the same content is not imported again.
  • The covering commands that gbrain access rule prints for open twins now keep --allow-without-owner when the rule leaves the owner out, so they are accepted as printed.

gbrain doctor finds the facts that purges left behind before #8859 was fixed, and gbrain repair purged-page-facts retires them after you approve the preview.

Fixed

  • Facts left active by older purges can be found and retired (#8859 follow-up). Before the #8859 fix, purging a page left the facts extracted from it and filed on no page active, so recall kept returning claims from a removed page. The new doctor check purged_page_facts counts them, and gbrain repair purged-page-facts previews each one with its class and expires exactly the previewed rows on --apply --expect <hash> --yes. Rows of a page the purge ledger records are retired by default; rows of a page gone with no purge record need --include-ambiguous. Nothing changes on upgrade.
  • Purges match those facts through an index. A new index on the head of a fact's context (migration v315) keeps a purge from scanning every fact in its source. The doctor check and the repair still read every page-extraction fact filed on no page.
  • The repair expires rows but writes no purge record, so the same claim could be saved again later.

Itemized changes

  • src/core/repair/purged-page-facts.ts, src/commands/doctor/checks/purged-page-facts.ts: the repair kind (explicit-only, preview-bound, destructive consent) and the read-only check.
  • src/core/facts/gone-page-facts.ts: the context match uses contextHeadSql, the expression v315 indexes.
  • src/core/schema-migrations/v315-facts-context-head-index.ts: idx_facts_context_head, built concurrently on Postgres.
  • Docs: docs/guides/repair.md#purged-page-facts.
  • Tests: new test/repair-purged-page-facts-8859.test.ts (managed PGLite and Postgres, unmanaged PGLite).

GitHub's "Latest" release is now the stable release, with notes covering everything since the previous stable; every batch release is published as a pre-release. Installs and gbrain upgrade already followed the stable tag, so nothing changes for anyone on the documented install. The Releases page and its Latest badge now show that same release, about once a day, instead of every batch release. Anyone who downloads from releases/latest by hand now gets the stable release. People watching the repository for releases are still notified of pre-releases, because GitHub has no setting to skip them.

Changed

  • .github/workflows/release.yml publishes each release with prerelease: true and make_latest: false.
  • scripts/promote-stable.ts ends every promotion by marking the release Latest. A forward promotion also replaces its notes with every CHANGELOG entry since the previous stable; the oldest entries drop, with a link, past GitHub's size limit. Latest never moves back to an older release on a forward promotion, so until stable passes the newest batch release published before this change, Latest stays where it is.
  • A hand re-run of the release workflow that would republish the current stable release as a pre-release refuses.
  • Alpha binary self-update reads VERSION at the latest-stable tag and downloads that exact release instead of releases/latest.
  • The upgrade matrix counts pre-releases as releases.
  • The macOS 26 signature check takes its tag from latest-stable.

Added

  • scripts/release-channel.ts has three commands. backfill brings releases published before this change in line. undo takes the rule back out. preview-promotion shows what a promotion would write. backfill and undo only print their plan unless given --apply.
  • Docs: docs/RELEASING.md, "Which release GitHub calls Latest".

gbrain git commit-push now commits pages named through a symlinked source path, says why for each path it leaves out, gbrain embed stops reading a source id as a page, and sources writer git-durability --enable names the exact command to commit the AGENTS.md it writes.

Who is affected: anyone with a scheduled writer (a cron job, hook or agent that runs gbrain git commit-push) whose source path is a link into a Git repository, or whose source folder contains a .git entry of its own: a leftover folder or file, or a link to the repository's .git. Those writes were refused or reported as skipped, so the pages stayed on disk and never reached Git or the remote. After upgrading, run the writer once and check git status in the repository: anything it left uncommitted before is committed on that run.

Fixed

  • A source whose path is a symlink to a folder inside its Git repository (for example a source at ~/source-a pointing to repo-a/brain) can be committed with paths spelled through the link: gbrain git commit-push --path ~/source-a -- ~/source-a/notes/daily/2026 commits the new and modified files in that folder. Before, a path spelled through the link was refused as outside the checkout, or failed with "beyond a symbolic link" when the link sat inside the repository, so scheduled writes stayed out of Git. A named file that is itself a symlink is still committed as the link.
  • A source folder whose .git is a link to its repository's own .git (for example repo-a/brain/.git -> ../.git) is committed at its real place in the repository. Before, a commit-push run with --path pointing at that folder wrote the pages into a copy of the folder at the top of the repository, pushed that, and reported success, while the real files stayed uncommitted; run from the repository root, the same folder was reported as a nested repository and skipped. If a source of yours has such a link, check the repository for a stray top-level copy of the folder after upgrading.
  • An empty or broken .git folder, or a .git file pointing at a repository that no longer exists, left under a source path no longer makes commit-push report commit_paths_skipped or "not inside a git checkout" for a folder its repository already tracks. Only a folder that Git treats as its own repository counts as nested, and that one is still refused with its reason.
  • A folder whose tracked files are staged for removal (as after git rm --cached) while the files are still on disk is still refused, and nothing is added back or removed, but the refusal now names that cause as staged_for_removal and both ways out: gbrain git commit-push --untrack to stop tracking the files, or git restore --staged to keep them tracked. It used to be refused as files "already staged by another writer". The "staged for removal" explanation no longer appears for a folder that was skipped for any other reason; before, every skipped folder got it, which pointed people at the wrong fix.
  • gbrain git commit-push --dry-run now runs the same skipped-path check as the real run and fails with the same paths and reasons, so a dry run no longer reports success for a run that will fail.
  • Each path commit-push leaves out now says why, in the message and in the JSON detail (skipped[].why, next to reason), and the push log gets one [push] SKIPPED, NEEDS ATTENTION line naming every skipped path and its reason. A wrapper that keeps only the error code still leaves the reason in the log.
  • gbrain embed --status --source <id> no longer fails with "Page not found: ". Embed has no --status flag, so a single-page run now refuses any flag it doesn't take with unknown_flag (exit 2) before it asks for spend approval, and names gbrain embed --stale --dry-run --json as the free way to read embedding coverage. The page slug is now the first argument that isn't a flag or a flag's value, so --source, --batch-size, --priority, --max-usd and --pace-max-concurrency values are never read as the slug.
  • gbrain sources writer git-durability <source> --enable still commits nothing itself, but whenever the AGENTS.md (or RESOLVER.md) it owns is not committed it now returns the exact gbrain git commit-push command as agents.commit_argv, with a real commit message. That includes a copy an earlier version left untracked while the block is already current, which used to get no hint at all and stayed untracked.
  • When gbrain sync finds no readable repository for a brain folder and creates one, it now refuses instead if a folder above it (through any link) already holds a .git, so an unreadable outer repository (for example one Git refuses for its ownership) never gets a second repository nested inside it.

gbrain setup claude-code --shape personal (or codex) gives a new brain a starting shape: three empty starter pages about you, your notes folder imported read only, and a short block that tells the agent to read those pages. The pages are me/facts, me/voice and me/how-i-decide. Each is an empty form marked status: stub, so nothing about you is invented, and a re-run never overwrites a page you have edited. --notes DIR imports that folder's markdown without writing into it; above 500 files setup prints the import command instead of running it. Setup then checks that keyless search finds one of your notes and prints a question to try in a new session. A brand-new brain set up by an agent without --shape gets a brain_shape question in its first-run decisions, and nothing is written until it is answered. --shape none removes the block and keeps the pages. Every command setup hands back keeps the run's --scope project and --source, and the import and page writes go to the pinned source. The draft-in-voice skill now treats a stub page as missing.

Itemized changes

  • src/commands/setup-brain-layout.ts (new): the starter pages, the managed gbrain:brain-layout block, the read-only notes import and search check, and the brain_shape decision. A page is written only when gbrain get reports page_not_found; any other read error is reported as pending instead of overwriting. The notes count skips symlinked folders.
  • src/commands/setup.ts: new --shape personal|none and --notes DIR flags for claude-code and codex; --shape company is refused for now. --no-instructions also takes out a layout block an earlier run wrote.
  • skills/draft-in-voice/SKILL.md (and the plugin copy): a status: stub page counts as missing.
  • Docs: new docs/guides/brain-layout.md.
  • Tests: new test/setup-brain-layout.test.ts.

Imported conversations no longer keep cookie values from Cookie: and Set-Cookie: headers or curl's -b/--cookie, one-line .netrc passwords, or a password sent as its own message after a password: line. gbrain transcripts ingest copied these into the searchable page verbatim, because no detector knew a cookie value or a label in one message whose value came in the next. The transcript importer now redacts them. On pages imported earlier, gbrain transcripts audit-secrets finds the cookie values and .netrc passwords (doctor asks for a fresh audit), but not a password sent as its own message: a stored page no longer marks where one message ends. The sources push gate and the compiled-context scan are unchanged. Three cases still reach the page verbatim, as they do today: a .netrc entry written over several lines (machine, login and password each on its own line), a -b cookie on its own continued line of a multi-line curl command (what Chrome's "Copy as cURL" produces), and a quoted cookie value inside escaped JSON (sid=\"…\").

Itemized changes

  • src/core/secret-scan-cookie.ts (new): the cookie_header rule. Values after Cookie: / Set-Cookie: headers and after curl's -b/--cookie are redacted; cookie names, Set-Cookie attributes, prose, placeholders and values under 8 characters stay. Cookie values have their own echo cap (ECHO_MAX_UNIQUE_COOKIE), so a pasted cookie jar never uses up the shared cap for bearer, assignment and labeled values.
  • src/core/secret-scan-labeled.ts: transcript-only .netrc password form, and a key or token label ending its line (here is my API key:) whose one-token value (12+ characters, a letter and a digit, not a file name or version) is alone on the next line.
  • src/core/secret-scan.ts: new ScanOpts.transcriptRules and ScanOpts.leadingDangling, both off by default; TRANSCRIPT_LANE_SCAN_OPTS and trailingDanglingLabel for the transcript lane.
  • src/core/transcripts/render.ts: redactSession uses the transcript-lane options and carries a label left dangling at the end of a message into the first non-blank line of the next message only.
  • src/core/transcripts/secret-audit.ts: audit-secrets uses the same options; TRANSCRIPT_SECRET_AUDIT_VERSION 5 → 6, so transcript_secret_exposure asks for a fresh audit. src/core/transcripts.ts (the recent-transcripts read) uses the same options.
  • Tests: test/transcripts-secret-cookie-header.test.ts, test/transcripts-secret-cross-message.test.ts, test/transcripts-secret-corpus.test.ts and test/secret-scan-push-gate-unchanged.test.ts (a golden proving the shared scanner's other option sets are byte-identical), over the synthetic corpus test/fixtures/transcripts/synthetic-credentials.ts.
  • Docs: docs/guides/data-ingestion.md#credential-redaction and the key-files entries.

[0.60.723.0] - 2026-11-15

gbrain access rule add no longer lets a hand-typed reader list lock the brain owner out of a folder. When an owner is set, a rule whose --readers leaves the owner out is refused with readers_omit_owner, the same check and message gbrain repair visibility --readers already uses, and nothing is written. Pass --allow-without-owner when the owner really should not read it. A brain with no owner set is unchanged, and it works the same on a repo-less source (rules in the database) and a git source (.gbrain/access.yml).

Itemized changes

  • src/commands/access-rules.ts: rule add checks the owner before previewing or writing; new --allow-without-owner flag, named in gbrain access --help.
  • src/core/access/owner.ts: the owner lookup and the readers_omit_owner error now live here and are shared with src/core/repair/readers.ts.
  • Docs: docs/guides/access-designations.md and the setup skill name the flag.
  • Tests: new test/access-rule-owner-gst2698.test.ts (refusal, override and no-owner case on a repo-less and a git source, PGLite and Postgres); two existing tests that rule a page away from the owner on purpose now pass --allow-without-owner.

A squashed landing on master reuses the pull request's passed full test run instead of running the whole suite again.

Changed

  • scripts/ci-reuse.ts: a one-parent push to master (a pull request landed as one squashed commit) now reuses the landed pull request's own run: the complete, passed full run of Test or E2E on a head GitHub names for the commit with the same tree, under the same artifact, job and repository checks a merge commit gets. Every run on those heads is read and the newest one decides: if it failed or is still running, the push runs in full, and a cancelled run or one that skipped its full job is passed over. A one-parent push no pull request names (a hand revert, a direct push), one whose lookup fails, one whose tree matches no full run, and a root commit still run in full. docs/ci-event-parity.md, "Master-push reuse", says so.

Outside contributors now get an accurate picture of what happens to their pull requests, and issues our own automation files are labelled so community reports are easy to find. The comment posted on an outside pull request and the new "How an outside pull request is handled" section of CONTRIBUTING.md describe what really happens: docs-only changes can be merged as they are after a read, small plainly correct fixes are rebuilt with credit to the author, every PR lands as one squashed commit with outside authors credited by a Co-authored-by: trailer, and the three-open-PR limit never applies to anyone who had a PR taken in the last 90 days. Every issue the repo's scripts and CI open now carries the agent-filed label, and the new-issue page links to a "Community issues" search that leaves those out.

Itemized changes

  • scripts/fix-wave-gate.ts, scripts/fix-wave-closeout.ts: the gate and close-out comments describe the current rules and link to the new CONTRIBUTING section.
  • scripts/lib/gh-issue.ts: createIssue always adds agent-filed; scripts/upgrade-matrix.ts and scripts/lib/master-red.ts keep it on every issue they open.
  • .github/ISSUE_TEMPLATE/config.yml: a "Community issues" contact link.
  • Docs: CONTRIBUTING.md, README.md, docs/RELEASING.md and the CI key-files page.
  • Tests: new test/scripts/gh-issue-agent-filed.test.ts; the nightly-issue, fix-wave-gate and fix-wave-closeout tests updated.

A second folder of notes (say, your company's beside your own) shows up in search without Git and without guessing why it doesn't.

Changed

  • On a brain only you use, a source you add without --federated or --no-federated joins default search, so gbrain query finds its pages right away. On a shared brain it stays out, and gbrain sources add and gbrain sync --source <id> say so with a source_not_in_default_search notice whose fix is gbrain sources federate <id> (also under notices in --json). Sources you already have keep their setting.
  • gbrain sources add <id> --path <folder> accepts a plain folder that is not in a git repository, without --force, and says that git history is off for it. A git repository with nothing committed is still refused. gbrain sources inspect on a plain folder now points at sources add.
  • gbrain setup codex lists the provider key names (OPENAI_API_KEY, VOYAGE_API_KEY and the rest) under env_vars, so Codex passes keys you already exported to gbrain and search stays semantic. Only names are written.
  • The "sole non-default source ... routing to 'default'" line no longer prints on every command when that source is already in default search.
  • gbrain sources federate on a keyless brain says embedding is off instead of telling you to run gbrain embed --stale, and sources --help and pages --help no longer show an old version in their header.
  • README: the voice line points at agent-voice (the Twilio recipe is deprecated), says how voice notes and chat transcripts get in today, and shows how to add a second folder.

Two tests that failed CI runs on unrelated changes now pass on a busy runner. Neither changed what it checks.

Tests

  • test/proactive-brief.test.ts: the two tests that check a delivered brief writes its context_volunteer_events rows went through turnContextWithBrief, which gives the brief only 300 ms of wall clock from the start of the request. On a slow runner the base assembly used up that time, the brief came back empty, nothing was delivered or logged, and the row check saw []. turnContextWithBrief now takes an optional clock (it still defaults to the 300 ms cutoff), and those two tests pass the same wide test clock the file's other brief tests already use. A probe that slows every query by 60 ms fails the old test and passes the new one.

  • test/skill-resolver-hook-lane.test.ts: the 5 MB Hermes history test compared one timed run of a small turn with one timed run of the 5 MB turn against the 50 ms lane deadline. Each run varies by 20 to 30 ms on an idle machine and more on a loaded one, so one descheduled turn could fail it (it failed 1 run in 5 under load locally). It now times three runs of each and compares the fastest of each, with the same 50 ms bar.

  • A member's connection can no longer use sources_resolve_path to learn that a page it may not read exists. A guessed path for a restricted or owner-private page now gets the same answer as a path with no file and no page, even when an older readable copy of that path was deleted. The brain owner's own connections read exactly what they read before. (#8761)

  • The gbrain access rule add commands printed by gbrain doctor, the labelled-only create refusal, gbrain access serves and the federation exposure check now include the owner in --readers, so they are accepted on a brain with an owner set. Re-running gbrain access serves after applying its printed fix now reports the rule as already in place.

  • sources_resolve_path on a plain directory with no Git history now refuses with sync_not_applicable (reason plain_directory) and a next step, instead of a raw Git error. Contributed by @Kkkkkuro.

A first install prints no Bun error and ends with memory connected, on every agent harness.

Changed

  • gbrain no longer ships a postinstall script, so bun install -g github:garrytan/gbrain#stable no longer ends with "Blocked 1 postinstall", and a clone's bun install no longer prints "postinstall skipped". The script did nothing on a fresh install; migrations already run in gbrain setup, gbrain upgrade and on the first command.
  • After reinstalling over an older brain, commands print a migrations_pending notice naming the upgrade migrations still to run, with gbrain apply-migrations --yes as the fix.
  • gbrain init treats connecting the agent as an install step (its own harness_wiring notice), not a setting awaiting approval. The first-run settings are reported as in effect, with how to change them, and no "Reply 'defaults'" question; automatic saving shows as off (recommended salient) instead of "default: salient".
  • The install docs no longer tell agents to stop and ask about search mode: connect first, then report the settings and the cost matrix in the install reply.

A test of the OpenClaw skill plugin's daily log no longer fails on a slow runner. It didn't change what it checks.

Tests

  • test/openclaw-skill-plugin.test.ts: the plugin writes each daily-log line through a chain of awaited appends, and the test's Node driver exited a fixed 300 ms after its last step. On a slow runner, the 164 appends in "the daily log keeps event order when a skill read follows a turn at once" were still queued when the driver exited, so the tail of the log was lost and the check failed. The driver now waits, for at most 30 s, until the log holds the number of lines the test expects, and then exits as before. The other test that reads this log (six lines) waits the same way. A probe that slows every append by 5 ms fails the old test (109 lines lost) and passes the new one.

The README now opens on one brain for you and your company.

Changed

  • The first screen says who GBrain is for, what you have a few minutes after installing, and how to add your own notes and your company's docs (a facts page, a decisions log, call transcripts) as two searchable folders with three commands.
  • It says plainly that GBrain is not a data warehouse or analytics store.
  • The install block and the hosted option are unchanged; the sections that used to sit above the install now follow it.

gbrain repair redacted-facts now restores facts the conversation extractor wrote, which keep a row number on a page with no Facts fence (#8820).

Fixed

  • src/core/repair/redacted-facts.ts: a fact with a row number is treated as a Facts-fence row only while its page's fence holds an active row with its text at that number. A conversation-extractor fact with no such fence row is now restored in place, keeping its row number, by the same row-exact rewrite as a database-only fact; its page is not changed. Before, the preview listed these facts as evidenced and the apply left them damaged with changed_since_preview. Any other fact whose fence row is gone is now listed as unproven (fence_row_missing) instead of evidenced, so the preview no longer promises a restore the apply cannot make. The repair is still preview-first: --apply --expect <hash> restores only the previewed set. Reported by @tarush1989.
  • Docs: docs/guides/repair.md (redacted facts class table and apply paragraph).
  • Tests: test/repair-redacted-facts.test.ts (PGLite, and Postgres through test/e2e/repair-redacted-facts-postgres.test.ts): a managed brain restores an extractor fact and leaves its page revision unchanged; an unmanaged brain restores an extractor fact and keeps a non-extractor fact whose fence row is gone. Both failed before the fix on both engines.

gbrain repair visibility --propose drafts who should read each open page, so a brain shared by several people can be labelled without typing every reader list by hand. It reads the participants, attendees and to, from and cc addresses already on open pages, matches each address to an enabled member's declared email, and groups the pages by reader set with the owner always first. For each set it writes a private slugs file (mode 0600) and prints the existing gbrain repair visibility --slugs-file ... --readers ... command, so you preview and apply with the tools you already have. It labels nothing itself. A page with an address that matches no member, or a slug a slugs file cannot carry (one starting with #, padded with spaces or holding a line break), is listed as unmatched instead of guessed. DM, chat and mail folders with no address list get an access rule add command to fill in. On a brain that serves one person it says there is nothing to propose and reads no pages. It works the same on a repo-less brain and a git brain, and refuses to write its files inside a git work tree.

Itemized changes

  • src/commands/repair-propose.ts (new): builds the proposal, writes each slugs file through an exclusive temp file and a rename, and shell-quotes every printed command.
  • src/commands/repair.ts: new --propose and --out flags on gbrain repair visibility.
  • src/core/access/open-pages.ts, src/commands/access-serves.ts: share the open-page query and git-root lookup with the new command.
  • New error code repair_propose_dir_in_git, listed in docs/guides/error-codes.md.
  • Docs: docs/guides/repair.md (proposing readers) and docs/architecture/key-files/access.md.
  • Tests: new test/repair-visibility-propose-gst2670.test.ts (PGLite and Postgres).

Purging a page retires the facts extracted from it, gbrain migrate embeddings stops refusing on some databases, a growing transcript is no longer re-synthesized on every dream pass, legacy embedding labels no longer wipe valid vectors, and search says how far an upgrade's index rebuild has got.

Fixed

  • Facts extracted from a purged page are retired with it (#8859, reported by @ocscdsa). Facts the extractor filed on no page and that name the page only in their context stayed active after the page was retired or purged. Retire and purge now match them too. Facts left behind by earlier purges are not changed here.
  • gbrain migrate embeddings no longer always refuses with "Source census changed while acquiring locks" on databases whose collation sorts source ids differently from code-unit order (#9307, reported by @ozabramov; first fixed by @Masashi-Ono0611 in #9310).
  • A growing transcript is re-synthesized at most twice a day (#6753, reported by @andreineacsu). A transcript synthesized within dream.synthesize.cooldown_hours is held, so repeats drop from up to 24 a day to 2. The daily spend ceiling is unchanged ($120 worst case). Explicit --input and --date runs are never held.
  • Chunks stamped with a bare model label keep their vectors (#8533, reported by @babyspectrum). The false-stamp guard, the projection preserve check and the stale-signature invalidation now share one rule: a bare label counts when its model name and width match the configured model; a different explicit provider prefix is still invalidated. Before, the last two nulled valid vectors and forced a paid re-embed.
  • Search says how far an upgrade's index rebuild has got (#9192, reported by @jonahberg). While the rebuild runs, the partial-results notice gives pages done out of total, the measured rate and the command to watch progress, and the upgrade prints it once.

Itemized changes

  • src/core/facts/gone-page-facts.ts, src/core/facts/subject-infer.ts, src/core/persistence/page-purge.ts: extracted-from-page facts are expired and purged with their page.
  • src/commands/migrate-embeddings.ts: the locked census is sorted the same way as the requested source list.
  • src/core/cycle/synthesize-completion.ts, src/core/cycle/synthesize.ts: per-transcript re-synthesis hold.
  • src/core/embedding-invalidation.ts, src/core/engine-sql/chunks.ts, src/core/page-state/projections.ts: one chunkModelMatchesSql rule.
  • src/core/search/retrieval-coverage.ts, src/core/ops/search.ts, src/core/interop-notices.ts, v153's notice handler: rebuild progress.
  • Tests: one new test per fix, on PGLite and Postgres where the bug is engine-specific; the #5051 test changes one assertion and its count.

The skill resolver now leaves its skills block out when no skill fits the message, instead of showing one on every turn.

Fixed

  • gbrain skills resolve, the hook lane, and the OpenClaw and Pi plugins apply top_gate by default (0.6): a block whose best skill only matches weakly is left out (block_reason: weak_top_match), and a skill named in the message still gets through. On harness-evals' public set P with its 351-skill catalog, blocks on requests that need no skill drop from 61 of 61 to 7 of 61, and the acceptable skill stays in the block for 174 of 228 skill requests (176 before). skills.resolver.top_gate 0 restores the old behavior.
  • gbrain skills resolve and the OpenClaw and Pi plugins now read skills.resolver.top_gate (and per_harness.<harness>.top_gate); before, only the hook lane did. --top-gate still wins on the command line.

Two slow test files are split along their own seams, so no file comes near CI's 300-second per-file limit. test/enforced-member.serial.test.ts took 185 to 191 s on a 4-vCPU machine, and CI killed it at 300 s on slow runners with no assertion failing. It is now three serial files that take at most 77 s each. test/setup-serves-several.test.ts took 193 to 240 s and is now three unit files that take at most 78 s each. The tests and assertions are unchanged (27 and 645, and 27 and 270), and no product code, timeout, budget or retry count changed.

Tests

  • test/enforced-member.serial.test.ts keeps the reads and refusals. Writes and the git commit-push suite move to test/enforced-member-writes.serial.test.ts, and the access.enforced_member_writes case moves to test/enforced-member-writes-setting.serial.test.ts. Each file seeds its own brain through test/helpers/enforced-member-brain.ts.
  • The spawned gbrain setup cases split out of test/setup-serves-several.test.ts into test/setup-serves-stdio-member.test.ts and test/setup-serves-upgrade.test.ts, with the shared box in test/helpers/setup-serves-box.ts.
  • The new serial files get rows in scripts/serial-files.tsv, test/postgres-unit-arms.txt and scripts/postgres-arm-weights.json. The unit and serial weight maps get no hand-added rows, because they must match the miner's metadata, so new files use the p75 estimate until the next re-mine.

docs/RELEASING.md now says how a pull request lands on master: as one commit titled with the pull request's title and number.

Changed

  • docs/RELEASING.md, "Releases per batch", has a "How a PR lands" paragraph: each pull request, release pull requests included, lands as one squashed commit whose subject is its title followed by (#<number>), with other authors credited by Co-authored-by: trailers, while branches keep bringing master in with merge commits and are never rebased.
  • The contributor audit's offline fixture now includes a squashed (#<number>) commit next to its "Merge pull request" merges, so the audit's tests cover both commit shapes it reads.

On a thin client, gbrain serve now connects a stdio-only agent harness to your remote brain instead of refusing.

Install the gbrain CLI on the agent's machine, run gbrain init --mcp-only with an OAuth client from the brain host, and register gbrain serve as the harness's stdio MCP server, the same entry a local brain uses. Every call goes to the brain host with that client's grant and comes back unchanged, refusals included; the bridge opens no local database, stores no page text and adds no access. Reported and first proposed by @jel405 (#8543).

Itemized changes

  • Bare gbrain serve on a thin client is a stdio bridge. It accepts --surface, --advertise and --source-guard (the argv gbrain's plugins register); request_tools {surface} widens only the session and writes nothing on the host. serve --http, --access and the other host flags still refuse with requires_local_engine, with a fix that keeps the flags (--access points at registering a narrow read-only client instead).
  • Bridge failures are v1 agent envelopes with a bounded recovery and a fix that never adds a scope. A write whose answer was lost keeps its request_id (write_outcome_unknown) so it can be checked or resent without writing twice. URLs in envelopes and stderr lose their userinfo, query and fragment.
  • A call that names an audience (_meta.on_behalf_of) is relayed only to a host that enforces it (gbrain 0.60.543.0 or newer); otherwise the bridge answers host_upgrade_required.
  • gbrain doctor on a read or read+write thin client reports the missing admin scope as expected instead of advising a re-register with admin.
  • gbrain agent register --harness openclaw prints the lines to run on the OpenClaw machine: gbrain init --mcp-only … then openclaw mcp add gbrain --command "$(command -v gbrain)" --arg serve ….

gbrain extract timeline --source db no longer fails on a large page with many stored versions on a managed Postgres brain.

Retraction used to read every stored version of the page in one statement inside the publication transaction. With hundreds of versions that statement ran past the 5 second statement timeout on every retry, so the page's timeline extract failed with write_interrupted on every maintenance run. The version scan now runs before the transaction and reads one version at a time. Inside the transaction gbrain only checks that the versions it read are still the page's, with the same text. A row whose versions changed in between stays until the next run. Reported in #9362.

Fixed

  • gbrain extract timeline --source db and the maintenance timeline pass no longer time out with write_interrupted on a page with hundreds of stored versions (#9362).

Guest notes now refuse writes from a database role that cannot see which sources are guest sources, and several smaller guest-note gaps from the first release are closed.

A brain whose gbrain process connects to Postgres as a scoped row-level-security role used to read the guest source list as empty, so a write into a guest source passed as if the source were ordinary. Such a role is now refused with writer_coordinator_required until it can read guest_sources (one SELECT policy, added to the maintainer recipe in docs/ENGINES.md). The owner, a role that bypasses row-level security and PGLite brains see no change.

Fixed

  • Schema migration v314 re-creates the guest guard in the trust trigger and the guest-record guard so they refuse a role that row-level security keeps from reading guest_sources; guest admission, remember, links and remote deletes refuse the same way instead of treating the source as ordinary.
  • gbrain guests promote no longer answers already_promoted when the kept page was purged, so promoting again writes the page and keeps one link. v314 adds a unique index on the promotion link (a brain that already holds duplicate links keeps them, gets no index and a warning), and two promotions of the same note to the same page that race each other both succeed with one link.
  • A guest note whose mark writes no row now fails the save instead of committing a guest page with no mark.
  • gbrain guests init counts the source's pages under the source row's lock, so no page can land between the check and the change.
  • The repair gbrain doctor suggests for a source whose config says guest but has no guest record, gbrain guests init --source <id> --yes, now works when the source holds pages.
  • A cycle phase skipped because the guest-source lookup failed now says the lookup failed (reason: guest_source_lookup_failed, with the error), instead of reporting a plain guest-source skip.
  • A read the guest-note stamp cannot copy is refused, instead of marking the handler's own object.
  • A page write on a brain whose schema is behind the guest-notes migration is refused migrations_pending instead of failing with a missing-table error.
  • The owner-credential read behind gbrain guests grant-owner-clients and doctor's guest_source check no longer loads token hashes.
  • gbrain auth rescope --client <id> with no grant flag now prints the client's operations and slug prefixes, and gbrain delete --help says --expected-revision is required unless --force.
  • docs/guides/guest-notes.md gives the setup order that works (init, register the broker client, rescope its operations, then add it as a writer), and says that deletes need a revision or --force and that a put with attribution but no audience is refused readable_by_invalid.

On a PGLite brain held by gbrain serve or the job worker, gbrain wakeshift emit cron.finished now records the run instead of dropping it, and revoking a credential always ends its watches.

Fixed

  • A cron run that finished while gbrain serve or gbrain jobs work held the PGLite brain was lost. Emit now saves it as a small private file (0600) under the brain's own gbrain home, and the process holding the brain records it within seconds with its original finish time, so a watcher still wakes. Two processes reading the folder at once never record a run twice. A file that is malformed, oversized or has an impossible time goes to rejected/ with one stderr line and never holds back the runs after it. At most 1,000 runs wait per brain. Postgres brains and brains with reporting back off never use the folder.
  • The next direct emit also records any leftover runs, and it records its own run even when the leftovers can't be read.
  • gbrain auth revoke --id <id> with a legacy token id written in uppercase now ends that token's watches and asks Wakeshift to forget its user.
  • When removing an access member failed to ask Wakeshift to forget that member's user, the wake drainer now asks again. A member who is still active keeps their Wakeshift user.

gbrain can now report back when a pending write settles or a cron run finishes, on brains shared with access members, and to OpenClaw and thin-client agents. Reporting back stays off until the brain owner turns it on.

Two new events join job completion. watch_write_request wakes the conversation when a write_pending write commits, fails, conflicts or is cancelled. It uses the same check as get_write_request, so someone else's request reads as not found. watch_cron wakes it after the next run of a cron the owner registered with gbrain wakeshift cron add <name>. The cron's own script ends with gbrain wakeshift emit cron.finished --name <name> --status ok|failed, which records the run and never blocks or fails the script, even when it can't open the brain. On a PGLite brain, a run that finishes while gbrain serve or a job worker holds the brain isn't recorded. Every wake is still a pointer (ids, status and times), never page text or cron output. A brain with access members can now report back too. The owner gives their terms and privacy links to gbrain wakeshift enable, and each member's agent shows the agreement step and records acceptance before its first watch. Removing a member, revoking a token or deleting an OAuth client ends that principal's watches and asks Wakeshift to forget its user. Guide: docs/guides/wakeshift.md.

Itemized changes

  • New MCP tools watch_write_request, watch_cron and accept_wakeshift_terms, listed only while reporting back is on. list_watches and unwatch now cover every kind of watch.
  • New gbrain wakeshift cron add | remove | list and gbrain wakeshift emit cron.finished. enable takes --terms-url and --privacy-url (https), which a brain with access members needs. status and doctor say when gbrain's published events are older than the installed gbrain; run gbrain wakeshift enable again to publish the new ones.
  • New gbrain wakeshift watch job <id> | write <request_id> | cron <name>, which runs on a thin client (gbrain init --mcp-only). It prints Wakeshift's instructions as the last line and never saves a session key.
  • New gbrain setup openclaw --wakeshift. After consent, it installs and enables Wakeshift's OpenClaw plugin, pinned at @wakeshift/[email protected]. On a gateway that listens beyond the machine, it prints Wakeshift's question about the hooks endpoint and never answers it. A new doctor check wakeshift_openclaw warns when the installed plugin differs from the pin.
  • gbrain setup claude-code and gbrain setup codex show a "Report back (Wakeshift)" row in the first-run settings. After a remote submit_job or submit_agent, a once-per-session notice (wakeshift_report_back, muteable) says how to have the conversation woken. With reporting back on, the MCP instructions gain one sentence about it. With it off, the instructions are unchanged.
  • Fixed: Wakeshift's instructions in a watch result were cut at 2,000 characters, which dropped the step that starts Claude Code's listener. They now pass through whole, up to 16,000.
  • Schema migration v313 adds wake_crons, wake_cron_runs, wake_terms_acceptances, wake_forgets and wake_subscriptions.target_ref. All are empty unless reporting back is on, and repo-less brains get them unchanged.

[0.60.722.0] - 2026-11-14

The brain owner's own remote agents can now edit, append to and relabel the private pages they already read, and everyone else is still refused.

Changed

  • A credential bound to the brain owner, with write scope and speaking for itself, may change a visibility: private page it reads with put_page, edit_page or append_page, or with put_pages and staged uploads (begin_page_upload and commit_page_upload), and relabel it (for example to visibility: world). A put_page that leaves out visibility or readable_by keeps the stored ones, an empty visibility counts as left out, and a page private only through a privacy alias, a pinned question, a derived type or a private origin is saved as visibility: private (keeping its pinned_question) by any of these writes that states no visibility; the page revision and the journaled request name the owner's credential. Other members, unbound tokens and OAuth clients, attested calls, the owner's credential without write scope and writes to pages whose readers or folder rule leave the owner out get page_not_found (or their earlier refusal) as before, on repo-less and repo brains alike, and so does any write to a soft-deleted private page and a credential fenced to slug prefixes. A staged upload over the single-string ceiling is now held to the same readable_by check as a smaller one.

Fixed

  • Purging a private page now also labels a life-chronicle event private when the event's own derived_from names that page; if the label can't be written, the page stays soft-deleted instead of being purged.
  • The purge's in-transaction check for derived pages reads only the pages that name the purged page instead of scanning the whole source, so single purges on large brains are much faster.
  • gbrain pages purge-deleted --dry-run now lists exactly what --yes would purge, including parked pages a consented run would release, and nothing is written before consent.
  • A purge held to keep derived pages private now tells the owner to finish it with gbrain delete <slug> --purge --source-id <id> as a new request, which writes the purge tombstones.

CI runs the unit suite in 34 shards instead of 30, so every shard fits its 600-second budget again.

Changed

  • Shard timing weights are refreshed from the median of five recent nightly Test runs (unit and serial) and the latest full E2E nightly. With those weights the heaviest of 30 shards packed to just over 600 seconds; with 34 shards each packs to about 529 seconds.
  • The test job in test.yml and the coverage-full-unit job in e2e.yml both run 34 shards, and their coverage manifests expect all 34. A full run starts four more unit jobs than before.

Fixed

  • bun run release:unstamp no longer refuses a branch because master edited a published migration after the branch was cut. An edit the branch itself made to a published migration is still refused.

Guest notes: a group member's note is saved, marked as a guest's, warned about on every read, and becomes your own page only when you say so. People in a group chat with your agent sometimes say things worth keeping. A guest source keeps those notes apart from your own pages, with a record of who said them. Nothing changes until you run gbrain guests init: a brain without a guest source reads, writes and costs exactly as before, and no grant, privacy label or search scope changes on upgrade.

A guest source has four rules the database enforces, whatever program writes. Only writers on the source's list (an OAuth client or legacy token, for example the broker your agent's host runs) may create a page there or change its text, and only as a new page; your own CLI, gbrain import, sync, remember, add_link and put_raw_data are refused source_writer_restricted, while delete, restore and purge from your CLI still work. No page there is ever stored above the external_untrusted trust tier, even after an owner rewrite or a trust confirmation. Guest notes feed no facts, takes, timeline entries or links (guest_page_immutable). And a guest source is database-only, so attaching a local path is refused. gbrain guests init also refuses an existing source that already holds pages, deleted pages not yet purged included, so your own pages never turn into guest notes.

When the writer saves a note, gbrain records who it was saved for, which group, when and which client sent it, in its own table that no later write can edit and that never comes from the page text. Every read that returns a guest note (search, query, get, list, find, backlinks, timeline, think and the rest) marks it as an outside guest's words: origin: "guest" and a guest object in JSON, and a warning line plus the external-data envelope in text. A contract test fails if any operation, present or future, returns a guest row without the mark.

You keep a note as your own page with gbrain guests promote <slug> (MCP promote_guest_note, owner only; anyone else is refused guest_promotion_owner_only), which writes a normal page and records the link; the broker can no longer delete a promoted note (guest_page_promoted). Promotion reads the note through the caller's normal read scope, and promoting again finishes a promotion whose link was not recorded (an interrupted run or a write still pending) without writing the page twice. Your own tokens and OAuth clients read the guest source only once it is on their read list: gbrain guests grant-owner-clients previews that and --yes applies it. gbrain doctor --only guest_source checks the whole setup, including that the source's two records agree.

Added

  • gbrain guests init --writer <principal> (preview, then --yes) creates or marks the guests source; gbrain sources writers <id> [--add|--remove|--clear] changes who may write.
  • gbrain guests promote <slug> and MCP promote_guest_note: the owner keeps a guest note as their own page.
  • gbrain guests grant-owner-clients: adds the guest source to the owner's own clients' read lists, preview first.
  • Doctor check guest_source.
  • Error codes source_writer_restricted, guest_page_immutable, guest_promotion_owner_only, guest_page_promoted and their recovery paths (docs/guides/error-codes.md, docs/guides/write-refusals.md).
  • Guide: docs/guides/guest-notes.md.

Itemized changes

  • The guest-notes schema migration (src/core/schema-migrations/v<NNN>-guest-notes.ts; GUEST_NOTES_MIGRATION in src/core/guests/migration.ts names its number): the guest_sources, guest_marks and guest_promotions tables under row-level security, the per-source trust ceiling in the trust trigger, and the writer backstop trigger.
  • src/core/guests/: admission (writer list, new-page-only, allowed frontmatter and readers), the mark and its warning, derived-row refusals, promotion, owner credentials and source setup.
  • Read paths (src/cli/page-render.ts, src/cli/search-row-format.ts, src/cli/guest-lines.ts, src/core/access/page-items.ts, src/core/think/index.ts, src/core/search/transcript-quote.ts) carry the guest mark and warning.
  • A guest writer can delete only the guest notes it saved itself; deleting a note another writer saved is refused permission_denied.
  • Stale extraction, propose_takes (scoped or not) and the cycle's fact, atom and enrichment phases leave guest pages out.
  • Reads on a brain whose schema is still behind that migration keep working, because the guest check falls back to the source config when the guest_sources table is missing.
  • Goldens regenerated: test/fixtures/goldens/sql-text/pages.json (stale-extraction query excludes guest pages), test/fixtures/goldens/exports/types.json (guest types and error codes), the doctor goldens (the guest_source row and the three new tables under row-level security) and test/fixtures/goldens/cli/tools-json.json (the new operation).

[0.60.721.0] - 2026-11-14

Deleting a private page no longer opens the pages derived from it: each page that was private only because of the deleted page now gets its own visibility: private label before the deleted page's row goes.

Fixed

  • A summary whose derived_from names a private page, an atom extracted from a private page, and a synthesized concept built from such an atom read as private to remote callers even when their own label says world, because their origin is private. Once the origin's row was removed (an explicit gbrain delete --purge, or the purge after the 72-hour recovery window, including after a remote delete or a git rm), nothing kept them private any more, and every remote caller that is not the owner could read them through get_page, get_versions, list_pages and search.
  • Now gbrain gives each such page visibility: private first, as a label-only write. The page keeps its trust tier, no lowering proposal is filed, and on a brain with Git only the label line of its file changes. A soft delete (local or over MCP) writes the label once the delete committed, as a maintenance write (on a managed brain its batch names the delete's request id); if that write fails, the delete still goes through and the purge writes the label later. The delete's response is unchanged and names none of the labelled pages. Every purge writes the label again for each origin expired or expiring within a day before it removes any row, so an origin removed outside gbrain is covered from the next cycle.
  • Pages that are already private, pinned questions and pages that do not derive from the deleted page are never changed. A page that names readers (readable_by) and was hidden from them only through its private origin gets the label too and keeps its readers, so its listed readers still can't read it. The owner, a granted_sources credential and a listed reader read exactly the same pages as before.
  • When a label cannot be written (for example a file in the checkout that gbrain may not write), the deleted page is kept soft-deleted instead of purged. An explicit purge returns purge_held, the purge phase lists the page as derived_origin_parked and warns, and the new doctor check derived_origin_parked names the page with the derived page that blocks it and the reason. Once the cause is fixed, the next purge writes the label and frees the page, and the purge after removes it (a consented gbrain pages purge-deleted --yes does both, so its preview lists it). A page derived from it that is written while a purge runs is caught inside the purge's own transaction and holds the deleted page the same way, and a derived page that is itself soft-deleted holds it until it is purged or restored, so restoring it never brings it back readable.
  • A deleted page that is restored and later made world leaves its derived pages private until the owner relabels them. (GST-2689)

Searches stay within their four-connection limit, the extraction-lag check saves a database round trip, and four tests that failed now and then for reasons unrelated to the code under test now pass reliably.

Fixed

  • One search or query can no longer hold a fifth database connection. Reads that turn off the query planner's JIT compiler or set a statement timeout used to take their own connection without waiting for one of the search's four slots. The projection readiness check is one of these reads, so a search running while any write was open on the database server could use five connections. These reads now wait for a slot like every other read. A read whose cancel signal fires while it waits now stops at once instead of waiting for another read to finish.
  • The extraction-lag check (used by gbrain doctor and the graph ranking) reads whether managed persistence is on at the same time as its config, not afterwards. On a remote database that saves one round trip.

Tests

  • The read round-trip test now holds each wave of queries until it settles, so it counts dependent round trips the same way however fast the database answers.
  • The status-mode server test no longer fails when the lock holder's process id happens to appear inside the server's random instance id. A process id in any other field still fails it.
  • Four HTTP-server end-to-end tests now remove the access members they add. Leftover members made a later test file's connection check fail.
  • The search connection limit test waits for background writes from its warmup search before it starts measuring, and checks that JIT-off and statement-timeout reads respect the limit.
  • The lock-wait sync test prints the rows that did not commit when it is about to fail, so a future failure names its cause.

On a source whose files live in Git, two paths no longer save a fact only in the database or drop it: a held fact the owner releases while managed persistence is off, and a commitment fact the Google loops job meets a passing database error on.

Fixed

  • The Google open-loops job (loops_extract) now retries when saving a commitment's fact hits a passing database problem: a cancelled statement, a lock or serialization conflict, a dropped connection, or the write door giving up after its own retries (write_interrupted). Before, the job saved the open loop with no fact attached and moved on, so the fact for that commitment (for example "send alice-example the invoice summary by Friday") was lost. Now the job stops, the queue runs it again, and the retry saves both the fact and the loop that points at it. A write the door reports as still pending keeps the old behaviour, because it may still commit and a retry could save it twice.
  • On a source marked git-canonical (canonical: git, for example source-a), gbrain trust release of a held fact while managed persistence is turned off now refuses with git_home_required (reason persistence_disabled) and leaves the fact held, instead of writing it to the database with no file in Git. The refusal's fix previews turning managed persistence back on (gbrain sources writer activate --confirm-quiesced --dry-run --json); run the release again once it is on. Sources that are not git-canonical release exactly as before, and with sync.write_through off every write on a source stays database-only, the release included, on purpose.

Facts that were dropped, or released database-only, before this fix are not recovered by upgrading. A dropped commitment fact comes back only if its email thread is extracted again; a fact already released database-only stays that way.

[0.60.720.0] - 2026-11-14

On a brain that serves several people, gbrain now counts the pages with no label, which every member's key can still read, and names the commands that close them; --labelled-only keeps members to the pages labelled for them.

Added

  • Open pages count (GST-2670). A key bound to a member reads a page that has no label like any other key, so a chat export or mail import loaded before anyone named its readers is open to every member. gbrain now counts those pages, read-only, in the gbrain access serves several preview and receipt (also shown by gbrain setup <harness> --serves several), in gbrain access serves, in the new doctor check open_shared_pages and in gbrain post-upgrade. Each shows how many members' keys read them, the top folders, and the commands that label them, with the owner already in every reader list. A visibility: private page counts as open only when remote private pages are on for this brain or a reading member's grant opens its source. Doctor warns while DM, chat, conversation or mail pages are open and reports other open folders as information. The count changes nothing, never changes an exit code, and doesn't run on a one-person brain. gbrain access serves --no-open-pages skips it. See docs/guides/access-designations.md#pages-every-members-key-can-read.
  • --labelled-only (GST-2670). gbrain access serves several --members <ids> --labelled-only and gbrain setup <harness> --serves several --labelled-only keep each member's key to the pages labelled for that member. It is opt-in and never applies to the owner, so with the owner in the roster the run is BLOCKED before anything is written. Without the flag, no existing key reads anything different.

A storage manifest: agents can register what they archived to object storage and find it again later. A new storage source holds one validated card per archived item: what it is, where every copy lives (store, bucket, path, restic repository and snapshot, size, file count, hash and its kind, when it was verified), where it came from and which skill restores it. gbrain storage-register adds a card (from stdin or --card) and never overwrites one; gbrain storage-update <id> --expected-revision <rev> patches a card bound to the revision it read, keeps the page's access labels and records restores newest first; gbrain storage-find "<words>" ranks the exact id, then an exact alias, then the cards holding most of the words, and a miss lists the closest cards; gbrain storage-get and gbrain storage-list read. On a brain with a Git checkout, gbrain storage init hosts the shelf in its own repository first (preview, then --expect <plan_hash>; it needs managed persistence), so each card is a commit before it is indexed. A repo-less brain gets a database-only shelf on the first register.

Cards never carry credentials or commands. A card is refused when any field looks like a secret, an environment assignment, a credential variable, a URL with user info or a presigned URL, or when any key is named command, cmd, script, exec, shell or instructions; object keys are checked too, a value under a key named for a credential refuses wherever the word sits (password, restic_password, token_value, token_count, application_key) unless the key ends in a suffix that names where something lives (_env, _file, _path, _name, _ref, _id, _var, _variable), and the refusal says to rename a key that holds no credential that way, and the refusal never echoes the value. A first shelf creation that stopped part way is finished by the next register or storage init, and an archived shelf refuses until it is restored. A restore is { skill, args } with flat scalar arguments. gbrain stamps registered_by and origin itself, and holds no store credentials.

The shelf is written only through the storage verbs. put_page, remember and every other page writer are refused storage_source_managed there, a request parameter cannot open the door, register and update run only from the trusted local CLI, and a dream cycle or enrichment run scoped to the shelf skips it. The source is not federated, so a remote connection reads cards only when its grant names storage.

Agents find the shelf, and older archive records move onto it. The brain-first lookup table every harness reads gains one row, "Find something archived to B2, Supabase or S3" (gbrain storage-find / storage_find), and a new storage-manifest skill says when to register, to find before restoring, to restore through the agent's own restore skill with the card's restore.skill and restore.args, and to record the restore with gbrain storage-update. gbrain storage adopt moves interim records (stored-object pages under storage-manifest/ holding their record as JSON under ## Record) onto the shelf: the preview shows each mapped card and every field a card refuses with its reason; --apply --expect <plan_hash> registers each card with origin: adopt and tags the original storage-manifest-adopted, deleting nothing; a record whose page a remote (MCP) client last wrote is blocked and never registered, so only records written on this machine or by sync become cards; each card takes its original's access labels (visibility, access, confidentiality, readable_by, with readable_by narrowed to the readers of every folder rule over the original), so a card is never read wider than its original, and an original whose readers and folder rules share no reader, or whose rules are still being applied, is blocked; and refuses until each refused field is named with --drop-field (a credential-shaped value always refuses); --finalize previews, then soft-deletes the adopted originals only when every card matches its original (a restore recorded since is not a difference); the preview names any other section an original holds and flags one longer than the 64 KiB it reads, and an original edited after the preview, even just before its delete, refuses. gbrain storage export --jsonl and gbrain storage import --jsonl <file> (preview per line, then --apply --expect) rebuild a shelf from a file, registering missing cards with origin: import and never overwriting a differing card; each line carries its card's access labels under labels and import restores them (a card on the shelf with other labels is a conflict); export leaves out and names a stored card that fails the card rules. Applies are idempotent and report side_effects and, when a run stops part way, the command that finishes it; a write gbrain accepted that has not committed yet is listed under pending with its receipt and the gbrain write-request <id> poll (exit 10), never reported as no change. The new doctor check storage_manifest warns about cards that fail the card schema, interim records not adopted yet and a hosted shelf whose directory is gone. On a brain with a Git checkout, adopt and import wait for gbrain storage init.

Itemized changes

  • src/core/storage-manifest/adopt.ts, src/core/storage-manifest/jsonl.ts, src/commands/storage.ts: gbrain storage adopt (preview, apply, finalize) and gbrain storage export|import --jsonl, each preview-first and bound to its plan_hash (a stale one refuses preview_changed). schema.ts exports the credential-shape rules, the command-key test and the card keys adopt maps onto, plus cardContent/cardDigest for comparing copies; source.ts gains shelfWritable (the refusal a storage write would get, without creating anything); find.ts gains readShelfCards.
  • src/commands/doctor/checks/storage-manifest.ts, src/commands/doctor/registry.ts, src/core/doctor-categories.ts: the read-only storage_manifest check (brain category), not_applicable with no shelf and no interim record. Doctor goldens regenerated for the new check.
  • src/core/error-registry.ts, docs/guides/write-refusals.md, docs/guides/error-codes.md: storage_adopt_field_refused (consent: the user agrees to leave refused fields out) and storage_adopt_unverified; storage_record_invalid gains reasons no_record, duplicate_id, access_unpreservable, access_pending, labels and remote_written.
  • skills/conventions/brain-first.md (and its generated copies in AGENTS.md, INSTALL_FOR_AGENTS.md, skills/RESOLVER.md, the bootstrap template and the plugin trees), skills/storage-manifest/SKILL.md, skills/RESOLVER.md, skills/manifest.json, openclaw.plugin.json: the lookup row and the skill.
  • docs/guides/storage-manifest.md: the guide (verbs, repo-less and Git-checkout brains, adopt, export and import, doctor, privacy).
  • Tests: test/storage-manifest-adopt.test.ts, test/storage-manifest-jsonl.test.ts and test/doctor-storage-manifest.test.ts (PGLite, repo-less) and test/storage-manifest-adopt-hosted.test.ts (managed PGLite with a Git checkout).
  • src/core/storage-manifest/ (schema.ts, door.ts, source.ts, find.ts) and src/core/ops/storage-manifest.ts: the card schema and secret checks, the write door, the source lifecycle (storage init), keyless ranked lookup (exact id, page_aliases, word matches, pg_trgm closest cards) and the five ops storage_find, storage_get, storage_list, storage_register, storage_update.
  • src/core/persistence/page-mutations.ts, src/core/persistence/memory-mutations.ts: the page door and remember/forget refuse writers other than the storage verbs on a source with role: storage_manifest; the owner's import of commits already in the hosted repository still syncs.
  • src/core/cycle.ts, src/commands/enrich.ts: a cycle scoped to the storage source returns skipped with reason storage_manifest_source; enrichment considers no card there.
  • src/core/source-resolver.ts: the hosted shelf never counts as the sole non-default source, so after storage init bare put, capture and sync keep routing where they did before. src/core/persistence/preadmit-cache.ts caches the page door's source read with its new role column.
  • src/core/error-registry.ts, docs/guides/write-refusals.md, docs/guides/error-codes.md: storage_source_managed, storage_manifest_missing, storage_record_invalid, storage_record_secret, storage_card_exists and storage_revision_conflict, each with its recovery path. Remote register and update refuse with the shared trusted_local_only.
  • src/commands/storage.ts: gbrain storage init; the help separates the tiers report (status) from the shelf verbs.
  • Goldens: test/fixtures/goldens/cli/tools-json.json gains the three read tools; test/fixtures/goldens/exports/types.json gains the new error codes in the canonical union and lists ThinkSynthesisStatus in a new order (the new schema module declares the not_json literal first; the type is unchanged).
  • Tests: test/storage-manifest.test.ts (PGLite, repo-less) and test/storage-manifest-hosted.test.ts (managed PGLite with a Git checkout).
  • src/core/schema-pack/base/gbrain-base-v2.yaml: version 1.6.0 declares stored-object, so cards never read as an undeclared type and unify-types leaves them alone (test/schema-pack-stored-object-type.test.ts; doctor goldens regenerated for the new pack identity).
  • src/core/engine-sql/sources.ts: the local-path source list (autopilot's fan-out, cycle_freshness) leaves out the hosted shelf, whose cycle is always skipped. A register that loses a same-id race is refused storage_card_exists. storage-update rewrites a card whose stamped fields were removed by hand instead of failing.

A summary page built from a private note now stays hidden from other people after that note is renamed and another page takes its old name.

Fixed

  • Concepts from renamed private notes (GST-2687). gbrain writes short claim pages (atoms) from your notes, then groups atoms into topic pages (concepts). When a private note was renamed and a later rename gave its old name to a different, public page, the concept built from that note's atoms could be read by people other than the brain owner through get_page, search, query and list_pages. The atoms themselves were already hidden in this case; the concept is now hidden the same way, because gbrain follows the note's rename history to the private page it moved to. The brain owner, trusted local commands and a credential allowed to read private pages in that source still read it, exactly as before. Nothing is rewritten: the rule is checked when a page is read.
  • Doctor and repair. gbrain doctor's derived_visibility check now counts such atoms (looser_atoms) and concepts (looser_concepts). gbrain repair visibility now follows the same rename history, so it previews them and stamps them private; before, it judged them by the public page that took the old name and could stamp an unlabelled one world.
  • Schema migration v311 adds the database function gbrain_concept_inputs_private, which holds this check so search stays fast on large brains. It changes no rows. On a brain where the migration has not run yet, reads keep working and still hide these concepts.

[0.60.719.0] - 2026-11-14

A brain owner can now make member CLIs read-only: with access.enforced_member_writes set to refuse, every write under GBRAIN_ENFORCED_MEMBER refuses.

Added

  • gbrain config set access.enforced_member_writes refuse makes enforced member mode refuse put, link, unlink, tag, untag, timeline-add, remember, edit_page, append_page and git commit-push with named paths, with exit 3 and reason writes_off. refuse_pages refuses the same page writes and leaves named-path git commit-push to its usual checks, for a host whose member turns still commit files. The refusal comes before any page, path or source lookup, so it reads the same for a new, readable or hidden target and names none. The key lives in the brain's database, so it works on a brain with no repository, and a member can't change it because config is refused under the mode. Unset or checked keeps today's checked writes; any other value, or a config read that fails, counts as refuse. Reads, a CLI without the variable, git commit-push --pull and --dry-run, sync, sources refresh and write-request are unchanged. See docs/guides/enforced-member.md#writes-off.

Upgrading from an older release (about v0.60.150 to v0.60.248) no longer ends in exit 1, and doctor no longer reports a stopped serve as a second running build.

Fixed

  • gbrain upgrade run from an older release finishes with exit 0 again. Those releases load one more file from the new install into the old process after the swap, and that file pulled in code needing functions the old process didn't have, so the upgrade printed an internal_error and exited 1 even though the new binary and its migrations were in place. That file now loads nothing but the line it prints (GST-2688).
  • gbrain doctor's runtime_build_skew check no longer counts a process that has already exited as a running build. An older serve stopped with SIGTERM can leave its heartbeat row looking fresh for up to 30 seconds; a row whose process the check can prove is gone (same machine boot, pid gone or reused) is now skipped, while a row it can't prove dead still counts (GST-2688).

On a source whose pages are canonical in Git, a fact the Google loops extraction saves for a person or company with no page is no longer thrown away when the database has a brief hiccup while gbrain prepares the write.

Fixed

  • On a canonical: git source, a fact for an entity with no page is kept on a holding page (memory/unattributed, or its folder's holding page under a folder access rule) until the entity's page exists. Choosing that page takes one read of the access rules. If that read hit a passing database error while the write was being prepared (a cancelled statement, a stale cached plan, a lock or serialization conflict, a dropped connection, a pooler's client limit), gbrain treated it as "the access rules cannot be read" and refused the write for good with access_policy_pending. For the Google loops extraction, the only writer on that path, the commitment fact was dropped while the loop row stayed, so the open loop pointed at no fact.
  • Now that error is handed back to the write queue, which releases the write and prepares it again, the same way it already retries every other passing database error. The fact lands on the holding page with its entity, committed to Git, and the loop row points at it. A statement that keeps getting cancelled still stops after persistence.max_transient_retries tries and ends write_interrupted with nothing written.
  • Real refusals are unchanged. An exact page rule on the entity still refuses git_home_required, and an access state the database role really cannot read (for example permission denied on the access tables) still refuses access_policy_pending.
  • A hiccup on the same read earlier, while the loops job first hands the fact to gbrain, still loses that one fact as it did before. The loops job keeps the loop row and logs fact projection failed.
  • Facts dropped this way before the fix are not recovered by this change.
  • On the same kind of source, an owner's gbrain trust release of a held fact for an entity with no page used to report released while saving the fact only in the database, with no row in Git, whenever no holding page could carry it: an exact page rule on the entity, or access rules the database role could not read. Now the release refuses with nothing written and the fact stays held: git_home_required names the entity's page to create, and access_policy_pending asks you to retry. Creating that page does not unlock the release, because a page under its own rule counts as private for a release (page_not_found); drop the hold and remember the fact on the page instead. Facts released this way before the fix are not changed; gbrain repair facts-home lists them as folder_restricted. (GST-2676)

When a shallow checkout hides a folder's access-rule history, the first import now refuses with a code and the exact command to fix it, gbrain access rule migrate moves rules between git and the database, renamed pages keep their atom lineage, connection strings lose their password whatever the scheme, and a policy-only mirror no longer trips the git convergence check. Community wave W-A1, second part: bug reports verified on master and fixed with our own code.

Fixed

  • First access-rule import (GST-2638). The first import of a source's access rules refused when it could not read their git history. It now tells a missing history apart from an unreadable one. An unreadable history refuses with the new code access_policy_history_unreadable, and the fix it names is git -C <source root> fetch --unshallow and then the same command again. A plain sync on a shallow checkout whose rules do have committed history still fails closed. gbrain access rule migrate --to git|database moves the rules from one store to the other, the preview shows when only the owner can read pending pages, and the command exits 1 when the result would be invalid.
  • Policy-only mirrors (GST-2634). gbrain doctor's git_convergence check no longer flags a source whose mirror holds only access policy.
  • Ingest refusals (#8982). A refused POST /ingest now carries a code, a fix and a docs link.
  • Renamed pages, config secrets and root metadata deletes (#8920). When a later rename onto an old slug drops its alias, an atom is still judged by the private page its origin moved to. A concept is still judged as before: by its input atoms and their origins' rename aliases only. Config secret keys are redacted, and the password in a connection URL is removed for any scheme. Deleting a root metadata file publishes the change to the database only.

Name the room on the CLI, and think answers and writes carry room marks too

--room <list> now works on every gbrain operation command, and think, put_page and remember answers are marked for the room like reads are. A harness that runs gbrain as a command, or a person checking what a group chat could see, gets the same marks an MCP caller gets, and the agent writing into a room can tell which parts of a synthesized answer came from pages the room cannot read.

  • gbrain <command> --room alice-example,group:team-a prints exactly what the command prints without it. With --json every item carries room_access and the document carries _meta.room (a list becomes { results, _meta }); without it stderr names what the room cannot read. The local CLI marks the brain owner's view, or the member --as names; with no owner set, every item is unknown and the notice says to run gbrain access owner set.
  • A trailing --room with no value is never dropped. The command runs, every item is marked unknown (invalid), the degraded stamp shows and it exits 0.
  • It works through a running gbrain serve that owns a PGLite brain (the owner marks the answer) and on a thin client (sent as _meta.room). A host or owner that predates room marks leaves every item unknown (not_supported).
  • A think answer is marked at the strictest class of everything its synthesis retrieved or was shown (the anchor and the graph pages its prompt names, each under its own source); an answer built partly from takes or a calibration profile, or with no recorded evidence, is unknown. put_page and remember mark the page or fact they wrote and are never stamped degraded for it, so a committed write never reads as one to retry.
  • The shipped access-policy and agent templates now say how an agent acts on its owner's own turn in a group chat (at the owner's scope, written for the room, with truly private material offered in a private chat and never a refusal; everyone else is answered at the group's level) and that a send the principal asks for in their own verified message goes out as asked, with no approval phrase or draft-first step. Existing installs keep their own files.
  • Guide: docs/guides/rooms.md.

[0.60.718.0] - 2026-11-14

Three ways a commit could reach the remote past the push secret scan are closed, and the scan can no longer be skipped. In a managed root that sits in a subfolder of its repository, a commit could change that root's own .gbrain-scan-allow (for example to allow every path) and be pushed along with the secret it hid; it is now held for review like one at the top level. A strict (canonical: git) source's commit is now trusted only when its files are exactly what gbrain wrote, so anything a Git clean filter added is scanned. Git replace refs can no longer slip files into a page-write commit the scan trusts. commitAndPush also lost its unscanned mode, so no caller and no push handed over from a busy publish lock can publish without the scan; it reads only the commits the remote lacks, fetching a missing remote-tracking branch first.

Fixed

  • In a managed root below its repository's top level, the allowlist's path globs match paths relative to that root.
  • A strict-source commit made before this release that is still unpushed is scanned once when it is pushed.
  • A page-write commit whose staged file another process reset at the same moment is no longer reported as committed with the old content; the write fails and is retried.
  • The shared-checkout guide says that an uncommitted .gbrain-scan-allow is honoured, because it is the user's own answer to the scan.

On a source whose pages are canonical in Git, a fact the Google loops extraction finds for an entity with no page that an exact page rule names is now refused before anything is stored, the way remember already refused it.

Fixed

  • On a canonical: git source, a commitment fact for an entity with no page under an exact page rule (gbrain access rule add --slug people/bob-example --readers alice-example) used to be stored in the database only, with no file or commit holding it, so a rebuild from Git lost it. Once memory/unattributed existed, the same write was refused instead, with a message that named memory/unattributed rather than the entity. The write is now refused git_home_required (reason facts_lane) before any row is written, and the message and fix name the entity's page to create. An access state the database role cannot read refuses access_policy_pending (retry), and that message now names the entity too.
  • The loops job still records the loop. Its log line reads [loops_extract] fact projection failed … (git_home_required) (or (access_policy_pending)) and names the entity's page, so an operator can see which page to create.
  • Sources that are not strict, brains with no Git repository, and fact writers that do not keep a fallback entity behave as before.
  • Facts an earlier version already stored this way are untouched: gbrain repair facts-home --source <id> lists them as folder_restricted (page_rule) and its apply leaves them; creating the entity's page lets the extract_facts upkeep step fence them.

[0.60.717.0] - 2026-11-14

Saving a fact to a holding page is faster and keeps less history

Saving a fact to a holding page (memory/unattributed, or a folder's holding page) no longer slows down and fills the database as the page grows (GST-2497). Each save used to read the page's whole facts table about seventeen times and store another full copy of the page, so a page with thousands of facts took over a second per save and every save added a copy of it.

  • One save now reads the table at most three times. A parse of a given table is reused, and every caller gets its own copy.
  • A holding page keeps its newest five saved versions, the way large pages already do. Every fact row stays, the page still holds every row, and Git keeps every state where the brain has a repository. An older holding page sheds extra versions in batches of at most 100 on its next saves. get_versions and gbrain revert see only the five; an as_of read of the page says its history is incomplete, and gbrain repair unwithdrawn-facts lists that page's candidates as ambiguous (--include-ambiguous restores them).
  • After a sync, the held-fact move no longer plans a move on every sync when a folder rule covers a holding page but not the entity of a row on it. A rule that narrows a held row below its holding page's readers still moves it right after the sync.
  • On Postgres, the held-fact move's plan reads the write-through setting under the same deadline as its other reads, so a session holding the config table no longer keeps a sync past its preparation budget.

The owner can ask a shared agent in a group chat and still get the whole answer

A harness can now tell gbrain which group chat an answer goes into, and gbrain marks every page and fact with whether the people in that chat could have read it. The owner of a brain who asks their agent something in a team chat should get the same work as in a direct message, but the agent also needs to know what not to repeat in front of the room. Declaring the room does exactly that and nothing more. The read is never refused, shortened, reordered or failed because of it, and a call with no room returns what it always did.

  • Declare the room with the X-Gbrain-Room header (gbrain serve --http) or request-level _meta.room; a harness that can set neither may pass a room tool argument once whoami shows features.room: true. Entries are member ids, group:<id> or guest.
  • Each item in get_page, search, query, list_pages, find_pages, get_chunks, get_timeline, get_backlinks, recall and entity answers, each hot-memory fact and each saved fact a search returns (also shown on its line of the saved-facts block) carries room_access: {readable: true}, {readable: false, reason} (readers, folder_rule, labelled_only, private_label, derived, personal_source) or {readable: "unknown", reason}. _meta.room sums it up. A link, a timeline row or an entity-card fact is readable only when the room can read every page it depends on.
  • The room_above_level notice tells the model to use those items for the work, keep them out of the reply unless the ask needs them and offer them in a direct message. room_hints_unavailable says when the room could not be judged within 3 seconds or at all (the answer still goes out whole and on time, every item marked unknown, with the degraded stamp), and room_conflict says when a room argument disagreed with the harness; an argument can only add people.
  • The owner's own credential needs no may_act_for to declare a room, because a room only adds marks; any other member's credential needs may_act_for covering the room. Each marked call writes a room_marked row to the audience audit log with the room's principals and the counts, never page names or content.
  • No setting, no migration, and it works the same on a repo-less brain. Guide: docs/guides/rooms.md.

[0.60.716.0] - 2026-11-14

Model-written Life Chronicle events no longer read as your own notes after a trust claim, legacy bearer tokens keep their tools when the default MCP surface is narrowed, every way a page is hard-deleted now clears what the page left in six side tables, the secret scan catches passwords in escaped JSON and URL-encoded values, and the drift report records where it came from. Community wave W-A1: bug reports verified on master and fixed with our own code.

Fixed

  • Trust claim and legacy chronicle events (#9284). gbrain trust claim-sources lifted legacy Life Chronicle event pages (captured_via: life-chronicle:auto) to "your notes", even when the page they summarise is an imported third-party email, so remote agents got a model summary of a stranger's text unquoted. The chronicle capture kind is now capped at agent_written, and a claim never puts an event above its depth page's tier, including a depth page the same claim classifies for the first time. --relower --source <id> can lower what a claim lifted even after sources set-trust ended that claim. Lowering never raises a row and never changes who can read a page. Rows already lifted change only when you run the relower; gbrain doctor counts them and names the dry-run command.
  • Legacy bearer tokens and mcp.default_surface_dcr (#9266). Setting the DCR default surface also narrowed legacy bearer tokens, which have no client row to pre-seed, so they lost tools their grants allowed. Legacy tokens now keep the server surface (still bounded by their scope, source and operation grants, and still clamped by GBRAIN_MCP_FORCE_SURFACE); OAuth clients with no surface row still get the default. Doctor and the advisor already reported it this way.
  • Purge leftovers (#9146). Only delete --purge removed a page's proposed takes, open loops, core-edit notice text, aliases, redirects and rename rows. Tombstone expiry (including after a managed connector delete), the batch delete connectors use and a restore that replays a purge now remove them too, as does source removal for the two tables that did not cascade. Write-gate holds go with the page, except a hold that arrived, or was re-opened by the same content arriving again, after the soft delete, which is a new pending write and stays until its own disposition. Works the same on brains with no repository.
  • Secret scan (#8265). The high-entropy assignment detector missed a password whose key quote is escaped (JSON inside a JSON string) and one after a URL-encoded %3D/%3A separator. Both are caught, ordinary query strings and encoded URLs are not, and an encoded URL that carries a username and password still is. New transcript imports redact both shapes; pages imported before are not rewritten, so gbrain doctor asks once for a read-only, model-free transcript re-audit that lists them.
  • Drift report provenance (#9208). The dream drift phase wrote reports/drift-<date> with no derivation, so it was stored at tier unknown with no origin. It now records a derive:drift origin and takes the lowest tier of the pages, takes and timeline entries the judge read, capped at agent_written, with each of them recorded as an input.

[0.60.715.0] - 2026-11-14

Skills and harness: skills that match the CLI and update safely, OpenClaw and claude-cli fixes, and skill resolver additions. This release carries three PRs: #8918 (with #8190 and #8353), #8931 (with #8947) and #9095.

Skill checks, safe skill updates, the OpenClaw route and provider fixes (#8918)

Skills can no longer drift from the CLI or grow without a reviewer seeing it.

Every gbrain ... invocation a bundled skill shows now runs through the dispatcher's own subcommand and unknown-flag checks in CI, and every operation-shaped tools: entry must be a real operation. A PR that renames a command or flag fails until it updates the skills that name it, with the file, line, nearest valid name and the edit to make. Each SKILL.md over 16 KB has a size ceiling in skills/skill-budgets.tsv (today's size plus 10%), new skills stay within 16 KB and descriptions within 1 KB. CONTRIBUTING.md gains "Changing a skill", the update ladder adapted from skill-diet (MIT): fix it below the skill first, edit the existing line instead of appending, update the skill in the same PR as the code, and record Skill-size: before and after in the commit message.

What changed

Situation Before Now
A skill shows gbrain config sett ... or gbrain doctor --jsn passed CI; an agent following it got an error bun run check:skill-health fails with the nearest valid name
A skill declares a tools: entry that is not an operation passed fails with the nearest operation
A SKILL.md grows past its ceiling, or a new one passes 16 KB passed fails; raising a ceiling is a reviewer-visible edit
skills-commit-gate.sh four checks also runs the skill health guard and prints Skill-size: lines for staged skills

Installed skills get the same health check at runtime.

gbrain skillpack health [<skill>] checks the skills an agent actually loads (the harness or workspace skills folder, else the bundled skills) for size against budget, growth past 10% over the last diet, description size, missing triggers, and gbrain commands, flags or tools this gbrain does not have. Each finding names what it saw, what it expected, the fix (for an unedited copy of a bundled skill, gbrain skillpack reference <skill> first) and a verify command. It reads files only, so a brain with no git repository gets the same report. Folders without a skills manifest or skills.lock.json are checked only for copies of gbrain's bundled skills.

Situation Before Now
An installed skill still shows a flag a newer gbrain removed an agent following it got an error gbrain doctor warns (skill_health) with the line and the fix
An installed skill grew well past its last diet nothing reported gbrain skillpack health reports it; doctor shows it without warning
gbrain maintain, gbrain skillpack-check no skill information one-line skill health summary (skill_health in skillpack-check's JSON)

Installed skills you never edited can take gbrain's current version whole, and that update never overwrites a skill you edited.

gbrain skillpack reference <name>|--all --update-untouched replaces a skill only when every one of its files still matches what gbrain installed. Scaffold, sync and harness installs now record the sha256 of each skill file they write, per skills root. A skill with any edited file, a pinned skill (gbrain skillpack pin), or a file with no recorded base that matches no released version is skipped whole and its diff is printed. Installs from before bases were recorded can opt in with --accept-historical when every file equals a released version listed in the new skills/skills.lock.history.json. Every update writes a journal and backups to ~/.gbrain/skillpack-backups/<id>/ before it changes anything, rehashes each file right before replacing it, and runs the conformance, trigger and resolver checks; a failure rolls back only the files that still hold the update's bytes. gbrain skillpack restore <id> undoes an update and refuses files you changed since, unless you name them with --skip. It never prompts (--dry-run is the preview), works the same on a brain with no git repository, and gbrain upgrade writes no skill files.

Skill updates

Situation Before Now
You want gbrain's new version of skills you never touched --apply-clean-hunks applied hunks two-way and could not tell your edits from gbrain's --update-untouched replaces only skills whose every file still matches the install
A skill you edited or pinned --apply-clean-hunks could overwrite it hunk by hunk skipped with local_edits or pinned and its diff
An update fails a check or the process dies mid-update — rolled back per skill; doctor reports skillpack_update_incomplete with the restore command
Exit code of an update — 0 nothing failed, 1 with --fail-on-skipped when a skill was skipped, 2 on a failure
~/.gbrain/skillpack-state.json schema v1 schema v2 with per-skill bases and pins; a v1 file still loads

OpenClaw route

On OpenClaw, every skill gbrain installs now loads in one read, and setup and the install docs name the per-turn rule as the primary route, with skills as the fallback.

OpenClaw serves a skill only whole. A SKILL.md larger than one read (64,000 characters on a model with 200k+ context, 32,000 at 100k to 200k, 16,000 below that, and a 32,768-byte page when the context window is unknown) gets "cannot be partially served" and nothing else. company-brainify (37 KB) is trimmed to 30 KB with the same steps, gates and checks. A unit test now fails when any SKILL.md that gbrain skillpack scaffold --harness openclaw --all installs passes 32,000 bytes, or when anything gbrain writes for OpenClaw names skills_read, OpenClaw's deferred skill tool, which fails when called directly. In harness-evals, a listed skill was followed on 1-3% of turns and a per-turn rule on 97-100%. gbrain setup openclaw already installs that rule; it now says the rule is the primary route, skillpack scaffold --harness openclaw says its skills are the fallback, and INSTALL_FOR_AGENTS.md and docs/mcp/OPENCLAW.md lead with the rule.

Providers and skills

Six provider and skill bugs fixed (#8142, #8143), and Particle joins as a chat provider.

  • Particle is a chat and expansion provider. Set PARTICLE_API_KEY and use particle:particle-1 (the id Particle's responses report, particle-ai/particle-1, works too). It runs over Chat Completions at https://api.particle.ai/v1, with tools, the subagent loop, prompt caching and a 1,048,576-token context; point provider_base_urls.particle at https://us.api.particle.ai/v1 to keep traffic in the US. Spend caps price it at $0.25 input and $1.00 output per million tokens, with reasoning billed as output; cached input (listed at $0.03) is priced at the input rate until that rate is verified. A thinking: 'off' call such as an eval judge sends reasoning_effort: 'none' and keeps its requested output cap, and reasoning text never becomes answer text. No default model changed.
  • A System One provider that refuses every call no longer reads as "budget exhausted" (#7272). A decide call the provider refuses with a 4xx (402 billing, 401/403 auth, 400/404 model) never ran, so it now costs $0. Before, each refusal was charged its estimate, the daily budget filled with calls that never ran, and decide status, doctor and every slot reported budget_exhausted instead of the provider error. A 429, 5xx, timeout or transport failure keeps its charge.
  • An embedding migration keeps a provider URL set with gbrain config set (#7666). After the switch committed, the migration rebuilt its AI gateway from file and env config only, so a provider_base_urls.<provider> override in the database (a local Ollama on a non-default port) vanished and every re-embed failed with Cannot connect to API. The gateway now keeps the database overlay; file and env still win per key, and nothing from the database is written to the config file.
  • A deepseek: model pointed at another OpenAI-compatible endpoint gets room to think (#7662). With provider_base_urls.deepseek on another origin (Ollama Cloud, a proxy), gbrain still sent DeepSeek's thinking switch and the small judge cap that assumes it works, so a model that kept thinking used the whole cap and every facts relink judgement came back model_unparseable. That route now gets no switch and the thinking-model output headroom, priced the same way. The native DeepSeek origin, any path on it included, is unchanged.
  • openrouter:voyageai/voyage-* embeddings are priced (#7034). OpenRouter names Voyage's models voyageai/<model>, which missed the voyage: price rows, so a capped gbrain embed stopped with no_pricing. They now price from those rows (OpenRouter's rates are the same).
  • The bootstrap interview saves an OpenRouter key as openrouter_api_key (#7977). A sk-or-v1- key matched the OpenAI sk- prefix and was saved as openai_api_key.
  • daily-task-manager names its page by the slug ops/tasks (#7932). Its tool calls used ops/tasks.md, a different slug, so get_page missed the existing task list and the first-run rule could start a second one.

OpenClaw, Hermes, Windows and claude-cli harness fixes (#8931)

OpenClaw setups stop breaking each other, upgrades stop breaking OpenClaw, and four other harness fixes land. Six fixes for people running gbrain with OpenClaw. Setting up one agent's workspace no longer rewrites or removes the per-turn brain rule every other agent on the same gateway relies on. Background turns (heartbeats, cron, finished commands) no longer get told to answer. gbrain upgrade no longer reinstalls gbrain in a way OpenClaw refuses to load as a plugin. gbrain doctor stops failing on skills another tool put in the workspace. The OpenClaw guide now leads with an install route that works and documents OpenClaw's own remote MCP connection.

gbrain's background claude-cli calls no longer read your personal Claude Code instructions. On Claude Code 2.1.314 or newer the claude-cli provider runs its child with --setting-sources "", so your ~/.claude/CLAUDE.md, memory and settings.json stop steering gbrain's extraction and summary calls (and stop adding about 42k cached tokens to each). The trade-off: settings.json's env, apiKeyHelper and permission rules don't reach the child either; the subscription login is kept. To load them again, set GBRAIN_CLAUDE_CLI_USER_SETTINGS=1 in your shell or ~/.gbrain/.env. Older CLIs are unchanged.

Four more harness fixes: a per-turn hook you paste into Hermes by hand now counts, gbrain setup claude-code --instructions-only writes the per-turn rule hook on Windows, request_tools stops promising tools a stdio session cannot call, and gbrain schema show --source reads that source's pack.

  • gbrain setup openclaw --workspace <dir> leaves the gateway-shared turn-rule plugin alone when it would name another server, and --no-instructions there no longer removes it, or the brain-first block in other workspaces' AGENTS.md, for every agent; it says how to change it for all agents instead and exits 0. A plugin an older gbrain wrote for the same server is rewritten (reported by @Masashi-Ono0611, #8401).
  • OpenClaw heartbeat, cron and exec-completion turns get the brain-first rule without its closing "answer and finish" sentence, so they stop inviting replies nobody asked for (reported by @Masashi-Ono0611, #8440).
  • gbrain doctor reports skills in an OpenClaw workspace that have no resolver rows and no manifest.json as warnings, not errors, the way it already treats a foreign skills/ found from the current directory (reported by @Masashi-Ono0611, #8629).
  • gbrain upgrade reinstalls a Bun global install with --backend=copyfile, and its recovery commands say the same, so an install OpenClaw loads as a plugin keeps loading after an upgrade (reported by @Masashi-Ono0611, #8663).
  • The OpenClaw guide marks the ClawHub bundle route as not available: the ClawHub package named gbrain is an unofficial old build that fails to install. It leads with the stdio route (reported by @Masashi-Ono0611, #8470).
  • gbrain agent register --harness openclaw no longer says OpenClaw has no remote MCP client, and the OpenClaw guide documents connecting over streamable HTTP with openclaw mcp login (reported by @Masashi-Ono0611, #8747).
  • When your config.yaml already has a hooks: section, gbrain setup hermes prints the pre_llm_call entry for you to add. Once you paste it, gbrain doctor now sees it (and still reports it untrusted until Hermes approves it) instead of calling it missing (reported by @Masashi-Ono0611, #8857).
  • gbrain setup claude-code --instructions-only, doctor's turn-hook repair and the hook-launcher refresh now write the per-turn rule hook for a Windows gbrain binary (C:\...\gbrain.exe) instead of skipping it with "no absolute gbrain binary was found". The hook names its paths in Git Bash's /c/... form, which is how Claude Code runs hooks on Windows. Codex runs hooks on Windows without a POSIX shell, so it still gets the printed step, now with that reason (reported by @Masashi-Ono0611, #8451).
  • On a stdio session started at a narrower surface (gbrain serve --surface starter), request_tools {tools: [...]} lists only the tools the session can call. A tool outside it comes back with its schema, listed: false and a note to call request_tools {"surface":"full"} first, instead of a promise that it is in your tool list (reported by @Masashi-Ono0611, #8734).
  • gbrain schema show --source <id> shows that source's own pack (its per-source override, else the brain-wide pack) instead of failing with Unknown pack: <id>. A pack you name still wins (reported by @Masashi-Ono0611, #8900).

Itemized changes

  • src/core/setup/openclaw-turn-plugin.ts: the plugin's package.json records the access it was written for (gbrain.access); syncOpenclawTurnPlugin({ workspace }) returns print (unverified, naming the --remove command) for an installed plugin whose rule would change, falling back to the server or launcher in its RULE literal when it records none, and removeOpenclawTurnPlugin({ workspace }) skips; syncBrainFirst's opt-out no longer strips other OpenClaw workspaces' blocks; src/commands/setup.ts passes --workspace (not on --remove). The plugin's before_prompt_build returns renderBrainFirstTurnRule(access, 'openclaw', { reply: false }) when ctx.inputProvenance.kind is internal_system.
  • src/core/check-resolvable.ts: the low-confidence unreachable downgrade also applies to openclaw_workspace_* detection.
  • src/commands/upgrade.ts: bun update, bun add -g (stable, pinned and restore) and bunReinstallCommand / reinstallHintLines pass --backend=copyfile.
  • src/core/mcp-registration.ts: openclawThinClientBlock names OpenClaw's native remote MCP route.
  • Docs: docs/mcp/OPENCLAW.md, docs/guides/upgrades-auto-update.md, TODOS.md.
  • Tests: test/brain-first-turn-rule.test.ts, test/setup-brain-first.test.ts, test/check-resolvable.test.ts, test/mcp-registration-blocks.test.ts, and the upgrade suites that pin the child bun argv.
  • src/core/ai/providers/claude-cli-language-model.ts: runClaude adds --setting-sources "" when a cached claude --version (once per binary per process) reads SETTING_SOURCES_MIN_VERSION (2.1.314) or newer. GBRAIN_CLAUDE_CLI_USER_SETTINGS=1 skips the flag and the probe, and a cwd .env cannot set it (src/core/env-trust.ts). Reported and first fixed by @Masashi-Ono0611 (#8350).
  • Docs: docs/ai-providers/claude-cli.md (the trade-off), docs/guides/skillopt.md, docs/architecture/key-files/core-ai-continued.md.
  • behavior-changes.d/claude_cli_setting_sources_8342.md: the upgrade-notice row.
  • Tests: test/claude-cli-setting-sources-8342.test.ts (including the opt-in); the test/claude-cli-output-cap.test.ts stub answers --version.

Skill resolver: session cleanup, extra skill folders and top_gate (#9095)

The skill resolver's hook lane now cleans up old session files before it exits, each harness can add its own skill folders, and a new experimental top_gate setting can hold back weak suggestions. The skill resolver is the optional feature that offers your agent the right installed skill on each turn. It is still off by default, top_gate is off too, and none of this blocks a turn.

Before, the hook lane exited right after its turn, so idle session files waited for the next CLI run to be cleaned up. Now the lane runs one cleanup pass of at most 64 files before it exits, and never runs past its deadline. Before, Claude Code and Codex catalogs were built only from each harness's default skill folder. Now you can list extra folders per harness, and setup keeps them on every later run.

How to use it

  • Extra skill folders: gbrain config set skills.resolver.per_harness.claude-code.dirs /abs/dir,/abs/other (or .codex.dirs, at most 32 absolute paths), then run setup again. Setup builds the user catalog from the default folder plus every listed folder that exists, and the catalog's automatic rebuild keeps them. Unset the key and run setup again to drop them.
  • top_gate: gbrain config set skills.resolver.top_gate <0-100>, --top-gate on gbrain skill-resolver, or top_gate in the turn input. It empties the block when the best skill the block would show scores below the threshold, after dividing the score by the catalog's maximum idf so the number means the same thing at any catalog size. A skill named in the turn bypasses it. skills.resolver.per_harness.<harness>.top_gate overrides it for one harness.

Things to watch

  • top_gate stays off because no threshold has yet removed suggestions on no-skill turns without hiding many right ones. On held-out turns written to break length-based gates, threshold 1.33 kept false suggestions at 0.17 to 0.19 but cut hit@5 on skill turns from 0.515-0.611 to 0.26-0.28. The guide shows both measurements; treat it as a setting to measure, not one to turn on.
  • An empty block can now carry block_reason: "weak_top_match".
  • min_match behaves as before. Its docs now say it filters candidates and is not a gate.
  • A bad top_gate or dirs value makes the lane log skill_resolver_config_invalid and deliver nothing, like the other turn settings.

A keyless brain keeps itself tidy, and doctor says when upkeep last ran

gbrain maintain --unattended runs the free upkeep a brain needs, and is safe to run from a scheduler. It runs the same steps as gbrain maintain --safe (link and timeline extraction, stale-page checks, the free repair lanes) plus a new bookkeeping cleanup step. It never deletes a page and never starts a paid model call, with one exception you can switch off. On a brain with an embedding key, it embeds pages and facts that are missing embeddings, capped at $1 a day per brain (maintenance.embed_daily_usd, 0 turns it off). Two runs that overlap embed once, because the second skips that step. A run that is killed while embedding cannot spend that day's cap a second time, because the step books the rest of the cap before it spends and settles it afterwards. It works the same on a brain with no git repository and no API keys. A run that has nothing to do, or that finds the brain busy, degraded, behind on migrations or already looked after by gbrain autopilot, exits 0 and says why. A step that fails three runs in a row is retried once a week instead of every run, and doctor names it until the step succeeds or stops being part of upkeep.

  • gbrain maintain --status shows when upkeep last ran, what each step did, what failed, and when the next run is due.
  • A new doctor check, unattended_maintenance, reports the same thing. It warns when upkeep hasn't run for 48 hours or a step keeps failing, and its fix is gbrain maintain --unattended. A brain where maintenance.auto is false, or where autopilot runs upkeep, reads as ok.
  • gbrain maintain --purge previews which soft-deleted pages are past their 72-hour window. --yes --expect <hash> purges exactly the previewed set, matching each page by source and slug, and refuses with preview_changed when the set moved since the preview. Upkeep never purges pages on its own.

Smaller changes

  • The old bookkeeping cleanup inside the cycle's purge phase is now its own phase, bookkeeping_gc, with the same retentions. It never touches a page, so upkeep can run it while purge stays manual. purge's summary now counts only pages, sources and orphaned clone folders.
  • gbrain autopilot --install on a PGLite brain or a brain with no repository now refuses with a structured error (autopilot_pglite_daemon, autopilot_no_repo) that points at gbrain maintain --status, instead of a plain message.
  • doctor's cycle_freshness fix no longer asks for spend consent on a brain with no chat or embedding key, since nothing it runs there costs money.
  • The cycle's skip message for a phase that needs a filesystem no longer claims the reason is a Postgres or remote engine.

Itemized changes

  • src/core/maintenance/run.ts (new): the --safe step list (moved from src/commands/maintain.ts, which re-exports it), unattendedPlan, and runUnattendedMaintenance. The run record is maintenance.last_run in the database config, so repo-less brains have one. Brain-wide steps run once per run, then one item per source.
  • src/core/maintenance/status.ts (new): readUpkeepStatus, the one view behind --status and doctor.
  • src/commands/doctor/checks/unattended-maintenance.ts (new): the unattended_maintenance check.
  • src/core/cycle/bookkeeping-gc.ts (new) and src/core/cycle/phase-table.ts: bookkeeping_gc, and a tier (free, paid, deletes, filesystem) on every phase, which the upkeep step list is tested against.
  • src/core/cycle.ts: CycleOpts.stamp (full, source_only, none) so a partial run can't claim a full cycle, and deferFactEmbedding so fact extraction stays free on a keyed brain.
  • src/commands/autopilot-refusals.ts (new) and src/core/error-registry.ts: new codes autopilot_no_repo, autopilot_pglite_daemon, cycle_already_running, engine_degraded, maintenance_disabled and maintenance_owned_by_autopilot.
  • Tests: test/maintenance-unattended.test.ts (PGLite and Postgres) and test/e2e/maintain-unattended-keyless-pglite.test.ts.

Upkeep runs while your agent is connected, and --schedule covers the rest

An open gbrain serve now does the free upkeep for you, so a default install needs no service, cron job or second command. When the agent connected to gbrain serve (stdio or HTTP) goes quiet, serve runs gbrain maintain --unattended in slices of about 3 seconds, at most one full round a day per brain, and stops a slice the moment a tool call arrives. It writes the same record that gbrain maintain --status and doctor read, with runner serve. It never deletes a page, and on a keyless brain it never spends anything. When two serves share one brain, only one runs it and the other skips. A source too big to finish within 30 slices is parked: the rest of the round finishes, serve stops doing upkeep until the next round (it tries the parked source again each round and after each restart), and gbrain maintain --status and doctor warn and name it; gbrain maintain --unattended finishes it at once. A tool call that stops a slice never counts toward the 30. Turn it off with gbrain config set maintenance.auto false (or GBRAIN_SWEEP=0, which also stops the idle sweep).

  • gbrain maintain --schedule is for a brain with no agent session open most days. It installs a job that runs gbrain maintain --unattended every 6 hours (launchd on macOS, a systemd user timer on Linux, cron otherwise). It asks before installing, --dry-run shows exactly what it would write, and --unschedule removes it. Nothing installs it by default. A scheduled run that finds a live gbrain serve on the brain exits 0 with maintenance_owned_by_serve and leaves the work to serve.
  • gbrain init and gbrain setup <harness> now say how upkeep works in their first-run settings block, and how to turn it off.
  • doctor's overdue upkeep warning and gbrain maintain --status now also offer gbrain maintain --schedule as the second fix.

Itemized changes

  • src/core/maintenance/idle.ts (new): runIdleUpkeep (the stdio serve calls it from its idle tick, after the sweep) and armIdleUpkeep (the HTTP serve's new idle timer), with a 3-second budget per run. src/core/foreground-activity.ts gains onForegroundRequest, which aborts a run in flight, and src/commands/serve-http.ts marks each /mcp POST as foreground work.
  • src/core/maintenance/run.ts: a PGLite brain runs CHECKPOINT after a finished round, recorded as the checkpoint step. A budgeted run counts each item's unfinished slices and parks the item at UPKEEP_PARK_AFTER_SLICES (30; a fresh 10,000-page brain needs 24 quiet slices for its first round), recorded under parked. Without this, a source whose facts walk alone outlasts a 3-second slice (past about 12,000 pages) restarted forever and the round never finished.
  • src/commands/extract.ts, src/core/cycle.ts and src/core/cycle/connector-extract.ts: the cycle's stale link drain now stops between batches of 25 pages when the cycle is aborted. Before, a first upkeep run on a 10,000-page brain with no repository spent 64 seconds in this one step, past its 3-second budget, and tool calls made meanwhile took about twice as long.
  • src/core/cycle/extract-facts.ts: the facts walk now gives the event loop a turn every 50 pages. On PGLite, queries finish without ever letting timers or stdin run, so on a 10,000-page brain the walk froze the serve for about 2 seconds and a tool call sent during it waited until the walk ended (p99 1.7 seconds). With the fix, the freeze is under 150 ms, a call stops the walk, and tool calls during upkeep stay at p99 241 ms against 139 ms with upkeep off.
  • src/commands/maintain-schedule.ts (new): --schedule, --unschedule and the live-serve check, routed before the engine opens. New codes maintenance_owned_by_serve, schedule_unsupported (Windows and containers) and schedule_install_failed.
  • Tests: test/serve-idle-upkeep-gbra438.test.ts, test/maintenance-idle-gbra438.test.ts (PGLite and Postgres, including two serves on one brain), test/maintenance-park-gbra438.test.ts, test/extract-stale-managed-signal-gbra438.test.ts (PGLite and Postgres), test/maintain-schedule-gbra438.test.ts and test/e2e/serve-idle-upkeep-keyless-pglite.test.ts.

[0.60.714.0] - 2026-11-13

Background extraction no longer loops forever on a default brain, and seven dream and maintenance jobs that stopped or double-charged now finish cleanly.

Fixed

  • The stale extract job stops chaining itself without end (#9295). On the default config (frontmatter link sweep off), a page whose frontmatter named a page written later stayed "stale" after every sweep, because the check counted frontmatter references the sweep never re-resolves. The extract job then queued a new continuation every run. The sweep and doctor's extraction-lag count now count only the references the configured sweep re-resolves, and a continuation chains again only when it left fewer pages stale than the run before it. The frontmatter reference is left unchecked, not marked done, so turning autopilot.incremental_extract_include_frontmatter on still finds and links it. A brain already looping simply stops at its next run; nothing to run on upgrade. Managed brains re-resolve both kinds of reference and were not affected.
  • A released managed-atoms run can be resumed (#9079). Resuming a release after an interruption reused the original run's request id for a different write and failed with "This request_id was used for a different write". The release now has its own run key.
  • An edited managed page is resubmitted, not kept, when its activation is replayed (#9275). Activations now record the page's prior content, so a replay can tell an operator's edit from the run's own write. Activations recorded before this release settle the old way once.
  • A nested AI call that is refused before anything is sent no longer charges the outer budget (#9086). The outer hold now settles at zero usage when the inner guard refuses, and the reported remaining budget subtracts live holds, so a group no longer reads full while it still has room. A call that was actually sent and then failed still books its usage.
  • gbrain repair agent-session-facts works on segmented pages (#9027). It reads the page whole (up to 24 MiB) and writes it back bound to that revision, instead of reporting changed_since_preview on every apply. Larger pages are declined with page_body_segmented.
  • A Claude CLI reply that uses native tool-call markup is retried instead of read as a final answer (#9085). Tool calls the model wrote in a form gbrain cannot run used to end the turn as plain prose. The reply is now treated as a transient failure, so the job's normal retry limit (at most two more attempts) applies. Markup quoted inside a code span or code block still reads as an answer.
  • An extract job left with only blocked or held pages chains nothing (#8672).
  • doctor's stale-mention check reads segmented pages (#9121). It used to report links on segmented pages as stale because it never read their body.

[0.60.713.0] - 2026-11-13

A write killed in the middle no longer holds up its source

When a gbrain process dies while it holds the write at the head of a source's queue, the next write clears it at once instead of waiting. A crash or kill -9 used to leave that write's claim running with nobody renewing it, and every later write to the source waited behind it (write_pending) for the rest of its lease: 30 s while the write was preparing, about 6 minutes while it was publishing. On PGLite that wait happened on every such kill.

  • On a PGLite brain, the next gbrain process that opens the brain releases every claim written before it started and runs those writes again. The brain admits one process at a time, so their owners have exited. A process that cannot open the brain releases nothing.
  • On Postgres, a process killed on the same machine was already released at once; that is unchanged. A process on another machine, in another container or from before a reboot still waits out its lease, because gbrain cannot prove it is gone.
  • gbrain sources writer status now says so. A claim whose owner has provably exited reads owner_exited: true, resumes_on_its_own: true and owner_live: false, and its next is claim_owner_exited (nothing to do) instead of claim_running ("the owner is still renewing").
  • docs/guides/troubleshooting.md lists how long each case takes and what status prints meanwhile.

A write receipt says when its Git commit landed

A receipt read after the write (gbrain write-request <id>, get_write_request, list_write_requests) now reports where Git is. For a page on an ordinary (not canonical: git) checkout, persistence.git_state used to stay queued forever, even after the background Git effect committed the file, and no commit was named anywhere. It now comes from the Git effect: committed with the commit in persistence.git_commit, or superseded, skipped, failed or retrying with the cause in persistence.git_reason, and queued only while the effect is queued. Nothing stored is rewritten, so a Git effect from an earlier build, or one whose commit landed but whose push had to be retried, reads committed with no commit. Receipts of canonical: git sources are unchanged.

Reported by @ethanbeard (#9227).

Label repairs no longer refuse pages over gbrain's own stamps

gbrain repair visibility and gbrain repair private-labels now label a page whose file differs from the database only by stamps gbrain keeps in the database, and they hold the pages they can't edit instead of failing the run (GST-2603). A page retyped by a schema pack, a page with the validate: false grandfather stamp, or a page with a legacy_type retype record used to refuse every label repair, either as "cannot be made to the page's file line by line" or as source_changed. Enough of them in one window stopped the whole run as item_failures_systemic.

  • Those stamps no longer count as a difference. The page is labelled and its file still gains only the visibility line. A real difference (a date, the tags, the title, the body) still blocks the edit.
  • The preview checks every page in the window. A page whose file differs in a real key, or whose label line can't be edited as one line, is listed as held under details.held with the key that blocks it and its next step: gbrain sources reconcile <id> <slug> --preview for a file edited outside gbrain, or the line to fix. Held pages stay out of the preview hash and the approved set, so the rest of the window applies.
  • A page that starts refusing after the preview is declined with the new label_edit_refused code and nothing written. It never counts toward item_failures_systemic, the run exits 0, and the receipt names the pages that applied.
  • A brain without a repository writes the label to the database only, as before.
  • Labels and folder rules apply per source, so a page with the same slug in another source stayed readable after you labelled or ruled one copy, and a reader with no source filter found it by slug and in search. The previews of gbrain repair visibility, gbrain repair private-labels --class restore and gbrain access rule add / import now list those copies under open_twins: each same-slug page in another source that isn't private and that a caller outside the change could read, with their count and the command that covers them (the same repair or rule with --source <other>). The list is read-only and bounded, isn't part of the preview hash, and changes no copy for you.

Captured session files that mention a large entity page are extracted again instead of being stopped and later deleted.

Fixed

  • The session sweep no longer fails every window whose facts name an entity page stored in segments (a page over about 1 MiB). Facts the extractor left without an entity and that mention such a page were checked against the page's whole body, which a segmented page refuses (page_body_segmented) after the model call; the window failed five times, stopped, and the session file was deleted at the un-ingested retention limit. The check now reads the page the way its writer does, so the window commits its facts onto the page's ## Facts table, on a Git source and on a repo-less brain. Reported by @dvasselin (#9279).
  • A refusal that retrying cannot change now stops its window after the one paid attempt instead of five, retention keeps the session file it holds, and gbrain doctor (memory_writeback) counts these files as held. The first sweep after an upgrade retries each such window once, so a brain with files stopped by this bug recovers with no command.
  • A Gmail thread you purged while its import was still waiting to commit no longer leaves its open loops behind. gbrain waiting could keep listing the thread's subject and evidence quote after the purge; the sync now drops loops for threads it refused as purged in the same run when no page for them exists (#9011).

[0.60.712.0] - 2026-11-13

A page left behind when its file moved can be folded into the moved page

gbrain repair moved-pages folds a page whose file moved into the page the moved file imported as (GST-2584). When a file moved while sync held the old page's deletion (file_gone), the new path imported as a fresh page and the old page stayed at its old slug with its links, facts, timeline and saved versions. Neither side of gbrain repair file-drift was right for it: --prefer file dropped what only the old page held, and --prefer database wrote a stale copy of the file back.

  • The new repair pairs the two pages only on proof read from Git. The old path's last committed content must be what one new path took in the commit that deleted it, or the content of that path's first commit, made between the old file's add and its deletion. The old page must still hold exactly that content, so later edits to the moved file don't matter, and a page edited in the database after the move stays held as both_changed. A frontmatter id alone never pairs two pages, and the file name only explains why nothing matched.
  • The preview lists every page whose file is gone, with a reason when it is kept: several candidates, another old page claiming the same moved page, segmented storage, a difference in who may read the two pages, a database edit after the move, or a history too shallow to read. --apply --expect <hash> --yes folds exactly the previewed pairs with destructive consent. Each pair's proof and both pages are re-checked under their page keys, including that the old file is still gone. The fold moves the old page's links, relationship dates, timeline, raw data, attachments, user tags, aliases, facts and saved versions to the moved page, makes the old slug an alias, clears the hold and soft-deletes the old page, which gbrain restore brings back within 72 hours. No file is written. A brain without a repository has nothing to fold, and the preview says so.
  • gbrain repair file-drift --held now lists such a page as moved_to <slug> and prints the fold command instead of a recreate or a delete. The file_gone hold says when the page simply has no import record, instead of naming a database edit.
  • A file that still names the type gbrain-base 1.0.0 gave its path, equal to the copy an older gbrain stored, is no longer counted as a database edit when a later pack retyped the page. The import keeps the retyped type and records the rename as a legacy_type retype stamp, so later edits keep it too.

On a strict source, a fact for a page-less entity under a folder access rule now lives in Git on its folder's own holding page, with exactly the readers the entity's page would have. On a source whose pages are canonical in Git (canonical: git), a fact saved for an entity that has no page yet is held as a row of memory/unattributed until the page exists. An entity under a folder access rule was the exception. Naming it on that shared page would tell every reader the entity exists, so its fact stayed in the database only, and a rebuild from Git lost it. Now it is held on the holding page of the entity's own folder (partners/memory/unattributed for partners/acme-example). Every rule that covers the entity covers that page by path, so the fact has the same readers as the entity's page, nested and pending rules included, and it is stored the way the entity's page would be, either committed to Git, or database-only under storage.db_only or a path Git ignores. A rule added later moves the held facts it covers off the shared page as part of its own apply, and gbrain access rule add lists them in its preview first. An entity under an exact page rule (a rule on its own slug) has no such page, so remember now refuses with git_home_required (reason facts_lane) and a fix that creates the page; when the database role cannot read the access rules, it answers access_policy_pending with next: wait. Sources that are not strict and brains with no Git repository work as before.

Itemized changes

  • src/core/facts-fence.ts: isHoldingSlug (the shared memory/unattributed or any <folder>/memory/unattributed), its SQL twin and folderHoldingSlug. A folder's holding page reads only entity cells of its own folder; any other is a fence problem.
  • src/core/facts/held-fact-move.ts: holdingHome replaces holdsInGit and names the holding page, or refuses (page_rule, access_unreadable). moveHeldFacts reads every holding page, moves rows into entity pages, and re-homes shared-page rows whose entity a rule now covers; the shared page's row is removed, never struck with the destination. planHeldMoves is the one plan the move and the rule preview share, covering every holding page, in the order the move runs. An entity's struck (withdrawn) held rows leave the holding page with its live rows, removed rather than moved, and an entity with only struck rows gets its own removal. A move still pending when the write wait ends is reported (pending, with its request id) and not counted as moved.
  • src/core/persistence/memory-mutations.ts, memory-prepare.ts, facts-prepare.ts, facts-maintenance.ts, prepared-maintenance.ts, forget-prepare.ts, single-prepare.ts, page-identity.ts, model.ts, src/core/facts/*, src/core/trust/release-publish.ts: every fact writer, forget, purge and release treats any holding page as a holding page.
  • src/core/persistence/git-declared.ts, verb-errors.ts: the facts_lane refusal for an exact page rule carries a put_page fix; through remember it keeps the frozen invalid_params (and unavailable for access_policy_pending) with the canonical code, reason and fix.
  • src/core/persistence/coordinator.ts: a held fact on a database-only folder's holding page reports write_through: {written: false, skipped: db_only, held_for} instead of no_repo_configured.
  • src/core/search/private-visibility.ts: a held fact is also judged by its holding page as a page, so a raw-only rule (covers_derived: false) hides it as it hides the page; a fact still on a holding page for an entity a rule now narrows below that page (a folder rule over memory/unattributed, an exact page rule over a folder's holding page; a rule synced from Git, until the move) reads only for that rule's readers, and for nobody while the rule is pending.
  • src/core/entities/resolve.ts: a holding page under an entity directory (people/memory/unattributed) is never an entity page itself.
  • src/commands/access-rules.ts: the rule preview lists held_facts_moving and the apply moves them (held_facts: {moved, unmovable}), for the database and git stores; gbrain access apply and gbrain access resume move them too. Each listed fact's to comes from heldRowDestination, the resolution the move itself uses, so a row saved under a bare name shows the page that name resolves to. The preview now runs the move's own planner. It lists every row the apply moves (from and to, and because: its page exists for a row that moves only because its entity's page now exists), the struck rows that leave with them (held_struck_rows_leaving), and per page the rows and versions the receipts will record (held_versions_editing.pages: the versions carrying them now plus the one each move saves). The apply reports a pending move as incomplete with its request id.
  • src/core/facts/fence-fact-append.ts, src/core/persistence/prepared-maintenance.ts: a held row that leaves a holding page (the shared page or a folder's) for a page not every reader of it can read also leaves that page's saved versions, with the entity's struck rows, in the same transaction; every other byte of each version is kept, Git history is not rewritten, and the move's write receipt records it (held_versions_edited: the rows' fact ids, the row and version counts, where they went). The rule preview counts the rows and versions (held_versions_editing). A saved version stored in segments is not edited. The preview names how many per page (segmented_versions_not_edited, and a line saying readers of that page's saved versions can still read the moved rows there), and the apply's held_facts says complete: false with the same words (segmented_pages). A saved version with no timeline keeps a NULL timeline.
  • src/core/facts/purge.ts: purge's saved-version redaction moves, unchanged in SQL and transaction, into rewriteVersionBodies (its keepNullTimeline flag, which purge does not pass, writes a NULL timeline back as NULL for the held-row edit), which the held-row edit above also calls, so the page body reader ratchet still lists one saved-version reader.
  • src/core/persistence/prepared-maintenance.ts: the row hash a fact-fence adoption or held-fact move pins counts valid_from and valid_until to the second, the precision a fence cell keeps. A move accepted while the checkout is busy now commits under the request it named after the facts reconcile rewrites a new row's time to its cell's; before, that request ended revision_conflict and a second request made the move.
  • src/commands/sync/perform.ts, src/core/cycle/extract-facts.ts: after a managed sync releases its lease, held facts that a rule it applied narrows below their holding page's readers move at once (one query decides, and nothing is due when nothing changed); the maintenance cycle stays the fallback. That query and the move's plan, the entity resolver's reads for a bare-name row included, share one deadline, the sync preparation budget, and each of the move's own page reads ends at a budget of its own, so a lock another session holds on pages or page aliases no longer stalls the sync without end. A plan cut off at its deadline leaves the move to maintenance. Running it after the lease only spares a follower sync the wait, since the lease is re-entrant and the move's writes are not sync rows. Either path names the page when the move left a segmented saved version unedited (HELD_VERSIONS_NOT_EDITED in the cycle) or is still pending (HELD_FACT_PENDING).
  • src/core/repair/facts-home.ts, src/core/facts/facts-home.ts: gbrain repair facts-home writes a ruled entity's fact to its folder's page; folder_restricted now means only page_rule or access_unreadable, with sync.write_through off too (holdingPageFor).
  • Docs: git-canonical.md, entity-scoped-facts.md, access-designations.md, repair.md, write-refusals.md, memory-boundaries.md, MEMORY_VERBS_v1.md, error-codes.md (regenerated).
  • Tests: new test/folder-holding-page-gst2492.test.ts (held in Git before the reply and kept by a rebuild from a clone; three nested rules plus a pending one; no existence leak on get_page, search, query, recall, list_pages and find_pages; a raw-only rule; the rule apply's preview and move; the owner with the opt-in on and off; the page-rule and unreadable refusals locally and over MCP; database-only folders; background writers and repair; non-strict and repo-less unchanged; a segmented saved version named by the preview and reported incomplete by the apply, the sync path and maintenance; the preview's destination for a bare-name row equal to where the apply puts it; an exact page rule over a folder-held row, by CLI and by sync; the preview equal to the apply and its receipts; struck rows and a struck-only entity; a pending move; a NULL timeline kept); updated GST-2492 suites; PGLite and Postgres.

[0.60.711.0] - 2026-11-13

Upgrade and maintenance stability for self-hosters: upgrades are tested from real releases every night, health alerts page only on what needs someone, routine maintenance noise heals itself, an upgrade adds nothing you didn't turn on, the serve restart after a self-upgrade runs once per build without freezing the autopilot, a macOS brain on an external volume starts under launchd, a new doctor check counts session connections against the pooler limit, and twenty dream-cycle, doctor and repair fixes stop phantom work and false failures.

An upgrade could break a brain an older release created with no one noticing before its users did, a weekly health alert could read "score 30" because of the agent's own skills folder, a nightly cycle could end partial for reasons that needed no one, and one upgrade path installed a daemon on a brain that never asked for one. This release fixes all four.

On a brain with no code, resolve_symbol_edges used to stamp up to 2,000 chunks per source on every run, each one a non-HOT row update on the largest table. It now skips sources with no symbol edges and walks only chunks that own one, and code-graph readiness waits on the same chunks. Several phases also misreported their own results. A drain window closing at the wrong moment read as a non-retryable atom_drain_no_progress, an accepted concept publication still committing failed the whole concept phase, and the managed conversation-facts backfill printed "0 facts inserted" while its facts were still publishing. Each now reports what actually happened.

doctor embedding_state warned that thousands of deleted-page chunks were overdue and recommended a brain-wide irreversible purge that would remove almost none of them, because the purge holds derived retirements for 30 days. It now counts overdue chunks with the purge's own predicates and reports held chunks as information. autopilot --status and doctor's global_maintenance_stalled could keep a stalled verdict for up to a day after a clean pass; the pass now refreshes it.

Serve restart, macOS launchd and the session lane (#8144)

The serve restart after a self-upgrade now tries once per installed build, leaves alone a serve service installed for another brain, and no longer freezes the autopilot while it runs. #7478 had the relaunched autopilot restart the serve its installed service runs when that serve was still on the old build. Three gaps came out of its review (#8013). A restart that came back unverified was tried again by every autopilot relaunch, because the attempt was only remembered in memory. On macOS, where the autopilot doesn't filter processes by GBRAIN_HOME, one brain's autopilot could restart the com.gbrain.serve installed for another brain on the same OS user. And the restart command ran synchronously for up to ten minutes, which stopped the daemon's timers and signal handling while systemctl --user restart waited for the old serve to stop.

Now serve-restart-attempt.json under the brain's .gbrain directory remembers the build the restart was tried on, so each build gets one attempt across relaunches. The autopilot restarts the serve service only when its plist or unit starts this brain's own serve wrapper, and audits any other case as skipped / manual. The restart command runs asynchronously with a 120-second cap.

On macOS, an autopilot for a brain on an external volume now starts under launchd instead of exiting 78, and an autopilot pause no longer makes the worker restart itself every ten minutes or makes doctor call the held queue wedged. launchd opens a job's program and its log files before the job runs. When those sat on a volume other than the home folder, as they do for a brain on an external GBRAIN_HOME, launchd ended every spawn with exit code 78 (EX_CONFIG) and no pid, even though the same wrapper ran fine by hand. gbrain autopilot --install now detects a brain home on another volume, has launchd start the wrapper through /bin/bash, and writes the job's logs to ~/Library/Logs/gbrain/<label>.log and .err. The wrapper and every brain file stay under GBRAIN_HOME, and --status reads the new log path. A brain on the home volume gets the same plist as before. Re-run --install to move an existing job onto the new layout.

An autopilot pause (gbrain autopilot pause, or the marker a migration sets) stops the worker's job claims. The worker's own no-progress watchdog didn't know that. With jobs waiting and nothing completing, it emitted stalled and the autopilot restarted its worker about every ten minutes for as long as the pause lasted. Now the watchdog restarts its clock while either marker is present, so a long pause can't trip it the moment the pause clears. The database liveness probe keeps running, and a real stall after resume is still caught. Doctor's wedged_queue reports a queue held by a pause as a warning that says which pause holds it and how it clears, instead of failing with advice to restart the worker.

Contributed by @Masashi-Ono0611 (reports #8017 and #7966, and the approaches in #8018 and #7968).

New doctor check session_lane_budget counts this database role's session connections across every gbrain process against the pooler's limit. Every gbrain process opens its own pools (serve's ordinary pool, each process's direct pool, a sync's import workers), and a pooler's per-role session limit (Supavisor pool_size) is shared by all of them. Nothing counted them together, so a host running a serve, a cron sync, a jobs worker and an MCP stdio server could exhaust the lane (EMAXCONNSESSION) with every process inside its own pool and no warning before the first refusal. The check reads pg_stat_activity once (no query text), attributes each gbrain <kind>:<pid>:<nonce> connection to a process kind and host, counts pooler and other connections apart, and warns at 80% of the limit with the kinds holding the most. Behind a pooler, set GBRAIN_SESSION_LANE_LIMIT to the pooler's per-role limit; with no pooler it compares with max_connections less the reserved slots. PGLite: not applicable. No default changes.

What changed

Situation Before Now
A migration fails, or an upgrade rewrites page content, on brains an older release created found by users after the release a new nightly Upgrade matrix workflow installs older releases (one, ten and fifty back, about 30 days back, and stable) the way the README says, upgrades each with its own gbrain upgrade on PGLite and Postgres, keyed and keyless, with and without a Git repository, and checks install, serve restart, schema head, doctor, preserved pages, facts and sources, new writes and searches, and the old binary reading the upgraded brain
A health alert from gbrain doctor a score that counted every check, the agent's own skills folder included gbrain doctor --alert runs only the page-class checks, fresh, under a 30 s budget, and exits 0 (none), 17 (act, with each check's own fix) or 18 (unknown, with alert.consecutive_unknown)
A skills folder doctor found by walking up from the current directory that gbrain did not install graded against gbrain's conventions and lowered the score reported as "not a gbrain skills folder (
A dream purge or consolidate write accepted but still pending after the publish wait the phase failed and the cycle ended partial the phase warns dream_write_deferred with the receipt id and the next pass reconciles it; after 3 passes still pending, doctor dream_write_deferred names the receipt and pages
Every transcript in the synthesize window is already synthesized triage, the manifest context and the child queue still ran the phase returns no_candidates before any of that
A transcript grows between runs that keep dying each run had a new content key, so the 3-deaths breaker never tripped deaths also count per transcript file; the same 3-in-24h limit applies; gbrain dream reset-key accepts the dream:phase: keys --list prints
Pages whose chunks already hold current-model vectors but carry no or an old signature read as another model's space until a paid drain stamped them the embed phase stamps them first, up to 500 pages a run, with no provider call; doctor embedding_restamp_stalled pages when that keeps failing
A Postgres brain with sync.repo_path set and no autopilot, on upgrade the v0.11.0 migration's Phase F ran gbrain autopilot --install Phase F installs nothing (reason not_installed), prints the install command and doctor shows it as a minions_migration line; a brain that already runs autopilot gets it reinstalled as before

For operators

  • The page class (PAGE_CHECK_LEVELS): connection, migrations, write failures and stalls, build skew, managed sync stopped, maintenance stalled, corrupt JSONB rows, RLS exposure, the paid-loop breaker, dream_write_deferred and embedding_restamp_stalled; the opt-in autopilot install suggestion that minions_migration shows after an upgrade never pages. Every doctor report carries alert; health_score, brain_checks_score, category_scores and status are unchanged. Cron recipe and exit codes: docs/guides/troubleshooting.md.
  • New switches, on by default: dream.write_deferral.enabled and embed.restamp.enabled; false restores the old behavior. Every heal writes a row to the heal.audit config record (last 200 rows, 30 days). Nothing heals on upgrade, and repo-less brains get the same behavior.
  • The upgrade matrix runs nightly against the newest release, opens one upgrade-break issue on a red run and closes it on a green one; gh workflow run upgrade-matrix.yml -f candidate=<tag> runs it on demand, and pull requests touching migrations, the schema or the updater run a subset. Upgrading a v0.59.2.0 brain now passes the preservation check (#7814), so the matrix carries no known exceptions. Details: docs/ci-upgrade-matrix.md.
  • INSTALL_FOR_AGENTS.md has one Opt-ins list (keys and embeddings, autopilot, dream schedule, connectors, auto-upgrade, release channel), and gbrain setup <harness> prints one line pointing to it. The default keyless install starts no daemon, writes no service, connects nothing and keeps self-upgrade in notify mode, now locked by test/boring-default.test.ts.

Itemized changes

Upgrade restart, autopilot and doctor (#8019, #8024, #7302, from #8144):

  • src/commands/upgrade-restart.ts: the per-build attempt marker in restartServeAfterSelfUpgrade; the ownership check before restartStaleServeService collects processes; hostExecutor spawns asynchronously and kills a command past its timeout (120 s for the autopilot's serve restart, 600 s for gbrain upgrade --restart-only as before).
  • src/core/serve-service.ts: serveServiceRunsWrapper reads the launchd plist or the systemd unit and checks that it starts this brain's wrapper.
  • src/core/upgrade-restart.ts: RestartExecutor.run may return a promise, and runRestartSteps awaits each command.
  • Docs: the Restart section of docs/guides/upgrades-auto-update.md and the upgrade-restart.ts key-files entry.
  • Tests: test/autopilot-serve-restart-8013.test.ts.
  • src/core/autopilot-paths.ts: autopilotLaunchdHostPaths compares the brain home's volume with the home folder's; on macOS resolveAutopilotJob takes its log paths from it and sets launchdViaBash for a brain on another volume.
  • src/commands/autopilot.ts and src/commands/autopilot/jobs.ts: generateLaunchdPlist writes /bin/bash before the wrapper when launchdViaBash is set; installLaunchd creates the log directory; plistWrapperPath reads the wrapper after /bin/bash, so job ownership, status and uninstall keep working.
  • src/core/minions/worker.ts: the non-supervised stall watchdog treats autopilotPaused() as a hold.
  • src/commands/doctor/checks/queue-jobs.ts: wedged_queue returns warn with details.paused (operator or cooperative) while a pause marker is present.
  • Docs: docs/guides/live-sync.md (log location for an external brain on macOS), docs/guides/queue-operations-runbook.md (a paused queue), and the worker.ts key-files entry.
  • Tests: test/autopilot-launchd-external-home-8017.test.ts and test/autopilot-pause-watchdog-7966.test.ts.
  • src/core/session-lane.ts (new) and src/commands/doctor/checks/session-lane.ts (new): the session-lane census and the session_lane_budget check, registered after persistence_consumers; docs/ENGINES.md#pool-sizing names the remedies. The Postgres doctor goldens mask that check's live census.
  • test/cli-socket.test.ts classifies GBRAIN_SESSION_LANE_LIMIT (doctor only) and GBRAIN_NAME (a regex constant in core/session-lane.ts) as unable to change a socket read.
  • Tests: test/doctor-session-lane-gbra133.test.ts (fixed rows, real PGLite, real Postgres), added to the Postgres unit arms with its measured weight.

Dream cycle (#8326, from #8137):

  • Managed gbrain extract --stale counts pages waiting on an attendee's person page (skipped_attendance_incomplete) apart from pages edited during extraction (#6820).
  • Extract receipts no longer write the reserved type key into frontmatter, so doctor's stale_frontmatter_keys stops gaining a page every run (#7058).
  • resolve_symbol_edges stamps only chunks that own symbol edges and returns at once for a source with none; no migration and no bulk stamping (#7119).
  • The atom drain reports a window that closes before a batch starts as a window stop (exit 11, retryable), not atom_drain_no_progress (#7826).
  • The nightly reports/drift-<date> page counts as a derive-phase database-only page, so doctor's ghost_pages and undeclared_db_only_pages stop flagging it and repair ghost-pages stops proposing its deletion. The rest of reports/ is unchanged (#8054).
  • The nightly LongMemEval probe fixture anchors its relative-date questions with a question_date, so "last month" and "this year" are judged against the haystack, not today's date (#7308, #7501).
  • The nightly probe runs its questions one at a time, so worst-case cost reservations fit the run cap; each question's judges still run together and no dollar cap changed (#7534).
  • cycle.grade_takes.auto_resolve.enabled: true now takes effect in the scheduled cycle, as documented; it stays off by default and the 0.95 threshold is unchanged (#7542).
  • A concept whose member links were lost gets them back on the next run, with no model call and no duplicate links (#7548).
  • On a managed brain, a concept publication still pending after its wait is reported as publication_pending (warn) instead of failing the phase; the next run reuses it (#7547).
  • The managed conversation-facts backfill reports how many pages are still publishing instead of a bare "0 facts inserted" (#7500).
  • propose_takes drops a claim whose exact text an operator already rejected on the same page, so an unrelated edit no longer puts it back in the queue; reworded claims are still proposed (#7299).

Doctor, repair and maintenance (#8327, from #8137):

  • doctor raw_provenance exempts a generated page only when raw_trace_exempt is true; an explicit false no longer hides missing provenance (#8039).
  • doctor --fast --only <check> marks the report incomplete when the early stop skipped a selected check (#8114).
  • doctor maintenance_write_drift reports unknown (warn) instead of ok when candidates exist but none can be compared on this host (#8115).
  • gbrain repair <kind> --apply keeps its approved set when the last approved item fails, so the printed --expect command resumes instead of refusing preview_changed (#8058).
  • A clean brain-wide maintenance pass refreshes the verdict autopilot --status and doctor read, so a stalled warning clears with the pass (#7854).
  • doctor embedding_state warns only on deleted-page chunks gbrain pages purge-deleted would remove; deleted_page_chunks gains purgeable and held (#7825).
  • Maintenance preflight (dream lint and friends) recovers from a device-only physical-root change under the writer lock, as ordinary writes already did; any other identity change still refuses and a busy lock still fails closed (#7760).
  • Purging a deleted page drops that page's held fence generations, which no later write could settle, so doctor generation_held no longer warns about a page that no longer exists (#7774).

Gmail threads you purged stay purged, forwarded emails keep their body, live facts stay in recall, the MCP request log can be capped without losing its totals, a remote think runs on the host's own model, a moved brain folder names its way back, and more fixes across connectors, transcripts, facts, takes, links and code search.

Gmail's change feed reports a thread whenever anything on it changes, a label or a read flag, however old the mail is. gbrain imported every one of them, so a mailbox set to 90 days of history could gain years of old threads. Those threads now stay out, the same way the first backfill keeps them out. A purged Gmail thread that the change feed delivers again no longer writes its subject and quoted text back into open loops, on managed and unmanaged sources alike. Forwarded emails used to be saved with only the note typed above the "Forwarded message" line, or empty, because that line was read as the start of a quoted reply. The forwarded message is now kept. Two unmanaged Google cases that ended as held items only a person could clear now sync cleanly. One is a page whose checkout file was missing only gbrain's Facts table after one of its facts was forgotten, and the other is an item whose page the owner purged while Google kept listing it (it is skipped and stays purged).

On a GitHub source with a webhook, a refresh waiting for one issue swallowed the event for any other issue of the same source, and deleting an issue right after editing it refetched the page instead of removing it. Each item event now gets its own refresh or delete.

Fact extraction and capture now record who said what, and where a page came from, more faithfully. Lines you quote with > in a turn are no longer saved as your own facts, a Claude Code Remote Control session is yours rather than an agent's, facts that open with a log word ("Notes from", "Watched") link to the one entity they name, a facts batch refused for a stale embedding config lands once the write owner restarts with the fixed config, transcripts ingest --facts --json reports a budget stop instead of reading complete, a remote timeline entry, link or tag no longer overwrites how its page entered the brain, and a live fact whose context note happens to read "superseded by #N" is no longer hidden as superseded. The speaker labels "User" and "Assistant" no longer block a fact from linking to the person it names.

Paging through takes no longer repeats or skips takes that were written together, and rebuilding takes from a page keeps the "superseded by" link on struck takes. gbrain takes propose --accept no longer needs --dir and prints the accepted claim, and accepting a proposal that supersedes a fact works on a segmented entity page. gbrain reconcile-links --source <id> now links the pages of that source instead of reporting ok with no edges on a brain with named sources, and a write failure other than a missing code page now fails the run. Timeline entries built from inline [Source: …] citations keep the filenames, IDs and hashes written in backticks. Old timeline rows that an earlier parser split at a hyphen (foo-bar.de stored as foo / bar.de) are now removed by sync and gbrain extract timeline --prune-orphans instead of sitting next to the correct row forever. One "portfolio" in a Facts fence or Timeline row no longer types every person-to-company link on the page as an investment.

Hermes transcript snapshots no longer fail while Hermes is writing, and gbrain transcripts ingest --all keeps each file's discovered format. A managed gbrain import with sync.write_through off refuses once before admitting anything, gbrain connectors sync --dry-run writes no state, and a connector re-render keeps the trust_tier marker that gbrain repair provenance stamped, so doctor stops listing those pages again after every import. Open loops no longer gain a twin when the model rewords a loop's text over the same quoted line, withheld_noise lists every sender in JSON, more no-reply sender spellings count as noise, and gbrain google status probes Contacts with a call the read-only Contacts permission covers. Doctor's atom_provenance_drift no longer flags an edit that left the page body unchanged.

Ten contributor bug fixes from community wave 1 (#8149) and the #8128 follow-up waves A (#8323) and C (#8328) are rebuilt on current master: atoms follow a renamed origin page, a tombstone purge never removes a root sync metafile, nested config credentials stay hidden, a very long tag gets a clear refusal, the DB link path keeps links to child pages, and code search finds top-level callers and Elixir definitions. Each fix keeps its regression test, and the contributor credits stay with the people who found and diagnosed each problem.

You can now cap the MCP request log's size without losing the dashboard's totals or the audit trail. gbrain serve --http writes one mcp_request_log row for every call, tools/list included, and nothing ever deleted them, so a busy brain's log grew without limit. Retention is opt-in and off by default, so nothing changes until you run gbrain config set cycle.purge.mcp_request_log_retention_days 90 (30 to 3650 days; 1 to 29 is refused because the built-in usage reports look back 30 days). Then the nightly cycle deletes request rows older than the window, and a PGLite brain held by a live gbrain serve prunes itself once an hour in short batches. The surface_change, memory_setting_change and access_admin audit rows are never deleted, and each pruned row is counted per client first, so the Agents page's total requests and last-used times stay exact. A usage report whose window reaches back past deleted history says so: gbrain auth clients --usage names the --days that is complete, the advisor skips its "starter op unused for 90 days" half, and scripts/derive-starter-ops.ts refuses. Deleted request detail cannot be restored. gbrain doctor reports the log's size and suggests turning retention on once it passes about 1M rows or 1 GiB, asking you first. Proposed and first implemented by @lubosxyz (#6770).

A remote think now runs on the brain host's own think model unless the host allows the model the caller asked for. Every think answer is a paid call on the host's key, and before this change any MCP or HTTP client (or a thin-client gbrain think --model) could pick the most expensive model the host could reach. Now the requested model runs only when it is the host's think model (models.think, then models.default) or the host lists it in the new think.remote_models key (comma-separated, empty by default). Any other model, or a listed one the host cannot run (no key for its provider, or no chat support), no longer fails the call: the host model answers, warnings carries MODEL_OVERRIDE_IGNORED:not_allowed or MODEL_OVERRIDE_IGNORED:unusable, and the result carries a model_override_ignored notice that says how the owner can allow it. Remote rounds runs one round. The local CLI --model is unchanged and still exits 1 on an unusable model. A keyless brain answers a remote think the same way with or without model.

Moved your brain folder? gbrain now tells you how to bring it back. When a managed checkout was moved to a new path, or a folder above it was renamed, gbrain sources set-path and gbrain sources add used to refuse with a generic "identity changed" or "checkout missing" error that offered no way out, and their --dry-run said everything was fine. Now the preview and the real command refuse the same way, name where gbrain recorded the folder and where it is now, and give the exact commands that get you back without losing the folder's ownership or Git history. Nothing new can rebind ownership. The commands are the ones that already worked, and gbrain still checks everything they do.

A moved brain folder: how to get it back

Run the command you were going to run, with --dry-run:

gbrain sources set-path notes /Users/alice-example/Brain --dry-run

The refusal (recovery_required, detail physical_root_moved) carries its exit as a fix the agent asks you about first, because it moves your folders:

What happened The exit it names
Folder moved, its source still registered ln -s <new> <old>, gbrain sources set-path <source> <new>, rm <old>
Folder moved after gbrain sources remove mv <new> <old>, then gbrain sources add <id> --path <old>, which reattaches the folder's history
A parent folder renamed the same, on the parent folder
Copied to another disk and the original deleted restore the original, or switch to classic mode (gbrain sources writer deactivate --dry-run); a copy cannot be reattached

Things to watch

  • Existing pages of forwarded emails keep their cut body until the thread changes again. Nothing is rewritten on upgrade.
  • Old Gmail threads already imported through the change feed stay in the brain and keep updating when they change. Shrinking the window or a full sync does not remove them.
  • At 50 waiting item jobs for one GitHub source, a GitHub webhook for an item event that no waiting job covers now gets HTTP 429 webhook_backlog_full (Retry-After: 30), where it used to get 202 while the event was dropped. GitHub does not redeliver a 429 on its own: redeliver it from the webhook's Recent Deliveries, or let the next sweep or gbrain sync --source <id> catch it.
  • A live fact row an older build linked as superseded is cleared by a full reconcile (gbrain dream --phase extract_facts --source <id>); the nightly cycle clears it only when the page changes.
  • Duplicate open loops written before this change and person-to-company links already typed from a machine-written section stay as they are until their thread or page is processed again.
  • gbrain takes list --status <x> used to be ignored silently and now exits 1, pointing to --expired and gbrain takes propose --json.
  • A POST /ingest capture whose credential's write source is archived or removed before the capture lands now fails its job with source_changed (detail: webhook_source_unavailable) and writes nothing, where it used to be written under default, a source the credential may not be granted. Restore the source and run gbrain jobs retry on the job. Captures that do not come from a webhook keep the default fallback.
  • TypeScript, TSX, JavaScript, Python, Ruby and Elixir code files re-chunk once on the next sync (#5001, #7080). On a brain with an embedding provider that is one paid re-embedding pass over their changed chunks; gbrain embed --stale --dry-run counts them first.
  • MCP request log retention is off until you set cycle.purge.mcp_request_log_retention_days. A remote think that asks for a model the host has not listed in think.remote_models now gets the host model and a model_override_ignored notice instead of the model it asked for.
  • A moved brain folder: gbrain sources add --dry-run and gbrain sources writer claim --dry-run now refuse a folder that holds a copied ownership stamp, and gbrain sources set-path --dry-run refuses a missing checkout, as their real runs always did. They still write nothing.
  • A moved brain folder: a remote agent whose page write reaches such a folder sees only the detail, never a path.
  • A moved brain folder: a brain without a Git repository is unaffected, and the default keyless brain's gbrain sources set-path stays a plain path update.

Itemized changes

Connectors (#8941)

  • src/core/google/google-source.ts: processThread takes a cutoffMs; the history delta and held-thread retries pass now - g_history_days, and a thread whose newest message is older is settled without a page and without a hold unless it already has a page, which keeps updating (#8482). Reported and first fixed by @Masashi-Ono0611.
  • src/core/google/google-render.ts: trimQuotedReply keeps the whole body when a Gmail ---------- Forwarded message ---------- or Apple Mail Begin forwarded message: line comes before any reply marker; reply trimming, including a forward quoted inside a reply, is unchanged (#8719). Reported and first fixed by @javieraldape.
  • src/commands/serve-http-webhooks.ts: the item idempotency key carries delete or refresh, the per-source waiting cap is 50 (was 1), and a cap hit answers 429 webhook_backlog_full with Retry-After: 30 unless the waiting job it returns covers the event (the same item and intent, or an explicit source-wide sync for an edit; an automatic freshness sync, which can skip, covers nothing). In src/core/minions/handlers/sync.ts an item job that finds another sync of its source running is deferred 30 s at a time without burning an attempt, instead of completing as a source-wide skip, so its refresh or deletion is not lost (#8799). Reported and first fixed by @javieraldape.
  • src/core/facts/file-body-guard.ts, src/core/google/sweep-shared.ts: replaceableFile's alsoHeldAs receives each stored state, so the connector puts back that state's own Facts fence (the live row without the withdrawal overlay, versions with it) (#8645). Reported and first fixed by @Masashi-Ono0611.
  • src/core/google/sweep-shared.ts: the unmanaged importRendered screens the render against page purge tombstones (intentCarriesPurgedContent) before any drift check or temporary write and skips a match with one log line naming gbrain pages unpurge; it returns no slug, so the Gmail sweep runs no loop detection or extraction on that thread and drops its grace hold (#8644). Reported and first fixed by @javieraldape.
  • src/core/persistence/connector-sync.ts, src/core/google/sweep-shared.ts: on a managed source, an import the connector session skips as purged content returns purged: true on its skipped result, and importRendered returns no slug for it, so the same Gmail early return applies there (#8938).
  • Docs: docs/guides/github-source.md (backlog refusal, busy-source retry), docs/architecture/key-files/google-and-loops.md, new docs/architecture/key-files/google-sweeps.md.
  • Tests: test/google-gmail-delta-window-8482.test.ts, test/google-render-forward-8719.test.ts, test/serve-http-github-items-8799.test.ts, test/google-withdrawn-fence-carry-8645.test.ts, test/google-classic-purged-item-8644.test.ts, test/sync-handler-github-item-busy-8799.test.ts.

Facts and capture (#8971)

  • src/core/facts/writeback-gate.ts: an accepted ambient-writeback turn is banked and hashed without its > quote lines outside backtick or tilde fences (a line like >90% of traffic is still yours); the quoted-material check recognizes tilde fences too. Reported and first fixed by @Masashi-Ono0611 (#8410, #8412).
  • src/core/persistence/facts-maintenance.ts: a page-triggered or input-keyed facts batch refused embedding_configuration by the write owner's consumer is re-admitted from its retained facts (no new model call, retained visibility, unattributed facts included) once no full consumer that was running at the refusal is still live; until then a retry replays the stored refusal and spends nothing, a claim forgotten in the meantime stays forgotten, and a real model change still admits nothing. Reported and first fixed by @babyspectrum (#8526, #8554).
  • src/core/facts/subject-infer.ts: capture-stream openers (captured, watched, reviewed, logged; notes, capture, log, review, watch as labels) no longer veto subject inference as a competing name. Reported and first fixed by @mgunnin (#8656).
  • src/core/transcripts/ingest-facts.ts, src/commands/transcripts.ts: transcripts ingest --facts returns budgetExhausted, budgetReason, pagesSkippedTooLarge and pagesLockSkipped in --json and prints a line for each in the human summary. Reported and first fixed by @harjothkhara (#8722, #8757).
  • src/core/context/session-speaker.ts: a headless session that a claude remote-control server (or its rc, remote, sync, bridge aliases) hosts over --sdk-url is the owner's; a worker that session launches is still agent-launched. Existing speaker sidecars keep their first record. Reported and first fixed by @exGeni (#8080, #8892).
  • src/core/persistence/page-mutations.ts: only remote writes that supply page text (put_page, edit_page, append_page, capture) stamp mcp:<operation> provenance; other remote page writes keep the stored source_kind, ingested_via and ingested_at. Reported and first fixed by @zsimovanforgeops (#8667).
  • src/core/facts-fence.ts: only a struck (~~claim~~) row takes a superseded by #N link from its context cell, the way only a struck row reads forgotten:; a live row keeps the text as context and stays in recall. A live row an older build linked is cleared by a full reconcile, which gbrain dream --phase extract_facts --source <id> runs; the nightly cycle clears it only when the page changes, and on a managed brain an earlier reconcile receipt for an unchanged page can replay until then. Reported and first fixed by @alanshurafa (#9069, #9073).
  • Docs: docs/guides/ambient-writeback.md (Remote Control sessions under automatic detection), docs/guides/fence-format.md (the context column) and the KEY_FILES entries for the touched files.
  • Tests: test/facts/writeback-gate.test.ts, test/facts-embedding-refusal-heal-8526.test.ts (PGLite and Postgres), test/facts-subject-infer.test.ts, test/extract-conversation-facts.test.ts, test/agent-session-speaker-8064.serial.test.ts, test/put-page-provenance.test.ts, test/facts-fence.test.ts, test/extract-facts-phase.test.ts, test/fence-write.test.ts.

Takes, timelines and links (#8936)

  • src/core/engine-sql/takes.ts: listTakes ends its ORDER BY with t.id DESC, so LIMIT/OFFSET pages over takes sharing created_at are stable. Test: test/list-takes-tie-order.test.ts. Reported and first fixed by @Masashi-Ono0611 (#8883).
  • src/core/cycle/extract-takes.ts: the reconcile and rebuild walks write rows with takesPreparation.toCanonicalBatchInput, the projection page writes already use, so an inactive fence row citing superseded by #N keeps superseded_by = N. Test: test/takes-reconcile-superseded-pointer.test.ts. Reported and first fixed by @Masashi-Ono0611 (#8391).
  • src/commands/reconcile-links.ts: the CLI resolves one source (--source, then GBRAIN_SOURCE, .gbrain-source, the registered path and the brain default); the markdown scan, page reads and both edge directions stay in it; __all__ is refused (invalid_params); an addLink failure other than a missing target rethrows. Help line updated. Test: test/reconcile-links-source-scope.test.ts. Reported and first fixed by @babyspectrum (#8532).
  • src/core/timeline-citations.ts: a citation summary keeps the text of each same-line inline code span (backticks dropped, never emphasis-unwrapped); fenced code and spans holding citation or comment markup stay out. supersededInlineCitationEntries also returns the code-blanking reading, so rows stored by it are retired on reconcile. Test: test/timeline-citations-inline-code.test.ts. Reported and first fixed by @Masashi-Ono0611 (#8895).
  • src/core/timeline-extract.ts: supersededTimelineReadings (renamed from supersededCitationTimeline; persistence/canonical-projections.ts follows) adds the pre-#3280 bullet split, read by a linear scan, so retractRemovedTimelineEntries and the managed projection retire those rows when their detail is extractor-written. Test: test/timeline-legacy-split-7949.test.ts. Reported and first fixed by @Masashi-Ono0611 (#7955).

Transcripts, import, connectors, loops, facts and doctor (#8319, #8332)

  • src/core/transcripts/hermes.ts: when the store has both WAL sidecars, the Hermes adapter opens it read-only and runs VACUUM INTO a temporary file, so a checkpoint during the copy no longer tears the snapshot (SQLITE_IOERR_SHORT_READ) (#7067). A store at rest, or one in a directory gbrain cannot write, is copied into the temporary directory as before, and nothing is created beside the store. Any other snapshot error, a busy writer included, is reported instead of falling back to a copy that could tear. Tests: test/hermes-snapshot-7067.serial.test.ts, test/hermes-snapshot-readonly-dir.test.ts, test/hermes-snapshot-busy-rethrow.test.ts.
  • src/core/transcripts/ingest.ts, src/commands/transcripts.ts: transcripts ingest --all carries each file's discovered format instead of detecting it again, which failed on Claude Code logs whose first record is over 2 MiB; an explicit --format still wins (#7700). Test: test/transcripts-all-discovered-format-7700.test.ts.
  • src/commands/import.ts, src/core/persistence/import-mutations.ts: a managed import with sync.write_through off refuses owner_unavailable (detail: disabled_by_config) once, before admitting any file, and names the supported path (commit, then gbrain sync --source <id>) (#7085). Tests: test/import-write-through-off-7085.test.ts, test/e2e/import-write-through-off-7085-postgres.test.ts.
  • src/core/connectors/sync.ts: connectors sync --dry-run writes neither last_sync_at nor auth_error_at, and a preview whose listing stopped early returns partial with listComplete: false and wouldFetch (#7689, part 2). Test: test/connectors-dry-run-7689.test.ts.
  • src/core/persistence/connector-sync.ts, src/core/frontmatter-recovery.ts: a connector re-render carries the stored trust_tier marker forward only when it equals the tier the import assigns (external_untrusted), and trust_tier is a protected frontmatter key, so a marker can never raise trust (#8052). Tests: test/connector-trust-marker-8052.test.ts, test/e2e/connector-trust-marker-8052.test.ts.
  • src/core/google/loops-extract.ts: the extracted-loop dedup key is thread, direction and the normalized verbatim quote, and same-quote items in one response collapse; rows written earlier keep their key when their quote matches (#7289). Test: test/loops-extract-run.serial.test.ts.
  • src/core/ops/loops.ts, src/commands/loops.ts: withheld_noise returns every sender in JSON; only the human render caps at 10 and points to gbrain waiting --json (#7285).
  • src/core/google/api-access.ts: the google status Contacts probe calls people/me/connections?pageSize=1, which contacts.readonly covers; no scopes are added (#7490). Test: test/google-status-contacts-probe-7490.test.ts.
  • src/core/google/google-render.ts: isNoiseSender matches separator and localized no-reply local parts (no_reply, no.reply, do_not_reply, noresponder, notificaciones, notificacoes, ne-pas-repondre, keine-antwort) on the part before the @ only (#7856). Test: test/loops-noise-5012.test.ts.
  • src/core/facts/subject-infer.ts: the extractor's speaker labels no longer count as competing names, and names that start or end with a digit count only when they resolve to an existing page (#7827, items 1 and 3). Tests: test/facts-subject-infer-7827.test.ts, test/e2e/facts-subject-infer-7827-postgres.test.ts.
  • src/commands/takes.ts, src/core/take-proposals.ts: takes propose --accept resolves its own root (repo-less brains accept through the database) and prints the accepted claim; takes list --status exits 1 (#7271). Tests: test/takes-propose-cli-7271.test.ts, test/e2e/takes-propose-cli-7271-postgres.test.ts.
  • src/core/facts/proposal-supersede.ts: the managed decide_proposal preparer reads through wholeBodySnapshot, so a supersede_fact proposal on a segmented entity page within PAGE_MAX_BYTES no longer refuses page_body_segmented (#8060). Tests: test/trust-supersede-segmented-8060.test.ts, test/e2e/trust-supersede-segmented-8060-postgres.test.ts.
  • src/core/machine-sections.ts, src/core/link-extraction.ts, src/core/page-body/segment-links.ts: roleTriggerText drops the ranges rolePriorSuppressedRanges already excludes, for inline and segmented extraction; no extractor or pipeline version bump (#7882). Test: test/link-role-prior-trigger-7882.test.ts.
  • src/core/mentions/pass.ts: on a belief-covered source the mention pass locks the page keys and journals its link writes through unmanagedBeliefPoint(..., 'mention_pass'), so the drift check no longer reconciles a freshly published page as drift (#7031). Tests: test/belief-stale-links-7031.test.ts, test/e2e/belief-stale-links-7031-postgres.test.ts.
  • src/commands/doctor/checks/atom-body-recheck.ts (new), src/commands/doctor/checks/extraction-sync.ts: atom_provenance_drift finds the page_versions edition whose hash matches the atom's source_hash and compares its body with the live body, fences removed; equal bodies count under a new body_unchanged detail (at most 200 groups, 20 versions each). The doctor JSON goldens gain "body_unchanged": 0, and scripts/page-body-readers-baseline.tsv gains three INLINE rows for the re-check's body reads (#6825). Test: test/doctor-atom-provenance-drift-6825.test.ts.

Webhook captures

  • src/core/minions/handlers/ingest-capture.ts: a webhook capture (untrusted_payload, source_kind: webhook) whose write source is archived, unregistered or removed mid-write refuses source_changed (detail: webhook_source_unavailable) instead of writing under default, on managed and unmanaged brains. Test: test/ingest-capture-no-default-fallback.test.ts.
  • src/core/minions/handlers/ingest-capture.ts: the job normalizes its slug once (validateSlug) before its snapshot, applied-record and writer lookups, so a mixed-case X-Gbrain-Slug naming an existing page on a managed brain replaces that page instead of dead-lettering with revision_required. A trusted local tombstone for a mixed-case slug now deletes the lowercase page instead of skipping it. Test: test/ingest-capture-mixed-case-slug.test.ts.

Ingest and facts salvage wave (#8921)

  • Atoms follow a renamed origin page, and the privacy filter follows it too (#7462). After a sync rename, stale-atom repair, provenance drift, derived visibility, the private-visibility filter and visibility repair used to treat an atom's origin page as gone. They now follow the rename alias. An atom is private when either the old or the renamed page is private, so this only tightens; source_slug is never rewritten.
  • A tombstone purge never removes a root sync metafile (#7613). AGENTS.md and the other sync metafiles at the root of a checkout could match a tombstone and be deleted.
  • A second extract_atoms run of the same origin waits instead of failing (#7923), and the atom session waits out the previous batch's Git effects instead of failing on a holder about to release (#7461).
  • An interrupted conversation-facts backfill keeps its committed counts (#7068), and doctor's timeline_history recount shares the run deadline (#7038).
  • Claude Sonnet 5.5 and Haiku 5.5 get thinking-off controls they accept (#7567). Sonnet 5.5 rejected disabled with a 400. Opus 5.5 is tracked separately in #8317.
  • Credentials nested in config objects stay hidden (#7346). gbrain config get and config show used to print an S3 secret key inside the storage object in clear text, and even config get storage.secretAccessKey did. Both are redacted now, and --raw still prints the value. files upload-raw reports the storage backend it actually used, and on a managed brain it refuses before uploading instead of leaving an unrecorded object behind. The v0.9.0 migration guide shows how to configure S3, R2 and MinIO storage in ~/.gbrain/config.json.
  • gbrain agent run refuses a value flag written after the prompt (#7672). gbrain agent run "do X" --tools read used to treat --tools read as prompt text and run the job with every tool. Write the flag before the prompt, or put -- before the prompt to keep it as text. gbrain edge-proposals accept --limit=5 12 and its sibling subcommands now say to drop the flag, instead of advising --limit 5, which would have acted on proposal 5.
  • --flag value and --flag=value mean the same thing in more places (#7624). gbrain doctor --scope brain and gbrain models doctor --skip <provider> work with a space as well as with =, an unknown scope is an error, and gbrain jobs watch --refresh-ms 250 changes the refresh rate instead of being ignored.
  • A very long tag gets a clear refusal (#7052). A tag over 2688 bytes used to fail a page write with a confusing storage_error. put_page, add_tag, capture and import now refuse it as tag_too_long, naming the tag and its size, and sync holds the file and keeps going.
  • gbrain extract links --source db keeps links to child pages (#6983). A markdown link like [note](notes/first.md) resolves from the linking page's own folder first, then its parent folders, the way --source fs does. If no page exists at that path, gbrain still falls back to the root-relative page, and the write-time link check agrees with the graph.
  • code-callers finds calls from top-level code (#5001), and Elixir works with code-def, code-callers and code-callees (#7080). TypeScript, TSX, JavaScript, Python, Ruby and Elixir files re-chunk once on the next sync. On a brain with an embedding provider that re-embeds the changed chunks of those code files once, so a brain with large code sources pays one embedding pass over them; gbrain embed --stale --dry-run counts them first.
  • Runtime and storage fixes. gbrain jobs supervisor start --detach works with the compiled binary (#7532). Supabase Storage sends publishable and secret keys only in the apikey header (#7205). gbrain brainstorm and gbrain lsd cache the slug-prefix list in memory instead of in the config table that every command reads (#7838).
  • Advisors and repair give advice that works (#7281, #7282, #7137). The stale-sync advisor skips archived sources. The timeline-coverage advisor counts meeting, conversation and calendar pages by their own date. An auto-tier gbrain db-repair --yes that can't reach a database that keeps refusing connections tells you to check the database server instead of running the same repair again.
  • Smaller fixes. Link reads count hidden former relationships in SQL instead of fetching them (#7075). Autopilot logs auto_drain_window_clamped when it shortens an atom drain (#7076). loops list and open_loops say when a list was cut short and how many matched (#7291). A null api_error_status counts as no status in the Claude CLI output-cap match (#7273). The Hermes turn rule names gbrain's tools mcp__gbrain__recall (#8068). gbrain init --help names starter as the default registration surface (#7983). The 0.53.0 migration's refusal for an unclaimed or inactive source names the preview to run first (#7915). The ci:local runner installs curl (#7765). Docs: the topologies guide (#7733) and the AlphaClaw/OpenClaw onboarding known issue (#7965).

Request log retention (#8952, from #9145)

  • src/core/mcp-request-log-retention.ts (new): pruneMcpRequestLog deletes rows past the cutoff in batches of 2,000 (at most 200 per run, 400,000 rows), each batch one statement that picks the oldest rows in created_at order with FOR UPDATE SKIP LOCKED, skips AUDIT_LOG_OPERATIONS, deletes them and adds them to mcp_request_log_rollup, so a failed tally deletes nothing and two prunes at once never double-count. runMcpLogRetention reads cycle.purge.mcp_request_log_retention_days and records each run under cycle.purge.mcp_request_log_last_prune. startMcpLogPruneTicker is the hourly PGLite prune with a 3-second budget that yields between batches. lifetimeUsage adds kept rows and the tally.
  • access_admin, the owner access administration audit row (GST-1094), is in AUDIT_LOG_OPERATIONS too, so retention never prunes it and usage reports never count it.
  • Migration v310 adds mcp_request_log_rollup (token_name, requests, last_used_at, pruned_through) with RLS on. It copies and deletes nothing.
  • src/core/cycle.ts: the purge phase runs the prune when retention is set and reports purged_mcp_request_log_count when it removed rows.
  • src/commands/serve.ts: both gbrain serve and gbrain serve --http start the hourly prune on PGLite, paused while a delegated sync runs.
  • src/core/mcp-usage.ts: readClientOpUsageWithCoverage returns each window's coverage (truncated, covered_days, pruned_before) and truncatedWindowNote the fix and recovery sentence; AUDIT_LOG_OPERATIONS names the never-pruned rows. readClientOpUsage is unchanged.
  • gbrain auth clients --usage prints window truncated: ... with the complete --days; --json adds window_truncated, covered_days, pruned_before and window_note. scripts/derive-starter-ops.ts exits 1 with the note on stderr for a truncated window. The advisor's mcp-client-fit drift finding skips its unused-member half and says why.
  • Admin dashboard: /admin/api/agents totals and OAuth last-used read the log plus the tally; /admin/api/requests adds retention_days and pruned_before, which the Request Log page shows.
  • gbrain config set cycle.purge.mcp_request_log_retention_days refuses values other than 0, off or a whole number from 30 to 3650, and writes nothing.
  • Doctor mcp_request_log_retention (new, ops): log rows and size; past about 1M rows or 1 GiB with retention off it warns, and its fix asks the user before turning retention on and is never a doctor --remediate step. With retention on it reports the last prune and the pruned-before date, and warns when no prune ran for 48 hours.
  • Docs: docs/mcp/ADMIN.md ("Request log retention"), docs/operations/mcp-surface-runbook.md, docs/architecture/system-of-record.md.
  • Tests: test/mcp-request-log-retention.test.ts (PGLite and Postgres, repo-less: off by default, every row class tallied exactly, audit rows kept, capped runs resume, a failed tally deletes nothing, a missing tally table skips, config set refusals, the PGLite ticker under a live serve --http, two prunes at once on Postgres), test/serve-http-admin-request-log-retention.test.ts, test/doctor-mcp-request-log-retention.test.ts, test/mcp-request-log-retention-cli.serial.test.ts (auth clients and derive-starter-ops through the real entry points), and new cases in test/mcp-usage.test.ts and test/advisor-mcp-client-fit.test.ts.

Remote think model (#9050, from #9145)

  • src/core/ops/takes.ts: remoteThinkModel resolves the requested model, the host think model and each think.remote_models entry through the same alias resolution and normalization, then runs the requested model only when it is the host model, or when it is listed, passes probeChatModel and has its provider's key. For remote callers, the #1698 explicit-model error no longer applies and rounds is 1.
  • src/core/interop-notices.ts, src/core/error-registry.ts: the model_override_ignored info notice (per session for each reason); for not_allowed its fix is gbrain config set think.remote_models <models> on the brain host (paid consent).
  • src/commands/think.ts: the thin client turns the server's MODEL_OVERRIDE_IGNORED:* warning into the same notice, and --model help says what a thin client gets.
  • src/core/config-runtime-keys.ts: think.remote_models is a registered key; no remote operation writes config, so only the host sets it.
  • src/core/behavior-change-notice.ts: one row for the remote think model rule. Docs: docs/guides/remote-mcp.md security posture.
  • Tests: test/think-remote-model.test.ts (PGLite and Postgres, remote MCP dispatch on a repo-less brain): an unlisted model, a listed model, a listed model the host cannot run, a listed model whose provider key the host lacks, the host model named explicitly, remote rounds, the local operation keeping #1698, and a keyless brain with and without model; test/think-remote-model-thin-client.serial.test.ts: the thin-client --model notice; test/think-pipeline.test.ts pins the #1698 error on the local caller and no error for a remote one.

A moved brain folder (#8950, from #9145)

  • src/core/persistence/physical-root-record.ts: describeMovedPhysicalRoot (read-only file evidence: the stamp's recorded root is gone, and the reservation of that root, beside it or beside the new root after a parent rename, carries the stamp's token, brain and worktree; sameDirectory from device, inode and birth time), physicalRootMovedError (path-free) and physicalRootMovedRefusal (the local CLI refusal with an ask_user fix chain in argv form, paths shell-quoted in the prose). An ownership stamp other users can read at the new path now refuses ownership_record_not_private instead of the generic error.
  • src/core/persistence/physical-root.ts: movedPhysicalRoot accepts the evidence only for this brain, a worktree this host owns and a host binding still at the recorded root; reservePhysicalRoot raises the path-free refusal on its copied-marker branch.
  • src/core/persistence/source-lifecycle.ts, topology-locks.ts, administration.ts: add, claim, rebind and writer claim build the full refusal (never for the automatic first-write claim); their dry runs predict it, and a rebind dry run of a missing checkout refuses source_checkout_missing like the real run. When the automatic first-write claim finds a missing or moved checkout under the source's folder, it now refuses path-free too (physical_root_moved, or the path-free source_checkout_missing). Before this, a remote page write in that layout was told the checkout's host path.
  • src/core/persistence/verb-errors.ts, src/core/error-registry.ts: the physical_root_moved reason and its explanation; docs/guides/error-codes.md regenerated.
  • Docs: docs/architecture/topologies.md#moved-root, docs/guides/write-refusals.md#moved-root.
  • Tests: test/physical-root-moved-8625.test.ts (PGLite and Postgres: leaf and parent moves, bound and removed sources, Git and plain directories, a copy on a new inode, two sources in one worktree, writer claim --dry-run, five lookalikes that keep the generic refusal, a readable stamp, the classic keyless set-path, and a remote put_page whose refusal carries no path or worktree id, before activation and on an activated brain with a moved checkout under the unbound source's folder).

Proposed and first implemented by @Kkkkkuro (#8625).

[0.60.710.0] - 2026-11-13

Cursor can keep a project's pinned skills in front of the agent

gbrain setup cursor --skill-loading resolve, run in a project, gives Cursor the skills you pin for that project on every request. Cursor gets no per-turn suggestions: none of its hooks can add text to the model's request.

  • Setup builds a catalog for the project from the skill folders Cursor loads there (.agents, .cursor, .claude and .codex skills in the project and your home directory) and prints its name. gbrain skills pin add <id> --catalog <name> checks the id against that catalog.
  • The pins go in their own always-applied rule, .cursor/rules/gbrain-pinned-skills.mdc, with each skill's SKILL.md path. In a git repository its path goes in .git/info/exclude, so it is never committed. Your own rules are never touched, and with no pins there is no file.
  • gbrain doctor reports Cursor as partial: pinned only, no per-turn hook, unverified. Nothing on this machine can see Cursor's request to confirm the rule reaches the model.
  • The pins sit beside the skills index that --skill-loading spring writes into gbrain's brain-first rule: on Cursor, resolve leaves that index in place, and --skill-loading spring leaves the pinned-skills rule in place.
  • gbrain setup cursor --skill-loading off removes the rule and its exclude line; the pins stay in config.json. Nothing opens a brain, so a keyless or brainless install stays as it is.

Private pages stay private in more places, search finds the right page more often, and gbrain warns when one credential would let an agent show one person's chat to another. This release brings together four changes to access and search privacy. Database functions can no longer be steered by a temporary table another session creates. Trust review shows exactly what you are confirming. A token with the private-pages grant stops seeing an old private marker through page history. Setup and doctor flag a single shared credential behind an agent several people talk to. Search and recall fix eight answers that were quietly wrong or quietly costly, search.mode tokenmax brains can see and fix pages that never reached the synopsis tier, and two opt-in switches let a long alias or a name inside a longer query find its page.

What changed for you

One credential behind an agent several people use is now flagged. A credential reads the same pages whoever talks to its agent: an unbound one reads everything its grant reaches, and one bound to a member reads what that member may read. In a host-harness pilot, one bot token over a brain holding many channels' chat made agents answer people with other people's messages. The new unscoped_multi_user_token safety notice fires from gbrain setup openclaw|hermes|grok-bot|muse, from setup claude-code when a chat-channel plugin is enabled, from mcp grant and connect for a bot harness or a grant that reads chat from 2+ channels, from skillpack scaffold into an OpenClaw or Hermes workspace, and from auth create and auth register-client when the new credential reads chat from 2+ channels. It fires only on a brain that serves several people (gbrain access serves); on a one-person brain the routes stay silent and the doctor check is not applicable. Its fix leads with binding the credential (gbrain auth rescope --token <name> --member <id>, --may-act-for <members> for one agent several people share, or gbrain setup <harness> --serves several for a local agent), and names a source rescope, sources unfederate or GBRAIN_SOURCE only as narrowings that leave the credential unbound. The read-only doctor check unscoped_multi_user_token runs the same detection over every active token, OAuth client and wired local agent. No grant, default or exit code changes.

Database functions keep pg_temp last on their search path. Every gbrain function definition now pins pg_catalog, public, pg_temp itself, so a schema replay or gbrain reindex-search-vector no longer drops what migration v277 added, and bun run check:search-path refuses a definition without it. The end-of-migration hardening retries a concurrent catalog update and skips itself after a lock refusal instead of waiting a second lock timeout. Doctor function_search_path points at gbrain apply-migrations --yes below v277, and past it prints the ALTER FUNCTION statements to run as each function's owner.

Trust review shows what you confirm. Release and drop previews quote every free-text field and strip invisible characters. gbrain trust review --accept-all skips, and never applies, an item it cannot preview in full. A page the write gate flagged as instructions is listed by trust review (flagged_page) and is never part of accept-all; confirm it on its own with gbrain trust confirm after reading it.

The private-pages grant closes a history gap. get_versions hides an older version that carried the pinned_question marker from a token with the private-pages grant, even after the live page dropped the marker. Rotating a harness token still never carries the grant, and now prints the gbrain auth rescope command that grants it again. docs/guides/multi-source-brains.md lists which reads the grant opens and which keep hiding private pages.

Search and recall answers that were quietly wrong now behave.

You ask for Before Now
search with match: "keyword" on a slow brain "Keyword match count unavailable" the total
keyword search for acme-example, a company with an alias the alias matches, not the company page both, counted
get_links to districts/pa-<long name> districts/<REDACTED:voyage> the slug
think over a take held Sep to Jan a claim with no dates since=… until=… source=…
gbrain eval trajectory with an annual then a monthly MRR a 92% "regression" nothing flagged
repeating a query in brain B after brain A closed mid-embed a second paid embedding call the cached vector
gbrain import as a non-owner role after the owner ran ANALYZE "Projection planner statistics were not refreshed" every time quiet; a large pass still warns once
open_loops when the sources table cannot be read "no google source", coverage complete no_google_sources: null, coverage partial, a notice to retry

Recall also drops bidi and control characters from evidence titles, a saved fact a correction superseded no longer comes back through entity facts, and sources_not_readable counts only pages search returns.

search.mode tokenmax brains can see and fix their synopsis coverage. Import and sync embed tokenmax pages at the title tier and leave per-chunk synopses to a job that nothing ever queued, so a brain could search on title-tier vectors for weeks while reporting tokenmax. Doctor synopsis_tier_coverage now reports chunk and page coverage and warns below 50%. gbrain reindex --synopsis previews the pages, tokens and dollars and spends nothing; --yes [--max-usd <usd>] runs the sweep under one spend cap and resumes where it stopped. gbrain config set search.mode tokenmax names this command instead of reindex --markdown. Nothing is scheduled automatically and no default changes.

Two opt-in switches let a long alias, or a name inside a longer query, find its page. With gbrain config set search.qualified_aliases true, a 3 to 6 word phrase inside a query of up to 14 words that is the alias of exactly one page you can read brings that page to the top. With gbrain config set search.title_qualified true, a person or company page whose title appears inside the query gets the title boost once. Both are off by default; with them off, results and cache keys are unchanged. Neither marks a page create_safety: exists. A common three-word alias can lead a query about something else, so check your own queries before turning them on. Proposed and first implemented by @calebhicks.

All of this works the same on a brain with no git repository.

Things to watch

  • takes_search rows now carry since_date, until_date and source, as takes_list already does.
  • A trajectory regression now needs two observations with the same unit and period at different dates.
  • On a brain that never ran v277, or whose functions another role owns, doctor function_search_path warns until the printed ALTER FUNCTION statements run as the owner.

Itemized changes

  • src/core/function-search-path.ts, src/core/migrate.ts, src/core/schema-migrations/v277-function-search-path-pg-temp.ts, scripts/check-search-path.sh, src/schema.sql and every schema module that defines a plpgsql function now pin pg_catalog, public, pg_temp in each definition; hardenFunctionSearchPaths runs v277 and the end-of-run step in one short transaction under the migration timeouts, retries a concurrent pg_proc update and turns a lock it cannot take into schema_lock_blocked. src/commands/doctor/checks/function-search-path.ts names each function still on a legacy path and the ALTER FUNCTION to run as its owner.
  • src/core/trust/review.ts, owner-actions.ts, release-publish.ts, src/commands/trust.ts, src/commands/doctor/checks/trust-release-unpreviewed.ts: quoted previews with default-ignorable characters stripped, accept-all skips an item it cannot preview, flagged_page items, the index-previewed release mode, and the doctor check matching a release by request id.
  • src/core/engine-sql/pages.ts, src/core/token-mint.ts, src/core/bootstrap/harness.ts, src/commands/sources-add-next.ts: the private-pages grant hides a version that carried pinned_question, and a harness token rotation prints the rescope that grants the private pages again.
  • src/core/ops/search.ts, src/core/search/evidence-date.ts, exact-lookup.ts, keyword-statement.ts, projection-readiness.ts, saved-facts-match.ts, src/core/sources-not-readable.ts, src/core/last-retrieved.ts: evidence titles without bidi or control characters, superseded saved facts kept out of entity facts and matched on normalised figures and stemmed names, sources_not_readable counting only returned pages and active facts, exact lookup in keyword order, and keyword page search that limits before joining chunks.
  • Tests (search path, trust and private pages): test/function-search-path-harden.test.ts, test/function-search-path-pglite.test.ts, test/migrate-search-path-best-effort.serial.test.ts, test/doctor-function-search-path.test.ts, test/scripts/check-search-path.test.ts, test/e2e/function-search-path-postgres.test.ts, test/trust-owner-actions.test.ts, test/trust-release-fence.test.ts, test/trust-release-preview-binding.test.ts, test/private-pages-grant.test.ts, test/auth-private-pages-output.test.ts, test/harness-rotation-grants.test.ts; recall: test/search-saved-facts-match.test.ts, test/search-identifier-lookup.test.ts, test/mcp-sources-not-readable.test.ts, test/projection-readiness.test.ts, test/search/keyword-pages-limit-first.test.ts, test/last-retrieved-writers.test.ts and their Postgres arms under test/e2e/.
  • src/core/synopsis-sweep.ts, src/commands/reindex-synopsis.ts, src/commands/reindex.ts, src/cli/commands/reindex.ts: gbrain reindex --synopsis previews and, with --yes, queues one contextual_reindex_per_chunk job per page under one worker-enforced spend cap, runs them in-process and resumes on a rerun; a failed job-status read stops the worker and fails the run with that error, and --json keeps the worker's log lines on stderr. src/commands/doctor/checks/synopsis-coverage.ts (synopsis_tier_coverage, also in the remote doctor, confined to the caller's sources), src/core/search/modes-report.ts (the coverage line in gbrain search modes), src/core/search/mode-switch-ux.ts (the tokenmax banner and the TTY prompt). Tests: test/synopsis-sweep-8744.serial.test.ts, test/e2e/synopsis-sweep-8744-postgres.test.ts, test/mode-switch-ux.test.ts.
  • src/core/grants/shared-exposure.ts (new): the detection (bot harnesses, Claude Code channel plugins, the chat-channel census over a credential's or stdio server's read set), the notice and its fix.
  • src/commands/doctor/checks/unscoped-multi-user-token.ts (new), registered in the doctor registry and categories; doctor goldens regenerated for the new check.
  • The install routes: src/commands/setup.ts, mcp.ts, harness-connect.ts, skillpack/scaffold.ts, auth.ts. Warnings go to stderr and into notices or warnings of a --json result.
  • src/core/error-registry.ts: the notice code. docs/mcp/ADMIN.md#one-token-several-people, skills/mcp-access/SKILL.md.
  • Tests: test/shared-exposure.test.ts, test/shared-exposure-routes.test.ts (PGLite brains with no repository), test/shared-exposure-serves.test.ts (a one-person brain stays silent, a bot harness on the owner's brain included; the several-people fix binds first).
  • src/core/search/keyword-paging.ts: readKeywordPage passes its row read to countKeywordMatches, so the count waits for the rows plus the grace, as on the hybrid path. Reported and first fixed by @akinduroifedayo in #8500.
  • src/core/search/keyword-handles.ts: slugHandles judges each alternative of a bare OR on its own; a negation or a non-handle phrase only stops its own alternative's handles, and quote marks split words and an empty phrase is ignored, as in websearch_to_tsquery. Reported and first fixed by @javieraldape in #8633.
  • src/core/search/output-redaction.ts: from_slug, to_slug, origin_slug and their *_source_id are identity fields; link context is still scanned. Reported and first fixed by @calebhicks in #8545.
  • src/core/engine-sql/takes.ts, src/core/engine.ts (TakeHit), src/core/utils.ts, src/core/think/gather.ts, src/core/think/sanitize.ts: both take searches select since_date, until_date and source; think renders until=, and all three attribute values, plus any holder outside the holder grammar, are neutralized and escaped like the claim. Takes SQL-text and export-surface goldens regenerated. Reported and first fixed by @calebhicks in #8546.
  • src/core/trajectory.ts (comparablePairs), src/core/trajectory-format.ts: regressions and the (superseded prior) annotation pair only same-metric, same-unit, same-period observations at a strictly later date, and a point with no finite date is left out instead of dropping the entity's whole trajectory block. Reported and first fixed by @calebhicks in #8547 and #8541.
  • src/core/ops/loops.ts, src/core/error-registry.ts, src/core/notice-ledger.ts: a failed sources read answers no_google_sources: null, completeness: "partial" and a loops_source_read_failed notice with no database error text; the digest says so. skills/google-loops/SKILL.md names the case. Reported and first fixed by @javieraldape in #8718.
  • src/core/ai/query-embed-cache.ts, src/core/ai/gateway.ts, src/core/ai/engine-gateway.ts: an engine's close (dropQueryEmbeddingsTagged) no longer bumps the process-wide generation. It marks that engine's gateway state released, and embedQuery skips the store when the state it started under was released, so only the closed engine's in-flight embeds are dropped. A cache_namespace engine keeps its in-flight embed for the reopened brain, and a full clearQueryEmbedCache still drops every one. Reported and first fixed by @javieraldape in #8639.
  • src/core/search/projection-statistics.ts: can_inspect applies pg_stats_ext_exprs's own rule (pg_has_role(relowner, 'USAGE') outside row security), so a SELECT-only role reads the projection statistic as uninspectable instead of uncollected; missingSearchStatistics and doctor planner_stats_stale stop listing it as absent (doctor says the role cannot inspect it and adds projection_statistics: "uninspectable" to its details, instead of calling it fresh), and statisticsStillCurrent lets a pass below the threshold skip the refresh when the statistic is collected or not inspectable. Reported and first fixed by @Masashi-Ono0611 in #8630.
  • Tests: test/search-keyword-paging.test.ts, test/search/keyword-handles.test.ts, test/search/keyword-handles-or-alternatives-8618.test.ts, test/search/output-redaction-link-slugs-8545.test.ts, test/takes-search-validity-8546.test.ts, test/trajectory.test.ts, test/trajectory-format.test.ts, test/loops-source-read-failure-8452.test.ts.
  • Tests (query embeddings and planner statistics): test/engine-gateway-state.test.ts (closing engine A while engine B's query embed is in flight leaves B's vector cached, so repeating the query makes no provider call; an engine's own in-flight embed is still not cached after it closes); test/engine-cache-namespace.test.ts (a namespaced engine's in-flight embed is still there for the reopened brain); test/projection-statistics.test.ts (a SELECT-only role over owner-analyzed statistics: nothing missing, cannot be inspected, two small passes refresh nothing and print nothing, a large pass warns once); test/e2e/planner-stats-postgres.test.ts (doctor run as a SELECT-only role reports the projection statistic as uninspectable).
  • src/core/search/alias-hop.ts: qualifiedAliasCandidates builds the 3 to min(6, n-1) word windows (3+ content words, longest then leftmost, at most 42) and adds them to the same single alias read; the first window with a readable, non-excluded claimant decides, one claimant leads at top-of-organic + epsilon (moved or injected through the private re-check) stamped qualified_alias_hit, two suppress it. Order is full-query alias, then qualified, then the token hop. A query over 6 words probes only the windows, so the exact-alias hop, search.alias_token_hop and the rerank gate's alias grades keep their 6-word cap.
  • src/core/search/intent-weights.ts, src/core/search/hybrid/request.ts: applyQualifiedTitleBoost runs last in applyIdentityBoosts on every path; it boosts a person or company title (2+ content words) mentioned in the query or a given-name variant by title_boost, stamps title_mention_boost, and skips forward title matches and rows that already carry exact_match_boost.
  • src/core/search/hybrid/rank.ts, src/core/search/remote-page-cap.ts: autocut and the remote page cap keep a qualified_alias_hit row; classifyEvidence, CRAG and System One protection do not read it. gbrain search --explain prints title-mention.
  • src/core/search/mode.ts, src/core/search/modes-report.ts, src/core/config.ts, src/core/search/config-values.ts: qualified_aliases and title_qualified mode knobs (off in every bundle), append-only qa=1 / tq=1 cache-key parts only when on, gbrain search modes rows, and gbrain config set refuses a value other than true/false (exit 2, invalid_params, nothing written).
  • Docs: docs/architecture/RETRIEVAL.md, KEY_FILES entries for alias-hop.ts, hybrid.ts, explain-formatter.ts, title-match.ts and remote-page-cap.ts.
  • Tests: test/search/alias-hop.test.ts, test/search/qualified-alias-visibility.test.ts (PGLite and Postgres, readable_by members, visibility: private, MCP search with remote: true; Postgres arm also through test/e2e/qualified-alias-visibility-postgres.test.ts), test/search/title-boost-stage.test.ts, test/search/title-retrieval-arm.test.ts (keyless keyword-only path), test/search-mode.test.ts, test/config-set-unread-prefixed-key.test.ts, test/search-remote-page-cap.test.ts, test/search/explain-formatter.test.ts.

[0.60.709.0] - 2026-11-13

A gbrain serve that recovers from status-only mode can sync again, and the brain owner's remember works on an entity page stored in segments.

Fixed

  • Syncs after a recovered gbrain serve (#9224). When gbrain serve starts while another gbrain process holds a PGLite brain (a second agent session or a cron job), it answers only gbrain_status until the brain opens, then becomes the brain's owner. After that, every gbrain sync handed to it that had pages to write failed with storage_error, and its own sync_brain failed with "Unknown engine kind for db-lock: pglite", until the agent session restarted. Those syncs now commit. The stand-in engine that status-only serve runs on now passes through every property of the live engine, not only its methods, and a synchronous method returns its own value instead of a promise. A Postgres gbrain serve that started while its database was unreachable gets the same fix. A serve that opens the brain when it starts was never affected.
  • A clear error when the owner itself faults (#9224). An unexpected fault inside the process that owns the brain now reaches the caller as internal_error (detail owner_internal) with an incident ID, instead of a storage_error whose gbrain doctor fix could not see the fault. The error says the request was not at fault and that restarting the owner and rerunning the command recovers; a write also names its request ID to check first with gbrain write-request. Its fix, gbrain sources writer status --json, is read-only and names the owner process. The owner's stderr records the error class, the gbrain code location and the database error text under the same incident ID, with connection strings redacted, and none of it crosses the local socket.
  • The owner's remember on a large entity page (#9213). With a token bound to the brain owner, remember onto an entity page stored in segments was refused with page_body_segmented, while the same call from an unbound token committed. It now commits. The owner's visibility checks no longer read the page body, which on such a page cannot be read as one string. The member scope-page check and the shared-origin retention purge of a page over 24 MiB had the same mistake and are fixed too. The guard that refuses whole-body reads of a segmented page is as strict as before.
  • Better advice when a write is refused on a page stored in segments. The refusal no longer suggests get_page body: "auto", which only reads the page. remember is told to save the claim without an entity (it is filed under memory/unattributed), and any other write over MCP names the edits that work on such a page: edit_page, append_page, add_tag and remove_tag.

Reported by @thiagodalvarenga (#9224) and @andreineacsu (#9213). First fixed by @Masashi-Ono0611 (#9231, #9215).

Sync no longer holds pages over gbrain's own frontmatter stamps, and file-drift repairs can compare before you pick a side

A page that differs from its file only by what gbrain itself wrote is no longer held as file drift. A schema-pack retype changes a page's type in the database and keeps the old type as legacy_type (plus an optional subtype) without touching the file. A file that still named the old type then read as a different page, and a later mass rewrite of the files, or a page imported before import records existed, turned that into a file_drift hold that needed a person. Sync now reads such a file with the stored retype stamp kept, so nothing differs. A real database edit (a label, a body change) is still held.

  • Existing holds of that kind release by themselves. The next sync after upgrading re-screens each held file_drift page that is not a deletion once, releases the ones that now agree with no commit and no change to any page's body, type or stamps (a retype stamp releases with no page write at all; a stale stored copy of the type key is normalised away), and says how many (Released N held file(s) that no longer need a decision., JSON holds_released). On a seeded source of 3,000 holds, 99% of them normalisation-only, one sync released 2,970 and left the 30 real differences held, in about 27 s on PGLite and 61 s on Postgres; the next sync took about 0.1 s.
  • A later commit over a page that only gained derived keys imports. When the page side changed only by keys gbrain derives (a stored copy of the reserved type key, stated or inferred from the path, or the retired grandfather's database-only validate: false) and someone commits a body edit to the file, sync imports the file instead of holding it, and the stamp stays.
  • A write that names the type a retype replaced still takes it. Only a file reading keeps a retype's stamp; a put_page with an explicit type: is stored as written.
  • A page whose file is gone under a gbrain.yml db_only folder is no drift. An existing deleted-side hold there now releases.
  • gbrain repair file-drift compares without a winner. Any selection without --prefer lists, per page, what differs, which side is newer (the file's last commit against the page's updated_at) and what a repair would do, and writes nothing. A --prefer preview with no side names that comparison as its fix.
  • --prefer file never drops a newer page in bulk. A --held, --all-drifted or --request preview lists a page edited after its file's last commit as kept_newer_page and leaves it held; naming the page with --slug still resolves it, with a warning naming the edit it drops. A page whose file is gone is never deleted by a comparison (GST-2568).
  • A --prefer preview says how much it drops. Each page's listing line counts the lines of the side the apply overwrites that the result doesn't carry ("drops 3 line(s) the file holds"), a warning gives the total for the run, and --json carries details.pages[].lost_lines and details.lost_lines_total, so whoever agrees to the apply sees what the losing side held. The consent the apply asks for repeats the total.
  • An unbound collision can keep its database side, with consent. A page an unbound write left database-only, with a different file at its slug path, is drift to gbrain repair file-drift --slug <slug>. sources reconcile refuses to take its database side over that file and names this step: --prefer database names the file's commit, its apply needs the user's consent, and it records the file as the page's origin.
  • The kept-newer-page text prints previews, not runs. It tells you to preview either side, and says a database-wins apply needs the user's consent.

[0.60.708.0] - 2026-11-13

CI checkouts on Ubicloud runners retry a stalled fetch instead of hanging until the job times out.

Fixed

  • Every checkout in a CI job on a Ubicloud runner now has a time limit (3 minutes, 6 for a full-history fetch). A checkout that runs past it is stopped and a second checkout step runs at once, so a stalled git fetch costs about three minutes instead of a timed-out job and a rerun of the whole Test run. This changes only CI; nothing users install changes.

A source delete can no longer leave its pages behind, a purge keeps the conversation-facts outcome rows, an older gbrain can no longer roll back a newer schema, and the commit door's dry run refuses whatever the real run refuses. This is the persistence/sync salvage wave: five fix PRs built in late October (#7853, #8237, #8152, #8320, #8330) rebuilt on current master, so the fixes they carry that master does not already have finally ship.

Data you could lose before

  • A source delete never strands its pages (#7072). On a brain without managed persistence, an expired-archive purge, gbrain sources purge or sources remove now refuses with source_pages_remain and changes nothing when the database would not cascade the delete to the source's pages (the pages-to-sources foreign key is missing, does not cascade, or its trigger is disabled). The nightly purge lists that source under blocked and moves on. Brains whose foreign key is intact purge as before, and pages an earlier purge already stranded are not touched. Managed brains already delete a source's pages before its row.
  • Older binaries can't overwrite a newer schema (#7776). An older gbrain that runs gbrain init, apply-migrations (including --force-schema) or upgrade against a brain a newer gbrain migrated now refuses with client_behind before anything runs. Before, its schema replay reinstalled its own older function bodies, write guards included. New doctor check schema_function_bodies compares the installed bodies with this binary's.
  • Facts purge keeps outcome rows (#7831). gbrain forget <id> --purge on a claim whose text is exactly EXTRACTION_COMPLETE or EXTRACTION_NOT_APPLICABLE no longer deletes the outcome rows with the same text, so those pages aren't extracted (and paid for) again.
  • Secret audit (#8034). audit-secrets and re-scans again report a password glued to a redaction marker by punctuation. Import-time redaction is unchanged.
  • Ghost code pages (#7967). gbrain repair ghost-pages retires a code or image page at its recorded file instead of at <slug>.md, and on a git-canonical source the door commits the file's removal.
  • Failed repairs change nothing (#7960 item 1). When gbrain repair provenance --apply fails in the database after writing the page file, it puts the file back and reverts its commit.
  • Dream write-through (#8092). After a refused commit, rewrites of existing atom and concept files and removals of retired atoms' files are retried on the next run, and a busy commit lock no longer freezes the worker's event loop.
  • Read-only mirrors stay read-only (#7098). On a brain without managed persistence, a source marked gbrain sources mirror-readonly is left alone by page write-through and delete, timeline entries, fact fences, forget, the extract cycle, sync --full and brainstorm --save; owner trust repairs list its pages needs_review.
  • Dream runs leave read-only mirrors alone too (#8935). On a brain without managed persistence, the dream cycle no longer writes, rewrites or removes page files in a source marked gbrain sources mirror-readonly (or a clone gbrain owns), and no longer puts a fallback file for another source under .sources/ inside such a checkout. The pages stay in the database, and the mirror's git pull --ff-only keeps working. Reported and first fixed by @javieraldape in #9212.
  • An agent's sessions are not saved as you (#8080). Dream synthesis and OpenClaw sub-agent compaction honor the agent speaker marker. When an agent continues a session already saved as yours, gbrain saves nothing from it and reports speaker_mismatch. gbrain repair agent-session-facts --apply reads the session corpus once per run and keeps in-page rows attributed correctly after a slow write.

The commit door

  • Dry run (#7752). gbrain git commit-push --dry-run runs the real run's path checks (source root, a --base that moved, a sparse-checkout exclusion, another writer's staged file, a staged rename, a deleted or modified file under an ignore rule) against a copy of the index and exits 1 with the same error. It still writes nothing and takes no lock.
  • Literal paths (#7792, #7793). The brain-commit-push.sh fallback treats glob characters as literal, so committing [ab].md no longer also commits a.md, and the fix commands the door prints for another writer's staged file name literal paths.
  • Extract receipts (#7620). Receipts are sealed in the transaction that writes them, so remote search and query no longer report safe_index_pending after every extraction. Receipts written before this release still need gbrain repair safe-chunks --apply once.
  • Strict door index refresh (#7755, #7756, #7757, #7758, #7717). The refresh removes index copies a killed gbrain left behind, retries a briefly refused rename on Windows, and a page committed with a legacy file mode (100664) gets no mode-only commit.

Clearer refusals and next steps

  • Protocol guard refusals (#7777) fail with the guard's own code and remedy instead of an "unexpected owner exception" storage_error. Contributed by @javieraldape (#7913, rewritten on master).
  • unreadable_state (#7780) says the role cannot fully read the writer state and asks before suggesting a role change.
  • Stored source paths (#7646) are judged absolute by one rule everywhere (sync, config set, write-through, set-path --clear, doctor). Writes on a relative root record relative_root and name gbrain sources set-path.
  • Git-canonical conversion (#7956) refuses a root it can't read with source_root_unreadable, and an apply whose source was switched back to mixed while it ran reports incomplete.
  • git-durability enable (#7634) checks its approval before writing AGENTS.md, wiring a credential or retrying parked pushes.
  • Parked push retry (#7110) keeps going past a request whose page became private and ends partial.
  • Embedding stop signal (#7772) retries as embedding_aborted instead of failing for good.
  • Retired grandfather receipts (#7837) read as history in doctor's persistence_terminal_failures.
  • Synthesis holds (#7849) each name the fix that releases them, and doctor counts swept holds.
  • Copied brain (#7671) physical_root_directory_changed names the relocated-root self-transfer as its fix. Nothing is re-stamped automatically.
  • Repair wording (#7715, #7910, #7881): stuck-recovery on a git-canonical source says commit-then-sync; repair url-clone-fences lists a held file once; a streamed put/capture refusal uses the page-write wording.

Brainstorm message idea from @javieraldape (#7113, rewritten on master). Every change behaves the same on a repo-less brain except the door and conversion items, which need a checkout as they always have.

A forget interrupted after its Git commit still withdraws the fact

Git is the parent, so the database now follows a commit that already landed. On a source configured canonical: git, a write is committed to Git first and its database row second. When something failed between the two (an unexpected error after the Git commit, before the row committed), the write ended failed with storage_error and the reply said "nothing was saved", although Git held the change. For a forget that meant the fact stayed active until the next sync, no withdrawal was recorded, and the same claim could be remembered again as a new fact (GST-2517).

  • The owner now replays such a write. The replay finds its change already committed and records it without a second commit. A forget turns the fact inactive and records its withdrawal, and a remember or put_page gets its one row. A shared fact forgotten by the owner and a member's own-page fact behave the same. A killed process in the same window was already recovered this way.
  • If the replay fails too, the request ends failed with the landed commit on its receipt (landed_before_row). A remember or forget reply then names that commit and says the database has not recorded it, instead of "nothing was saved". A forget submitted again with a new request_id records the withdrawal without a second commit.
  • gbrain repair failed-writes no longer says a refused forget "will be produced again by the next dream cycle": only callers make forgets, so it lists one as a write to resubmit.

For contributors

  • No git call under src passes git's global --literal-pathspecs switch any more; read paths name each path as :(literal)<path> through literalPathspecs, test/git-env-guard.test.ts refuses the switch anywhere in src, and literalPathspecs('') throws.
  • The crash robot reports a failed page read as read_failure instead of a false projection_drift (#7651).

Hermes can turn on the skill resolver with one setup command

gbrain setup hermes --skill-loading resolve now wires the skill resolver into Hermes the way it already does for Claude Code and Codex. Before, Hermes users built the catalog, edited config.json and pasted the hook by hand.

  • Setup builds the hermes-user catalog from $HERMES_HOME/skills (nested category folders included) and adds the resolver's pre_llm_call entry to gbrain's marked hooks: block in $HERMES_HOME/config.yaml, approving it in Hermes's shell-hooks-allowlist.json. The block also carries the brain-first turn rule. Each feature adds and removes only its own entry: --no-instructions keeps the resolver, --skill-loading off keeps the turn rule, and the block goes when its last entry does. An existing install's turn-rule block is left byte for byte as it was.
  • The command runs without a shell, the way Hermes runs hooks: /usr/bin/env GBRAIN_HOOK_LANE=skill-resolver <launcher> <gbrain> hook user-prompt --harness hermes.
  • If your config.yaml already has its own hooks: section, setup refuses and prints the entry to paste under it. Once you paste it, the next run approves and keeps it. --skill-loading off withdraws that approval but leaves your own section alone, so delete the pasted entry yourself. A Hermes with no SOUL.md still gets the resolver.
  • Hermes compacts conversations in place and tells hooks nothing about it. The lane now starts a session's dedupe fresh on a first turn, or when the conversation history is shorter than it was on the last turn, so skills offered before a compaction can be offered again. It reads the history's length in the same streaming pass, so a large history still costs about one JSON parse.
  • gbrain config set skills.resolver.per_harness.hermes.owner_sender_ids <id>,<id> limits suggestions to turns from those senders, using the same rule as the OpenClaw plugin. A turn from anyone else, or with no sender, gets no block.
  • Pinned skills for Hermes (gbrain skills pin add <id> --catalog hermes-user) go in a marked block in $HERMES_HOME/SOUL.md while the resolver is on. With no pins there is no block, and SOUL.md is never created.
  • gbrain doctor's skill_resolver check now covers Hermes, and warns when Hermes's allowlist no longer approves the hook.

The OpenClaw skill resolver no longer suggests skills the agent cannot use

gbrain setup openclaw --skill-loading resolve now leaves out of each agent's suggestions the skills OpenClaw hides from that agent's model. Before, the catalog came from the agent's skill folders alone, so a suggestion could name a skill that is not eligible on the host, that the agent's skill filter excludes, or that OpenClaw dropped at load, and the model could not open it.

  • Setup reads openclaw skills list --json --agent <id> for each agent and adds every catalog skill it does not show the model (modelVisible false, or not listed) to that agent's exclude_ids. It names those skills, and tags the ones OpenClaw did not list with skill_resolver_host_load_cut.
  • When the command fails, or lists none of the catalog's skills, nothing is filtered and setup says why. With no openclaw on PATH nothing changes.
  • Run setup again after changing which skills an agent can use; the plugin reads the new list from its state file without a gateway reload.

Pi can turn on the skill resolver with one setup command

gbrain setup pi --skill-loading resolve installs a Pi extension that names the few installed skills that fit each request. Pi lists every skill in its system prompt; with many skills the model picks from a long list, and now it also gets a short ranked one.

  • Setup builds the pi-user catalog from $PI_CODING_AGENT_DIR/skills and ~/.agents/skills, and writes the extension to $PI_CODING_AGENT_DIR/extensions/gbrain-skill-resolver/. It runs one turn through the extension under Node and, when pi is on PATH, checks that pi --help lists --gbrain-skill-resolver; if Pi does not load it, setup exits 3 with skill_resolver_plugin_not_installed.
  • Before each user run, up to 3 suggestions go to the model as one hidden message with each skill's SKILL.md path. A skill already suggested in the session is not offered again until Pi compacts the conversation, also across pi -p -c runs. Pinned skills (gbrain skills pin add <id> --catalog pi-user) go in a system prompt section that stays the same from run to run. A pinned skill, or one Pi does not show the model, is never suggested.
  • With Pi 1.5.2 and a fake model server, a cold pi -p run waited 70 to 100 ms for suggestions (the budget is 150 ms on a cold run and 50 ms after), and a pi -ne -p run got nothing.
  • Pi started with extensions off (pi --no-extensions, or a program that embeds Pi with noExtensions, such as QM) gets no suggestions; setup and gbrain doctor say so. Project skill folders are not in the catalog.
  • gbrain setup pi --skill-loading off removes the extension and its state. Nothing opens a brain, so a keyless or brainless install stays as it is.

A remote caller's hybrid search no longer times out its keyword arm on a common word.

Fixed

  • Hybrid search ranks at most 50,000 matching chunks, but for a remote (MCP) caller the planner still read every matching chunk before that bound applied, because the private-page checks made the page join look expensive. On a large Postgres brain a word in most chunks then ran past the 8 s keyword statement timeout, the search answered from vectors alone (keyword_arm_failed), and the serving process skipped the keyword arm for that query for the next 10 minutes.
  • When a remote caller sets no type, frontmatter or date filter and its source scope covers every live source (the usual MCP scope), the keyword statement now reads matching chunks straight from the full-text index and checks each one's page by key, with every visibility, member-access, quarantine, deletion and archived-source filter applied before the bound. It stops reading once it has 50,000 candidates.
  • Measured on a synthetic 2.1M-chunk Postgres brain (warm, median), the remote keyword statement for a word in 95% of chunks went from 9.6 s to 0.89 s, a word in a third of chunks from 1.4 s to 0.72 s, and a word in 5% of chunks from 0.88 s to 0.64 s. Rare terms are unchanged (about 40 ms), with identical results. A whole MCP search for the 95% word now answers with keyword hits in about 1 s, where it used to time out and fall back to vectors.
  • Results are identical whenever a query matches fewer than 50,000 chunks. Above that, the 50,000-chunk sample is drawn in index order instead of another plan's order, so which pages of a very broad term rank can change; the result was already a sample there (keyword_candidates_incomplete).
  • A remote caller scoped to only some sources runs the old page join through the same statement, now with sequential scans off as the OR retry already ran. A word in 95% of chunks went from 9.1 s to 1.7 s there, and its other results and timings are unchanged. Local callers and typed or dated searches run exactly the statement they ran before.
  • For a word in nearly every chunk, the statement turns sequential scans off, as the OR retry already did. The planner otherwise picked a seq scan, which does not keep the pages it reads in cache and took about 5 s to reach the bound after a restart, against under 1 s for the index scan.
  • The keyword-mode page statement and the strict page count now run with the same 32 MB work_mem as the keyword arm, so a common word's bitmap no longer goes lossy: 5.8 s to 3.7 s for keyword paging and 0.68 s to 0.26 s for a count on the same brain.
  • scripts/bench/retrieval-scale/keyword-broad.ts times the keyword statements for broad, medium and rare terms and can diff two checkouts' results; run.ts takes --queries and --no-query-cache.

A brain that serves several people refuses reads from connections bound to nobody

On a brain that serves several people, a token, OAuth client or stdio server bound to no member can no longer read every unmarked page for whoever is asking. After a 14-day grace for connections that already existed, such a read is refused with requester_scope_required and nothing is read. A brain serves several people when two or more members are enabled, or when config access.serves is several. A brain pinned access.serves one never refuses, however many members it has, and a credential bound to a member, the owner's included, reads exactly what it read before. Everything is database state, so repo-less and keyless brains get the same refusal and the same grace with no new config.

Connection Before Now, on a brain that serves several people
Bound to a member (--member, the owner included), naming an audience per call (--may-act-for), or attested reads as that person unchanged
Bound to nobody, existed when the brain started serving several people read every page no rule restricts reads for 14 more days, each answer carrying _meta.access.requester_scope with the date and the fix; then refused
Bound to nobody, created later read every page no rule restricts refused at once (requester_scope_required, reason created_after)
Marked --reads-unscoped --yes by the operator (no such mark) reads as before, on purpose, audited
Any connection on a one-person brain, or the local CLI unchanged unchanged

Itemized changes

  • src/core/access/requester-scope.ts (new): the gate, run in dispatchToolCall after access resolves and in every agent job tool call. Every read op and every read-scope op (think included) is gated unless it reads no page: whoami, get_brain_identity, get_status_snapshot, search_modes, get_health, the caller's own join_brain / leave_brain / sync_brain_skills / request_tools, and a non-admin credential's own get_job / wait_job. When the serves policy cannot be read it is read once more, and a second failure refuses retryably (requester_scope_unknown) rather than reading as one person; a bound credential is never refused that way. A connection whose reads are refused gets no _meta.brain_hot_memory on the calls the gate lets through (the exempt ops above and writes), so it cannot read recent facts through whoami.
  • access.serves is read without regard to case or surrounding spaces (One pins one person). access.serves_since is set once and never cleared, so a brain that served several people once and does again later gets no new grace. A connection counts as existing before that date only when it was created in an earlier second, so one created in the same second is refused like a later one.
  • Migration v309: reads_unscoped on access_tokens and oauth_clients, and a trigger that keeps config access.serves_derived and sets access.serves_since once, the first time the brain serves several people, so a request reads both from its config snapshot. Disabling and re-enabling a member does not restart the grace. Nothing else is rewritten.
  • gbrain auth create, gbrain auth register-client and the admin dashboard's key mint refuse an unbound credential on such a brain (credential_scope_required). gbrain auth rescope ... --reads-unscoped --yes (none removes it) is the explicit opt-out, written to the access-reads-unscoped audit file; gbrain access check reports it.
  • The fix every refusal, notice, whoami note and gbrain access owner hint gives for stdio is the per-harness binding, gbrain setup claude-code|codex --stdio-member <id>, on PGLite and Postgres alike, or a per-person key over gbrain serve --http for a harness several people talk to. Nothing suggests config set mcp.stdio_member, and no upgrade, doctor or setup path writes it.
  • Doctor: unscoped_shared_reads fails per unbound credential with its state, grace end and bind command (warns while access.requester_scope_enforce is false); shared_stdio_binding fails when mcp.stdio_member is set on a host where OpenClaw or Hermes is installed, with gbrain config unset mcp.stdio_member as its ask-first fix, and is not applicable on a brain that serves one person. gbrain post-upgrade lists the unbound credentials.
  • src/commands/setup-serves.ts: a hermes on PATH counts as the Hermes agent only when it is a script, so React Native's compiled hermes engine no longer makes setup refuse --stdio-member or doctor fail shared_stdio_binding.
  • Hosted brains keep access.requester_scope_enforce false until their control plane binds each connection; every unbound read then answers with the notice and nothing is refused.
  • Docs: docs/guides/write-refusals.md (the three codes), docs/guides/setup.md, docs/guides/multi-source-brains.md, docs/tutorials/company-brain.md, docs/architecture/key-files/access.md; a behaviour-change notice row explains the change on upgrade.
  • Tests: test/requester-scope.test.ts (PGLite and Postgres, repo-less and a repo brain: the refusal, the grace, the pinned one-person brain and the owner-bound key unchanged, every gated op refused before its handler, the hosted switch, agent jobs, the CLI), test/shared-stdio-binding.test.ts.

A flaky timer test is now deterministic, and every test file shows up in CI run evidence.

Fixed

  • The db-lock heartbeat test no longer fails when a busy CI runner pauses the process. It now waits until the heartbeat fires, up to 5 seconds, instead of checking after a fixed 40 ms.
  • Two export test files no longer share cases by importing one test file from the other, so each file reports its own tests when both run in one process. A new test blocks that import pattern.
  • The CI health report counts a failed checkout retry step as an infrastructure failure, not a test failure.

These change only tests and CI scripts; nothing users install changes.

[0.60.707.0] - 2026-11-13

A pull request that is behind master's newest release no longer fails verify.

Fixed

  • check:changelog-headings (in bun run verify) refused every branch whose CHANGELOG.md lacked a release heading that master has. Since pull requests carry no version and may land without containing current master, every such branch failed verify as soon as a release landed. The check now skips the release headings newer than the branch's own VERSION, which the merged tree carries, and still refuses a branch that dropped a heading at or below its VERSION, so a cut-off changelog is still caught.

A labelled-only member no longer reads and writes as an unbound guest when the database hiccups. When gbrain looks up which member a token, OAuth client or local server is bound to, a dropped connection, a timeout or a failed statement used to look the same as "bound to nobody". A credential bound to nobody reads and writes every page that names no readers, so during that hiccup a member kept to its labelled pages could read, overwrite or create unlabelled pages. Now the request fails with access_identity_unavailable, a retryable error that tells the agent to wait a few seconds and send the same request again; nothing is read or written. A queued write that meets a lock or statement timeout there is still retried by itself, as before. A shared agent's write for a labelled-only member, or a trusted system's write for a person it vouches for, stays labelled-only when the lookup of who it writes for fails. Credentials that really are unbound read exactly as before, and so does every caller on a brain whose database predates member access.

Itemized changes

Member lookups fail closed

  • src/core/access/caller.ts: liveIdentity (behind memberIdentityFor and stdioMemberIdentity) no longer returns the empty identity when its lookup fails. A statement the database rejects throws access_identity_unavailable. A transient failure (transientDatabaseFailure or transientInterruption: a dropped connection, a lock or statement timeout, a stale plan, a publication past its ceiling) and any failure that is not a statement error (a preparation deadline, an abort) are rethrown unchanged, so the persistence consumer still releases and retries the write. An unbound credential, an unknown member and a disabled or removed member stay empty. A missing access_members or access_group_members table or member_id column (isMissingSchemaObject) still reads as unbound: all three arrived with the member-access migration, before the labelled-only column, so no member of any kind can exist there; any other missing table or column throws.
  • src/core/persistence/page-visibility.ts: a stored write's labelled-only check also reads the grant of the members its stored audience names and of its attested actor's member, so a re-resolution that fails and reads as the guest identity never makes a labelled-only member's write unlabelled. A database failure while re-resolving the attested actor is rethrown (the write is retried) instead of reading as the guest; a coded refusal (an audience no longer permitted, an actor no longer linked) still reads as the guest.
  • src/core/persistence/transient-retry.ts: transientDatabaseFailure and connectionCapFailure move here from coordinator.ts, which re-exports them, so the access lookup can use them without an import cycle that fails at module load.
  • New error code access_identity_unavailable (src/core/error-registry.ts, src/core/error-catalogue.ts): class retryable, retry_after_ms=2000, and a fix whose next is wait that calls whoami to confirm the binding reads again (the error-code table's Verify column now says gbrain whoami --json too). Documented in docs/guides/write-refusals.md and docs/guides/error-codes.md.
  • src/mcp/dispatch.ts: a coded error raised while resolving the caller's access comes back as an MCP error result, as every other refusal does, instead of a thrown call.
  • src/mcp/server.ts: gbrain serve with a stdio member still starts when that member's lookup fails at startup with access_identity_unavailable or a transient database failure; its writeback instructions treat the server as bound to a member, which keeps its writes private. Any other error is thrown.
  • src/core/persistence/page-visibility.ts: the labelled-only create check drops a shortcut that skipped the check when the request's own policy said the brain had no restrictions. The check runs inside the persistence consumer, where that policy is never set, so the shortcut never fired; the check now always asks the database whether any labelled-only member exists.
  • Tests: test/access-labelled-only-fail-closed.test.ts on PGLite and Postgres. A labelled-only credential whose lookups fail gets access_identity_unavailable on get_page, list_pages, search and the stdio binding, and cannot update or create an unlabelled page; a missing member table or member_id column still reads as unbound, and any other missing object throws. A transient lookup failure is rethrown unchanged, and an owner put_page whose stored-write lookup meets one lock timeout still commits. A stored write for a labelled-only member through an audience or an attested actor stays refused when that lookup fails. It also pins the create refusal when the grant reaches a brain that served unrestricted requests. It covers one long-lived engine, a member that already wrote, a second serve process (Postgres only, because PGLite locks its data directory to one process) and access.restrictions_present deleted by hand. A failed visibility probe never lets a write through inside a transaction, and outside one it counts as labelled-only.

[0.60.706.0] - 2026-11-13

A database-wins file-drift repair no longer discards a newer file edit

gbrain repair file-drift --prefer database and gbrain sources reconcile --apply could overwrite a change to a page's file that sync had not imported yet, or bring back a file another writer had just deleted, with nothing in the preview saying so. A repair now checks each page's file against sync's import record before it takes the database copy.

  • src/core/persistence/unimported-file.ts (new): the database copy may win only when the database itself held the version the file records: the bytes sync's import record names (or the blob a sync adopted into it), or gbrain's own write of a revision the page has held (live or in its version history). Anything else is newer, whoever committed it, including gbrain's own writer commit of a revision whose database row never landed and a page with no import record at all; a wrong "keep" costs a hold and a wrong "take" costs an edit. A database edit made after the file's last commit still wins. The preview names the last commit that touched the file (sha, author, date).
  • gbrain repair file-drift --prefer database with --held or --all-drifted leaves such pages held and lists them kept_newer_file; a file that another writer's delete commit removed is kept_newer_delete instead of being recreated. Naming one page with --slug still previews it, and the preview and the consent text name what it would discard.
  • Every gbrain repair file-drift apply now asks for the user's consent (consent: 'destructive' in the repair registry), and the consent text lists the changes it would discard and the files a tombstone_edited or purged_differs --prefer database apply removes. A script that applied without --yes now exits 3.
  • gbrain sources reconcile (preview and --apply, and the writer_reconcile_preview / writer_reconcile_apply operations) refuses source_changed when its result takes the database copy over such a file.
  • Hold texts, fix steps and the docs that print a database-wins repair now give the preview as the step and say the apply needs the user's consent.
  • docs/guides/repair.md explains the new behaviour and how to find edits an earlier repair overwrote, from the reconcile backups it kept.
  • Tests: test/db-wins-unimported-gst2579.test.ts on PGLite and Postgres.
  • The @hono/node-server override (pulled in by the MCP SDK) moves from ^2.0.5 to ^2.1.6, now resolving 2.1.8, which fixes GHSA-qv5m-9pm6-hqp3 (a reset connection during a WebSocket upgrade could crash gbrain serve --http) and GHSA-jpvv-r467-5p4w (a backslash in a request path could get past a static-path check).

The MCP server's HTTP framework, Hono, moves to a release that fixes five published advisories.

Security

  • The hono override (Hono comes in through the MCP SDK's HTTP transport) moves from ^4.13.11 to ^4.13.16, now resolving 4.13.18. Hono 4.13.13 had five advisories, all fixed in 4.13.16: GHSA-3gg2-335m-rxc6 (the css helper could let HTML through when a closing tag was split between the template and a value), GHSA-46wx-v44x-9hh2 (csrf() did not check requests sent by hyperlink auditing), GHSA-h7mg-cq63-vhpq (the RegExpRouter could skip wildcard middleware for some routes and for paths with an encoded line terminator), GHSA-4h7h-rv7c-7m2v (validator('form') skipped validation for multipart/form-data requests with extra parameters) and GHSA-4m76-xfj5-7975 (mount() could forward a path that bypassed middleware scoped inside the mount prefix). Four are rated medium and one low. bun audit and the OSV scan now report nothing.

[0.60.705.0] - 2026-11-12

Pull requests stop picking a version: they carry a changelog fragment, and a release PR numbers each batch

Releases are cut per batch instead of per pull request. Until now every PR stamped the next PATCH version, wrote its own CHANGELOG section and appended its upgrade-notice row to one shared array. So every landing on master made almost every queued PR conflict in the same three places, and each one had to be restamped and tested again at its turn. From the next release on, a PR writes its entry to changelog.d/<branch-slug>.md, writes any new upgrade-notice row to behavior-changes.d/<id>.md, and touches no version. A release PR, opened a few times a day and right after an important fix, gives everything that landed since the last release one PATCH version and writes every stamp. Measured on 59 open PRs, keeping notice rows in their own files takes conflicts in the notice table from 22 to 0 and in the module size table from 27 to 1. This release is the last one a PR stamps by hand; docs/RELEASING.md#releases-per-batch describes the new flow.

  • .gstack/ship-policy.json turns on merge: "direct" with changelog_fragments in changelog.d. The release stamp runs bun scripts/release-restamp.ts --stamp-only, and bun scripts/build-behavior-changes.ts writes the one release output, src/core/behavior-changes.generated.ts.
  • scripts/build-behavior-changes.ts dates each behavior-changes.d row by the release section that holds its marker comment (the format is in behavior-changes.d/README.md), and BEHAVIOR_CHANGES spreads the generated table after its released rows, which stay where they are. It refuses an id that starts with a digit, a marker that appears twice and a fragment whose bump: line says anything but patch. A row whose text is computed stays in the table as since: releasedIn('<id>').
  • scripts/release-restamp.ts gains --stamp-only <version> (the release PR's stamp, with no fetch, merge or commit), bun run release:unstamp (turns a PR stamped the old way into one commit holding only the stamp reset, its fragment and its notice-row files) and bun run migrations:renumber. Plain release:restamp refuses once master's policy says merge: "direct", and unstamp refuses to overwrite a fragment that already exists.
  • scripts/check-behavior-since.ts refuses a literal since: 'TBD' row. With --stamped, which the Release workflow runs, it also refuses an unfolded fragment, a stale generated table and a missing or unreadable ship policy.
  • The Release workflow no longer follows a cancelled master run to a later commit with the same VERSION, because many commits share one VERSION between releases; it fails and names the run to re-run. A manual Release dispatch publishes only while master's head is the release commit, and otherwise refuses and names that commit's Release run.
  • test.yml and e2e.yml key master push runs by commit, so a newer push never cancels the release commit's run. Every landing on master now runs full Test and E2E to the end.
  • Docs: docs/RELEASING.md, CONTRIBUTING.md, CLAUDE.md, AGENTS.md, docs/TESTING.md and docs/contributing/version-recovery.md describe the new flow, and changelog.d/README.md and behavior-changes.d/README.md give the file formats.

[0.60.704.0] - 2026-11-12

QM can read a room's context pack, and the owner sets how a client reads named sources from another machine

context_pack reads a room as its audience. A QM client that reads named sources as the scope used to be refused when it asked context_pack for a room's source, so QM had no session brief in shared conversations. Now a call that names source_id reads only that source, which covers its entities and entity_refs, the core block, the counterparty brief and hot facts. Nothing from the person's own member page or from pages kept to named readers comes back, and privacy is never relaxed.

  • A scope call is refused with invalid_params before anything is read when it sends entity_refs without source_id, names an entity_refs entry in another source, or sends session_id (a session's thread rows are the person's).
  • A client set --attest-read person keeps reading as the person, and its pack is narrowed to the named source the same way.
  • Contract docs/contracts/qm-memory-provider/v5.json marks context_pack scope_read: true. v1 to v4 stay byte for byte as published, and v4's markings are frozen in the builder.

gbrain mcp admin attest <client> --attest-read person|scope|none sets the same attest_read that gbrain auth rescope --attest-read sets, over the owner session (attestRead on /admin/api/access/attest). It refuses what the local command refuses, previews with --dry-run, and logs attest_read in the access_admin row.

[0.60.703.0] - 2026-11-12

A missing link target written while another page's links are being extracted now wakes the page that mentions it. A [[Name]] reference to a page that did not exist yet was saved with the time of the save. If the target page was written during the extraction, after the extractor had looked for it but before the save, the target looked older than the saved reference, so the stale link sweep never went back to the mentioning page and the link stayed missing until one of the two pages was edited again. v0.60.679.0 named this narrow race as open. Name and basename references are now stamped with the database time taken just before the extraction reads the pages it resolves against, so a target written after that read wakes the mentioning page once on the next sweep, and a target the extractor already saw never loops.

  • src/core/wanted-links-store.ts: wantedReadAt returns the database clock; replaceWantedLinks stamps name rows with the readAt its caller took before reading, and refuses name rows that arrive without one. Slug rows keep their v0.60.679.0 behaviour.
  • Every caller that writes name rows takes the read time first: gbrain extract (full and --stale, on every run, including a page that turns stale after the pre-flight count), the wanted-link sweep, link reconciliation, and link preparation on page writes.
  • Still open: on Postgres, a writer transaction that starts before the read time and commits after the extractor's read keeps an updated_at older than the stamp, so its mention waits for the next edit of either page.
  • Tests: nameRowPublishedAfterItsTargetWasWrittenWakesItsOrigin and basenameRowPublishedAfterItsTargetWasWrittenWakesItsOrigin in test/wanted-links.test.ts and test/e2e/wanted-links-postgres.test.ts (both red first), and test/wanted-links-stale-count-race.test.ts. Idea from @akinduroifedayo's #9149.
  • Also carries the nightly shard-weight refresh: scripts/e2e-weights.json and its metadata file, exactly as the weight miner wrote them from a full E2E nightly.

[0.60.702.0] - 2026-11-12

A member can now be kept to pages labelled for it. A group chat bot or a shared assistant bound to a member read every unlabelled page its source grant opened, and every summary, atom and fact made from them. gbrain access member labelled-only <id> on (off by default, unchanged on upgrade) keeps every credential of that member to pages that name it: the member or one of its groups in readable_by, or a folder or page rule listing them. Unlabelled pages, and derived pages and facts that name nobody, are hidden on every read path; with access.members_read_lineage on, a derived page passes only when every origin names the member. Writes follow the same rule. Only the operator on the brain host sets it, and never on the owner. It works the same on a repo-less brain and with rules in access.yml.

Itemized changes

Labelled-only members

  • The labelled-only migration (src/core/schema-migrations/, numbered at merge): access_members.labelled_only (default false) and the SQL function gbrain_access_origin_labelled_for(), which passes a derived page only when it has an origin and every origin names the caller. Nothing is rewritten.
  • src/core/access/labelled-only.ts: gbrain access member labelled-only <id> [on|off], local operator only (remote callers get permission_denied), never on the owner (access_labelled_only_owner, exit 3; gbrain access owner refuses a labelled-only member the same way).
  • src/core/search/private-visibility.ts: CallerAccess.labelledOnly adds one conjunct to every caller page, version and fact filter; derived pages and facts that name nobody fail closed, and under member lineage pass only when every origin names the caller. Facts on the member's own member page stay readable, lineage on or off. src/core/access/caller.ts, actor.ts and src/cli/as-preview.ts set it for a labelled-only member, and labelledOnlyFor never reports the configured owner, on any door.
  • The remote read floor: remoteReadFloorSql and remoteProvenanceFloorSql (private-visibility.ts) are the plain page and fact fragments plus the labelled-only conjunct, and exactly the plain fragments for every other caller. Every remote read gate that is not caller-aware uses them: recorded-time reads (get_page and get_versions with known_as_of), the as_of amendment list, fact lookups by id (remember replaces, forget), belief history, proactive briefs, entity resolution and the sources-not-readable check. test/remote-read-floor.test.ts lists the write-side, maintenance and classification sites that keep the plain fragment, each with its reason.
  • Attested reads: a read that names sources for a labelled-only person is refused (permission_denied, src/core/access/scope-read.ts), since the scope read drops the member identity and the grant with it.
  • Writes: src/core/persistence/page-visibility.ts holds a labelled-only writer to the same read rule and refuses a create that would not name it (access_labelled_only_unnamed, checked in page preparation, streamed uploads and capture ingest); the enforced member CLI (src/cli/enforced-member.ts) does the same.
  • The semantic cache key (xlo=1) and the per-caller key (labelled_only) separate a labelled-only caller.
  • gbrain access member list, access check and who-can-read report it; doctor access_labelled_only_members.
  • Docs: docs/guides/access-designations.md (labelled-only members), docs/guides/error-codes.md, key-files entry.
  • Tests: test/access-labelled-only.test.ts (every read path, derived pages and facts with member lineage off and on, older versions, recorded-time and as_of reads, a fact by id, the member's own facts, the attested scope-read refusal and owner, writes over MCP, streamed uploads, capture ingest and the enforced CLI, the owner and remote refusals, cache keys, check, who-can-read and doctor; PGLite and Postgres, rules in the database and in access.yml); test/remote-read-floor.test.ts; migration, catalog and doctor goldens.

[0.60.701.0] - 2026-11-12

Flaky and slow tests fixed at their cause

Four tests that failed at random in CI now pass reliably, one E2E file is back under its time limit, and one of the fixes is a real bug: a cycle could re-read pages it had already handled. Each fix was proven with a forced probe that reproduces the CI failure before the fix and passes after it. No assertion, timeout or retry limit was loosened.

  • src/core/cycle/source-window.ts: the cycle's page-listing cursor steps back to before the oldest transaction still open on another connection, so a page written by that transaction can't be missed. That count included Postgres autovacuum workers, which never write pages. An autovacuum pass open while the cycle listed pages made the next cycle read the pages written since that pass began a second time. Autovacuum workers no longer move the cursor (test/sync-disabled-cycle-window.test.ts).
  • test/facts-entity-moved-7602.test.ts: the test created and killed the page write's facts-absorb job before the write's own background effect had queued it. When the effect ran later, it found a dead job under its key and queued a fresh one, so the preview reported a waiting job. The test now waits for the effect to finish before it touches the job.
  • test/persistence-index-lock-reclaim-6932.test.ts: cleanup killed a helper shell whose rm -rf kept running after the shell died (dash forks a script's last command instead of replacing itself with it), and then ran chmod -R over the tree that rm was still deleting. Cleanup now waits for each helper to exit.
  • test/e2e/worktree-refresh-postgres.test.ts ran six refresh suites in one file and took 139 to 156 s on CI against the 180 s per-file limit. Its crash-recovery, orphan and diverged-publish suites moved to test/e2e/worktree-refresh-recovery-postgres.test.ts; all 95 cases still run, and each file takes about 100 s on a 4-vCPU machine.
  • scripts/postgres-arm-weights.json: the two PostgreSQL arm weights that were estimates now hold measured medians.
  • test/docs-repo-paths.test.ts (the scripts/ and test/ paths named in docs must exist) runs in bun run verify as check:docs-repo-paths. Before, it ran only in the full suite and in scripts/ci-doc-checks.sh, so a pull request's own checks never ran it, and a doc edit could name a missing path and pass.

[0.60.700.0] - 2026-11-12

A gbrain commit no longer reverts someone else's commit of the same file

When a script, an agent or a person commits their edit of a page file while gbrain is committing the same file, their commit now stays. If that commit landed between gbrain reading HEAD and moving the branch, gbrain's one retry used to rebuild its commit on top of it with the bytes gbrain had staged earlier. HEAD silently went back to gbrain's version, the next sync reported success with nothing imported, the index kept the other version staged, and later edits to the page could be refused as an uncoordinated local edit. The other writer's bytes survived only in their own commit.

  • A managed brain's Git effect now checks, before each commit attempt, that no commit since it staged its files changed any of them and that the index still holds what it staged. When one did, it commits nothing over it. A file whose edit the other writer committed completes superseded (reason committed_by_another_writer), the next sync imports the other version, and gbrain's write stays in the page's history. Every other file in the group is put back and retried on the new HEAD, so a file that still holds gbrain's bytes is committed on top.
  • A strict (canonical: git) source's write refuses git_target_changed in the same case instead of rebuilding over the other commit, so recovery handles it the way it handles an edit made during the write.
  • To find edits a brain already lost this way, docs/guides/shared-checkout-git.md has a read-only loop over the history.

[0.60.699.0] - 2026-11-12

The Ubicloud gate VMs run a current git

The test suite passes on a gate VM again. Gate VMs ran stock Ubuntu 24.04's git 2.43. GitHub's runners and dev machines are on 2.5x. Three test files depend on newer git:

  • test/effect-git-secret-gate.test.ts needs git maintenance run --detach, which arrived in git 2.47.
  • test/sync-partial-clone-2331.test.ts (both the unit and PostgreSQL arms) needs newer partial-clone behaviour.

Because of that, every VM gate (bun run ci:ubicloud, scripts/pregate-remote-gate.sh when it can't reuse a CI run) failed those files no matter what the change was.

  • scripts/ubicloud/setup-ci-vm.sh now installs git from the git-core PPA, as GitHub's runner images do. It requires at least git 2.47.0 (GIT_MIN) and fails setup with the version it found when git is older. A prebuilt image whose git is older than that no longer counts as ready, so it runs the apt step too. On one standard-16 VM, setup took 97 s and installed git 2.56.1. All three files passed there: effect-git-secret-gate 29/0, sync-partial-clone-2331 unit 12/0, its PostgreSQL arm 23/0. sync-disabled-cycle-window's PostgreSQL arm also passed, 7/0.

[0.60.698.0] - 2026-11-12

A stalled background sync of a large source is retried with a budget that can finish it, and the retries no longer stop silently

  • A large source that the background syncer could not finish no longer retries forever with the same short budget. A resident gives its first run 30 seconds, and every resident restart or change of the process holding the notifier counted as "first" again, so each retry stopped at the same point having saved nothing. That short budget now goes only to a source with no failed attempts on record, and a retry after a run that saved nothing gets the full sync budget.
  • A source the background syncer marked stalled (notifier_stalled, "retrying every 300 s") now really is retried every five minutes. With sync.on_commit off (the default), the retry only happened while the original sync request was still on record, and a run that stopped early could clear that request, so the source sat stalled with nothing retrying it and nothing in the log. The retry now continues until a run completes, also after a restart or when another resident takes over the notifier, which now reads the state the previous one recorded. Turning sync.on_commit off, or excluding the source, still ends a retry that a HEAD move started.
  • gbrain sync --skip-failed, and the automatic skip after three failed syncs, no longer skip files that were refused because of how the source is set up rather than because anything is wrong with them (strict_requires_managed: a canonical: git source on a brain without managed persistence). Skipping them moved the sync checkpoint to the latest commit with no pages imported, after which sync status said synced and later syncs found nothing to do, even once the setup was fixed. The sync now stays blocked, says why, and names the two fixes: turn on managed persistence, or remove canonical: git, after which the next plain sync imports every page. A brain that was already skipped past this way needs one gbrain sync --source <id> --full. A scheduled gbrain sync --skip-failed && … on such a source now stops at the blocked sync until the setup is fixed.
  • A background job that went back to the queue because its worker was killed while running other jobs no longer waits for as long as the next worker stays busy. Such a job only runs alone, so a memory-watchdog stop can be blamed on the right job, and a worker that always had another job running never picked it up: a sync job requeued this way after a restart waited more than ten minutes. Once it has waited a minute, a worker now takes no new jobs until the ones it is running finish, then runs it first. A worker checks for such a job every five seconds, so a queue with none claims exactly as fast as before.

Every push gbrain makes from an unmanaged checkout now scans the commits it publishes for secrets, and a long first push can finish. gbrain git commit-push scanned only with --push-only --scan-secrets, so its commit mode, --pending, the post-commit hook gbrain sources harden installs and scripts/brain-commit-push.sh pushed earlier local commits unscanned; they now refuse publish_secret_found or publish_scan_unreadable like the four pushes below, and the refusal is logged in brain-push.log. The helper's fallback for a machine without gbrain commits and no longer pushes, since it cannot scan. A scan that runs out of its 300 s budget refuses naming the same command with --scan-timeout <seconds>, which gbrain git commit-push, gbrain sources push and gbrain bootstrap repo now take; the scan reads every commit either way, prints its progress on stderr and keeps its temporary disk use to one batch of commits. No push follows tags any more, so push.followTags cannot publish an annotated tag's message unscanned. gbrain sources push reads origin's branch tip from origin itself when the branch tracks another remote (it used to fail the push, never skip commits), prints its refusal with --json and exits 5, and the access-policy commit reports a refusal as committed_not_pushed instead of failing.

Itemized changes

  • src/commands/git.ts: every mode passes scanSecrets (--scan-secrets is still accepted), --scan-timeout <s> (whole seconds, not with --pull), progress lines on stderr, and a scan refusal logged as LOCAL-ONLY, NEEDS ATTENTION ... the push was refused (<code>).
  • src/core/git/range-secret-scan.ts: commit messages read with one cat-file --batch and patches with one diff-tree per 256 commits (the patch file holds one batch, not the whole history), progress lines, PublishScanOptions (budget, progress, retry command), and a timed-out scan refused with an agent fix that reruns with four times the budget (at least 1200 s).
  • src/core/git/commit-push.ts, src/core/workspace-push.ts, src/core/bootstrap/repo.ts: --no-follow-tags --recurse-submodules=no on every push, so no tag or submodule commit goes out unscanned; gbrain sources push fetches origin's branch before it scans and returns the scan refusal as blocked_history.
  • src/commands/git.ts: gbrain git scan [--from <rev>] [--to <rev>] [--json] runs the push's scan over a range and changes nothing (no fetch, lock or push), printing what a push would refuse and each finding's fingerprint for .gbrain-scan-allow; commit-push --dry-run names the managed-worktree marker, or none.
  • src/commands/sources.ts, src/commands/bootstrap.ts: --scan-timeout; sources push prints a blocked_history refusal and exits 5.
  • src/core/access/policy-git.ts: a scan refusal returns committed_not_pushed with the refusal in its note.
  • src/core/brain-repo-durability.ts: the post-commit hook and the helper pass --scan-secrets (so an older gbrain still scans), the helper uses only a gbrain whose git commit-push --help names --scan-timeout, and its fallback commits, logs the commit as local-only and exits 4 (0 when the branch is already on the remote).
  • Tests: test/unmanaged-push-every-site-gst2456.serial.test.ts (each push site against a local bare remote, tags, another upstream remote, the --json refusal, the budget and progress); the #8902, fallback and bootstrap argv tests follow the new contracts.

[0.60.697.0] - 2026-11-12

A quarantined fact or take for a page that already exists is now held for review on a strict source, instead of failing. With trust protections on (write_gate.external_mode = quarantine), a remember or takes_add of tool output that reads like an instruction is held for the owner's review (write_held, a ref such as h1). On a source whose pages are canonical in Git (canonical: git), that write failed when its page already existed (for a fact, also when it had no entity and memory/unattributed already existed). Nothing was saved and nothing was held, so the owner had nothing to review, and the caller was told storage failed: remember answered unavailable with write_error: storage_error ("The write did not commit"), and takes_add answered git_home_required with write_error: storage_error. Now the write is held there exactly as on any other source. Nothing of the held text reaches a fact or take row, a page or Git until the owner decides. Releasing a fact commits it to the page's file before its database row, so a database rebuilt from Git keeps it; a held take still cannot be released on any source, and the owner drops it (gbrain trust drop h1). Sources that are not strict and brains with no Git repository answer as before.

Itemized changes

  • src/core/persistence/memory-prepare.ts, takes-prepare.ts: a held remember or take is a no-op publication (noop: true), as a duplicate fact already was. It writes no page content, so the strict-source check no longer reads it as a page stored without a file in Git (git_home_required), and it no longer queues an embedding effect for the unchanged page.
  • docs/guides/write-refusals.md: the git_home_required row says a held fact or take is not refused this way.
  • Tests: test/gated-hold-strict-page-gst2535.test.ts (a held fact on a strict source with the page, local and over MCP, and with memory/unattributed present: nothing stored before the release, the release committed to the page in Git before its row and kept by a rebuild from a clone of HEAD; a held take on a strict source: nothing stored, the release refused as on every source, the owner's drop; a source that is not strict and a brain with no Git repository unchanged for both); PGLite and Postgres.

[0.60.696.0] - 2026-11-12

gbrain mcp verify now passes its write check on a credential bound to a member. The verifier's test fact was always sent with visibility: "world", which remember refuses for a member credential, so every member-bound connection showed "write" and "write_readback" as failed even when it was set up correctly. A member-bound credential now saves the test fact the way a member's own saves go by default. It is kept private on that member's page, where no other member reads it. The verifier reads the fact back and withdraws it, as before. A credential bound to no member still shares its test fact explicitly, so a brain whose facts.default_visibility is private still reads it back. Page mode (coding-agent profiles and credentials with direct-write prefixes) had a second mistake: its put_page call passed a title parameter the tool doesn't declare, so every page-mode write was refused. The title now goes in the page's frontmatter. The same fixes apply to the verify step inside gbrain connect … --install.

Itemized changes

  • src/core/harness/verify.ts: the fact-mode write reads the credential's bound member from whoami (access.member) and sends visibility: "world" only when there is none. The page-mode put_page drops the undeclared title parameter for a frontmatter title.
  • Tests: test/harness-verify-member-credential.test.ts covers a non-owner member, the owner and an unbound credential, in fact mode and page mode, on PGLite and Postgres (repo-less). It also checks that a member's fixture is private on its member page and unreadable by another member, and that an unbound credential passes under a private facts.default_visibility.

[0.60.695.0] - 2026-11-12

An upgrade no longer adopts another application's OAuth table and then stops halfway, and rerunning gbrain apply-migrations --force-schema now recovers a brain that stopped. gbrain creates oauth_tokens, oauth_clients and oauth_codes with CREATE TABLE IF NOT EXISTS, so when another application sharing the Postgres database already had a table with one of those names, gbrain took it over, added its own columns to it, and a later upgrade stopped at v210 with operator does not exist: timestamp with time zone > bigint. That error named neither the table nor the column, and it came after earlier migrations in the same run had already applied. The documented recovery, rerunning the same command, then could not create gbrain's own table.

Itemized changes

  • Schema setup refuses a table gbrain did not create, before changing anything. Before the schema replay (gbrain init, every engine start that applies migrations) and before any migration in a run, gbrain checks that each OAuth table that already exists has every column gbrain has always given it, with the same type. When one does not, it stops with the new foreign_table error, which names the table, the column, the type it found, the type gbrain needs and the fix: give gbrain its own database, or move the other table with ALTER TABLE <table> SET SCHEMA <other_schema>, then run gbrain apply-migrations --force-schema --yes. A correctly shaped brain runs exactly as before, and v210 is unchanged. Guide: foreign table.
  • gbrain apply-migrations replays the schema before the pending migrations on every path. With the migration lease held (every PGLite brain, and every Postgres brain whose schema has the lease table) it ran the pending migrations without the schema replay, so a table the replay creates that was missing stayed missing and the next migration failed with relation "<table>" does not exist. It now goes through the same schema setup as gbrain init --migrate-only on every path. A brain an older release stopped at v210 recovers with gbrain upgrade and the steps the guide lists.
  • Tests: test/upgrade-foreign-oauth-table.test.ts runs on PGLite and Postgres: the refusal on an upgrade (a foreign oauth_tokens) and on a fresh install (a foreign oauth_codes or oauth_clients), the rerun that recreates a missing table, and the full recovery of a brain stopped at v210.

Reported by @jonahberg (#9190, #9191); first fixed by @Masashi-Ono0611 in #9195 and #9193.

[0.60.694.0] - 2026-11-12

Labels you set on an agent's page now hold, and an agent can no longer take you out of a page's readers. When you made an agent-written page private with gbrain repair visibility (or set its readers with --readers), the page kept its agent tier, so a later agent write at that tier could drop the label again with nothing held. Your label repairs now raise the page's label guard to your tier, so that write is held the same way as one over a page you wrote: it publishes with your labels kept and waits for you in gbrain trust review. Separately, a local write below your tier whose readable_by would leave you, the brain owner, out of a page you read now keeps you in the list (the rest of the narrowing still applies), and its lower_page item records the readers it asked for, which gbrain trust confirm tp<id> applies. Confirming such an item on a page whose own tier is an agent's changes only its labels and leaves the page at the agent's tier, so saying yes to a label change never vouches for the body; on a page of yours, confirm still endorses the edit at your tier. Your own readers changes, synced edits, agent narrowing that keeps you, and brains with the owner opt-in to read restricted pages are unaffected. The guard lives in a new database table and is set only by label repairs run after the upgrade, so labels you set earlier, and a managed Git brain rebuilt from a clone, start unguarded: running gbrain repair visibility --slug <page> or --readers <list> again on a page that already has those labels raises the guard without changing the labels or writing the page (GST-2521).

  • A new schema migration: new page_label_guards table, one row per page, removed with the page. src/core/repair/label-write.ts and the maintenance door's keep_trust path (src/core/persistence/prepared-maintenance.ts) raise it to operator_curated in the label write's transaction when the page is left private or naming readers, and remove it otherwise.
  • src/core/trust/page-write.ts: the label guard reads the page's guard row as a floor, and names the owner as a reader the write must keep when the owner reads the page by its readers and access.owner_reads_restricted is off. The held readers are the requested ones with the owner added, or the stored ones when the request is malformed. An error reading the owner fails the write instead of dropping the owner. The lower_page item records the page's own tier before the write (before_state.page_tier).
  • src/core/trust/page-handlers.ts: accepting a held item whose page's own tier was below operator_curated writes only the requested labels, keeps the page's tier and resolves the item labels_applied (status labelled); on an owner page the confirm is unchanged. Undo of either restores the kept labels.
  • src/core/repair/visibility.ts, src/core/repair/readers.ts: a named page that is already private, or already names the requested readers, and has no guard row is listed (visibility private kept, label guard raised, readers [...] kept, label guard raised) and applied by raising the guard under the page's row lock, with no page write or commit.
  • Tests: test/owner-label-guard.test.ts covers both routes with unbound, other-member and owner-bound reads before, during and after, accept and reject, an agent's own page, a page that named no readers, accept on an agent page keeping its tier, a re-run of the label repair raising the guard on already-labelled pages (repo-less and a managed batch with no commit), the cases that must pass, and the committed file on a managed Git source, on PGLite and Postgres.

A less trusted write that makes a private fenced fact world no longer stores a world fact before the owner decides. An agent put_page that turned a more trusted private ## Facts row world (the same claim at its row number or restated at another, or a different claim at its row number) stored its row world at once, so remote recall returned the private claim while the owner's proposal waited, and accepting that proposal then failed as stale. The writer's row is now stored private and contested, its proposal records the widening, gbrain trust confirm tp<id> makes the claim world and gbrain trust drop tp<id> keeps it private, and later writes of that fence by the same writer, and the background facts reconcile, keep it private meanwhile. A claim the page already holds world is not treated as a widening, and an accept applies the widening only while the page's fence row still asks for world. A remember that restates a private claim world is not covered by this change. The page text still holds the visibility the writer wrote, so a remote page read shows that row until the owner decides (GST-2511).

  • src/core/trust/fence-guard.ts, src/core/trust/supersede-handlers.ts: a contested replacement that would make a guarded private row world, or a new world row restating a more trusted private claim of the page, is inserted private, its supersede_fact proposal carries visibility: {from, to}, the owner's accept sets it (undo restores it), and a lower-tier write keeps that contested row private while its widening is pending, rejected or undone (GST-2511).
  • src/core/cycle/extract-facts.ts: the fence reconcile keeps the visibility of a contested row whose widening proposal is pending, rejected or undone, instead of widening it to the page's cell (GST-2511).

An agent edit can no longer make a private page readable before you decide. A local put_page (or capture, edit_page, append_page, revert_version, restore_page) from a writer less trusted than the page used to publish at once even when it dropped or widened the page's visibility, access, confidentiality or readable_by. The only safeguard was a lower_page item in gbrain trust review, and until you dealt with it, remote callers and other members could read the page. The usual cause was an agent that rewrote a page and forgot its frontmatter. Now that write publishes with the page's earlier labels kept, in the database and in the file it commits, and its receipt names the kept keys (privacy_labels_kept). The lower_page item shows the labels the edit asked for. gbrain trust confirm tp<id> applies exactly those labels if the page hasn't changed since, gbrain trust drop tp<id> keeps the earlier ones, and undoing a confirm puts them back while the page is still as the confirm left it. The same hold covers a derived page publication by a maintenance job (for example a dream patterns run that rewrites one of your pages), a page with no trust tier recorded yet (unknown), which counts as yours, and an agent that deletes a page and writes it again: a delete no longer lowers the page's tier. A credential bound to you reads the page throughout. Unaffected: writes that only narrow a label, writes that leave the labels alone, writers at or above the page's tier, your own synced edits, and gbrain repair visibility (GST-2518).

Itemized changes

  • src/core/trust/page-write.ts, src/core/persistence/page-prepare.ts: a local content write below operator_curated, and below the page's label guard tier, keeps the stored labels it would widen, and so does a derived page publication. The guard tier is the page's tier (operator_curated lowered by its frontmatter caps when the tier is unknown), raised to the earlier tier recorded on every lower_page item of the page since the last one you confirmed. The lower_page item records the kept and requested labels, and a later write by the same tier replaces them with what it asked for, so a write that puts the label back withdraws the request. A write that changes only labels publishes nothing and still files its item.
  • src/core/trust/page-handlers.ts: confirming a held item writes the requested labels in the same owner page write that confirms the page. New undo for a confirmed held item restores the kept labels and the edit's tier, and only while the page is at the revision the confirm wrote.
  • src/core/persistence/page-prepare.ts: a soft delete keeps the page's tier (it writes no content). src/core/persistence/prepared-maintenance.ts: a derived publication whose labels were kept files one item, with the labels.
  • src/core/trust/owner-actions.ts, src/core/trust/review.ts: the confirm and drop previews and the review item name the held labels.
  • Tests: test/lower-tier-label-hold.test.ts covers each label dropped and widened, remote reads (unbound, another member, the owner), accept, reject, undo, a stale undo, a withdrawn request, the two-step and reject-then-lower bypasses, a delete then a write, an unknown page, a derived maintenance publication, the cases that must pass unheld, and the committed file on a managed Git source, on PGLite and Postgres.

[0.60.693.0] - 2026-11-12

The brain owner's own agents can now withdraw a private fact the brain host saved. A credential bound to the brain owner (gbrain access owner set) reads private facts, but its forget refused one the host had saved with visibility: "private" and said only the brain host could withdraw it. Now, speaking for itself, the owner's credential (or a stdio server started with --stdio-member <owner>) withdraws a private fact it reads when the page that holds it is one the credential can already write. The fact is struck in the page's file on a brain with a Git repository, and on a strict source a fact held for an entity with no page is struck in memory/unattributed, so a rebuild from Git keeps it withdrawn. Three things stay as they were. A fact on a visibility: private page, on a page whose readers leave the owner out or on another member's page is still withdrawn on the brain host (gbrain forget <id>). A fact the owner confirmed or curated still asks the owner first (forget_requires_owner). Member credentials, attested calls (x-qm-actor), audiences, subagent jobs, slug-fenced clients and unbound connections get exactly the answers they got before.

The owner's agents can also withdraw a fact on a page that lists the owner as a reader. Since 0.60.682.0 the owner's recall shows facts on a page whose readable_by names the owner, but its forget and forget_fact answered not_found for them. Now they withdraw such a fact, private or shared, when the owner's credential may also write that page and the page of the entity the fact is about, if it has one, is one the owner reads the same way. The owner opt-in that lets the owner read pages kept to other members (gbrain access owner reads-restricted on) does not extend to forget: those facts are still withdrawn on the brain host. Every caller that is not the owner speaking for itself gets the answer it got before.

A member's save no longer reveals that a claim it never read was withdrawn. When the brain host or the owner withdrew a claim, a member that later saved the same claim on its own member page was refused with fact_withdrawn, which told it that a claim it could never read had existed and been forgotten. Now that save succeeds and gets the same reply it would get if the other claim had never existed. The fact is the member's own, it stays active, and the member's recall returns it. A save is still refused when the member could read what was withdrawn, which means a fact it forgot itself, its own copy that the brain host's forget withdrew, or for the owner, a fact it reads anywhere. A schema migration changes the two database guards that mark a newly saved row as withdrawn so a row on a member page is marked only by that page's own withdrawals; it rewrites no rows. Brains with and without a Git repository behave the same way.

Itemized changes

  • src/core/persistence/memory-mutations.ts, forget-prepare.ts: the owner's forget reaches a fact on a page that names it as a reader (ownerReachesFact): that page, and the fact's entity page where one exists, must pass GST-2490's owner read rule (ownerReadPolicyFragment in search/private-visibility.ts, the opt-in off) and the read rule a member write is held to, for the owner's identity. Preparation re-resolves that identity from the stored principal and checks the fact again under the publication's locks, so a page that drops the owner from its readers after submission refuses not_found.

  • src/core/facts/single-prepare.ts: a remote save onto a member page meets the page's own withdrawals and a withdrawal of another subject only when a withdrawn row is one the page's member reads (memberPageWithdrawal). Callers in memory-prepare.ts and ops/qm-memory.ts pass remote.

  • src/core/facts/withdrawal-schema.ts, insert-guard-schema.ts and a new schema migration make the insert guards mark a row on a member page only by that page's own withdrawals; access/member-facts.ts (withdrawalCoversFactSql) and facts/embedding-identity.ts treat such a row as covered by its own page's withdrawals only. A withdrawal still expires every matching row, member pages included, when it is made.

  • Tests: test/owner-facts-d1-gst2346.test.ts (the owner's forget on a page that names it, the nine non-owner caller shapes and a member reader unchanged, the opt-in, preparation from the stored principal and under locks, the managed-brain file; the member's save compared byte for byte after a host and an owner withdrawal, and the refusals that stay); PGLite and Postgres.

  • src/core/persistence/memory-mutations.ts: forget and forget_fact let the owner speaking for itself through for a private fact off every member and scope page, on a page no remote writer is kept from, and mark the stored request. The refusal that names the brain host now tells the owner what its credential does withdraw.

  • src/core/persistence/forget-prepare.ts, src/core/persistence/page-visibility.ts: preparation judges the owner again from the stored principal (storedOwnerSpeaksForItself), and once more under the publication's locks. A writer that is no longer the enabled owner, or that named an actor, an audience, a subagent or a slug fence, reads the fact as any remote caller does and gets not_found.

  • Guards kept: the same claim active on a page kept from remote writers refuses claim_linked and withdraws nothing; a more trusted fact files the forget_requires_owner proposal, which stores the fact's id and tier only.

  • Docs: access designations, MEMORY_VERBS_v1.md, the access key-files entry; behavior-change notice.

  • Tests: test/owner-facts-d1-gst2346.test.ts (the owner's forget and forget_fact, every other caller shape unchanged, the trust proposal and its accept, preparation from the stored principal, the managed-brain file) and test/remember-strict-held-gst2492.test.ts (a held private fact on a strict source, with a rebuild); PGLite and Postgres.

[0.60.692.0] - 2026-11-12

On a source whose pages are canonical in Git, every fact writer now commits a fact with no page of its own to Git, and doctor and a new repair find the ones an older gbrain left in the database only. In the last release only remember did this. Google loops, fact extraction, the managed facts writer and an owner releasing a quarantined fact now also commit a fact with no entity, or for an entity with no page, as a row of memory/unattributed. Before, they kept it in the database only, where a rebuild from Git lost it, or were refused with git_home_required once that page existed. Sources that are not strict, and brains with no Git repository, behave as before.

  • A fact saved under a bare name moves when its page appears in a folder. remember keys a new name such as "Kofi Example" as kofi-example. Once exactly one page with that name exists, such as people/kofi-example, the maintenance cycle moves the held fact into it in one commit and re-keys it to that page. If two pages could be meant, the fact stays held, and the doctor check and the repair preview name both pages.
  • New doctor check: strict_facts_outside_git. For each strict source it counts facts with no Git home, facts held more than 30 days for a page that never appeared, and held facts that cannot move. Its fix is the read-only preview of the new repair.
  • New repair: gbrain repair facts-home --source <id>. It is explicit-only and preview-bound: --apply --expect <hash> writes the listed facts to memory/unattributed, at most 500 per run, and it never runs on upgrade. It lists, without moving, the facts it cannot place. Those are facts for an entity under a folder access rule, a row the Facts table cannot show, an ontology observation, a row of a deleted page, and a fact whose entity page is deleted (that page may have been private, and purging it would not reach a row on the holding page, so the fact is listed until the page is restored or purged). The new writers skip such an entity the same way.
  • Known limit: a fact for an entity under a folder access rule that has no page yet still stays in the database only, because the holding page is readable by everyone. The doctor check counts these facts and the repair lists them as folder_restricted.
  • The entity_page_missing notice asks before it creates a page. While the fact is in the database only, the notice's step is now next: ask_user: its user_message says the fact is saved in the database only and asks whether to create a page for the entity, and the put_page call is the fix once the user agrees. When the fact is already held in Git, the notice only says where it is and that it moves once the page exists, with no step to run. This holds for single and batch remember, over MCP and in the CLI's [AGENT] block.
  • A sync into a strict source on a brain without managed persistence says why it imported nothing. The failure ledger, the summary and --json (failureCodes, plus a new failureAdvice) now carry strict_requires_managed instead of UNKNOWN. The advice no longer sends you to fix YAML or to use --skip-failed. docs/guides/git-canonical.md now shows how to rebuild a strict source's database from a clean clone.
  • The same claim held for two entities stays two facts. Fact extraction, and the repair, treated two memory/unattributed rows with the same text but different entity cells as one duplicate, so one was dropped from the index and the repair could not apply.
  • Purge readers now read memory/unattributed's entity column. The guard that drops a failed write's stored copy of purged content, and a page purge's claim inventory, used to see that page's held rows as a malformed table and miss them.
  • Docs: docs/guides/repair.md#facts-home, docs/guides/git-canonical.md, docs/guides/write-refusals.md, docs/guides/entity-scoped-facts.md, the remember section of docs/mcp/TOOL_REFERENCE.md and docs/protocol/MEMORY_VERBS_v1.md. Behaviour-change rows disclose each change after an upgrade.
  • Tests, on PGLite and Postgres: test/strict-background-writers-gst2492.test.ts, test/repair-facts-home-gst2492.test.ts, test/held-bare-name-move-gst2492.test.ts, test/sync-strict-unmanaged-gst2492.test.ts, test/held-fence-readers-gst2492.test.ts, test/facts-home-review-gst2492.test.ts and test/doctor-strict-facts-outside-git.test.ts. The entity_page_missing assertions in test/remember-pageless-entity-gst2446.test.ts and test/remember-strict-held-gst2492.test.ts now pin ask_user and the held form.

[0.60.691.0] - 2026-11-12

A privacy label repair on a database-only page now sticks, and a supported command stops tracking a folder's files without deleting their pages. Pages under a folder declared storage.db_only in gbrain.yml live in the database. Some of those folders still had files Git tracked from before the declaration, and on a git-canonical source every sync put the committed file back over the database, undoing a gbrain repair visibility that had made the page private. Sync now keeps the database version there, a commit to such a file can no longer widen who reads the page, and gbrain git commit-push --untrack takes the old files out of Git safely so the problem cannot come back.

Itemized changes

  • Label repairs under storage.db_only stick. On a strict (git-canonical) source a page whose file Git still tracks under a declared storage.db_only folder keeps a database change across gbrain sync and gbrain sync --full. A commit that changes the file still imports.
  • An import there never widens. On every managed source, a commit to a tracked file under storage.db_only keeps each privacy label (visibility, access, confidentiality, readable_by) the page has and the file does not state. A label the file states applies, a wider one included, so the owner can still widen on purpose. Files outside storage.db_only import exactly as before.
  • Ignored pages take label repairs. On a strict source, gbrain repair visibility on a page saved under a path Git ignores (git_ignored, no file) writes the label in the database instead of holding the page canonical_file_missing.
  • gbrain git commit-push --untrack -m <msg> -- <paths> removes the named files from Git and the index, keeps them on disk and commits only those deletions under the commit lock; --dry-run --untrack previews and refuses the same way. In a managed canonical worktree it only takes files under a storage.db_only folder that gbrain.yml at HEAD declares and .gitignore ignores (else untrack_not_db_only, nothing changed), never pushes, and every page stays live with its content. A plain commit-push of a removal staged by hand with git rm --cached now points at --untrack. Guide: untrack.
  • Tests: test/db-only-labels-gst2519.test.ts, test/sync-classify-strict-g3.test.ts and test/git-commit-push-untrack-gst2519.test.ts run on PGLite and Postgres, strict and non-strict sources.

[0.60.690.0] - 2026-11-12

Nightly checks stop failing on their own fixtures and sampling

Three nightly checks were failing because of how they measure, not because of gbrain. All three now test what users get. The Codex plugin check expected gated tools to be listed even when their publish gate was off. The pre-v0.18 upgrade fixture couldn't build its legacy database because a newer trigger depends on the column it removes.

  • test/e2e/codex-plugin-install-real.serial.test.ts: the expected tool list now resolves publish gates the way gbrain serve does for the seeded brain over stdio. Wakeshift's watch_job, list_watches and unwatch (#8670) and set_memory_writeback (#8481) are listed only when their gate is on. Before this change the check failed every night from 2026-11-06 on; it now passes against codex 0.147.0, the version the workflow pins.
  • tests/heavy/fixtures/down-mutate-pre-v0.18.sql: before dropping pages.source_id, the fixture drops every trigger that depends on that column. The first such trigger was the per-folder access stamp from #8134. The walk forward re-creates those triggers. tests/heavy/pg_upgrade_matrix.sh failed on the pre-v0.18 shape before this change and passes on both shapes after it.
  • scripts/bench/sync-freshness-gate.ts: the Postgres managed serve-watch cell is now measured 3 times, and the gate uses the median of the three p50s, as the PGLite managed cell already did. The baseline and the 25% limit are unchanged. The cell's nightly p50 moves in steps of the 50 ms probe. Single runs at 656.5, 659.2 and 706.8 ms failed the 627 ms limit on 3 of 19 nightlies, and each passed when rerun on the same commit. A regression that slows every run still fails.
  • scripts/ci-census-tests.sh: the pull-request census job now also runs test/scripts/run-unit-shard.test.ts (1.5 s, no database). It reads every test file's per-test timeouts, so a bad timeout now fails the pull request instead of a later full run.
  • scripts/sharding.ts, .github/workflows/test.yml, .github/workflows/e2e.yml: the unit suite now splits into 30 shards instead of 28, run two per job (15 jobs), and the shard weights are refreshed from the median of five Test nightlies and the latest full E2E nightly. On the new weights 28 shards came to about 601 s each, over the 600 s per-shard budget; 30 shards come to about 561 s. Both coverage merges now expect shards 29 and 30.
  • scripts/ci-typecheck.sh, .github/workflows/test.yml: when the typecheck job's tsc aborts without a type error (exit 134 or a signal, as when V8 hits the heap limit), the job deletes the restored tsbuildinfo and runs tsc once more from cold. A run seeded with a distant head's state could go past the heap cap where a cold run of the same tree fits, and because the aborted run saved that same state under its own head's key, every rerun failed the same way. A type error (exit 1 or 2) still fails at once with no retry, and a run whose tsc aborted or timed out no longer saves its state. test/scripts/ci-typecheck.test.ts pins the retry to aborts only. The macOS validation typecheck job (.github/workflows/macos-validation.yml) uses the same step, with two 15-minute attempts.
  • scripts/postgres-unit-arms.ts, .github/workflows/persistence-validation.yml: the PostgreSQL arms of unit-lane suites now run in 21 shards instead of 19. On current weights the largest shard planned at 554.6 s against its 570 s budget, so the next few arm files would have failed the budget test; at 21 it plans at 502.3 s. The budget and its assertion are unchanged.

[0.60.689.0] - 2026-11-12

Three opt-in ways to share what is derived from member-restricted pages: raw-only folder rules, audience members the owner reads, and member lineage reads. A folder rule kept the summaries, atoms, chronicle events and facts made from its pages as tightly as the pages themselves, so a transcript folder kept to one member hid every summary of those calls from everyone else; a group chat's own memory page was closed to the brain owner; and only the owner could read the derived pages and facts of pages it already reads. Each now has an operator setting, off by default, and nothing changes on upgrade: gbrain access rule add --raw-only keeps a folder's pages to its readers but lets whoever reads a derived page read it (use it for transcripts, never for a folder whose content is itself the secret: summaries of a raw-only page, and the facts extracted from it, are readable like any other derived page, so the facts pulled out of a fenced transcript are not fenced); gbrain access member kind <id> audience lets the owner read that member's own member page, versions, derived pages and facts; and gbrain access members-read-lineage on lets every member read the derived pages and facts of pages it reads, as the owner does. All three work the same on a repo-less brain and with rules in access.yml, only the operator on the brain host sets them, and no write widens.

Itemized changes

Access lineage designs (GBRA-108 designs 2, 3 and 4)

  • Migration v305 (src/core/schema-migrations/v305-access-lineage-designs.ts): access_rules.covers_derived (default true), access_members.kind (person | audience, default person), the SQL function gbrain_access_lineage_rules(), and new versions of gbrain_access_origin_restricted(), gbrain_access_origin_closed_for() (10 arguments) and gbrain_access_origin_hidden_for() (9 arguments); the v304 signatures stay as wrappers. Nothing is rewritten, and every existing rule keeps its digest.
  • Raw-only rules: src/core/access/folder-rules.ts (covers_derived: false in access.yml, the held widening derived_released and the narrowing derived_covered), src/core/access/policy.ts (stored, compared and previewed with lineage in gbrain access apply), src/commands/access-rules.ts (rule add --raw-only).
  • Audience members: src/core/access/members.ts (setMemberKind, member add --kind, refused for remote callers with permission_denied), src/core/search/private-visibility.ts (the owner's read of an audience member page, its versions, derived pages and facts), gbrain access member kind <id> person|audience; member list marks audience members.
  • Member lineage: config access.members_read_lineage (src/core/access/lineage.ts, gbrain access members-read-lineage [on|off], refused for remote callers); src/core/access/caller.ts and src/cli/as-preview.ts give it to a member speaking for itself, never to an audience of several or an unbound credential (an unbound granted_sources token included), while a member-bound granted_sources token gets its member's lineage; whoami adds members_read_lineage while it is on.
  • src/cli/enforced-member.ts: the enforced member CLI's output guard judges slugs under the record the read ran under, so with member lineage on it no longer withholds output naming a derived page the member reads.
  • The semantic cache key (src/core/search/mode.ts) and the per-caller key (src/core/access/member-facts.ts) carry the member lineage and the owner's audience members.
  • gbrain access check and who-can-read explain all three; doctor access_raw_only_rules, access_audience_members and access_members_read_lineage (src/commands/doctor/checks/access-lineage.ts).
  • Docs: docs/guides/access-designations.md (raw-only rules, audience members, members read lineage), key-files entry.
  • Tests: test/access-raw-only-rules.test.ts, test/access-audience-members.test.ts, test/access-members-read-lineage.test.ts (every read path for every caller, off and on, the held widening, remote refusals, cache keys and the write gate; PGLite and Postgres, rules in the database and in access.yml); shared fixture test/helpers/access-lineage-fixture.ts; migration, catalog and doctor goldens.

[0.60.688.0] - 2026-11-12

Sync no longer wedges on a source with thousands of old file-drift holds, and sync status stops calling a source synced while new commits wait

Since v0.60.677.0, every unknown-side file_drift hold written under the earlier drift screen is screened once more, and that screen reads each held path's earlier committed versions. It read them with one Git process per path and reported nothing while it ran. On a source with thousands of such holds, the pass stayed silent past the sync's 300-second no-progress check and was stopped (drain_stalled) before it stamped any hold, so every later pass repeated the same reads and the source never synced again.

  • The re-screen now reads every held path's history through one git cat-file --batch-check per 2,000 paths, with no file bodies, and reports progress as each path's history arrives. A Git read that hangs still reports nothing, so the no-progress check still stops it. On a generated repository with 80,000 files and 3,000 commits, 10,000 holds took 6.2 s instead of 141 s, and 80,000 holds took 66 s instead of 17 minutes. The longest silence was 36 ms instead of the whole read.
  • A pass stopped partway keeps the holds it re-screened, and the next pass reads only the rest.
  • gbrain sync status reads resumable / commits_waiting when nothing else is pending (or only holds that need no person) but the checkout's HEAD is past the last synced commit, with a next that runs the sync. It used to read synced until the next sync happened to run. A hold that needs a person keeps its own reason, and a company-brain source, which syncs to its approved revision, never reads it.
  • With sync.on_commit off, the notifier's fix now says so, instead of saying that no resident runs.

[0.60.687.0] - 2026-11-12

An upgrade can no longer run the same schema step twice at once. When two gbrain processes started on one brain together (a cron job and an agent, or two hosts after an upgrade), both could apply the same schema migration. On Postgres reached through a transaction-mode pooler such as PgBouncer or Supabase's pooler, the lock meant to keep them apart did not hold. The pooler hands each statement to whichever server connection is free, so the lock could be taken on one connection and left behind on another. The second process then ran the step too, which showed up as a failed migration (for example a duplicate key on pg_proc while a function was being created) or as a step applied twice, and a lock left behind could keep later upgrades waiting until they timed out. Every path that changes the schema now takes one lock that holds for the whole run, through a pooler or a direct connection, and on PGLite. There is nothing to do beyond upgrading.

Itemized changes

  • The schema setup a gbrain process runs when it opens the brain (which gbrain init and gbrain upgrade go through) and gbrain apply-migrations take the same schema-migration lock, and each migration step checks that the lock is still held before it runs.
  • On Postgres the lock is a transaction-scoped advisory lock held on one reserved server connection for the whole run, so a transaction-mode pooler cannot move it, and it is released when the run ends or the connection drops. A connection pool with only one connection keeps the previous session lock.
  • While one process holds the lock, another waits for it and then finds the steps already applied. If the wait times out, the message names the process holding the lock and, when that process is a running migration, says to let it finish rather than terminate it.
  • The idle-connection check that runs before a migration no longer counts the lock's own connection as a blocker.
  • Tests: test/e2e/schema-migration-lock-postgres.test.ts (direct Postgres and PgBouncer), test/e2e/schema-migration-lock-direct-pool-postgres.test.ts and test/schema-migration-lock-pglite.test.ts start two runners at once and check that each step runs once and that no lock is left behind.
  • For contributors: bun run typecheck raises its Node heap ceiling from 6,144 MB to 8,192 MB, because a cold tsc run now peaks at about 6.5 GB and one CI typecheck job aborted at the old ceiling.

[0.60.686.0] - 2026-11-12

The brain host can now set who reads a page without lowering its trust. gbrain repair visibility --slug <page> --readers alice-example,group:team-a (or --slugs-file <file> for a long list, --readers none to remove the field) sets a page's readable_by through the same label-only door: the page keeps its tier and provenance, no lower_page proposal is filed, and its file changes only on the readable_by line. Before, the only way was a put_page, which lowered the page to "written by an agent". Readers are checked first. An unknown or disabled member or an unknown group refuses with nothing written, and a list that leaves out the brain owner refuses unless --allow-without-owner is passed, because readable_by binds the owner too. A list of thousands holds no run-wide lock. On a managed brain each admission window of pages is admitted together (Postgres publishes them in groups of up to 8), the Git effect commits the ready files together (up to 100 per commit), and --json counts them by effect state in details.git. A managed page with no file outside a declared storage.db_only path is now held canonical_file_missing with nothing written (it was refused as a failed source_changed), for every repair visibility write.

A visibility repair now closes a fenced fact instead of leaving a world copy beside it. When gbrain repair visibility --class facts made a ## Facts row private and the stored row was more trusted than the repair, the fence guard kept the original world row live and filed a contested private copy, so remote callers could still read the fact while the page and its file read as fixed. A write that only makes a fence row private now makes the stored row private in place, keeping its id and trust tier, and files no proposal. On a brain without managed persistence the repair makes the fenced row private itself, because that import does not re-project the fence. A write that makes a more trusted private row world is still held for the owner.

Itemized changes

Set who reads a page, label-only (GST-2476, #9184)

  • src/core/persistence/label-edit.ts: the label edit also takes readable_by, written as a one-line list (readable_by: [a, b]); block lists and duplicate keys are declined. LABEL_KEYS is unchanged, so gbrain trust backfill --label-lowerings reads as before.
  • src/core/repair/readers.ts (new), src/core/repair/visibility.ts, src/commands/repair.ts: gbrain repair visibility (--slug <slug>... | --slugs-file <file>) --readers <ids|none> [--allow-without-owner], preview-bound (--apply --expect <hash>), up to 10,000 pages per run; reader validation (access_member_invalid, owner rule); details.git effect counts on apply; a managed run admits each window of pages together (applyBatch, publishMaintenancePages under the run's batch id); a managed page with no canonical file outside storage.db_only held canonical_file_missing.
  • src/core/repair/label-write.ts: a readers run publishes its pages under one maintenance batch.
  • Tests: test/readers-label-setter.test.ts (managed Git source with tracked files, a storage.db_only page and a page with no file; repo-less brain with remote reads by member, a 150-page slugs file, refusals and preview_changed).

A visibility repair closes a fenced fact (GST-2487, #9188)

  • src/core/trust/fence-guard.ts: a guarded row whose claim the writer kept and whose visibility it only narrowed (world to private) moves to the new row's position and turns private in place, instead of being detached and contested.
  • src/core/repair/visibility.ts: on a brain without managed persistence, a fenced fact row the label write made private in the fence is made private in the database too; fact_left_readable and fact_left_world stay for the rows that still cannot be closed.
  • Tests: test/label-only-repairs.test.ts (managed and repo-less, PGLite and Postgres) and test/trust-guarded-supersession.test.ts (narrowing applies in place; widening stays guarded).

[0.60.685.0] - 2026-11-12

remember now tells the truth about a fact saved for something that has no page yet, a source that lives in Git keeps that fact in Git, and the brain owner's own agents save facts shared by default. Three changes to how agents save and withdraw facts land together.

A fact saved for an entity whose page does not exist yet (remember never creates a page) used to get a reply that named the missing page, so an agent would report the fact as saved to a file. The reply now says the fact is stored in the database only and names the put_page call that creates the page. On a source whose pages are canonical in Git (canonical: git), such a fact no longer lives in the database alone, where a rebuild from Git would lose it: it is committed to Git as a row of memory/unattributed that names the entity, and maintenance moves it into the entity's page once that page exists. Other sources and brains with no Git repository store it as before.

A credential bound to the brain owner (gbrain access owner set) reads the whole brain, but its remember kept every fact private on the owner's member page unless the call passed share: true, its visibility: "world" was refused, and its forget refused a fact it had just shared. Those private facts also stayed out of hot memory, turn context and context packs, which carry shared facts only. Now, speaking for itself, the owner's credential saves a fact with no visibility on its subject's page for every reader of the source, keeps it private on its member page when the call passes visibility: "private", when the page or facts.default_visibility keeps facts private, or when it replaces a fact on that page, and withdraws a shared fact it reads. The trade-off: an owner agent that follows a planted instruction can now share a fact without passing a flag, so pass visibility: "private" for a fact that must stay with the owner. Member credentials, attested QM calls (x-qm-actor), audiences, subagent jobs and slug-fenced clients keep what they had.

The changes meet on a strict source. The owner's save for an entity with no page is held, shared, on memory/unattributed like any other shared save, the owner's forget strikes it there, and the brain host's purge removes it from the file.

Upgrade every gbrain host of a strict source together, because an older gbrain holds a memory/unattributed file that carries the new entity column as a malformed fence, with no data loss.

Reported by @andreineacsu (#9101). First fixed by @Masashi-Ono0611 in #9103.

Itemized changes

The reply for a fact on no page (GST-2446, #9165)

  • The reply names a page only when the fact is on one. For an entity with no page it used to carry page: "<the entity slug>", a page that does not exist; now the field is absent. A pending reply's target.slug follows the same rule, and so does a fact with no entity on a source that cannot hold the memory/unattributed page.
  • Each items entry of a batch carries persistence and write_through, the same fields the single form returns. Before, a batch item gave no sign of whether it landed in a page file or in the database alone.
  • When the source has a checkout, the reason no file was written is write_through.skipped: "entity_page_missing", in the reply and in the write receipt. It used to read no_repo_configured, which is false for such a source. A source with no checkout still reads no_repo_configured.
  • Where the brain's pages are in a Git repository, the reply adds one entity_page_missing notice per entity per session with the put_page call that creates the page. Once the page exists, the next maintenance cycle can move the facts saved for that entity into the page's Facts table and its file. A brain with no Git repository gets no notice, because there the database is the record.
  • gbrain remember prints note: the entity has no page, so this fact is stored in the database only, and a batch line ends (no page: database only).
  • Docs: docs/guides/entity-scoped-facts.md, the remember section of docs/mcp/TOOL_REFERENCE.md and docs/protocol/MEMORY_VERBS_v1.md describe the reply. One behavior-change row discloses the reply change after an upgrade.
  • Tests: test/remember-pageless-entity-gst2446.test.ts runs on PGLite and Postgres against a Git checkout, a content folder with no Git and a source with no checkout, single and batch, through the operation handler and through MCP dispatch. It also pins that the fact reaches the page's Facts table and its committed file once the page is created.

A strict source keeps a pageless fact in Git (GST-2492, #9179)

  • The reply, each batch item and the write receipt say where the fact is: page: "memory/unattributed", persistence: {mode: "filesystem"}, write_through: {written: true, held_for: "<entity>"}, and the entity_page_missing notice says the fact is saved and committed and when it moves.
  • The extract_facts maintenance phase moves held facts into a new entity page in one commit. The fact keeps its id, provenance and embedding, and the held row is struck (moved to <entity>); when the page is restricted the held row is removed instead, so the world-readable holding page keeps nothing of it. A page kept out of Git does not take the fact; the run warns HELD_FACT_UNMOVABLE and the fact stays held.
  • forget strikes a held fact's own row, and purge_fact removes it from memory/unattributed as it removes a fact with no entity, so a rebuild does not bring it back (a purge of a held fact also drops every copy of the same claim text on memory/unattributed, held for another entity or with no entity); remember … replaces supersedes a held fact from memory/unattributed or from the entity's page.
  • A client bound to slug prefixes may save a held fact for an entity inside its prefixes; a fact with no entity is still refused for it. An entity under a folder access rule is never held on the world-readable page, so its fact stays in the database, as before.
  • Only memory/unattributed reads the entity column, and an invalid entity cell makes its fence malformed. Every other page parses exactly as before. An older gbrain holds a memory/unattributed file that carries the column as a malformed fence, so upgrade every host of a strict source together.
  • Tests: test/remember-strict-held-gst2492.test.ts on PGLite and Postgres: the commit before the reply (handler, batch, MCP), rebuilds from a clone of HEAD, the move (one commit, interrupted after its commit and replayed), restricted and database-only pages, forget, purge, both supersede paths, a prefix-bound client, duplicates and the fence column.

The owner's own facts are shared by default (GST-2346 D1, #9178)

  • An explicit visibility: "world" onto a page whose facts are kept private is refused (share_target_restricted), as share: true is, and the reply still names the page and visibility the fact got. Works the same on a brain with no Git repository and on a managed one.
  • src/core/access/member-facts.ts: ownerSpeaksForItself (CallerAccess.owner, no attested actor, no audience, not a subagent, no slug fence).
  • src/core/persistence/memory-mutations.ts: remember takes the ordinary route for the owner and moves a default save the page or brain default keeps private to the owner's member page; forget takes the ordinary remote path for the owner when that path reads the fact. src/core/verbs.ts: no member_share_required for the owner.
  • src/core/facts/writeback-config.ts, src/core/ops/memory-settings.ts, src/mcp/http-transport.ts, src/commands/serve-http-mcp.ts: with writeback on, the owner's credential gets the instructions of a credential bound to no member.
  • Docs: access designations, the company-brain tutorial, TOOL_REFERENCE.md, MEMORY_VERBS_v1.md, the access key-files entry, skills/brain-ops/SKILL.md; behaviour-change row owner_facts_shared_gst2346.
  • Tests: test/owner-facts-d1-gst2346.test.ts (every caller shape, repo-less and managed, PGLite and Postgres); test/member-forget-refusal-9101.test.ts now expects the owner to withdraw a fact it shared.

Where the changes meet

  • test/remember-strict-held-gst2492.test.ts adds an owner-bound OAuth credential on a strict source, on PGLite and Postgres. Its save for an entity with no page is held, shared, on memory/unattributed with #9179's reply, its forget strikes the held row, purge_fact over MCP is refused trusted_local_only, and the host's purge removes the claim so a rebuild from Git does not bring it back.
  • docs/guides/memory-boundaries.md: the purge section says a purge of a held fact drops every copy of the same claim text on memory/unattributed.

[0.60.684.0] - 2026-11-12

An agent that runs gbrain for one member can now be held to that member's view of the brain. Set GBRAIN_ENFORCED_MEMBER=<member> in the launcher that starts the agent's gbrain calls, and every command it runs reads exactly what a remote credential bound to that member would read: pages kept to other members, private pages, folders the member is outside of and other members' personal sources leave no slug, title, body, fact or count behind. Writes still land where they land today, as trusted local writes, but the mode refuses a write to a page the member cannot read and every change to a page's readers or visibility. remember saves a fact only with --entity naming a page the member reads, and its refusal says so with the exact retry command. gbrain write-request <id> reads the member's own receipts, and any other id answers exactly like an unknown one. sync, sources refresh and git commit-push keep working with their output replaced by one line (sync and refresh take only --source <id>, --no-pull, --no-embed, --json and --dry-run, on a source the member reads), doctor prints only check names, statuses and counts, and hook succeeds silently so group turns never fail. Every other command is refused with one exact line, exit 3 and the code enforced_member_refused, --help included, and so is everything when the value is empty, names an unknown member or names the owner. Unset, nothing changes. The mode binds only a caller whose launcher pins the variable; a process that can unset it, or read the brain's files or database directly, is outside it. Guide: docs/guides/enforced-member.md.

Itemized changes

  • src/cli/enforced-member.ts, src/cli/enforced-member-classes.ts, src/cli/enforced-commit-push.ts: the gate, the allow-list (every CLI command and operation classified once) and the read, write and commit-push checks; src/cli/main.ts calls the gate before any dispatch.
  • src/core/cli-force-exit.ts: setStdoutCapture, so the mode also withholds what a command writes through the fd-1 chain (--json documents).
  • src/mcp/dispatch.ts: DispatchOpts.enforcedMember forces hidden privacy and no owner access on a dispatched read.
  • src/commands/recall.ts: recall reads through the member filter under the mode; src/cli/cli-socket-client.ts, src/cli.ts: the warm socket and the skill-resolver fast path are skipped under the mode.
  • src/core/error-registry.ts: new code enforced_member_refused (regenerated docs/guides/error-codes.md).
  • Known limits: a PGLite brain with a running gbrain serve refuses every enforced-mode read; a thin-client install refuses; creating a page at a slug the member cannot read refuses where a free slug would be created.
  • Tests: test/enforced-member.serial.test.ts (CLI subprocesses on PGLite and Postgres, repo-less and git), test/enforced-member-classes.test.ts.

[0.60.683.0] - 2026-11-11

Making a page private through an operator repair no longer lowers its trust or rewrites its file. On a managed brain, an operator who added visibility: private to one page through a local put_page, or ran gbrain repair visibility, saw the page drop from "your notes" (operator_curated) to "written by an agent", a pending lower_page proposal only a typed token could settle, and a re-rendered file: a slug key removed, a source_kind value overwritten with put_page, keys reordered, and provenance stamps and a trust_tier marker added. Both repairs and gbrain repair private-labels now write through one label-only door. The page keeps its tier, no proposal is filed, and its file changes only on the label lines, with every other key, its order and the body kept byte for byte. A new slug-bound form, gbrain repair visibility --slug <page>, is the operator's single-page way to make a page private (preview first, then --apply --expect <hash>). A plain put_page, local or remote, is still an agent edit and lowers the page as before. gbrain trust backfill --label-lowerings lists every lower_page proposal as label-only or content, counted by proposer, so an operator can see which lowerings came from label-only writes; it changes nothing, and raising a page back stays a trust decision.

Itemized changes

  • src/core/persistence/label-edit.ts (new): the label_edit intent (set and remove over visibility, access, confidentiality) and the line editor that applies it to a file's frontmatter, or declines a label line it cannot edit as one line.
  • src/core/persistence/page-prepare.ts: a managed_maintenance_page write that carries label_edit publishes the canonical file's own bytes with only its label lines edited (labelEditFile), keeping its body when it reads as the write's body, and refuses instead of re-rendering when no line edit reads back as the write.
  • src/core/persistence/prepared-maintenance.ts: maintenance page publications carry labelEdit.
  • src/core/repair/label-write.ts (new): the label-only write door both repairs use (the maintenance door with keep_trust on a managed brain; the import under the page lock and withTrustKeep otherwise); filesInCheckout and importAtRevision moved here from private-labels.ts.
  • src/core/repair/visibility.ts: page and fenced-fact items write through the label door instead of put_page; on a brain without managed persistence a page whose file sits in a checkout gets the same label-line edit in that file, imported with its tier kept, and a file it cannot write (sync.write_through off, a pulled clone, a write error) is held file_backed with nothing written; a segmented page is held (segmented_page); a page that changed after it was read is skipped (revision_conflict) instead of counted; new --slug selection (class pages), preview-bound, tighten-only; a fenced fact row it could not make private is reported fact_left_readable when remote callers can still read it, or fact_left_world when only its page's privacy hides it (NOT CLOSED in the text output for both), instead of a plain skip.
  • src/core/repair/private-labels.ts: clears and restores carry their label edit, so the file loses or regains only those lines.
  • src/core/trust/label-lowerings.ts (new), src/cli/commands/trust.ts: gbrain trust backfill --label-lowerings [--source ID] [--json], read-only.
  • Tests: test/label-only-repairs.test.ts (managed Git source and repo-less brain, PGLite and Postgres); test/owner-member-boundary.test.ts pins that the owner reads its private pages and pages that list it, and never another member's restricted page, on every read path.

[0.60.682.0] - 2026-11-11

The brain owner reads the derived pages and facts of pages it can already read, and the operator can let the owner read pages kept to other members. An atom, a synthesized concept, a Life Chronicle event, a page with derived_from or a fact made from a page under a folder rule, or from a page that names readers, was hidden from every remote caller, the owner included, even when the rule or the readers named the owner. Credentials bound to the owner now read them whenever they read every origin, with no setting. A new opt-in, gbrain access owner reads-restricted on, off by default and settable only by the operator on the brain host, lets the owner also read pages that readers lists, folder rules and page rules keep to other members, with their derived pages and facts, but never another member's own member page. Nothing changes for any other caller or for writes, and nothing changes on upgrade.

Itemized changes

Owner reads and the owner opt-in (GST-2490)

  • Migration v304 (src/core/schema-migrations/v304-access-owner-origin.ts): gbrain_access_origin_closed_for() and gbrain_access_origin_hidden_for(), the lineage arms of gbrain_access_origin_restricted() asked per origin for one caller. Only owner reads call them; gbrain_access_origin_restricted() is unchanged. Nothing is rewritten.
  • src/core/search/private-visibility.ts: ownerAccessSql for owner reads (read-only operations only: remoteCallerAccess in src/core/access/caller.ts, src/mcp/dispatch.ts, src/core/minions/tools/brain-allowlist.ts, and --as <owner> in src/cli/as-preview.ts), fact provenance and the readers checks for single rows; the search cache key and the hot-memory key carry the owner mode.
  • src/core/access/owner.ts, src/commands/access.ts: gbrain access owner reads-restricted on|off (refused for remote callers), shown by gbrain access owner, access check (owner_reads_restricted) and who-can-read; whoami adds owner_reads_restricted for every remote caller while it is on; doctor access_owner_reads_restricted (src/commands/doctor/checks/owner-reads-restricted.ts).
  • Unchanged on purpose: what every other caller reads, every write and publication check, member private facts, personal sources, shared-room audiences and attested calls.
  • Docs: docs/guides/access-designations.md (derived pages, member access, the owner opt-in), key-files entry.
  • Tests: test/owner-reads-restricted.test.ts (opt-in off and on, owner, other members and unbound keys, --as, writes, the remote refusal; PGLite and Postgres, rules in the database and in access.yml); doctor goldens.

[0.60.681.0] - 2026-11-11

The low-capacity warning now says it is not a refusal, and a quota refusal no longer reads a whole week of admissions while it holds the write counters. v0.60.686.0 made the write_capacity_low notice open with "This write was saved", but the notice also rides sync status, dry runs and failed batches, where nothing was saved. It now opens with "This notice is not a refusal: writes still work, and this call's own result says what it did", and over MCP its fix is the read-only get_brain_health call instead of a CLI command the agent would relay to its user. A write_quota_exhausted, write_quota_insufficient or background-pause refusal runs under the counter locks, and v0.60.686.0 had it walk the scope's whole week (up to 200,000 admissions) to match doctor's value. It now takes the 7-day mean rates the last forecast recorded, lifecycle changes included, and without a forecast from the last minute it samples the latest 1,000 admissions as before. Both apply only under a cap the operator set: since v0.60.686.0 the caps are off by default.

Itemized changes

  • src/core/persistence/capacity-snapshot.ts: write_capacity_low opens with "This notice is not a refusal" and never claims the call saved anything; its fix carries mcp (get_brain_health), and its why says nothing needs to change for this call.
  • src/core/persistence/capacity-forecast.ts, src/core/persistence/limits.ts: readCapacityForecast records each scope's 7-day mean gross and net rates for a minute (recordForecastRates, FORECAST_RATES_TTL_MS), and oneYearCapacity uses them, so the refusal names doctor's value, lifecycle changes included; without a recorded forecast it samples the latest 1,000 admissions and never walks the week. The sample's compacted size uses the forecast's SQL (receipt-bytes.ts).
  • src/core/behavior-change-notice.ts: one row for the new wording and the refusal's rate source.
  • Docs: docs/guides/write-refusals.md (write_capacity_low row), docs/guides/concurrent-writes.md, docs/architecture/key-files/core-persistence-capacity.md.
  • Tests: test/persistence-capacity-notices.test.ts (the notice rendered over HTTP and stdio runs get_brain_health; a week of lifecycle changes moves the brain's value to doctor's; a refusal reads at most the latest 1,000 admissions, and none after a forecast; a refusal after a forecast names doctor's value, on PGLite and Postgres).

[0.60.680.0] - 2026-11-11

gbrain now tells you when a privacy label it doesn't recognize leaves a page readable. A page with no visibility counts as private only when access is private, restricted or confidential, or confidentiality is confidential, secret, private or restricted. Any other value, such as access: personal or a confidentiality value naming a health record, was ignored without a word, so the page read as world to every remote caller and could be committed to the brain's git remote. Writes and doctor now say so. Who can read a page does not change.

You can now keep one page to some members, even when gbrain can't write the page. A page at the root of a source (one named after a folder, say) could not be covered by a folder rule, and a page whose file was removed from git outside gbrain refused a visibility: private edit with source_changed, whose receipt only offered the ghost-page repair, which deletes the page. gbrain access rule add --slug <page> --readers <list> now restricts exactly that page without writing it, and the refusal names that route. Removing the rule is the undo.

Itemized changes

Ignored privacy labels: a write-time notice and a doctor check (GST-2462)

  • src/core/search/ignored-privacy-labels.ts: one rule, in TypeScript and SQL, for a page whose access or confidentiality value gbrain ignores (no visibility, no recognized value, no other default-private rule), and the privacy_label_ignored safety notice. The notice names the key and the value, says remote callers read the page like an unlabelled one and, on a source with a git checkout, that the label does not keep it out of git (a repo-less source gets no git sentence), and its fix asks the user before adding visibility: private. It never quotes page content.
  • src/core/ops/pages.ts: put_page carries the notice for the page it stored. src/commands/import.ts, src/cli/commands/import.ts and src/commands/sync/ (perform.ts, full.ts, plain-directory.ts, run.ts): gbrain import and gbrain sync carry one notice per run that counts pages by value and names no page (--json under notices, otherwise on stderr); a managed sync reads the source's pages updated during the run.
  • src/commands/doctor/checks/privacy-label-ignored.ts, src/commands/doctor/wave-checks.ts: doctor privacy_label_ignored counts such pages by key and value. A value containing private, confidential, secret, restricted, personal, health or therapy warns, any other value is information, and the fix is the read-only gbrain list --where <key>=<value> --json. Bounded for large brains. The frontmatter GIN index serves the scan, which stops after 100,000 labelled pages, returns at most 50 values and runs under a 15 s statement timeout (an overrun reports the health as unknown).
  • Unchanged on purpose: the recognized values, visibility.honor_aliases, and every read and git placement decision. No repair rewrites labels.
  • Docs: docs/guides/access-designations.md (unrecognized label values), docs/guides/troubleshooting.md, skills/conventions/page-privacy.md, key-files entries.
  • Tests: test/privacy-label-ignored.test.ts (put_page, import, sync and a managed git source, doctor, and a pin that reads stay the same; PGLite and Postgres), test/ignored-privacy-labels.test.ts (the TypeScript and SQL rules agree on both engines).

Page rules: restrict one exact page (GST-2462)

  • src/core/access/folder-rules.ts: a rule can name one page with slug: instead of prefix: (gbrain access rule add --slug <page>, src/commands/access-rules.ts). It is stored in the same rulebook without the trailing / and covers only that slug, never a folder of the same name or a longer slug. A prefix: without / is still refused and now names the slug form.
  • Migration v303 (provisional number, access_page_rules): both prefix checks accept the page form, and gbrain_access_folder_rules / gbrain_access_pending_covers match a slash-less prefix only when it equals the slug; src/core/search/private-visibility.ts and the rule preview use the same matcher. Folder prefixes match exactly as before, so no page is restamped on upgrade. Upgrade every gbrain host that shares the brain before adding a page rule. Reads are decided in the database, so an older host still keeps the page restricted, but it refuses its own gbrain access rule commands and, on a git-backed source, stops syncing that source with stored rulebook holds a malformed rule until it is upgraded.
  • src/core/persistence/service.ts, terminal-classify.ts, verb-errors.ts, page-prepare.ts: a write refused because the page's canonical file is missing names the page rule as the way to narrow who reads the page, beside the ghost-page repair.
  • Docs: docs/guides/access-designations.md, docs/tutorials/company-brain.md, docs/guides/write-refusals.md#canonical_file_missing.
  • Tests: test/access-page-rule-fileless.test.ts (a root page named after its folder and a page whose file left git on a managed git source, and a repo-less brain, on PGLite and Postgres: put refuses and names the route, the rule hides each page from unbound callers while the owner reads it, nothing else is covered, no page file is written, and removing the rule restores the reads after approval).

[0.60.679.0] - 2026-11-11

Doctor, dream, enrich and maintain --safe fixes, landed together. gbrain maintain --safe no longer deletes pages or calls a paid provider, several doctor checks now report what they find accurately, dream phases stop and report correctly, and enrich, model-call errors, atom provenance, transcript dates and the stale link sweep get community-reported fixes. Each fix has a test that fails without it.

The unindexed_files maintenance phase no longer reports a complete scan when a source could not be scanned. A source whose Git checkout failed to read was counted as an error and the phase warned, but its details.complete still said true, so anything reading the phase result took the run as a full scan of every managed source. complete is now true only when no source stopped at its time budget and none errored. Reported by @Masashi-Ono0611 (#8887).

A source with automatic sync off now walks forward through its pages on a database whose DateStyle is not ISO. The per-source maintenance window saved its cursor in the session's date style. Under a day-first style such as German, DMY or SQL, DMY, a cursor dated after the 12th of a month was not a date the cycle handler could read, so it dropped the cursor and the next cycle started again from the source's oldest pages; a source over 2,000 pages never got past its first window. The cursor is now always written as a UTC ISO timestamp with microseconds. A source holding a cursor in the old form may start from its oldest pages one last time; from then on it advances. Reported by @javieraldape (#8771).

gbrain dream --drain stops as no progress when its batches only skip items the backlog keeps counting. The drain checked for no progress only after a batch that extracted nothing and skipped nothing. A page the phase skips on every batch while the backlog still counts it kept that check from ever running, so the drain ran all 1,000 batches, held the cycle lock for minutes and ended max_batches with exit 11, which tells the caller to rerun a stop a rerun cannot clear. A batch that extracts nothing now always rechecks the backlog, and when it did not shrink, the drain stops no_progress (atom_drain_no_progress, exit 1), and skips that do shrink it keep draining. Reported by @h6y3 (#8445).

gbrain dream --help --json prints its help as one JSON document. Under --json, stdout carries only a JSON document, and dream printed its help as plain text, which went to stderr. So stdout was empty, the run exited 0 and logged a missing-document event that gbrain doctor counts under its agent-contract check. It now prints {"command": "dream", "help": "<the help text>", "contract_version": 1}, and dream retriage --help --json and dream reset-key --help --json do the same. gbrain dream --help without --json prints exactly the same text as before. Reported by @Kkkkkuro (#8510).

Concept synthesis stops when its maintenance job times out or is cancelled. The synthesize_concepts dream phase never looked at the cycle's abort signal, so with a slow model it kept calling the model and publishing concept pages after its job had timed out, and an aborted model call could be written as a template fallback page. The phase now passes the signal to the model call and checks it before each concept and again right before each page is written, so nothing is published once the job is cancelled. The spend of the model calls that finished and the failed-attempt counts are still saved, so the recorded spend and the retry limit keep counting them. On a managed brain, concepts synthesized but not yet published when the job stops are synthesized again on the next run. No setting changed. Reported by @Nazim22 (#7986).

Doctor checks that misreported their findings now report them accurately. A community bug wave: each fix below was reported by a contributor, reproduced on master with a failing test first, and rewritten here.

gbrain maintain --safe no longer deletes pages or calls a paid provider. It used to run the whole dream cycle on the default source, so pages soft-deleted more than 72 hours earlier were purged for good, and the embedding, synthesis and other model phases ran too. The source cycle now runs only free phases that never purge: lint, backlinks, sync, the fence and content repair lanes, unindexed files, extract, extract_facts, symbol edges, emotional weight, consolidate, orphans and schema suggestions. It also holds back extract_facts when an embedding provider is set up or the decide conflict slot is on (its conflict sweep calls the decision provider), and the fence and content repair lanes when an embedding provider is set up (their publications embed) or fences.repair.llm has a repair model. The JSON report lists the phases that ran and the ones held back. Old soft-deleted pages are still purged by gbrain pages purge-deleted (preview it with --dry-run), gbrain dream and an installed autopilot, and the paid phases still run from gbrain dream.

Enrich write-ups from thinking models are no longer cut off. Enrich capped every write-up at 2048 output tokens, which thinking-by-default models such as Claude 5 spend on reasoning, so write-ups came back truncated or failed. Thinking models now get the gateway's thinking headroom; other models keep 2048.

An OpenAI account with no credits left is reported as a billing problem instead of being retried. OpenAI answers with HTTP 429 when an organisation has no credits, and gbrain read that as a rate limit and retried. A provider body whose error type or code is a billing code now decides billing wherever the error is wrapped.

Atom quotes keep their provenance when the model drops Markdown or punctuation, and a failed fence step no longer books a healthy facts round. Quote matching skips line-leading # and > markers, quotation marks and a spaced dash, never words or number signs; ambiguous or reworded quotes still fail closed. A fence write that fails with an I/O or database error now counts as failed and stops the facts.fence round, so doctor's extract_health warns.

A write-up that has its own dated timeline section round-trips. A page with no timeline whose body had a ## Timeline highlights section re-parsed that section as the page timeline, so enrich refused its own paid write-up with CONCEPT_REPUBLICATION_LOSSY.

Transcripts without a date in the filename take the date in their frontmatter, and the atoms extracted from a transcript take the transcript's date. Both used to fall back to the date of the run.

The admin dashboard's live feed keeps reconnecting after a dropped connection, and four smaller fixes land with it: the load throttle still applies its concurrency and memory limits when the machine reports no load data (Windows), a remember TTL past the representable date range is refused as invalid_params instead of throwing, a worker pool whose error callback throws now stops every worker before rejecting, and claude-cli tool loops read the previous turn's prompt cache instead of rewriting the whole transcript to it.

The stale link sweep no longer re-extracts one page forever when a bare-name frontmatter link matches a top-level page. A person page with founded: ["acme-toolkit"] resolves that name under companies/, so a top-level page named acme-toolkit does not satisfy it and the link stays wanted. The wanted-link row was still stamped as if its target existed, which marks its page stale on every pass, so the sweep queued the same page again and again (thousands of extract jobs an hour) and starved lower-priority jobs. Only a slug reference to an existing page is stamped that way now, and a row an earlier release stamped settles the next time its page is extracted, which the sweep does on its own. Editing the top-level page still re-checks the link once. One narrow case is left for a follow-up release. If the top-level page is created for the first time while that link is being extracted, the link can stay missing until either page is written again. Reported by @akinduroifedayo (#9140) and first fixed by @Masashi-Ono0611 (#9144); @akinduroifedayo's own fix (#9149) also found that narrow case.

Itemized changes

  • src/core/cycle/unindexed-files.ts: details.complete also requires errors.length === 0. Test: test/unindexed-files.test.ts (a source whose .git/HEAD is unreadable: phase warn, one error, complete: false).
  • src/core/cycle/source-window.ts: the listing instant and the last page's updated_at are formatted with to_char(... AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.US"Z"'). Test: test/sync-disabled-cycle-window.test.ts (the real queued handler under German, DMY advances window by window; German, SQL and Postgres styles under a +05:45 time zone give ISO UTC cursors with microseconds and the same windows), now also run on Postgres by test/e2e/sync-disabled-cycle-window-postgres.test.ts.
  • src/core/cycle/extract-atoms-drain.ts: the no-progress recount runs after every batch with extracted === 0, not only when skipped is also 0. Test: test/extract-atoms-drain.test.ts (a skip-only batch over a flat backlog stops no_progress after one batch; skip-only batches that shrink the backlog drain).
  • src/cli/help/render.ts: new printHelpText(command, text, args), the --help --json document for a handler that prints its own help text; src/commands/dream.ts, dream-retriage.ts and dream-reset-key.ts print their help through it. Test: test/dream-help-json-8510.test.ts (the real CLI: one document per help, no json_document_missing event, plain --help stdout equal to the document text).
  • src/core/cycle/synthesize-concepts.ts, src/core/cycle.ts: signal option, passed from the cycle like extract_atoms', handed to the narrative chat call and checked before each concept, right before each publication and before the final managed flush; an abort saves the retry counters and the finished spend to the rollup, then throws AbortError. Test: test/cycle/synthesize-concepts-abort-7986.test.ts (runCycle under the creator pack: a job aborted during the second model call publishes only the first concept and makes no third call; an aborted call is not published; an abort during the page read before the first publication publishes nothing; an aborted run still writes its spend and failed attempts).
  • timeline_history: on a brain with more than 10,000 pages carrying database-only timeline rows, a pass keeps only the first 10,000 for its final recount. It used to report that recount as exact, and a pass whose kept pages had been repaired was recorded as clean while the rest still held database-only rows. Such a pass now reports its counts as a lower bound ("at least"), keeps warning, and is never trusted as clean. A pass saved by an earlier release, which did not record whether it dropped pages, is treated the same way when its kept list is full. Reported by @Masashi-Ono0611 (#8511).
  • conversation_format_coverage: the sample now includes pages stored under the granular conversation types the extractor already reads (slack-thread, slack-dm-day, email-digest), so unparseable transcripts on a brain that never ran the type-consolidation pack are reported instead of silently skipped. Reported by @Masashi-Ono0611 (#8519).
  • minions_migration under gbrain doctor --fast: a migration the brain's config can decline (v0.53.0, shared_skills.migration.enabled) is no longer listed as not run. Without a database connection the check cannot read that setting, and gbrain apply-migrations --yes skips a declined migration, so the warning could never clear. The full gbrain doctor still reports it when it is not declined. Reported by @Masashi-Ono0611 (#8721).
  • stale_mentions: each stale example now carries its own gbrain unlink command with the link type the row is stored under, shell-quoted so a slug or link type with quotes or $(...) pastes as plain words, with the options first and -- before the two slugs, so a slug that starts with -- is never read as a flag. The hint used to say --link-type mentions for every row, which removes nothing for a typed link stored under a verb such as founded. Reported by @Masashi-Ono0611 (#8817).
  • schema_pack_consistency: pages gbrain itself wrote (extract receipts under extracts/, and pages from the code importer) no longer count as undeclared on a custom schema pack copied before the base pack declared their types. The exemption follows how the page was written, not its type name. A page of yours that uses one of those types, such as a Markdown note with type: code, still warns, as do your own undeclared types. Reported by @Masashi-Ono0611 (#8818).
  • embedding_state and gbrain embed --stale: a page whose body holds only a protected takes fence (which never reaches search chunks) is now counted empty instead of chunkless_healable. Doctor no longer warns about it, and embed --stale no longer plans a heal that can never produce a chunk. A page is treated this way only when the writer already chunked its current revision to nothing; a chunkless page that was never chunked, a page whose chunks were deleted after it was chunked, or an embed_skip page an older release stored without chunks, is still healed. Deleting a page's chunks clears that seal, in deleteChunks, an empty chunk replacement and purge_fact's chunk delete (which also empties pages whose body held no purged row), so such a page stays healable. Reported by @akinduroifedayo (#8502).
  • managed_sync_backlog: removing a source in the middle of a managed sync now removes its sync cursor too, whether the brain is managed or not. The cursor used to stay behind, and doctor warned about a stalled backlog forever with a resume command that failed unknown_source. A cursor an earlier removal already left behind is no longer reported, including after the source is added back under the same id. Reported by @omni-nordaware (#8741).
  • src/commands/maintain.ts: MAINTAIN_SAFE_PHASES is the allowlist of free, non-purging phases; maintainSafePhases intersects it with the source's own phase list and holds back provider-gated phases (extract_facts on an embedding provider or an enabled decide conflict slot, fence_repair and content_repair on an embedding provider or a resolved repair model); the dry-run and applied actions report phases and held_phases; the help text names gbrain pages purge-deleted and gbrain dream.
  • src/core/behavior-change-notice.ts: one row for the changed maintain --safe behavior.
  • scripts/module-size-limits.tsv: the notice file's ceiling grows by one for that row.
  • Docs: the maintain.ts row in docs/architecture/key-files/commands-4.md.
  • Tests: test/maintain.test.ts: the allowlist holds no purge or paid phase; each provider holds its phase; a non-default source keeps its subset; on PGLite a page soft-deleted 100 hours ago survives --safe and no unsafe phase runs; a keyless brain with the decide conflict slot on holds extract_facts.
  • src/commands/enrich.ts (defaultSynthesize): thinking models get the gateway's thinking output headroom; other models keep the 2048-token cap. A thinking-model write-up can reserve up to 32000 output tokens against --max-usd per in-flight call. Test: test/enrich/thinking-cap-7564.test.ts. Reported by @mattverlaque (#7564).
  • serializeMarkdown, edit_page's view and the timeline file splice: a body with no timeline sentinel whose own dated section would split under the bare-section fallback ends with an empty <!-- timeline -->; every other body keeps its bytes. Test: test/markdown-own-timeline-section-7738.test.ts. Reported by @mattverlaque (#7738).
  • src/core/ai/errors.ts: "no credits remaining" is a billing phrase, and a billing error.type/error.code decides billing through every wrapper layer. Test: test/ai/classify-billing-credits-8400.test.ts. Reported by @Masashi-Ono0611 (#8699, #8400).
  • src/core/ai/providers/claude-cli-language-model.ts: one content block per message over --input-format stream-json (rendered text unchanged, lone surrogates still become U+FFFD), so each turn reads the previous turn's cache. The subscription-only env scrub is unchanged. Test: test/claude-cli-prompt-cache-7352.test.ts. Reported by @spiky02plateau (#7352).
  • src/core/cycle/extract-atoms.ts (locateQuote): folds with the shared grounding core's markup option. Test: test/cycle/extract-atoms-quote-markup-8373.test.ts. Reported by @Masashi-Ono0611 (#8373).
  • Legacy fence step of extract_facts: an untyped fence write error counts in pagesFailed and halts the round with item_error; typed refusals, a busy source lock, segmented pages and files on another host stay holds. Test: test/extract-facts-fence-failure-health-8758.test.ts (PGLite and Postgres, managed and unmanaged). Reported by @javieraldape (#8758).
  • Transcript discovery and --input: an undated basename falls back to the frontmatter date; impossible calendar days are rejected and the date window filters on the resolved date. Test: test/transcript-discovery-frontmatter-date.test.ts. Reported by @spiky02plateau (#7267).
  • src/core/cycle/atom-source-date.ts: new transcriptAtomSourceDate (frontmatter date, else the path date that names the atoms/<date>/ folder); an undated transcript still gets none. Test: test/extract-atoms-source-date.test.ts. Reported by @spiky02plateau (#7268).
  • Admin dashboard: the EventSource onerror only sets the status from readyState and no longer closes the stream; admin/dist and src/admin-embedded.ts are rebuilt. Test: test/admin-dashboard-stream-reconnect.test.ts. Reported by @ramybarsoum (#8913).
  • Load throttle (shouldProceed), remember TTL parsing and the worker pool: see the paragraph above. Tests: test/backoff.test.ts, test/writeback-config.test.ts, test/worker-pool.test.ts. Reported by @ramybarsoum (#8913).
  • src/core/wanted-links-store.ts (replaceWantedLinks): checked_at is -infinity only for a slug row whose exact page exists; a name row gets now(), and the existing ON CONFLICT update re-stamps old rows on their origin's next extraction. Test: test/wanted-links.test.ts and test/e2e/wanted-links-postgres.test.ts (topLevelSlugNameReferenceSettles: the origin is not stale after one sweep, and a row forced to -infinity settles after one more; topLevelSlugNamePageEditWakesOnce: an edit of the top-level page lists the origin stale once and one sweep settles it). Reported by @akinduroifedayo (#9140); first fixed by @Masashi-Ono0611 (#9144) and also by @akinduroifedayo (#9149).

[0.60.678.0] - 2026-11-11

Pushes from an unmanaged checkout now scan every commit they publish for secrets, not only the files they just wrote. Four gbrain pushes from a checkout that is not a managed worktree carried earlier local commits to the remote unscanned: the access-policy commit, the durability-scaffolding commit, gbrain sources push (which scanned only its own staged files, then pushed the whole branch tip, even with nothing new to commit) and the first push of gbrain bootstrap repo after the remote branch was gone (which scanned only the current tree). A credential a script committed with plain git commit, including one a later commit removed or one in a commit message, went out with them. Each now runs the range scan managed publishes use over what the remote lacks (the whole history when the remote has no such branch) and refuses publish_secret_found or publish_scan_unreadable, with nothing pushed and the local commits kept. gbrain sources push pushes the exact commit it scanned, leased on the remote tip it read, so a remote rewound in between is scanned again. An allowlist edit that gbrain sources push or gbrain bootstrap repo commits itself, after scanning it, still publishes; one a script committed is left to the user. Remoteless checkouts, managed worktrees and repo-less brains are unchanged, and the policy command's retry hint keeps --scan-secrets.

Reported and first fixed by @javieraldape (#9159).

Itemized changes

  • src/core/access/policy-git.ts, src/core/brain-repo-durability.ts: the policy and scaffolding commits publish with scanSecrets; a scaffolding refusal names its code in the harden step.
  • src/core/workspace-push.ts: before the push, stagedScanPublishRefusal over origin/<branch>..HEAD, then git push --force-with-lease=refs/heads/<branch>:<scanned tip> origin <sha>:refs/heads/<branch>; a local branch that does not contain the remote tip is not pushed.
  • src/core/bootstrap/repo.ts: both first pushes scan the branch they push, with its whole history, and refuse if it moved during the scan (scanHistory option, a seam for runner fakes); src/commands/bootstrap.ts prints a refusal's code and next step.
  • src/core/git/commit-push.ts: a scanned publish whose remote has no such branch yet (no tracking ref) scans the whole history and leases the push on the branch being absent, where it used to refuse publish_scan_unreadable.
  • src/core/git/range-secret-scan.ts: stagedScanPublishRefusal, the range refusal with gate changes in workspace_push and bootstrap ledger commits exempt, and files over the range scan's cap exempt in workspace_push commits (its staged scan streams them up to 512 MiB).
  • Tests: test/unmanaged-push-history-scan-8902.serial.test.ts (each path against a local bare remote: a removed credential, a credential in a commit message, clean history, a remoteless policy commit, allowlist edits, a rewound remote).

[0.60.677.0] - 2026-11-11

You can now leave one source out of automatic publishing without turning off its sync. Set gbrain config set sources.auto_reconcile.source.<id> false, and neither gbrain sources git-reconcile (which autopilot runs each tick) nor gbrain sources refresh pushes that checkout's local commits. The source still syncs and fast-forwards. Before this, the only per-source switch was gbrain sources set-sync <id> off, which also stopped its sync. Nothing changes until the key is set.

  • src/core/persistence/worktree-refresh.ts: publishOptedOut reads the key for a worktree's members. A diverged refresh with any member off refuses refresh_diverged, naming the key, with nothing pushed.
  • src/commands/sources-git-reconcile.ts: a checkout with a source off reports unpublished with the reason and lists the sources under publish_off, in --dry-run too. A checkout only behind still fast-forwards.
  • src/commands/sources.ts: gbrain sources status prints the opt-out and adds auto_reconcile: false to --json.
  • src/core/bounded-child-exec.ts: with killTree, a child stopped at its time limit has its descendants stopped with it. The commit-push children of git-reconcile and refresh use it, so a timed-out publish leaves no git or ssh process running into the next reconcile.
  • src/core/config.ts: the key is registered, so gbrain config set accepts it.
  • Docs: docs/guides/live-sync.md, docs/guides/shared-checkout-git.md.
  • Tests: test/reconcile-publish-opt-out-gst2250.test.ts (PGLite and Postgres): an opted-out source stays unpublished, syncs and fast-forwards while another source publishes; a managed member opted out blocks refresh's publish and sync still imports; the global key, set-sync off and a key set to true behave as before; a hung child's grandchild is gone after the timeout.

[0.60.676.0] - 2026-11-11

Persistence, sync and job fixes, most from community bug reports and each reproduced on master: sync timeouts stop cleanly, held and deleted files release, Postgres pools and deadline reads stop leaking or hanging, derived effects stop reporting false recoveries, and fix commands print what a retry needs. gbrain sync --timeout N now stops as a partial run with exit 0 instead of being killed by its own watchdog, a held deleted file whose page was soft-deleted no longer stays needs_human, and gbrain git commit-push publishes a staged deletion under a wildcard-ignored folder. A Postgres pool that finishes opening after a reconnect is closed instead of leaking, a managed-sync preparation read gives up at its deadline instead of waiting on a saturated pool, and the sync cost check no longer scans every chunk when nothing has drifted. A links or facts-backstop effect that meets another write's publication in flight waits as writer_busy and retries within a second, instead of reporting recovery_required and waiting 30 s. A refused edit's fix command now prints the revision the retry needs, keyless init suggests an import that runs, cancelling a finished job changes nothing, and particle-1 on OpenRouter stops truncating. gbrain's own receipt bookkeeping no longer caps writes by default (the cumulative id and receipt-byte caps become opt-in, and housekeeping keeps receipt storage in check), gbrain repair file-drift no longer offers to copy an untracked or gitignored file into an origin-twin page, and a revision_backfill_pending refusal names gbrain apply-migrations --force-schema, the command that resumes the backfill.

Itemized changes

Sync timeout partials, deletion holds, ignored-dir deletions, git effect parking, snapshot generations (#8976)

  • src/core/sync-reconcile.ts, src/commands/sync-delegate.ts, src/commands/sync/plain-directory.ts: the --timeout N hard-deadline backstop arms one grace window after N, so the in-band stop always lands first; a delegated job keeps its own bound at N; a plain-directory source stopped by the timeout reports Sync PARTIAL (reason=timeout) with exit 0 instead of internal_error: import aborted. Reported and first fixed by @zkiihne (supersedes #7819).
  • src/core/persistence/sync-pass.ts: a concurrent_write hold whose file is gone joins the deletion candidates, so the existing no-live-page check (by id and slug) releases it; a hold whose page is still live stays held. Reported and first fixed by @Masashi-Ono0611 (#8751, supersedes #8753).
  • src/core/git/commit-push.ts: named paths that cover index entries stage with add -u, visible untracked paths with add -A, and a path already staged as deleted counts as staged; only the named paths are committed, a mistyped or ignored-only path still refuses, and a refusal restores the index. Reported and first fixed by @Masashi-Ono0611 (#7984, supersedes #7987).
  • src/core/persistence/effects.ts: a grouped Git completion failure is recorded against its effect's target, so per-target failures accumulate to the parking threshold instead of resetting to 1. Reported and first fixed by @Masashi-Ono0611 (#8606, supersedes #8631).
  • src/core/page-body/ingest.ts, src/core/persistence/segmented-apply.ts: materializedSnapshot reads the snapshot's own body generation; when that generation's rows are gone it refuses page_body_rows_unreadable instead of returning a newer body. Reported and first fixed by @javieraldape (#8492, supersedes #8501).
  • src/core/persistence/types.ts, src/core/persistence/verb-errors.ts: page_body_rows_unreadable is a write error code with its own non-retryable diagnostic instead of storage_error. Reported and first fixed by @babyspectrum (#8522, supersedes #8551).
  • Docs: docs/guides/shared-checkout-git.md, docs/guides/write-refusals.md, key-files entries.
  • Tests: test/sync-timeout-watchdog-order.test.ts, test/sync-concurrent-write-deleted-hold-8751.test.ts (PGLite and Postgres), test/commit-push-ignored-deletion-7984.test.ts, test/persistence-effect-git-group-failure-8606.test.ts, test/large-pages/snapshot-body-generation-8492.test.ts, test/memory-write-rows-unreadable-8522.test.ts; pinned deadline expectations in test/sync-hard-deadline.test.ts and test/sync-timeout-cli-dispatch.test.ts move by the grace window.

Late pool leak, preparation deadlines, type cache churn, stale-chunk scan, adoption recovery, timeline batches, export and init retries (#8980)

  • src/core/connection-manager.ts: disconnect() and retire() bump an epoch; a direct or session-fallback pool that finishes initializing after either is ended (bounded) and its waiters get CONNECTION_ENDED, and a late failure starts no retry. Reported and first fixed by @javieraldape (#8478, supersedes #8569).
  • src/core/postgres-engine/bounded-statement.ts, src/core/read-deadline.ts, src/core/persistence/bounded-reads.ts: a deadline read refuses once expired, reserves with an abort signal at the deadline, and discards a connection whose COMMIT is still pending; the refusal surfaces as preparation_deadline. Reported and first fixed by @javieraldape (#8651, supersedes #8698).
  • src/core/pg-type-cache.ts: re-describing a statement with the same parameter types no longer marks the cache dirty. Reported and first fixed by @Masashi-Ono0611 (#8647, supersedes #8657).
  • src/core/engine-sql/chunks.ts: with a signature, sumStaleChunkChars sums disjoint unembedded and drifted arms through UNION ALL, so the partial stale index serves the common case (measured 100 ms to 12 ms on a 200k-chunk fixture with no drift); totals are unchanged. Reported and first fixed by @Masashi-Ono0611 (#8811, supersedes #8813).
  • src/core/persistence/coordinator.ts, src/core/persistence/local-edit-adoption.ts: adoption recovery gathers its landed-commit proof before the recovery transaction and rechecks HEAD and the request's pin under the row lock, staying recovery_required on any change. Reported and first fixed by @javieraldape (#8869, supersedes #8914).
  • src/commands/extract.ts: sync's inline timeline extraction writes each page's rows through one batch projection (same trust gate and dedup), falls back to row-by-row only when the batch fails, and counts rows actually inserted. A missing or soft-deleted page still fails its slug with PageMissingError, checked under a page lock inside the batch transaction. Reported and first fixed by @troyhoffman-oss (supersedes #6805).
  • src/core/shared-skills/migration-export.ts: an exported or inventory-stage checkpoint with no bound source and no started topology claim can move to a new root, restarting at inventory with a fresh request; a bound or claimed checkpoint still refuses. A root refused for sitting inside a Git worktree is now checked before anything is staged, so no copy of the export is left in the repository; an earlier staging directory is removed only when it holds exactly the export's own files, and an earlier filled root is listed in the receipt (earlier_roots) and the printed next steps. Reported by @narco (#8528); first fixed by @Masashi-Ono0611 (supersedes #8628).
  • src/core/shared-skills/setup.ts: installPackagedSharedSkills resumes the caller's open or committed default-pack publication in the same source incarnation instead of submitting a second one. Reported and first fixed by @javieraldape (#8453, supersedes #8493).
  • Docs: key-files entries. scripts/structural-suites.tsv regenerated.
  • Tests: test/connection-manager.serial.test.ts, test/e2e/engine-open-postgres.test.ts, test/postgres-deadline-read.test.ts, test/e2e/persistence-preparation-pool-wait-postgres.test.ts, test/pg-type-cache.test.ts, test/sum-stale-chunk-chars-8811.test.ts (PGLite and Postgres), test/adopt-local-edit.test.ts, test/extract-timeline-for-slugs-batch.test.ts, test/shared-skills-export.serial.test.ts, test/shared-skills-packaged-key-7734.test.ts.

Derived effects wait as writer_busy behind a publication in flight (#8754)

  • src/core/persistence/effects.ts: the effect guard tells a publication in flight (its request running with a live claim, or already terminal with its record not yet cleared) from one that needs recovery. An effect that does not hold the worktree lock and meets only publications in flight is requeued writer_busy with the existing busy backoff (250 ms doubling, at most 60 s, never parked). A record whose publisher's claim lapsed, or a request in recovering, still requeues it recovery_required, and an effect that holds the lock (Git, withdrawal mirror) is unchanged. An embedding effect in the same wait records writer_busy too, on its unchanged retry schedule. Nothing was lost before (the next effect pass after the 30 s delay derived the links), but the receipt said a recovery was needed when none was.
  • test/persistence-derived-effect-publication-race.test.ts (PGLite, and PostgreSQL through test/e2e/persistence-derived-effect-publication-race-postgres.test.ts): drives a real put_page publication to its prepared and after_commit boundaries between a links effect's claim and its guard, and checks the effect waits as writer_busy with a delay under 5 s, then commits its link on the next pass; with the publisher's claim lapsed it stays recovery_required and still links once the record settles.

Fix commands print the revision a retry needs, keyless init hint, finished job trees, particle-1 headroom (#8909)

  • src/core/error-registry.ts, src/cli/page-render.ts: the fix for revision_required, revision_conflict and page_identity_changed is now gbrain get --include-content ..., and that read prints [gbrain] revision <rev>: put or delete this page with --expected-revision <rev>. on stderr; before, the suggested gbrain get -- <slug> printed only frontmatter and body. docs/guides/error-codes.md follows.
  • src/commands/init.ts: gbrain init --pglite --no-embedding now suggests gbrain import <dir> --no-embed, which runs on that brain, instead of an import that refuses because embeddings are off.
  • src/core/minions/worker.ts: a quiet-hours skip no longer rewrites an already-cancelled job's reason to skipped_quiet_hours.
  • src/core/minions/queue.ts: cancelling a job that already finished refuses and changes nothing (it used to cancel the job's still-live children while reporting that it could not cancel); submitting a child under a parent that already completed, failed, died or was cancelled refuses; both fail_parent updates clear the parent's lock. A behavior-notice row discloses the cancel and submit changes.
  • src/core/pglite-engine/code-edges.ts: getEdgesByChunk on PGLite binds the edge-type filter on every arm with one LIMIT per table, matching Postgres.
  • src/core/ai/recipes/openrouter.ts: particle-ai/ models get thinking headroom on OpenRouter, so thinking: 'off' calls no longer truncate at 2,048 tokens. No default model changes.
  • src/core/behavior-change-notice.ts, docs/guides/write-refusals.md: the refresh_replan_own_commits_8828 notice and the refresh_source_changed row now match the code (text only; the row's since is unchanged).
  • Tests: test/revision-fix-command-shows-revision.test.ts, test/init-keyless-import-hint.test.ts, test/worker-quiet-hours-skip-keeps-cancelled.test.ts, test/code-edges.test.ts, test/ai/recipe-openrouter.test.ts, test/minion-cancel-failed-parent.test.ts, test/minion-child-of-finished-parent.test.ts, test/minion-fail-parent-clears-lock.test.ts, test/write-receipt-remediation.test.ts.

Receipt bookkeeping no longer caps writes (#9021)

gbrain's own write bookkeeping no longer caps your writes. Every accepted write leaves a receipt, and gbrain used to cap how many request ids and how many receipt bytes one writer and one brain could hold (2,000,000 / 8,000,000 ids, 16 GiB / 64 GiB). Near those caps, background writers paused at 80% and every write receipt carried a low-capacity warning; at the cap, remember, put_page and sync refused with write_quota_exhausted until someone on the brain host raised a number. Nothing about your data needed that, because a cap only stopped writes. Both caps are now off unless the operator sets one. Receipt storage is kept in check by housekeeping instead. While receipts not yet compacted reserve more than persistence.housekeeping.receipt_target_bytes (default 8 GiB; the few KiB a compacted receipt keeps so its request id stays replayable never count, since nothing can shrink them), the compaction pass also compacts committed and cancelled receipts older than 7 days (except a subagent's own page writes, which a deferred synthesis postprocess still compares with the page), plus refusals that gbrain's own producers will write again from current content, and doctor persistence_capacity tells the brain host's operator. A refused write you made that gbrain repair failed-writes could still replay now keeps its intent at any age, until a later write resolves it or you drop it with gbrain repair receipt-compaction --include-failures; the retention pass used to drop it after 30 days. An explicit persistence.receipt_retention_days is left as set. On a brain that several hosts write to, these defaults hold once every one of them runs this version. An older gbrain still enforces its old caps and its 30-day retention, and the gbrain config set persistence.limits.* that its write_quota_exhausted refusal suggests would turn a cap on for every host, so upgrade the other hosts instead of raising it.

  • src/core/persistence/model.ts, limits.ts: principal_lifetime_ids, brain_lifetime_ids, principal_terminal_bytes and brain_terminal_bytes default to no cap; a value set with gbrain config set is enforced as before (refusal at the cap, background pause at its line). New key persistence.housekeeping.receipt_target_bytes (0 turns housekeeping and its warning off), measured by readCompactableReceiptBytes: terminal receipts not yet compacted that no recovery holds.
  • src/core/persistence/compaction-pass.ts: housekeepReceipts compacts up to 100 receipts older than 7 days per pass: settled page writes, and refused maintenance pages and managed sync, file import, reconcile and file repair page writes. Job receipts, a subagent's own page writes (restrictedNamespace authority), connector writes and every other refusal (a caller's write, conversation facts, atom members, connector items) keep their intent until the retention window while the brain is over the target and the retention is the default, with every exclusion of the retention pass; the pass record counts them as early. A housekeeping error is counted as housekeeping_error in the pass record and never skips the reservation sweep or the record.
  • src/commands/doctor/checks/persistence-capacity.ts: with no cap set, ok says so and warn fires above the target with the read-only gbrain repair receipt-compaction --json preview (actor host_admin); details.housekeeping reports target, stored bytes, the not-yet-compacted bytes the target counts and an explicit retention. The forecast, capacity notices, repair capacity rows and persistence_request_growth cover only caps that are set, so the /metrics gbrain_write_capacity_* gauges are exported only for a cap you set; an alert on them needs a cap.
  • src/core/persistence/journal.ts, terminal-classify.ts: the retention pass skips UNRESOLVED_CALLER_REFUSAL_SQL receipts (a writer-guard or source_changed refusal of a caller's put_page, add_timeline_entry or remember that no later request wrote, removed or repeated); gbrain repair failed-writes reads the same refusal predicates from there. A refusal with no recorded error code is not one of them and compacts at the retention as before.
  • src/core/persistence/diagnostics.ts: gbrain sources writer status --json lists only the limits that are set; an unset cap is left out instead of printing as null.
  • src/core/persistence/reconcile-backup.ts: the --backups listing leaves out max_retained_count while the brain id cap is unset (it serialized as null); the retained-backup bound is brain_recovery_bytes.
  • Docs: docs/guides/repair.md (refused writes keep their intent until resolved; no retention raise needed for a recovery), docs/guides/concurrent-writes.md (limits table, receipt housekeeping, growth model as sizing for a cap you set), docs/guides/write-refusals.md.
  • Tests: test/persistence-opt-in-caps.test.ts (counters at the old default caps admit background and interactive writes with no notice and no raise; a set cap still refuses; housekeeping above the target takes week-old settled page writes and refused maintenance and sync writes, and keeps a caller's refused write, a refused conversation-facts job, a committed job and young receipts; the retention pass keeps a 40-day-old replayable drift or guard refusal and compacts a resolved one and a non-replayable failure (red before: the drift refusal was compacted); a 40-day synthetic run never holds more than a week of uncompacted committed receipts; doctor's operator warning; compacted receipts alone above the target leave housekeeping idle and doctor ok, red before; the backup listing names no count bound while the id cap is unset, red before; a subagent's page write is kept from early compaction, a refusal with no error code still compacts at the retention, a housekeeping error still records the pass, and writer status prints no unset cap, each red before), test/managed-synthesis-postprocess.test.ts (an interrupted synthesis child output ten days old, with storage above the target and a maintenance drift on the page, is still verified and stamped by the next pass; red before: revision_conflict), test/persistence-capacity.test.ts (key count); the capacity forecast, notice and read-wave tests set the cap they exercise; the doctor goldens show the uncapped default and details.housekeeping.

A write that saved now says so when it carries a low-capacity warning, and the warning no longer asks the user to raise a write cap by tens of gigabytes. When a cumulative write cap (receipt bytes or request ids) was at 80% or forecast to run out within 30 days, every write receipt's write_capacity_low notice ended with actor: host_admin and next: tell_user_to_run: gbrain config set persistence.limits.principal_terminal_bytes <value>. An agent that read only the end of a successful remember took that for a refusal and told its user the memory was not saved. The value was also oversized, because it projected the last 24 hours for a whole year, so one bulk sync day could recommend several times the current cap. The notice now opens with "This write was saved and writes still work", its fix is the read-only gbrain doctor --only persistence_capacity --json, and the one-more-year value uses the 7-day mean rate. Days-to-exhaustion warnings still react to a bulk day. Doctor's finding keeps the gbrain config set command for the brain host's operator.

  • src/core/persistence/capacity-snapshot.ts: write_capacity_low says the write was saved and names doctor's read-only check (actor agent) as its fix, with no cap raise in the notice.
  • src/core/persistence/capacity-forecast.ts: recommended_limit (doctor persistence_capacity, the notice snapshot, gbrain sync status and /metrics) covers one more year at the 7-day mean gross and net rates instead of the larger of the mean and the last 24 hours.
  • src/core/persistence/limits.ts: oneYearCapacity (every write_quota_exhausted, write_quota_insufficient and background-pause refusal, and doctor's fallback) uses the scope's 7-day window over the forecast's window length while the scope admitted anything that week, and keeps the latest-1,000 sample for a scope idle all week; the compacted-size SQL moves to receipt-bytes.ts so both reads count the same net bytes.
  • Docs: docs/guides/write-refusals.md (write_capacity_low row), docs/guides/concurrent-writes.md (the notice, the mean-rate value, and returning a raised persistence.receipt_retention_days to the default before raising a cap), docs/guides/repair.md (unset the 90-day retention once a failed-writes recovery is done), docs/architecture/key-files/core-persistence-capacity.md.
  • Tests: test/persistence-capacity-notices.test.ts (the notice's fix and wording; a burst of 28 admissions in the last day over a 10-a-day week recommends the mean-rate value, ids exactly and receipt bytes under the mean-rate ceiling; after 1,000 admissions in one hour on a 100-a-week principal, a real write_quota_exhausted refusal names exactly the value doctor recommends, ids and receipt bytes, on PGLite and Postgres).

Repair file-drift leaves an untracked or gitignored twin file alone (#9054)

gbrain repair file-drift no longer offers to adopt an untracked or gitignored file into an origin-twin page. v0.60.629.0 taught the drift preview to read an origin twin with no recorded file against the file at its slug path, but only where sync admits that path. It didn't check that sync reads the file at all. Sync imports from the HEAD commit (and the working tree only with sync.include_working_tree or a detached HEAD, never an ignored file), so a slug-path file that Git doesn't track, or that .gitignore lists, is never imported or held. The preview still reported it as drift. It offered --prefer file to copy a possibly private, ignored file into the database, maintenance writes to the page were refused as drift, and doctor maintenance_write_drift counted it. Such a twin now stays database_only. An untracked file that sync does read from the working tree is still read as drift.

  • src/core/persistence/drift-conflicts.ts: twinReading reads adopt only when sync reads the slug-path file: committed at HEAD, or, when sync reads the working tree, not ignored by Git.
  • Tests: test/repair-file-drift-origin-twin-gbra108.test.ts (PGLite, managed source, real git): an untracked slug-path file and a gitignored one under sync.include_working_tree stay database_only with no sync hold, and an untracked file sync reads from the working tree still reads drift while sync holds it.

The revision-backfill refusal names the command that resumes it (#9053)

A revision_backfill_pending refusal now names the command that actually resumes the backfill, and the write-refusal guide says what a webhook capture of such a page gets (follow-up to GST-2112). The refusal's message, which is also the body of the 409 POST /ingest answers for a page the revision backfill has not reached and the error a dead ingest_capture job keeps, said gbrain apply-migrations --yes, which does not resume the backfill once the schema is current, and linked a section that does not cover it. It now says gbrain apply-migrations --force-schema on the brain host and links docs/guides/write-refusals.md#revision-backfill. Since #8940 that 409 stores and queues nothing, so the capture exists only with its sender, and an ingest_capture job already queued for such a page goes straight to dead; neither was written down. The guide's revision_backfill_pending row and a new upgrade notice row now say both, with the recovery, which is to resume the backfill, send a capture that got the 409 again, and re-run a dead job with gbrain jobs retry <id>. The E2E runner also gives test/e2e/source-retire-batched-postgres.test.ts (60,000 pages) the 4x per-file budget, because on a loaded CI runner it ran past the 180 s cap and was killed with a test in flight and no failed assertion.

  • src/core/page-state/types.ts: RevisionBackfillPendingError's message names gbrain apply-migrations --force-schema and the #revision-backfill anchor, matching the error registry's fix.
  • docs/guides/write-refusals.md: the revision_backfill_pending row covers the 409 at POST /ingest, the ingest_capture job that goes straight to dead, and the resend and gbrain jobs retry <job id> recovery.
  • src/core/behavior-change-notice.ts: one row for the changed message and the recovery.
  • scripts/run-e2e.sh: source-retire-batched-postgres.test.ts joins the 4x per-file budget list. No assertion in that file changed.
  • Tests: test/ingest-capture-keeps-readers.test.ts checks that the 409 body and the job's refusal name --force-schema and the anchor.

[0.60.675.0] - 2026-11-11

Tool calls through native OpenAI models no longer fill every optional argument with a placeholder. gbrain's tool loop (subagent jobs and SkillOpt rollouts) sends OpenAI models through the Responses API, and that API treats a tool definition with no strict field as strict: it marks every property required, so the model has to invent a value for each optional argument. A get_page call came back with a zero-line window and an empty byte range, which gbrain's own argument checks then refused. Native OpenAI tool definitions now say strict: false, so the schema's own required list holds and the model sends only the arguments it means. Anthropic, Google and the OpenAI-compatible providers (chat completions) send the same request as before. Reported and first fixed by @akinduroifedayo.

Itemized changes

  • src/core/ai/gateway.ts: toAISDKTools takes an optional strict value, and chatAdmitted(), which every chat() call goes through, passes false on the native-openai route only; every other route leaves the field out.
  • Tests: test/ai/native-openai-tool-strict-wire.test.ts drives chat() and toolLoop() against a local stub and checks the request body per route: native OpenAI carries strict: false on every tool and every turn with the schema's required list unchanged; native Anthropic, native Google and an OpenAI-compatible recipe send no strict field.

[0.60.674.0] - 2026-11-11

A managed sync's file freeze no longer holds timers through its Git reads

While a managed sync freezes a file, it now lets timers and other work run between its Git reads once it has run about 30 ms, so searches, lease renewals and heartbeats are no longer held behind them. Before, the freeze ran its reads of the file, its pinned blob, its content, its filter driver and HEAD one after another with no break. On a slow machine the first freeze of a large sync held timers past the 100 ms bound that test/sync-pass-batch-entries-6712.test.ts sets.

  • What a freeze records is unchanged. The reads of a committed file are of the pinned commit, or of HEAD checked by hash against the file's first read, so a write that lands during a break keeps the pinned content instead of importing the new bytes. A working-tree file is still read twice with no break between. The run is checked again after each break, so a cancelled run stops there.
  • test/sync-freeze-turn-write-gbra133.test.ts commits new bytes for the file during a break, between the file's first read and its HEAD read, on PGLite and Postgres. The frozen request keeps the pinned content and is not published, and a later freeze imports the new bytes under their own hash.
  • A fourth test in test/sync-pass-batch-entries-6712.test.ts adds 40 ms of CPU to each of the first freeze's reads. Before this change the longest stretch was 205 to 211 ms; it now stays under 100 ms.

[0.60.673.0] - 2026-11-11

Source checkouts now stay level with their remotes without help. A held page that changed only in Git imports on the next sync, and every way out of an interrupted refresh works.

Every source checkout now converges on its remote by itself: gbrain sources git-reconcile, run by autopilot each tick. One idempotent call brings every source checkout on the host level with its upstream. A checkout that is only behind fast-forwards. A checkout gbrain cloned for the source (gbrain sources add --url, with the source folder at the repository root) publishes its local commits merged with the upstream in memory: the same secret scan and rewritten-upstream refusal as refresh, bound to the upstream it checked so a force-push in between refuses, never a rebase or force-push. Any other checkout, such as a source folder inside a repository gbrain did not clone, reports its local commits unpublished unless sources.refresh_publish_diverged is true. A managed checkout goes through gbrain sources refresh and its Git durability rule. Sources with sync turned off are skipped unless named. A level checkout costs one fetch and reports in_sync, and callers on the same checkout take turns, so any agent, cron or hook can call it as often and as concurrently as it likes. Autopilot runs it each tick on every engine, its pulls and publishes in bounded child processes, so a slow push never stalls the daemon; autopilot.git_pull false skips it, and sources.refresh_publish_diverged false leaves local commits unpublished while checkouts behind still fast-forward. Exit 0 means level, 1 that a checkout needs the user, 4 that a retry can fix it.

  • src/core/git/reconcile.ts (new): one checkout under its own reconcile lock (outermost; no other path takes it, so it cannot deadlock a publish); fetch with the tracking ref's move logged, then a gbrain git commit-push --pull child, a --push-only --scan-secrets --expect-remote <checked upstream> child (a remote that moved in between is fetched and checked for a rewrite again), or upstream_rewritten; a fetch that loses the tracking-ref race to a concurrent fetch retries.
  • src/core/git/upstream-history.ts (new): the rewritten-upstream check, moved out of src/core/persistence/worktree-refresh.ts so refresh and reconcile share it.
  • src/commands/sources-git-reconcile.ts (new), src/commands/sources.ts: the command, grouping sources by checkout; a managed checkout waits out a running refresh up to --wait. A checkout that fails with an unexpected error (an unwritable .git, say) is reported for that checkout alone and the others still reconcile; only a failure of the whole run (the database, the disk) stops it. Its fetch carries the transport hardening every other remote git call has (no redirects, no external helpers, no submodules, never a credential prompt). On a managed brain a source with no checkout bound to this host is left alone, and a checkout with a recorded remote that gbrain did not clone keeps its local commits.
  • src/commands/autopilot-git-reconcile.ts (new), src/commands/autopilot-daemon.ts: the per-tick reconcile in the daemon, logged.
  • Docs and skills: docs/guides/live-sync.md, docs/guides/shared-checkout-git.md, docs/guides/cron-schedule.md (allow minutes for fetch and push calls), skills/cron-scheduler/SKILL.md (the managed cron line runs sources git-reconcile).
  • Tests: test/sources-git-reconcile.test.ts (level, stranded, behind, diverged, conflict, secret, rewritten upstream, a force-push between the check and the push, eight concurrent callers while another clone pushes, managed with durability on and off and publishing off, no checkout, publishing off, a checkout gbrain did not clone, sync off, a missing folder), test/autopilot-git-reconcile.test.ts (the daemon's event loop keeps ticking during a publish child; a refusal is reported with no value printed; a hung child is killed).

Held pages that only changed in Git import on the next sync. A Git edit to a page with no import record is held file_drift unless sync can show only the file moved. An earlier release could show that only by comparing the page with its file at the last synced commit; a later one also compares the path's earlier committed versions, but did not screen the holds the first rule had already written. Those holds are now screened once more. A page that still holds an earlier committed version of its file imports the Git edit and its hold clears; a page the database also changed stays held, and is not screened again.

  • src/core/persistence/sync-holds.ts: FILE_DRIFT_VERSION 2, so an unknown-side file_drift hold stamped 1 is re-screened once with the earlier committed versions.
  • Tests: test/sync-unknown-drift-history-rescreen.test.ts (a hold stamped 1 whose page holds an older committed version imports the Git edit; one whose page the database changed stays held at 2 and is not screened again); the existing rescreen and #7931 tests name the current screen version instead of the literal 1.

gbrain sources refresh --abandon no longer leaves the database ahead of the checkout, and every way out an interrupted refresh names now finishes it. A refresh stopped mid-sync with the checkout reset to its start commit could still be abandoned when a member's sync had imported part of the upstream commit but stopped (on its budget) before recording it: the imported pages stayed live with no file at HEAD. Abandoning while a sync pass that read the upstream commit was still running could also let that pass record the upstream commit after the fence lifted. Both now refuse and change nothing. The first names the fast-forward and --resume, and the second names sync_drain_lease_held until the pass ends. The other ways out are fixed too: --abandon at a HEAD past the upstream commit names --resume (which finishes in one run) instead of a reset; members synced to diverged lines get a refusal naming each commit and a merge of them instead of a fast-forward --resume would refuse again; a member commit git cannot place gets its own refusal and a fetch; writer status read off the owner host says the check needs the owner host; and the refusals name the revert route for an operator who rolled back on purpose (fast-forward, git revert on top, --resume). --abandon is never loosened and has no override.

  • src/core/persistence/worktree-refresh.ts: membersPastStart also counts a member with a managed-sync import or delete admitted after the row's barrier_sequence, and reads git merge-base --is-ancestor as exit 0, exit 1 or "cannot place". --abandon of a recovery_required row checks the members and aborts under the topology locks and every member's drain lease (underMemberLeases); a live holder refuses sync_drain_lease_held and an expired lease is taken over. keepPagesWayOut replaces keepPagesHead: a fast-forward to the newest member commit on the target's line, a merge of diverged member lines (ask_user), or a fetch when a member commit cannot be placed; --resume recomputes that refusal at any readable HEAD instead of repeating the stored text. adoptsHead is the one recovery-adopts-this-HEAD check, and --abandon reads it to name --resume at a HEAD past the target. refreshHolderReport names writer status on the owner host (actor: host_admin) when it has no checkout for a recovery_required row. The fast-forward refusal's suggestion and user_message name the revert route.
  • Docs: docs/guides/write-refusals.md (refresh_recovery_required, sync_drain_lease_held), docs/architecture/key-files/core-persistence-git.md.
  • Tests: test/refresh-recovery-followups-gst1465.test.ts (new, PGLite and Postgres through test/e2e/worktree-refresh-postgres.test.ts): a sliced pass's partial import refuses --abandon and the fast-forward then --resume finishes with every page at HEAD; a pass paused after its level read refuses --abandon sync_drain_lease_held; a fenced and a merged orphan past the target finish on one --resume and --abandon names it; diverged members refuse naming both commits and the named merge then --resume finishes; an unplaceable member commit refuses with the fetch at the start commit and at the target, and once fetched the fast-forward finishes; writer status off the owner host names the owner host; the revert route completes and the next sync removes the reverted page.

[0.60.672.0] - 2026-11-11

OpenClaw agents can now be shown the few installed skills that fit each message, the OpenClaw guide shows how one bot can serve several people over DMs with each person reading only their own pages, and running gbrain's own tests no longer writes into your brain. Three changes for people who run gbrain with OpenClaw and for contributors, landed together.

OpenClaw skill resolver (off by default)

  • gbrain setup openclaw --skill-loading resolve|shadow|off: a separate OpenClaw plugin, gbrain-skill-resolver, that puts up to 3 relevant skills in front of the model on each real user turn, ranked inside the gateway's Node process in a worker thread with a 50 ms deadline, so a busy gateway never waits on gbrain. It leaves OpenClaw's own skill list alone. shadow ranks and logs what it would have suggested without injecting anything; --live-fraction, --pins-route and --owner-sender stage a rollout. Switching modes or pins rewrites only gbrain's state file, so no OpenClaw command runs and the gateway does not reload. It reads top_k, min_match and pinned from skills.resolver the way the other lanes do. A partial install is never reported as done. It exits 3 with skill_resolver_plugin_not_installed and the commands still to run. Off by default.
  • gbrain skills pin add|remove|list --catalog <c>: operator-chosen always-on skills, shown as a pinned block and never suggested twice. gbrain ships none; pins must name a skill in the catalog.
  • gbrain doctor --only skill_resolver reports the OpenClaw plugin, its library, each agent's catalog and pins, and over the last 24 hours the outcomes, how often a suggested skill was opened and the deadline-miss rate; it warns when more than 5% of at least 20 turns missed the deadline.

One OpenClaw bot for several people, in DMs (docs only)

A new section in docs/mcp/OPENCLAW.md connects the bot over serve --http with a read-only token bound to no member, whose --may-act-for lists each person and guest. A small OpenClaw plugin uses the requester-scoped MCP connection resolver to send each DM sender's member as X-Gbrain-On-Behalf-Of; unmapped senders read as guest. The section limits the recipe to gateways with no group conversations, lists the gateway settings that keep people's sessions apart, and warns against the private-page relaxations that would show every person the owner's private pages. Those are a --private-pages granted-sources grant on the bot, the host hatches, reusing the owner's own token, and clearing access: private labels on pages that should stay the owner's. Docs only; no behavior changes. Proposed and first implemented by @exGeni (#8732).

  • docs/mcp/OPENCLAW.md: new section "One bot for several people, in DMs" (brain-host setup, the plugin, what each DM gets, refusals, revocation and the whoami check), and a pointer to it from the page's opening.

Tests stay out of your brain, and bun run verify passes on Windows

Three test and ratchet fixes reported by contributors and rewritten on master.

  • test/harness-install-ownership.test.ts, test/shared-skills-adapter.test.ts, test/shared-skills-native-router.test.ts: each install runs under its own scratch GBRAIN_HOME, so bun test from a shell that exports GBRAIN_HOME for a real brain no longer adds temp paths to that brain's agent-instructions.json (which doctor's harness_wiring then reported as broken installs). test/helpers/inherited-ledger-guard.ts fails a file whose tests change the inherited ledger. Reported and first fixed by @Kkkkkuro (#8434).
  • test/persistence-write-interrupted.test.ts: the test home goes through realpathSync, so the two worktree-backed cases pass on macOS, where the temp directory sits under a symlink that managed import refuses. Reported and first fixed by @babyspectrum (#8552).
  • .gitattributes, scripts/check-function-size.ts: *.tsv checks out with LF on every platform, and the function-size ratchet keys functions by forward-slash paths, so a Windows checkout no longer reports every baselined function as a new violation. test/scripts/tsv-eol.test.ts pins the attribute. Reported and first fixed by @mrsonord2240 (#8641).

[0.60.671.0] - 2026-11-11

A team member's forget now says who can withdraw a fact, instead of answering that the fact does not exist. A remote credential bound to a member, the brain owner's included, withdraws only the facts on its own member page. When it tried to forget a fact it can read but that lives elsewhere (one it saved with share: true, or a teammate's shared fact), it got not_found, so its agent believed the fact was gone while every reader still recalled it. It is now refused with a reply saying the brain host withdraws it with gbrain forget <id>, and the fact stays. A fact the caller cannot read still answers not_found, so the reply never confirms a hidden fact exists. An attested QM call (x-qm-actor) is unchanged and still answers not_found for any fact off the person's page, as QM's published contract says. The docs that said remember defaults to world now say that a member-bound credential saves privately on its member page unless it passes share: true. Reported by @andreineacsu (#9101).

An agent that passes source to remember is now pointed at provenance. The remember description said "provenance (required: source)", and agents read that as a parameter named source. The call was refused, and since v0.60.635.0 the refusal suggested source_id, the write source, which is not what the agent meant. In the nightly write evals, that one refusal caused every failed refresh-barrier episode on master. The description now says "provenance (required)", and an unknown source on remember suggests provenance.

A write that a refresh holds back now says which request_id to retry with. While gbrain sources refresh is updating a checkout, a write to it is refused as worktree_refreshing, and the refusal said "Retry the same request_id". A write sent without a request_id had no id to reuse, so an agent's retry went out as a new request. The refusal now names the request_id it was refused under, a new one when the call sent none, and over MCP its fix.mcp is the exact call to repeat with that id after retry_after_ms. A refused batch (put_pages, or remember with items) names the batch's own request_id, and memory_write, which replays by its idempotency_key, is told that key instead. Nothing is recorded for the refused call, as before. The retry commits under that id, and reusing the id later for a different write is refused (idempotency_conflict) like any reused id. A database-only (repo-less) source is never held back.

GPT-6.1 Sol's recorded-time page reads and typed graph chains now get answers instead of refusals, and two regressions that shipped without a test now have one. Sol fills in every optional argument of a tool call. A get_page with known_as_of (the page as the brain had recorded it at an instant) arrived beside include_content: true, include_timeline_entries: true, a one-line lines window, view: "read" and the rest, and gbrain refused the whole call because a recorded-time read does not serve those arguments. A traverse_graph with hops arrived beside depth, link_type and direction, and was refused the same way. Sol does not follow the refusal's fix, so in a replay of its recorded calls 178 of 651 get_page and traverse_graph calls failed on arguments alone. Both calls now run: a recorded-time read returns the whole page as recorded, a chain walks its hops, and the response's set_aside names each argument that was ignored and why. body: outline or window with known_as_of still refuses, because that names a read mode the ledger cannot serve.

The regressions: v0.60.58.0 sent one Postgres round trip per member of a bulk sync group (#5984), and v0.60.554.0 shipped a compiled binary whose admin UI could not load its JavaScript bundle. Both were caught by users, not CI. A Postgres unit test now holds a bulk group page to its statement budget, and the native and release builds now start the compiled binary and load the admin UI with every asset it references.

Itemized changes

  • src/core/persistence/memory-mutations.ts: when a member caller's forget misses its own member page, the fact is checked against the same visibility, privacy and eligibility filters recall uses for that caller; a readable fact is refused (scope_denied from forget, permission_denied from forget_fact, class host_only) and an unreadable one stays not_found.
  • Docs: docs/guides/access-designations.md (fact visibility), docs/mcp/TOOL_REFERENCE.md (remember), skills/brain-ops/SKILL.md (and the generated plugin copies), and the facts.default_visibility bootstrap check's detail now name the member-bound default.
  • Tests: test/member-forget-refusal-9101.test.ts on PGLite and Postgres with member- and owner-bound OAuth credentials on a repo-less brain.
  • src/mcp/validate-params.ts: remember's parameter hints map source to provenance ahead of the source-scope aliases; src/core/verbs.ts drops "source" from the description (the starter tool list stays inside its budget). Test in test/validate-params.test.ts.
  • src/core/persistence/journal.ts: the worktree_refreshing refusal's message, suggestion and fix name the request_id it was refused under and set it on the error (it already crosses the persistence owner's IPC hop as request_id); verb-errors.ts keeps it on the memory verbs' frozen unavailable refusal. src/mcp/dispatch.ts: a mutating call refused retryable with a request id gets fix.mcp = the same tool and arguments plus that request_id, and the fix still renders next: wait. Docs: the error registry suggestion, docs/guides/write-refusals.md and the generated docs/guides/error-codes.md. Tests: test/refresh-fence-retry-id-gst2353.test.ts (PGLite and Postgres, repo-less case included); test/refresh-retry-hint-gst1582.test.ts reads the new wording.
  • src/core/ops/known-as-of.ts: get_page {known_as_of} sets aside the arguments that shape a current read (window selectors, section, grep, context, include_content, content_only, include_timeline_entries, view, max_bytes, fuzzy and the rest) and returns set_aside {ignored, note}; only body: outline|window and include_quarantined: true refuse invalid_params, with the same keep-only-what-applies fix.
  • src/core/ops/links.ts: traverse_graph with hops walks the chain with depth, link_type and direction set aside and named in set_aside; malformed hops still refuse.
  • src/core/behavior-change-notice.ts: one row for the changed refusals.
  • Docs: docs/guides/multi-hop.md, the key-files rows for known-as-of.ts and links.ts.
  • Tests: test/tool-args-empty-optional-sol.test.ts replays Sol's shapes on every entry point (MCP remote and local, gbrain call, the SkillOpt rollout and write-capture registries, the subagent registry) against a ledger-enabled brain; test/traverse-graph-hops.test.ts and test/get-page-slices.test.ts assert the set-aside contract where they asserted the refusal.
  • test/managed-sync-bulk-round-trips-5984.test.ts (Postgres unit arm, registered in test/postgres-unit-arms.txt): a managed bulk sync of a multi-member group stays within 27 publication and 74 run statements per page (measured 24.2 and about 67); a probe adding 4 statements per member fails it.
  • scripts/native/compiled-admin-smoke.ts: starts the compiled gbrain binary's HTTP server, fetches /admin/ and every same-origin script and stylesheet the page loads, and fails unless each returns 200 with a JavaScript or CSS content type (the SPA fallback answered the missing bundle with text/html). It runs in .github/workflows/native-locks.yml and .github/workflows/release.yml, pinned by test/release-workflow.test.ts.

[0.60.670.0] - 2026-11-11

A nightly soak test for a busy personal agent, and a fixed convergence-nightly flake

The convergence nightly now also runs the managed-sync harness under a busy personal agent's workload, and its kill loop no longer fails by chance on the sync:passed_over round. This change touches only tests, the nightly workflow and docs. Nothing in src/ changed.

  • Busy-agent soak cell. test/busy-agent-soak.serial.test.ts runs six modes, (h) to (m), against a real Postgres behind transaction-mode PgBouncer: refresh ticks with writes inside the refresh barrier, doctor foreign_writers beside plain-git committers, sources git-gap on legacy rows, a strict conversion stalled between adopt groups, segmented staging that expires before publication, and the cron commit-push --base against MCP edit_page race. Every mode ends with the convergence oracle. The counts and cadences come from the committed generic profile test/fixtures/busy-agent-profile.json, which holds no content. Run it with GBRAIN_TEST_BUSY_AGENT=1 bun run test:busy-agent.
  • Informational nightly job. busy-agent in .github/workflows/convergence-nightly.yml runs the cell every night at 50 ms latency, with the run id as the seed. It is continue-on-error, so it never changes the convergence job's result. On failure it writes a summary that carries the seed, and it uploads the logs. A .github/nightly-known-red.tsv row parks its known red (#8975, tracked in TODOS.md), so a red in the convergence job itself still shows as unparked.
  • Shared harness. test/helpers/convergence-brain.ts now holds the real-process brain that both suites use. The convergence gate's oracle also counts deferred requests as open. When a recovery sync reports synced but HEAD has moved past the commit it synced, the harness runs its second allowed plain sync instead of stopping.
  • Kill-loop flake (#8838). The two kill-loop rounds whose fault point is sync:passed_over now sync with --no-bulk. Before, a bulk group could freeze the round's primed file and then drop it without admitting it. When sync froze the file again, it read the edited bytes and held it at once, so the pass-over never fired and the run failed "every planned fault fired". A forced probe (an edit at the freeze, then a kill at the next freeze) fails with bulk on and passes with --no-bulk.

[0.60.669.0] - 2026-11-11

Five flaky tests fixed at their cause

Five tests that failed by chance on some machines now pass or fail for a real reason. No assertion, bound or timeout was loosened, and nothing in src/ changed. Each fix comes with a probe that forces the failure every time on master and passes with the fix.

  • test/helpers/statement-counter.ts: the PGLite statement counter named a Bind by the last Parse it had seen. The engine's statement cache (src/core/pglite-statements.ts) prepares a statement the second time it runs, so a statement prepared before counting started was recorded under the text of the statement before it. test/fold-hook-noop-6712.test.ts then failed whenever background config reads had prepared SELECT value FROM config WHERE key = $1 before its window. The counter now reads a bound statement's text from the engine's cache, and records an unknown name as a placeholder. test/statement-counter-warm-cache.test.ts pins it.
  • test/persistence-consumer-idle.test.ts: the first tick's maintenance count read 8, 9 or 10. On a slow machine the next tick's idle probe landed inside the count, and the compacted-receipt sweep added a statement until an earlier test had finished its cursor. The test now walks that cursor to done first and holds later ticks until the first tick is counted. The count is still exactly 8.
  • test/e2e/evidence-long-page.test.ts: with provider keys in the environment, every search called the live reranker, whose scores for identical calls differ in the last digits, so the byte comparison failed. The file now runs keyless, as CI's Selected E2E runs it.
  • test/e2e/source-retire-batched-postgres.test.ts ran into its 180 s cap. It now retires 2,017 pages (four full batches and a partial one) instead of 60,000, and a statement trigger records every committed page delete. So each test asserts the batches themselves: their sizes, the per-batch bound and one transaction each. Removing the batch LIMIT fails the remove and purge tests. The file takes about 57 s.
  • test/config-file-plane-keys.test.ts, test/consent.test.ts: the memorable enable tests hung waiting for a [y/N] answer when stdin was a terminal. They now pin process.stdin.isTTY to false while the command runs, so they test the non-interactive refusal however the runner's stdin is set up.
  • scripts/ci-census-tests.sh: the pull-request census job now runs 15 more tests that read every file under src/ (or a whole src/ subtree), such as test/cli-exit-verdict-pin.test.ts. Before, a change that broke one of them passed its pull-request checks and failed only in the full run. test/scripts/ci-census-tests.test.ts pins cli-exit-verdict-pin in the list.

[0.60.668.0] - 2026-11-11

The skill resolver can now find skills from more places, stays fresh when your harness writes into its skills folder, and runs inside Hermes. The skill resolver is the optional feature that offers your agent the right installed skill on each turn. It is still off by default and never blocks a turn.

Before this release it only read skill folders. Now it also reads OpenClaw's skill list, the Claude Code plugins you have enabled and a repository's own .agents/skills for Codex. Before, a harness that wrote a hidden file into its skills folder, as Codex and Hermes both do, made the resolver think the folder had changed. It then rebuilt its index about once a minute. That no longer happens. Long-running hosts can also keep the index open between turns instead of reopening it every time.

How to use it

  • Claude Code plugins: gbrain setup claude-code --skill-loading resolve now also indexes your enabled plugin skills, named <plugin>:<skill>. --skill-loading off removes that catalog again.
  • OpenClaw: openclaw skills list --json | gbrain skills index build --catalog openclaw-main --from-json - --dir <skills folder> --harness openclaw builds a catalog from OpenClaw's own list, taking each skill's body from its folder.
  • Hermes: wire /usr/bin/env GBRAIN_HOOK_LANE=skill-resolver /abs/path/to/gbrain hook user-prompt --harness hermes as a pre_llm_call shell hook, build the hermes-user catalog, add hermes to skills.resolver.harnesses and set skills.resolver.enabled to true. Setup does not wire Hermes yet. The guide in docs/guides/skill-resolver.md gives the exact steps.
  • Codex: nothing to do. A repository's .agents/skills folders, from the repository root down to your working folder, are searched like Codex itself loads them.

What changes on a busy machine

Situation Before Now
Harness rewrites a hidden file in its skills folder before every turn Index marked stale, rebuilt about once a minute Stays fresh, 0 rebuilds in a 40-turn bench
Turn time in that bench (median / 90th percentile) Rebuilds on top 36.8 / 42.3 ms, against 33.7 / 38.1 ms with no file churn
A SKILL.md symlink pointed at a different skill Old skill offered until the next rebuild Seen as a change on the next turn
No catalog built yet for a Claude Code turn Logged as an error Logged as no_catalog, turn continues

Things to watch

  • Turn output gains block_reason, which says why a turn offered no skill (ack, no_match, filtered, deduped, paced, budget_exhausted, shadow or error). It is null when a skill was offered.
  • gbrain doctor warns when a harness has the resolver on its prompt hook but not on its session-start hook, and names the setup command that fixes it.
  • A null entry under skills.resolver.per_harness now reads as unset instead of failing.

Itemized changes

  • Catalogs: OpenClaw skills list --json input (hidden skills skipped), --from-json with --dir folder bodies, JSON catalogs up to 64 MB with bodies trimmed to 4 KB, when_to_use and disable_model_invocation in JSON entries so a JSON list and the same folder build the same hashes, and --claude-plugins for the claude-code-plugins catalog.
  • Freshness: dot-entries and node_modules are never read, and a moved directory mtime counts only when its list of subfolders and SKILL.md files changed. Symlink targets are part of the listing key.
  • Runtime (src/core/skills/resolver/): openCatalog keeps each catalog's receipt and open index and re-checks them once per turn, buildSkillIndexOffThread builds in a worker thread, and every session commit starts a bounded background session cleanup that never deletes a file a commit holds.
  • Hook lane: the Hermes reader takes extra.user_message, session_id and cwd with a bounded streaming read and answers {"context": block}. The codex-project-<hash> catalog covers each .agents/skills from the .git root down to the working folder.
  • budget_exhausted stays true on an acknowledgement or no-match turn once the budget is spent.

For contributors

  • Compiled builds pass --define GBRAIN_BUILD_COMMIT, so build receipts from a compiled binary record gbrain_commit.
  • Ubicloud gate VMs install CI's pinned Node 24.18.0 instead of the distribution's Node 18.
  • The tracked scratch file dbg-tmp.ts is removed.

[0.60.667.0] - 2026-11-11

A brain that serves several people can now be set up the safe way in one command: each person gets a member, a folder only they read and a key bound to them, and doctor fails when a person's folder has no owner. When one bot answers several people from one brain with one unbound key, it pulls one person's notes into another person's answers (garrytan/harness-evals#363 measured 15 to 47.5% of trials; bots reading through member-bound keys stayed at 0 to 0.5%). gbrain setup claude-code --serves several --members alice-example,bob-example writes access.serves = several, adds the members, adds a folder rule people/<member>/ per member through the folder-rule path (committed to .gbrain/access.yml on a repo brain, the database rule log on a repo-less one), mints one read-write key per member through the auth create mint (granted the labelled source) into a 0600 file under ~/.gbrain/member-keys/, and proves by reading every other member's folder as each member that none is visible. Re-runs keep everything and never replace a key; a partial failure finishes on the next run, and anything the preview marks BLOCKED stops the run before it writes. A one-person brain sees no change: --serves one stays the default and its output is unchanged.

Itemized changes

  • gbrain setup claude-code|codex --serves one|several, --members <id,...>, --members-file <file> (src/commands/setup.ts, src/commands/setup-serves.ts). Setup stays engine-free: the labelling, the serves read and the bound reads run as children of the launcher. An empty roster refuses access_serves_roster_empty.
  • On a brain that already serves several people (2+ enabled members, or access.serves = several; explicit wins), a plain setup refuses setup_serves_several and writes nothing unless an existing mcp.stdio_member naming an enabled member binds the connection; naming an owner no longer exempts it. Setup never writes mcp.stdio_member, because that key is brain-wide and would bind every stdio MCP session on the host, including the shared sessions of OpenClaw or Hermes. The per-harness --stdio-member binding described below takes its place. A brain setup cannot read refuses setup_serves_unreadable.
  • gbrain access serves (read-only view) and gbrain access serves several --members ... [--yes] (src/commands/access-serves.ts): the labelling on its own, preview first. Keys are idempotent on the member id (token member-<id>); a key whose file is gone is reported with the revoke step, never re-minted on its own. Key files never go inside a git work tree, symlinks resolved (access_serves_key_dir_in_git); a kept key file must hold the live key, and a minted key whose file cannot be written is revoked. An existing rule on people/<member>/ that anyone else reads is BLOCKED, and the receipt lists people folders no rule covers.
  • src/core/access/serves.ts (new): brainServes(engine), the one read of whether a brain serves several people.
  • src/commands/auth-mint.ts (new): the token mint gbrain auth create used inline, shared with the labelling; auth create output is unchanged.
  • src/commands/access-rules.ts: addFolderRules adds rules through the same preview-then-write path as gbrain access rule add.
  • Brain-first block and per-turn rule (src/core/setup/brain-first.ts): on a several-person install whose stdio server is bound to a member, the source_ids / "apply your own access rules" line becomes "This connection reads only what the person asking may read." (--instructions-only --serves several checks the brain first). One-person text is unchanged; doctor's harness_wiring repair names --serves several for such a block.
  • New doctor check unlabeled_people_folders (src/commands/doctor/checks/unlabeled-people-folders.ts): on a several-person brain, fails for each top-level people/<name>/ folder with pages that no folder rule (on that folder or on people/) covers; the fix previews the rule and asks the user who reads it. Not applicable on a one-person brain.
  • Docs: docs/guides/setup.md ("A brain that serves several people"), INSTALL_FOR_AGENTS.md, skills/setup/SKILL.md, skills/conventions/brain-routing.md.
  • Tests: test/setup-serves-several.test.ts (PGLite repo-less and Postgres in process with the minted keys verified by the serve --http verifier; spawned setup on PGLite; a repo brain with a real git repository).

A stdio connection can now be bound to one person for one harness only: gbrain setup claude-code --stdio-member <member> (or codex) writes the binding into that harness's own MCP entry, and setup no longer writes the brain-wide mcp.stdio_member. Before, binding stdio meant one config key in the brain, so every stdio session that opened it, from any harness on any machine, read as that member, private pages included. That is why setup refused the flag on Postgres. Now the member rides in the one harness entry that starts gbrain serve --stdio-member <member>, so another harness on the same machine, or another machine on the same Postgres brain, stays as it was.

Itemized changes

  • gbrain serve --stdio-member <id> (stdio only; refused with --http): that server reads as the live member, ahead of config mcp.stdio_member. A disabled or unknown member reads as nobody, with no fallback to the config key. Only the server's own argv sets it; nothing a caller sends (tool arguments, _meta, headers) does. whoami reports bound_by: stdio_registration.
  • Stored writes from such a server stamp the member on the write authority (a jsonb field, so no migration), and preparation and publication re-resolve it live. A stdio wake subscription uses principal id registration, so the drain re-checks that member.
  • gbrain setup claude-code|codex --stdio-member <id> no longer needs --serves several or a PGLite brain. The member must be enabled, or one of --members in the same run. It is still refused off claude-code and codex, with --instructions-only or --remove, and wherever OpenClaw or Hermes is installed; that refusal now names a per-person HTTP key as the route. --no-stdio-member removes it.
  • Re-runs, gbrain post-upgrade's launcher refresh and doctor's setup re-run keep an entry's existing --stdio-member and never add one; setupRerunArgv never names it.
  • Tests: test/access-binding.test.ts (PGLite and Postgres), test/setup-serves-several.test.ts.

[0.60.666.0] - 2026-11-11

QM, the AI helper that can use gbrain as one of its memories, gets a fourth contract version, and each fact it recalls now says whether the person can forget it. docs/contracts/qm-memory-provider/v4.json describes owner-approved proposals (memory_propose), marks which personal tools read named sources as a shared scope, and adds own_page to every recalled fact. v1, v2 and v3 are byte-identical, so QM keeps whatever version it vendors until it moves.

What changed

Area Before Now
memory_recall fact rows id, text, date, kind, arm also own_page: true when the fact is on the person's own member page in the connection's source, the same check forget_fact makes, so forget_fact accepts it on ownership; false when forget_fact for that attested person answers fact_not_found. arm is unchanged
get_write_request / cancel_write_request on a proposal request_id, state, retry_after_ms, operation, source_id, slug, proposal, and decision / write_error when closed also created_at and updated_at, as a write's receipt carries them
QM contract v3: no proposal, no scope marking, no own_page v4 adds memory_propose (arguments, caps, the held answer, the proposal receipt and its states, cancel_write_request as the proposer's cancel, how long a closed proposal is kept, and "a state not listed means unknown, not published"), scope_read: true/false on each personal tool, and own_page; v1-v3 unchanged

Itemized changes

  • src/core/access/member-facts.ts: ownMemberFactIds returns the facts on the caller's own member page. forget and forget_fact decide ownership with it, and memory_recall sets own_page from it, so the two cannot disagree.
  • src/core/ops/qm-memory.ts: each recalled row carries own_page.
  • src/core/persistence/memory-proposals.ts: the proposal receipt carries created_at and updated_at.
  • scripts/build-qm-memory-contract.ts, docs/contracts/qm-memory-provider/v4.json (new): contract v4, generated from the live registry; the actor_required fix row names memory_propose from v4 on.
  • Docs: docs/contracts/qm-memory-provider/README.md, docs/guides/qm.md, docs/architecture/key-files/access.md.
  • Tests: test/qm-memory-tools.test.ts checks on PGLite and Postgres that every own_page: true row is one forget_fact withdraws and every false row is refused fact_not_found, whichever arm found it; test/qm-memory-contract.test.ts pins the v1-v3 bytes and checks v4 against the registry; test/qm-memory-propose.test.ts checks the held answer, the receipt and the cancel against v4.

[0.60.665.0] - 2026-11-11

One file in a group of Git effects can no longer park the whole group, and a failed step no longer leaves files staged. A page write's Git effect commits in a group of up to 100 files that share one git add. When one of them was already tracked and a later ignore rule covered its folder (a rule ending in /), git add staged every file it was given and still exited 1. gbrain read that as the whole group failing and parked every effect as git_unavailable, including files whose bytes were already committed. Retrying parked them again, and their files stayed staged in the checkout. Every later gbrain git commit-push naming those folders then refused with commit_prestaged_paths. Now a tracked file is committed even when an ignore rule covers its folder, as plain git does. A new file an ignore rule covers is still never added and parks alone as git_target_ignored, naming the rule. A file git refuses fails by itself while the rest of its group commits. Whenever a group's add or commit fails, the index entries it staged are put back. A commit that already landed is never undone.

Itemized changes

  • src/core/persistence/effect-git.ts: in a group, tracked files stage with git add -u (which never adds an untracked file) and new files with git add -A (which refuses an ignored one). When the group add fails, each file is staged alone, and a refused file fails by itself, as git_target_ignored naming the rule when that is the cause, and what was staged of it is put back. When every file fails for another reason, the group fails as before. Every failed step, an abort included, restores the group's index entries and their assume-unchanged and skip-worktree flags. Entries already equal to HEAD are left alone, so a commit that landed before its step failed is never reversed.
  • src/core/git/commit-push.ts: a file another writer adds or edits under a named folder after the commit lands is no longer reported as skipped (commit_paths_skipped). The commit_prestaged_paths fix text also says that files a parked gbrain Git effect left staged are gbrain's own, and that gbrain sources writer retry-effects commits them.
  • Tests:
    • test/effect-git-failed-commit-unstages-gst2128.test.ts (failed commit, refused add, throwing step, landed commit, tracked file under an ignored folder, untracked ignored file whether status saw it as ignored or not);
    • test/effect-retry-already-committed-gst2128.test.ts (retried effects already in HEAD next to a tracked file under an ignored folder; a push the secret scan holds);
    • test/git-commit-push-skipped-paths-gst2116.test.ts (a post-commit add or edit under a named folder).

[0.60.664.0] - 2026-11-11

The Git effects of page writes no longer publish a script's secret. A script that commits into a managed checkout with plain git commit never passes the write path's secret scan. The refresh publish refused such a commit with publish_secret_found, but the next page write's Git effect pushed it anyway. Its plain push published every unpublished commit unscanned, and when the remote had moved its keep-checkout publish ran without the scan. Both paths now scan the exact commits they push, except the ones gbrain made itself for page writes (a Git effect's commit, recorded only after checking it holds exactly the bytes the effect wrote, or a hosted repository's door commit, as its commit ledger records them): page bodies are stored raw by design and their Git backups keep going, and the refresh publish now skips them the same way. A value that looks like a secret, added text too large to scan, a commit that changes .gbrain-scan-allow or a .gitattributes holds the effects with publish_secret_found; commits the scan cannot read in full hold them with publish_scan_unreadable. A first push to a remote that has no such branch yet scans the branch's whole history and creates the branch only when that scan passes. A commit that changes .gbrain-scan-allow or a .gitattributes is never published automatically, so the user reviews it and pushes it. Nothing is pushed, and the page writes stay safe in the database and in local commits. gbrain sync status and doctor name the user's choice, which is to remove the value from those commits or to allowlist a false positive by fingerprint. Then gbrain sources writer retry-effects <source> --parked-pushes runs the effects that parked. Every scanned push is also leased on the remote tip it scanned against. Before, a remote rewound past a pushed secret (an admin removing it with a force-push) left a stale tracking ref that hid the commit from the scan, and the next fast-forward put it back; now the scan runs against the remote's real tip. The in-process scan on the owner's event loop is capped at 10 seconds, inside the writer heartbeat's live window; a longer range goes to the keep-checkout publish, whose child process scans with the full budget.

Itemized changes

  • src/core/persistence/effect-git.ts: pushGitRoot reads the remote branch tip with git ls-remote, scans <tip>..<HEAD sha> and pushes that sha only as a fast-forward of the tip, leased on it; a tip that is not an ancestor, or a remote that moves before the push lands (stale info), goes to the keep-checkout publish, which now runs with --scan-secrets. The in-process scan gets 10 s; a range it cannot finish goes to the keep-checkout publish. A remote without the branch gets HEAD's whole history scanned and the branch created with a lease that it does not exist; one too large to scan in process holds with publish_scan_unreadable naming git ls-remote --heads. A refusal from either path holds the effects as publish_secret_found with the user's fix.
  • src/core/git/commit-push.ts: a scanSecrets publish scans <base>..<candidate> before every push and leases the push on that base, so each merge with a newly fetched remote tip is scanned again. It also scans <remote tip>..<candidate> before merging with a newly fetched tip, so a remote rewound past a pushed secret that also changed a file this machine changed refuses publish_secret_found rather than integrate_conflict, and a conflict's resolve command keeps --scan-secrets.
  • src/core/git/range-secret-scan.ts: rangeSecretRefusal, the one refusal both callers throw (publish_scan_unreadable for a scan that could not read the range, publish_secret_found otherwise; a gate-file change says the user pushes it). scanRange marks a scan that ran past its budget (timed_out) and scans a whole history when the base is empty. It skips commits the commit ledger records as page writes (effect, door, hosted_repo); gbrain git commit-push commits, adopted outside edits and commits the ledger no longer holds are still scanned, and an unreadable ledger means everything is scanned. A Git effect now makes its commit by plumbing (commit-tree over its staged paths, then update-ref HEAD <new> <old>, which refuses when HEAD moved, then git maintenance run --auto --detach as git commit does, without --detach on git before 2.47), so the ledger records a sha gbrain created before any ref names it. The scan skips an effect commit only by that sha, recorded verified when every staged file holds the bytes whose sha256 the effect recorded (the page walk passes the hash it checked too). An amend, reset or reword of it is a different commit and is scanned, a page file changed before it was staged leaves it unverified, and records written before this check existed are scanned. Its time budget is also checked while the diff is parsed.
  • src/commands/doctor/checks/git-convergence.ts, src/core/persistence/worktree-refresh.ts, src/core/persistence/git-push-holds.ts: every publish fix they name for a managed checkout (stranded, after committing its modified pages, diverged, refresh_diverged, and a held Git effect's conflict resolve) runs gbrain git commit-push --push-only --keep-checkout --scan-secrets, so a script's commit is scanned there too. This changes what the user sees. A commit that changes .gbrain-scan-allow or a .gitattributes now refuses publish_secret_found when the user runs the diverged fix or the refresh_diverged publish, where before their agreement published it.
  • src/core/git/commit-push.ts, src/core/persistence/git-conversion.ts, src/commands/git.ts: the publish commands the other messages hand back keep the scan too: a scanned publish's rerun hints (a push the remote keeps rejecting, a busy publish lock, commits outside the source root), the refresh's publish-by-hand line when it cannot find its CLI, the conversion's pending-push notice and a managed commit-push -m that leaves commits unpublished all name --keep-checkout --scan-secrets.
  • src/core/persistence/git-push-holds.ts: publish_secret_found and publish_scan_unreadable are held push codes, so doctor, gbrain sync status and retry-effects --parked-pushes cover them, with an ask-the-user fix (the refused commits, or git fsck).
  • Docs: docs/guides/shared-checkout-git.md (publish_secret_found and publish_scan_unreadable rows), docs/guides/live-sync.md.
  • Tests: test/effect-git-secret-gate.test.ts (an ahead checkout, a diverged one, a remote rewound past a pushed secret, a .gitattributes change and a remote without the branch whose history holds a secret each refuse with nothing pushed; a clean ahead checkout pushes, a clean history creates a missing branch, and a scan past the in-process budget still scans in the child; through the effect worker the page write's effect holds, sync status names the fix, and an allowlisted false positive releases it, parked or not; a page write whose body holds a credential-shaped value still backs up; a script commit, amend or reword racing the effect commit, a page file changed after the effect checked it, and an unverified ledger record each hold with nothing pushed; a source-scan walk's page with a credential-shaped body still backs up; a repaired unreadable commit releases its hold); test/persistence-git-removal-push-7089.test.ts and test/persistence-git-subdir-root-7091.test.ts (doctor's push fixes scan); test/guarded-publish.test.ts (a remote rewound past a pushed secret is not published again); test/git-literal-pathspec-hooks-7759.test.ts publishes its probe .gitattributes commit first, since the effect's push now refuses one; test/cli-flag-equals.test.ts lists range-secret-scan.ts as a git child argv reader.

[0.60.663.0] - 2026-11-11

The doc size budgets have room again

llms-full.txt is 751,119 bytes, 48,881 under its 800,000-byte budget, and the two key-files pages that most merges touch are under 54,000 bytes again. No budget was raised and no text was removed. Before, llms-full.txt sat 532 bytes under its budget and engines-2.md and core-services-2.md were within 8 bytes of the 60,000-byte page cap. So any change that added a sentence to those pages failed check:doc-history or the build-llms size test, and splitting a page cost bytes in KEY_FILES.md, which is 9,994 of its 10,000-byte cap.

  • scripts/llms-config.ts: the Grok Bot, Muse and in-agent setup guides are listed in llms.txt and no longer inlined in llms-full.txt. They only apply inside those hosted agents, and the inlined AGENTS.md routes those agents to them.
  • A key-files page at its cap continues on a new page linked from its last line, and the new page links back with [Previous entries](<page>.md). A page the index doesn't link is reached from its predecessor's last line. engines-2.md continues on engines-2-continued.md and core-services-2.md on core-services-2-more.md; the moved entries are unchanged.
  • test/docs-navigation.test.ts checks that every key-files page is reachable from KEY_FILES.md by following key-files links. A page the index does not link must name one predecessor whose last line links it, its chain must reach an indexed page, and no page has two continuations. The check:doc-history failure message and KEY_FILES.md state the rule.

[0.60.662.0] - 2026-11-11

QM, the AI helper that can use gbrain as one of its memories, can no longer write any page directly, and the brain owner can now review its proposals from another machine. Every page write QM makes for a person waits for the owner's approval, a fact it saves for a person stays private, and the owner's admin session reaches the proposal review over the running server. Unattested calls and member-bound credentials behave as before.

Every page write from QM waits for the owner

QM can no longer write any page directly; every page write waits for the brain owner's approval. QM is the AI helper that can use gbrain as one of its memories. Since v0.60.655.0 a call QM makes for a person (x-qm-actor) could not put, edit, link or add a timeline entry to a page; now no operation that writes a page is attestable, and a fact saved for a person can no longer be shared with everyone who reads the source. The person's own memory (memory_write, remember on their member page, forget, forget_fact) and owner-approved proposals (memory_propose) are unchanged. Unattested calls and member-bound credentials behave as before.

What changed

Area Before Now
Page writes from QM add_timeline_entry, add_link, put_page and edit_page were refused; every other page write could be listed in --attest-ops every operation that writes a page's content, frontmatter, tags, links, timeline, facts, raw data or uploads, or deletes or restores it, is refused, and so are the shared rows with no page (rate_answer feedback, log_ingest, loops_mute and loops_unmute) (actor_not_attested, reason op_not_attested), at --attest-ops, on every call and when a queued write publishes
A fact saved for a person share: true put it on its subject's page for every reader of the source refused (invalid_params, reason attested_share_refused), as is visibility: "world"; the reply points at memory_propose
Writeback instructions for a service acting for a person "share: true shares a fact about the team's work" never pass share: true; propose team knowledge with memory_propose
QM contract v2 and v3 deferred_ops named 11 operations they name every deferred operation; v1 is byte-identical

Itemized changes

  • src/core/access/actor-model.ts: ACTOR_DEFERRED_OPS gains every remaining page write, plus rate_answer (its feedback is keyed to the client, not the person, so one person could see another's restricted page in a refusal and move its ranking), log_ingest (text every reader of the ingest log sees) and loops_mute/loops_unmute (they close or reopen other people's loops). test/qm-attest-write-boundary.test.ts checks that every attestable mutating operation in the registry is either deferred or reviewed as not a page write, so a new page write cannot become attestable unnoticed.
  • src/core/verbs.ts, src/core/access/member-facts.ts, src/core/persistence/memory-prepare.ts: an attested remember refuses share: true and visibility: "world" at the call (a batch included), and a stored shared save made for a person is refused when it is prepared; a shared agent acting for an audience (on_behalf_of) still shares.
  • src/core/facts/writeback-instructions.ts: the instruction line for a service acting for a person.
  • docs/contracts/qm-memory-provider/v2.json, v3.json: regenerated deferred_ops (a published-contract amendment).
  • Docs: docs/guides/qm.md, docs/mcp/ADMIN.md, docs/guides/write-refusals.md, docs/tutorials/company-brain.md.
  • Tests: test/qm-attest-write-boundary.test.ts (new, PGLite and Postgres); test/access-qm-actor.test.ts calls every deferred operation attested; test/access-member-facts.test.ts and test/remember-share-gst2243.test.ts move their attested-write cases to the new boundary.

The owner reviews memory proposals remotely

The brain owner can now review QM's memory proposals from another machine. QM is the AI helper that can use gbrain as one of its memories; it proposes timeline entries and links with memory_propose, and nothing lands until the owner says yes. Until now that yes needed a shell on the brain host (gbrain memory-proposals). The owner's admin session now reaches the same review over the running server.

What changed

Area Before Now
Reviewing proposals remotely a shell on the brain host owner-session routes GET /admin/api/memory-proposals (?source=, ?cursor=, oldest first, 50 a page), GET /admin/api/memory-proposals/<request_id> and POST /admin/api/memory-proposals/<request_id> ({action: "approve"|"deny", reason?, dryRun?}), mirrored by gbrain mcp admin memory-proposals list|show|approve|deny; no OAuth token or MCP tool reaches them
Audit CLI decisions record decided_by HTTP decisions record decided_by: admin-api and one access_admin request-log row per applied decision, never the proposed text; a change that fails partway answers outcome: unknown
gbrain memory-proposals approve when the approved write fails exit 0 exit 1; the HTTP approve reports outcome: failed with the write's receipt

Itemized changes

  • src/commands/serve-http-access.ts, src/commands/serve-http-admin-api.ts: mountAdminMemoryProposals beside gap 4's access routes, on the same requireAdmin, dry-run, audit and outcome: unknown pattern; approve and deny call the CLI's library functions, so the one transaction, the re-checked actor and page revision and the writer-door publication are shared.
  • src/commands/mcp-admin.ts: the memory-proposals mirror; each decision takes exactly one request id, and deny needs a 1-500 character reason.
  • src/core/persistence/memory-proposals.ts: a keyset page cursor for the owner list and approveFailed, the one rule the CLI exit code and the HTTP outcome share.
  • Docs: docs/mcp/ADMIN.md, docs/guides/qm.md.
  • Tests: test/serve-http-memory-proposals.test.ts (new, PGLite and Postgres), test/mcp-admin-cli.serial.test.ts.

[0.60.661.0] - 2026-11-11

A take or a dated link no longer disappears when a page rewrite drops its section and a later write brings the section back. Two writers rebuilt a dropped section from scratch: the takes writers numbered their new take from 1 over the stored takes, and add_link wrote a links section holding only the new link. Both now start from what the brain still stores.

A take no longer disappears when think --take or gbrain takes extract --from-pages re-creates a takes section that a page rewrite dropped. When a page was rewritten without its takes section, its takes stayed stored, but the next think --take --anchor <slug> or gbrain takes extract --from-pages on that page built a fresh section numbered from 1: the new take replaced stored take #1 in place, with no record of the old one, and on a managed brain takes extract --from-pages --include-covered replaced the first stored takes and deleted the rest in the same write. These writers now render the stored takes, resolutions included, into the section they create and number their own takes after them, on repo-less, Git-backed and plain-directory brains. A take the user removed, or an explicit empty section retired, stays removed. A stored take whose resolution has no quality (recorded before takes had one), or whose text would not read back unchanged from the section (for example evidence holding the literal text <br>, which the section reads as a line break), cannot be written into the section, so the write is refused (invalid_input) instead of changing it. takes_add (gbrain takes add and the MCP tool) does not go through this yet. takes extract --from-pages skips a page whose takes markers are malformed (for example <!-- gbrain:takes:begin --> with two dashes) as fence_unparsed and goes on to the next page.

Each row of a page's ## Links section now keeps its own dates, and a section that a rewrite dropped comes back with them. A links row is keyed by target, type and source, but its valid_from and valid_until were stored per relationship (page, target, type) with no source. So two rows pointing at the same page with the same type shared one pool of dates, and after a page rewrite dropped the section, the next add_link wrote a section holding only the new link and deleted every other link of the old section, dates included. Dated statements now carry the row's source (link_transitions.link_source), and the identity index includes it. add_link on a page whose section is gone puts every kept row back with its own dates, remove_link deletes the removed row's own dates for any source (not only manual), and two rows that share a date both keep it. Dates stored before this change have no source. With one kept row they are that row's, and with several kept rows for one target and type the writer refuses link_dates_unattributed and asks for the section back through edit_page, instead of guessing. append_page now refuses text holding a links fence marker (append_links_fence); write links with add_link. A row of a dropped section whose target page never existed never became a link, so it is not brought back.

Itemized changes

  • Migration v302 (src/core/schema-migrations/v302-link-transitions-source.ts): adds the nullable link_transitions.link_source column (metadata only, under the schema lock timeout), builds the widened unique identity index online under a temporary name (CREATE UNIQUE INDEX CONCURRENTLY on Postgres), then swaps it in with one short drop-and-rename transaction. No row is rewritten. Requires Postgres 15, as gbrain already does since v011. The forward-reference bootstrap adds the column before the schema replay on an older brain.
  • src/core/links-fence-projection.ts: manual statements are written with their row's source and compared per row, so a row whose dates change replaces only its own.
  • src/core/persistence/manual-links.ts: add_link re-creates a dropped links section from the kept edges, each with its own dates; remove_link deletes the removed rows' own statements; both refuse link_dates_unattributed (src/core/error-registry.ts) when sourceless statements cannot be given to one row exactly.
  • src/core/persistence/page-append.ts: links fence markers in appended text refuse append_links_fence, whose fix names add_link.
  • src/core/takes-write.ts: withStoredTakesFence renders a page's stored takes (resolution cells included) into a new takes section when the body has no takes fence, refusing fence_unparsed when the body holds malformed takes markers, invalid_input when a stored resolution has no quality, and invalid_input when the render would not read back row for row with every stored cell unchanged. Used by addTakeToPage (think --take) and appendTakesToPageMdFirst (unmanaged takes extract --from-pages).
  • src/core/extract-takes-from-pages.ts: the managed takes extract --from-pages put composes from the same rendered section and dedupes new claims against those takes; a page the helper refuses is skipped with the refusal's code instead of ending the run.
  • src/core/behavior-change-notice.ts: one row for each changed behavior.
  • Docs: docs/architecture/key-files/hindsight.md, docs/architecture/key-files/core-persistence-continued.md, docs/architecture/key-files/core-services-2-continued.md, docs/guides/error-codes.md, docs/guides/temporal-edges.md.
  • Tests: test/link-dates-per-row-gst2228.test.ts (managed brain, repo-less and Git, PGLite and Postgres), test/link-transitions-source-migration-gst2228.test.ts (upgrade, interrupted runs, an invalid leftover index, writes between the column and the swap); test/recreated-takes-section-keeps-stored-takes.test.ts (think --take and takes extract on unmanaged brains, a plain directory and Git; managed extraction, repo-less and Git; removed takes, takes remove and empty sections stay removed; malformed markers skip one page; a resolution without a quality, or a cell that would not read back unchanged, refuses; PGLite and Postgres).

[0.60.660.0] - 2026-11-11

Agent reads answer "nothing here" instead of refusing, and say why a list came back empty. These fixes come from agent transcripts where models kept getting refused or got bare empty answers.

Itemized changes

  • get_backlinks with a type that other pages in the brain have, but no page linking to this one has, now returns the normal empty answer plus a note listing the types the linking pages do have. Before, it refused with unknown type, so an agent walking several accounts by type was refused once per account. A type no page has is still refused, and the refusal now also lists the brain's own page types. For a remote caller, both checks count only pages that caller can read. A type found only on private pages, pages whose readers leave the caller out, deleted pages or another source is treated as a type that does not exist and is never named.
  • list_pages with a type that no page has still returns the same empty list, but now adds a note naming the page types the brain does have, most common first (up to 12, with counts). Before, a guessed type such as account on a brain that files accounts as crm came back as a bare [], which reads like "no such pages". A remote caller's note names only types of pages it can read.
  • search with match: "keyword" now labels every row that matches the query as written keyword_exact, including rows from your own "A" OR "B" query or a one-word query. Their match scores sit below the cutoff the label used to need, so they were labelled weak_semantic, which told agents the batched lookup the instructions recommend had found nothing solid. create_safety (the "don't write a duplicate" hint) is unchanged for these rows. Rows found only by gbrain's own fallback of loosening a query to any-word matching stay weak_semantic. A query that is an entity's name, which keyword search also runs under the entity's other names, keeps the labels it had before, and default (hybrid) search is unchanged.
  • request_tools accepts tools and surface in one call. It changes the tool surface exactly as a surface-only call does (same limits and refusals), then returns the requested tool schemas, with the surface answer under surface_change. Before, the pair was refused, and some models send it every night. Calls with only one of the two answer as before.

[0.60.659.0] - 2026-11-11

A page purge removes the page's name from the volunteer log and the mention gazetteer

Purging a page now also deletes the rows that kept its name next to its slug. Before, delete_page purge:true and the retention purge left two stores behind, because neither has a foreign key to pages: context_volunteer_events kept the slug with a rationale that quotes the matched name (alias match "<name>") for up to 90 days, and mention_gazetteer_entries kept the page's normalized title and aliases until that source's next mention pass. The purge receipt called the page gone while both were still readable. Reported and first fixed by @alanshurafa (#9072).

What you did Before Now
delete_page purge:true, the expired-tombstone purge, or a restore that replays a page purge the slug and name stayed in the volunteer log and the mention gazetteer removed in the purge transaction, and the page receipt checks both stores
purge_fact on a claim that is also another page's name or quoted in a volunteer rationale the gazetteer was not checked both stores are probed and a match is listed as a residual
  • src/core/persistence/purge-name-stores.ts (new): scrubPurgedNameStores deletes the source's rows for the purged slugs from both stores. purgePageInTransaction, purgeDeletedPages, expireRecordedTombstones and the restore-time purge replay call it beside the retrieval-feedback scrub. A soft delete keeps both, so restore_page still brings the page back whole.
  • src/core/deletion-inventory.ts: both tables are swept (page_names); the fact receipt still probes them, since a fact purge sweeps neither.
  • Tests: test/purge-name-stores-9068.test.ts (PGLite, and Postgres through test/postgres-unit-arms.txt).

Removing a source, a connector delete and a late eval capture no longer leave a deleted page's slug behind

Four more ways a deleted page's slug outlived the page are closed, and the federation warning stops counting private pages. Before, only a page purge scrubbed the stores that name a page without a foreign key to it.

What you did Before Now
A query or search with eval capture on, whose capture runs after the response, racing a page purge the capture could write the purged slug back into eval_candidates.retrieved_slugs the insert keeps only slugs whose page still exists, checked under the same row lock the feedback writers take
A volunteered-context answer whose background volunteer-log write lands after a page purge (or while it is open) the write put the purged slug and its name-quoting rationale back into context_volunteer_events the write keeps only rows whose page still exists, under the same row lock
A connector, GitHub or Google sync deletes a page (deletePages), or a transcript re-ingest deletes a stale part retrieval logs, ratings, weights, the volunteer log, the gazetteer and captured eval slugs kept naming it scrubbed in the delete's transaction
gbrain sources remove, sources purge, the expired-archive sweep, or a managed remove/purge the source's retrieval logs, ratings, volunteer rows and captured eval slugs stayed deleted in the statement that deletes the source, and swept again right after for a write that raced it, unless a source with the same id was created in between
gbrain sources federate or sources restore on a source with a remote-written visibility: private page and no readers the warning named that page, though no-grant connections still cannot read it the warning applies the read paths' privacy rules, unless search.remote_private_pages or GBRAIN_REMOTE_PRIVATE_PAGES relaxes them
  • src/core/feedback/store.ts: insertEvalCandidate is the one eval capture insert both engines run. A facts-arm result (facts/<id>) names a fact, not a page, so it stays while that fact exists.
  • src/core/context/volunteer-events.ts: insertVolunteerEvents, behind the fire-and-forget volunteer-log sink, inserts only rows whose page exists, probed the same way.
  • src/core/engine-sql/pages.ts: deletePages runs scrubPurgedSlugs and scrubPurgedNameStores after the delete, in one transaction. The transcript importer's stale-part delete now calls deletePages. deletePage is unchanged: the page purge, its main caller, scrubs and records the stores in its receipt.
  • src/core/source-delete.ts: removedSourceScrubCtes runs inside deleteSourceRow and the managed lifecycle's source delete, then once more after it. deleteSourceRow skips that second pass when a source with the same id exists again, so a re-created source keeps its rows. retrieval_weights and mention_gazetteer_entries already cascaded from sources.
  • src/core/access/federation-exposure.ts: the exposure read adds privacyPartSql unless remotePrivacyRelaxed.
  • Tests: test/purge-scrub-followups-9068.test.ts and test/federation-exposure-private-9068.test.ts (PGLite, and Postgres through test/postgres-unit-arms.txt). test/eval-candidates.test.ts, the volunteer round-trip in test/migrate.test.ts and the fact-purge case in test/purge-name-stores-9068.test.ts now create the pages they log, since a capture or volunteer write keeps only live pages.

[0.60.658.0] - 2026-11-11

Pull requests now run the tests that check gbrain's lists, in about 4 minutes, instead of finding those failures in the 30-minute full run. Some tests fail whenever a change adds an operation, command, --flag, error code, doctor check or repair kind without updating the list that tracks it, or changes output a golden file pins. Before this release, pull requests skipped those tests and their first failure came in the full-ci run. A new check, enumeration and census tests, runs them on every pull request that changes more than docs. It is advisory: no merge rule reads it. bash scripts/ci-census-tests.sh --run runs the same tests on your own machine without Postgres.

Itemized changes

  • scripts/ci-census-tests.sh (new) prints the one file list: tests marked test-reads-source-ok[structural], every test that imports test/helpers/golden, the tests whose names match its enumeration pattern, the consumer idle-pass bounds (test/persistence-consumer-* files that mention idle), and the tests that read every file under src/ or test/ (its WALKERS list, such as the source-read smell test). It leaves out test/e2e/, serial and slow files, and test/setup.test.ts. On master that is 141 files. --run runs them all and --run <k> <n> runs one weight-balanced shard (scripts/sharding.ts).
  • .github/workflows/test.yml: the enumeration-census-tests job runs the list in two shards on 2-vCPU runners, only on pull requests. Trunk Dynamic CI gates it like the other cheap-tier lanes. The changes job adds a census output that skips a doc-only diff (scripts/select-e2e.ts --classify-only), and an unreadable diff still runs the job. cheap-tier, test-status, full-suite, the remote pregate and master-push reuse don't read the job, and contributor-gate stays the only required check. On this change's own run the shards took 3 min 40 s and 2 min 47 s, about 6.5 runner-minutes per push.
  • docs/ci-event-parity.md: new section "Enumeration and census tests" and a classification row.
  • Tests: test/scripts/ci-census-tests.test.ts (new) pins the list's shape and coverage. full-ci-workflow and native-only-workflow cover the new job's events and its doc-only skip. The PGLite snapshot cache-key census counts the job's key (16 keys).

[0.60.657.0] - 2026-11-11

A running gbrain serve, worker or autopilot no longer fails writes after another process changes a table's columns (GST-2260). When gbrain apply-migrations or an upgrade in another process added, renamed or dropped a column, every pooled connection that had already prepared a SELECT * or RETURNING * on that table failed its next run with "cached plan must not change result type". Outside a transaction the driver quietly prepared the statement again, but inside one the error had already aborted the transaction, so the caller saw it. A put_page came back internal_error, or was accepted and then failed at publication, once for each statement on each pooled connection. Version 0.60.646.0 made the write paths re-run the transaction it aborted. Now the first such error also makes every connection in the process prepare its statements again before its next query, and an OAuth token grant whose client read meets it and a minion job's completion or failure record are re-run too. Pools that do not prepare statements (transaction-mode pooler URLs, prepare: false) and PGLite are unchanged.

Itemized changes

  • vendor/postgres (patches/[email protected]): the first RevalidateCachedQuery error on a statement a connection had cached starts a process-wide plan generation; each connection drops its prepared statements before it builds its next query, and a statement described under an older generation is not cached in the new one.
  • src/core/retry-matcher.ts: retryOnStalePlan, one more run, on 0.60.646.0's isStalePlanError, for a transaction that does only database work.
  • src/core/oauth-grants.ts: the client-locked OAuth grant transaction runs once more when a stale plan fails its opening client read, before the callback writes anything.
  • src/core/minions/queue.ts: completeJob and failJob, whose transactions do only database work, run once more after a stale plan.
  • Tests: test/stale-plan-heal.test.ts (Postgres; listed in test/postgres-unit-arms.txt).

[0.60.656.0] - 2026-11-11

gbrain upgrade finishes on hosts where autopilot runs outside launchd and systemd, instead of refusing every time.

The upgrade restarted autopilot only through an installed launchd or systemd service, and refused (restart_refused, exit 3) for any other autopilot. That included gbrain's own Linux cron install (gbrain autopilot --install on a host without a user systemd bus), so a scheduled stable upgrade on such a host failed on every run. An autopilot whose launcher starts it again is now handed to that launcher. The upgrade stops it (only while ps still shows the same process; SIGTERM drains it) and waits up to six minutes for the launcher to start the installed build.

Situation Before Now
Autopilot started by this brain's gbrain cron line restart_refused handed to the cron line and verified back on the new build
Autopilot started by your own cron entry or wrapper restart_refused, fix kill <pid> still refuses, naming gbrain autopilot --install or gbrain config set self_upgrade.autopilot_launcher external (declare it only when the launcher starts autopilot again within 5 minutes); declared, it is handed over like the cron line
doctor self_upgrade_health silent until an upgrade failed runs the same pre-swap plan read-only: warns while gbrain upgrade would refuse, naming each process and its step (autopilot outside gbrain's service, a process from another install), and names a hand-off or a second Bun global install the upgrade replaces itself; a harness MCP server does not warn (--stop-mcp stops it), and a held or failed auto-upgrade is reported first

On a host whose autopilot runs from gbrain's cron line, the old binary still refuses with restart_refused on the way to this release, because it checks the restart before it installs. Run gbrain upgrade --no-restart, then gbrain upgrade --restart-only, so the new build does the hand-off.

The daily stable check no longer sticks on a rolled-back release. A release that was stable before (its stable-v<version> record tag exists) or that stable.json lists in known_bad is skipped and logged, so the newest release gets its soak instead of a refused promotion on every scheduled run (#8792).

A release can be marked not promotable when it soaked clean but is known bad from outside evidence. Dispatch promote-stable.yml for the current stable with not_promotable, not_promotable_reason and not_promotable_fixed_in. The stable record gains a not_promotable list, and no tag moves. The daily check never soaks or promotes a release on it, on the important path and the weekly floor alike, and a manual promotion of one is refused. Clients ignore the list, so unlike withdraw it moves no install to another release. not_promotable_remove takes a release off the list. A reason over 500 characters is refused, and so is a record larger than clients accept; nothing is pushed in either case.

[0.60.655.0] - 2026-11-10

QM, the AI helper that uses gbrain as one of its memories, can now read a person's own brain safely, read shared rooms only as the room's audience, and suggest edits that wait for the brain owner's yes. This lands the QM internal-provider chain (#8475, #8480, #8557, #8477, #8483, #8479, #8466) as one release. Nothing changes for a brain with no attested (x-qm-actor) client: unattested calls and member-bound credentials behave as before, and every new table is database state, so a repo-less brain works the same.

What changed

Area Before Now
Direct page writes from QM an attested call could run add_timeline_entry, add_link, put_page and edit_page refused (actor_not_attested, reason op_not_attested); --attest-ops refuses a list that names one, and a queued attested write of one fails when it publishes
Reads that name sources only memory_recall read named sources as their audience get_page, entity, whois, get_backlinks, get_timeline, find_experts, volunteer_context and query do too: restricted pages and member pages read as missing, and __all__ or all_sources is refused; an agent job submitted for the person reads the same way, and context_pack naming sources is refused on a client that reads them as the scope
How a client reads named sources always as the audience gbrain auth rescope --attest-read person|scope|none; scope also refuses any call that names no person; gbrain doctor warns (attest_read_shared_scope) when a person client reads a source a scope client reads
A person's own DM sources every attested caller with the grant read every DM source in it gbrain access member source add|remove|list binds a source to its member; an attested call reads its grant minus other members' personal sources, and a scope read reads no personal source at all
QM contract v1 and v2 v3 (docs/contracts/qm-memory-provider/v3.json) adds the person's own-DM tools with frozen result shapes, structuredContent on attested personal reads (re-measured against the result cap), an arm (own/ranked) on each memory_recall fact, and an opt-in memory_recall query_keep (head/tail) that adds query_truncated; v1 is byte-identical, and v2 changes only its deferred_ops list and its unknown_arguments wording
Suggested edits none memory_propose holds one timeline entry or one link for the owner; gbrain memory-proposals list|show|approve|deny on the brain host decides it, and approval publishes through the normal writer door with the member's id in its provenance
Members id and name an operator-declared email (gbrain access member add|set --email), and gbrain access member link --email <addr> --qm <id>
Remote owner administration a shell on the brain host gbrain mcp admin attest, member-add|member-set|member-disable|member-enable|members and link|unlink|links over the owner-session routes /admin/api/access/attest, /members and /member-links; no OAuth token or MCP tool reaches them

Itemized changes

  • src/core/access/actor-model.ts: the four page writes join ACTOR_DEFERRED_OPS; docs/contracts/qm-memory-provider/v2.json lists them in deferred_ops (a published-contract amendment).
  • src/core/access/scope-read.ts (new), src/mcp/dispatch.ts: SCOPE_READ_ACCESS, SCOPE_READ_OPS, attestedReadMode, also applied to agent job tool calls (src/core/minions/tools/brain-allowlist.ts); attest_read on access_tokens and oauth_clients (migration v298).
  • src/core/access/member-sources.ts (new): access_member_sources (migration v299) and attestedReadScope, also applied to agent jobs submitted for an attested actor.
  • scripts/build-qm-memory-contract.ts: v2 and v3 transport.unknown_arguments now say what every MCP read has done since v0.60.635.0: an undeclared argument named like a source scope or a filter (source for source_id) is refused with invalid_params and nothing is read, and any other undeclared read argument is only reported in _meta.warnings. Wording only; test/qm-memory-contract.test.ts pins it against the refusal check.
  • src/core/ops/qm-memory.ts, scripts/build-qm-memory-contract.ts: contract v3, qmMemoryToolResult for personal tools, recall arms, query_keep, and boundRecallChars trimming facts_groups and entity_unspecified to the facts it delivers.
  • src/core/persistence/memory-proposals.ts (new), src/commands/memory-proposals.ts (new): memory_proposals (migration v300), owner review, TTL and retention; a proposal's request id is scoped to its credential and member, and a proposed link re-checks its target and the approved page revision when it publishes.
  • src/core/access/members.ts, src/core/access/actor.ts, src/commands/access.ts, src/commands/serve-http-access.ts, src/commands/mcp-admin.ts: member email (migration v301), linkMemberByEmail, and the owner HTTP routes with an access_admin request-log row per applied change.
  • Inventories: access_member_sources, memory_proposals and the email column join the deletion and graduation inventories.
  • Docs: docs/guides/qm.md (reads that name sources, each person's own sources, proposed writes), docs/mcp/ADMIN.md, docs/architecture/key-files/access.md.
  • Tests: test/qm-scope-read.test.ts, test/qm-personal-sources.test.ts, test/qm-personal-tools.test.ts, test/qm-memory-propose.test.ts, test/access-members.test.ts, test/serve-http-access.test.ts, test/mcp-admin-cli.serial.test.ts (PGLite and Postgres where they touch the database).

[0.60.654.0] - 2026-11-10

A managed sync no longer fetches Git objects in a partial clone (git clone --filter). Every Git read a managed sync makes now runs with Git's transports and lazy fetching switched off. A file whose blob the clone never fetched is imported from your checkout when the checkout holds exactly that blob, and otherwise held alone as git_object_missing, with a message that names the partial clone and the git cat-file -p <blob> that fetches it. A blobless clone with a normal checkout syncs every file as before; an older object it never fetched (an earlier version of a renamed file or of the folder rules) is read only once you fetch it. A treeless clone (--filter=tree:0) that lacks a tree of the commit being synced stops the run with git_object_missing, naming the partial clone and the git ls-tree -r -d <commit> that fetches the trees; nothing is synced or deleted until then. The folder rules in .gbrain/access.yml are read the same way. A rule file or rule history the clone has not fetched makes the source read as nobody to remote members, as a shallow checkout already does, with a message that names the exact git command that fetches it. Run that command in the checkout and the next sync applies the rules and clears the mark.

Itemized changes

  • src/core/hardened-git.ts: NO_LAZY_FETCH_ENV is what a managed-sync read that runs under gitChildEnv adds beside HARDENED_GIT_ARGS: GIT_NO_LAZY_FETCH=1 (Git 2.45+) and an empty GIT_ALLOW_PROTOCOL, which allows no transport on any Git version.
  • syncGit and syncGitAsync (sync-discovery.ts), every read in sync-blobs.ts (ls-tree, cat-file --batch), the git show fallback, the delta and working-tree diff (sync-delta.ts), the rename-identity cat-file stream, the metafile head reads, the origin-evidence history walk and ignore check, the held-deletion read, the unindexed-file listing and the pending-import checks now run with both.
  • src/core/persistence/sync-entry-fault.ts: in a partial clone the git_object_missing hold names the partial clone and its fix is git cat-file -p <blob>; in any other checkout it still names git fsck --connectivity-only.
  • A partial clone with no checkout (--no-checkout) holds every committed file git_object_missing. When rename detection needs an older blob the partial clone lacks, the sync compares the whole tree instead, so the new path imports as a new page and the old page is deleted, and that rename does not keep its page.
  • src/core/access/policy-git.ts: the folder-rule reads every managed pass makes (git show and git log of .gbrain/access.yml and .gbrain/access.approved, and a rebuild's history replay) now run under HARDENED_GIT_ARGS and NO_LAZY_FETCH_ENV too; before, they were a managed-sync read that could still fetch in a partial clone. A policy file the commit names that the clone never fetched, or a history the replay cannot read, makes the source read as nobody, as a shallow checkout already does; the next sync that can read it clears that.
  • src/core/persistence/sync-discovery.ts: syncPartialClone and partialCloneTreeRefusal. The full tree walk and the reconcile level read (sync-pass-run.ts) stop with git_object_missing when a partial clone lacks a tree of the synced commit; any other listing failure is unchanged.
  • Git 2.40 to 2.44 ignore GIT_NO_LAZY_FETCH. There the empty GIT_ALLOW_PROTOCOL refuses the fetch instead, so a missing object reads as missing and nothing is fetched, the same outcome as on 2.45 and newer.
  • Tests: test/sync-partial-clone-2331.test.ts on PGLite and Postgres, with a marker on the partial clone's object fetches and a check that it never ran: a blobless clone that syncs a second commit after a pull, both blob read paths with and without GIT_NO_LAZY_FETCH (also with the file transport allowed in the environment), two treeless clones, the history walk that dates deletions, and folder rules whose committed file or history the clone lacks, which read as nobody until the command the problem names runs and then clear on the next plain sync.

[0.60.653.0] - 2026-11-10

Postgres connections to a database that isn't on your machine now use TLS unless the URL opts out. Before this release gbrain's Postgres driver connected in plaintext whenever the database URL had no sslmode, so the password and every page crossed the network unencrypted, even to a hosted database that supports TLS. Now every connection pool requires TLS (sslmode=require) for any host other than localhost, 127.0.0.0/8, ::1 or a Unix socket; private addresses and Docker service names count as remote. An sslmode in the URL is always honored, so ?sslmode=disable keeps a plaintext connection on a network you trust. A server that has no TLS no longer connects silently: it fails with postgres_tls_required, which says to turn TLS on at the server or add sslmode=disable. sslmode=require encrypts but does not verify the server certificate; sslmode=verify-full, with the provider's CA in NODE_EXTRA_CA_CERTS, verifies it too. A new doctor check, postgres_tls, warns when a brain on a remote host is connected without TLS. PGLite has no network connection and is unchanged.

If you self-host Postgres without TLS on another machine: after upgrading, gbrain refuses to connect with postgres_tls_required. Either enable TLS on the server (ssl = on with a certificate), or add ?sslmode=disable (or &sslmode=disable) to your database URL in ~/.gbrain/config.json or GBRAIN_DATABASE_URL/DATABASE_URL.

Itemized changes

  • src/core/pg-tls.ts (new): resolvePgTls / withPgTls, the one TLS resolver every postgres() call site in src/ uses (the module pool in db.ts, worker instance pools in postgres-engine.ts, the read and direct pools in connection-manager.ts, the graduation target probe and the schema-lock diagnostic connection). An ssl option, sslmode / ssl / sslrootcert=system in the URL, or the driver's PGSSL variable is explicit and left alone; a loopback host or socket keeps the driver default; anything else gets ssl: 'require'. After a failed default-required connect, tlsRequiredDefaultError sends the protocol's SSLRequest to the server and turns a "no" into postgres_tls_required (DNS, refused and timeout failures are not probed).
  • src/core/error-registry.ts: new code postgres_tls_required (host_only, actor user; fix asks the user, reads gbrain engine status --probe --json, verifies with gbrain doctor --only postgres_tls --json). docs/guides/error-codes.md regenerated.
  • src/core/pg-access-classify.ts: new reason tls_required_default (manual tier in gbrain db-repair, which never removes TLS on its own; fatal for persistence reads).
  • src/commands/doctor/checks/postgres-tls.ts (new): postgres_tls (ops) judges TLS from the client side, from the mode the driver used for the URL the engine connected with. It warns on sslmode=disable and on a sslmode=prefer connection that fell back or may have fallen back to plaintext (it never certifies prefer as encrypted), and is ok for a TLS client behind a pooler. The session's pg_stat_ssl row, which behind a pooler describes the pooler's hop, is detail only. It emits nothing on PGLite.
  • src/core/behavior-change-notice.ts: one row (postgres_tls_default) disclosing the new default and the opt-out.
  • gbrain init prints the Supabase URL format with ?sslmode=require; docs/ENGINES.md#tls-for-postgres-connections documents the modes, verify-full with Supabase's CA, and the opt-out; the personal-brain tutorial and the other documented Postgres URLs carry ?sslmode=require.
  • scripts/ci-local.sh, scripts/run-e2e.sh: the Docker CI gate reaches its untrusted-by-name service hosts (postgres-N, pgbouncer) with an explicit sslmode=disable, and the derived PgBouncer E2E URL keeps the base URL's query.
  • Tests: test/pg-tls.test.ts (loopback, remote, explicit modes, opt-out, a guard that every postgres() pool in src/ goes through withPgTls), test/pg-tls-default-refusal.test.ts (a fake server without TLS on a non-loopback interface gets the coded error; sslmode=disable and a TLS-capable server do not), test/doctor-postgres-tls.test.ts, test/e2e/pg-tls-postgres.test.ts (real servers: a TLS server through a non-loopback address connects with pg_stat_ssl.ssl = true on both pool kinds; an ssl=off server refuses, then connects with sslmode=disable and doctor warns; sslmode=prefer falls back and doctor says so; an opt-in TLS-client pooler with a plaintext upstream reads ok).
  • bun.lock: moves fast-xml-parser to 5.12.0 and fast-xml-builder to 1.3.1 (the minimum 5.12.0 declares) for a Low advisory in the parser's non-default stopNodes option, which gbrain's code does not reach; package.json ranges are unchanged.

[0.60.652.0] - 2026-11-10

The private-labels preview finishes on large sources. On a source with a few hundred thousand pages, gbrain repair private-labels ran past its 120-second statement bound while counting the pages a window of clears would expose, so the preview failed, and so did every --apply --expect, which repeats the check. The count used to pair every page with every cleared page. It now reads each page's references once and joins them to the cleared pages, so it returns the same pages and the same exposure map well inside the bound, which is unchanged. On a synthetic brain (scripts/bench/private-labels-dependents.ts: 300,000 pages in the window's source, 100,000 in another, 8 links per page) with a 1,000-page window, master hit the 120-second bound and this release took 0.33 to 1.06 seconds across the four shapes.

Itemized changes

  • The dependents statement in src/core/repair/private-labels.ts lists each page's source_slug, event.depth and derived_from slugs (with .md removed) in one pass, matches them to the cleared pages with equality joins, and applies the post-clear privacy rules only to the candidate pages. It runs with JIT off and is planned for its real parameters. No migration and no new index.
  • New equivalence tests on PGLite and Postgres run the old statement, kept verbatim in a test helper, next to the new one and require identical rows: a fixture with every kind of tie, each cleared page alone, a generated source and privacy aliases turned off.
  • A Postgres scale test requires a 50,000-page source with a 1,000-page window to finish inside 20 seconds. The old statement fails it.
  • scripts/bench/private-labels-dependents.ts builds the synthetic brain and times the preview, with --lineage for common derived_from chains and --class alias-and-visibility.

[0.60.651.0] - 2026-11-10

CI wastes fewer reruns: the macOS type check stops running out of memory, and a worktree test stops stalling for two minutes. The macOS type check now runs on GitHub's 14 GB Intel runner. On the 7 GB arm64 runner, a check that couldn't reuse a recent saved state ran short of memory, took 7 to 15 minutes and sometimes hit its 15-minute limit. On Intel a check with no saved state takes about 4 minutes. The stale-plan worktree test no longer waits out a 2-minute lease on PGLite: its wait for commits now lets timers run, so the background writer finishes its pass.

Itemized changes

  • The typecheck job in macos-validation.yml runs on macos-26-intel. It still fails on a real type error: a forced probe failed with TS2322 and exit code 2.
  • test/persistence-stale-plan-retry.test.ts pauses for timers after every pass of its commit wait, and a new probe pins that a timer-parked pass resumes during the wait. No assertion changed. The stall was test timing, and every page was committed.
  • actionlint moves from v1.7.11 to v1.7.12, which knows the macos-26-intel runner label.
  • The advisory harness check runs against harness-evals 7e8d872edb (v0.0.136). A hand run at that pin against this branch passed with 0 cross-scope pages, 36 assertions passed and 0 failed.
  • The shard weight maps for unit, serial and E2E tests are refreshed from the latest nightly runs.

[0.60.650.0] - 2026-11-10

Private-labels windows publish together and finish past declined pages. gbrain repair private-labels --apply --expect <hash> on a brain with managed persistence now admits a window's pages together and publishes them in groups, instead of one page and one wait at a time. On a 1,000-page window with Git commits and a second queue consumer (scripts/bench/private-labels-apply.ts), it went from 181.4 to 1365.3 pages a minute with no added database latency, and from 19.9 to 37.1 pages a minute with 20 ms added to each database round trip. Git commits for the window dropped from 990 to 22. The preview no longer claims a re-embed for every page: only pages stored under per-chunk synopsis, or with a chunk that has no vector yet, count.

Itemized changes

A page whose own state declines the write (its file changed outside gbrain, source_changed, or its body is stored in segments) fails alone and the rest of its group lands. A preview-bound apply of any repair kind that completes with only such declines now exits 0 and counts them under declined. The next private-labels preview lists pages an earlier apply tried and that did not change after every page not tried yet, and such a page declining again no longer counts toward the item_failures_systemic stop, so windows that carry declines forward keep finishing; pages not tried yet that all fail the same way still stop the run. A page that declined under a release before 0.60.649.0 has no ledger line, because those releases recorded only writes that landed, so the first apply after upgrading stops once with item_failures_systemic when five or more such pages head its window; that run leaves intent lines for its batch, so the next preview lists them last. The apply result names its ledger (details.ledger, details.ledger_rows_written).

A preview-bound repair apply now waits out a gbrain sources refresh for up to 10 minutes per step instead of 10 seconds. A refresh that runs longer ends the apply with worktree_refreshing, naming the refresh and how long it has run; rerunning the same apply resumes it (GBRA-443).

[0.60.649.0] - 2026-11-10

Access and privacy fixes. A member-bound credential now sees its own readable pages on every read surface, the brain owner's remote credentials get private facts on every fact surface, gbrain sources federate names the member-written pages it makes readable to everyone, and seven community-reported bugs in access, privacy, purge, repair and secret scanning are fixed.

Community-reported access, privacy, purge, repair and secret-scan fixes

Seven community-reported bugs in access, privacy, purge, repair and secret scanning are fixed. A member-bound caller now gets its own readable pages in core memory, entity lookups, reflex pointers and pull briefs, the same pages get_page returns. gbrain access who-can-read names the members a folder rule admits instead of saying everyone. Purging a page removes its name from the retrieval feedback tables and captured eval candidates. S3 and MinIO file URLs are signed for one hour, so a private bucket's file opens from another machine. The secret scanner now also catches a key right after a written \e[ terminal colour code. gbrain repair visibility now stamps a synthesized concept that has no visibility and no inputs, so the doctor warning about it can clear.

What you did Before Now
A member-bound credential reads core memory or resolves its own readable_by page the page was hidden the page is there, as in get_page
gbrain access who-can-read team-b/profile under a folder rule "every caller entitled to the source" the members the rule admits
delete_page purge:true, or the expired-tombstone purge the slug stayed in retrieval_* and eval_candidates removed, and the receipt checks each store
gbrain files signed-url with S3 or MinIO storage an unsigned URL, 403 on a private bucket a presigned GET that expires in an hour
gbrain repair visibility on a concept with no visibility and no inputs planned nothing, and doctor kept warning stamped private, and the warning clears
A browser MCP client sends X-Gbrain-Client from an allowlisted origin the preflight blocked the call allowed on both HTTP transports
Transcript text with a key right after a written \e[32m colour code the key was kept redacted

After upgrading, gbrain doctor asks for a fresh gbrain transcripts audit-secrets run (read-only) if your brain has imported transcripts, because the detectors changed.

Itemized changes

  • src/core/core-memory.ts, src/core/entities/resolve.ts, src/core/context/retrieval-reflex.ts, src/core/context/proactive-brief.ts: these reads filter with privatePagesFilterForCaller (and privateTimelineEventFilterForCaller for brief edits), the caller's member access plus the privacy rule with its grant and opt-out. Entity resolution and core memory take it only when asked (forCaller: the remote remember path, get_brain_context and the remote core budget message), so fence repair, subject inference and local core readers (compile-context --include-core, gbrain core, the session hook) keep the plain privacy fragment. Retrieval pointers and brief arms take the same forCaller option, which only the remote context_pack pull brief sets, so the per-turn hook, the OpenClaw context engine and gbrain watch keep the plain fragment even with the remote read hatch open. Reported by @javieraldape (#8442).
  • src/commands/access.ts: whoCanRead asks the read path's own predicate (accessAllowedSql with each member's satisfied folder rules) per enabled member, and the note names folder rules. Reported by @Masashi-Ono0611 (#8743).
  • src/core/feedback/store.ts: scrubPurgedSlugs deletes the source's retrieval_event_pages, retrieval_event_links, retrieval_feedback and retrieval_weights rows naming the slug (a page key or either end of a from|type|to edge key) and drops it from eval_candidates.retrieved_slugs where the results came from that source. Called in the purge transaction by purgePageInTransaction, the belief-ledger tombstone expiry and the bulk purgeDeletedPages. A federated capture keeps a slug that another of its sources still has a page for. The feedback writers insert rows only for pages that still exist, checked under a key-share lock on the page row, and the page purge scrubs after it deletes the page, so a write that races the purge on Postgres cannot bring the slug back. src/core/deletion-inventory.ts moves the four tables to swept; verifyPagePurge reports each store. A restore that replays a page purge (restorePgliteBackup, beliefs replay-purges) runs the same scrub on the restored brain (src/core/belief/purge-mirror.ts). Reported by @Masashi-Ono0611 (#8822) and @alanshurafa (#8823).
  • src/core/storage/s3.ts: getUrl returns a SigV4 presigned GET with a 3600 s expiry. New dependency @aws-sdk/s3-request-presigner 3.1028.0 (the client-s3 release), with @aws-sdk/util-format-url pinned to 3.972.9 from the same release. Reported by @liuwenyong1985 (#8793).
  • src/core/repair/visibility.ts: planLineage stamps private on a concept with no recoverable lineage and no visibility; an explicit value stays, and both still count in concepts_without_lineage. Reported by @Masashi-Ono0611 (#8856).
  • src/mcp/http-transport.ts, src/commands/serve-http-oauth.ts: X-Gbrain-Client joins the CORS allowed headers. Reported by @dajiaohuang (#8691).
  • src/core/secret-scan.ts: the escape boundary also accepts a written \e[…m colour code, next to the forms #8266 added. TRANSCRIPT_SECRET_AUDIT_VERSION is 4. Reported by @javieraldape (#8764).
  • Tests: test/caller-aux-page-reads.test.ts, test/access-who-can-read-folder-rules.test.ts, test/purge-retrieval-feedback-8821.test.ts, test/storage-s3-signed-url.test.ts, test/mcp-cors-client-header.test.ts, a \e[32m case in test/secret-scan-encoded-boundary-8266.test.ts, a new case in test/derived-visibility-repair.test.ts; test/storage-s3.test.ts checks the same URL bases on a presigned URL.

The brain owner's private facts, and a crash-safe private-labels repair

The brain owner's remote credentials (gbrain access owner set <member>) now get private facts beyond list-style recall: in the query facts arm, think trajectories, hot memory (_meta.brain_hot_memory), delta and context_pack hot facts. Facts taken from a page the owner cannot read (readable_by, folder rules) and facts in sources its grant does not read stay hidden, and the owner speaking for a shared-room audience gets world-only facts. Other members, unbound and public callers are unchanged, and the hot-memory cache keeps the owner's entries apart from every other caller's.

The owner's recall with a question now returns its private facts, as recall by entity or session already did. Before, ranking a question dropped them.

gbrain repair private-labels now writes its undo ledger ahead of each page. Every cleared or restored page gets an intent line before the page write and a committed line after it, so a run that stops between the two can still be undone with --class restore. When the ledger can't take the intent line, the apply refuses with storage_error and changes nothing (GST-2092).

--class restore treats ledger lines as untrusted input. It re-adds only access (a string) and visibility: private. A line with any other key, a non-string value or a page id that isn't an integer is counted as unreadable instead of failing the preview (GST-2092, GST-2128).

The private-labels preview now counts the pages each window would expose along with its own: derived_from pages, atoms whose origin is a listed page, synthesized concepts built from such atoms, and life-chronicle events extracted from a listed page or from a page whose only privacy is a derived_from naming one, each judged by the rules remote reads use. It reports them by category (details.dependents, no slugs) and warns when the count isn't zero. The preview hash covers the pages it counted, so an apply that would expose a page added since the preview refuses preview_changed (GST-2093). A ledger line a crash cut off no longer swallows the next one.

On a brain without managed persistence, a page whose file sits in its source's checkout is no longer cleared in the database only, because the next sync or re-import would put the label back from the file. The apply holds it with outcome file_backed and names the file to edit. Repo-less pages clear as before (GST-2128).

An apply that ends with a failure, a stop or a pending write keeps its approved preview, so rerunning the same --apply --expect <hash> resumes it instead of asking for a new preview (GST-2128).

Restore no longer lists a page whose label was removed by hand after a restore landed. The preview counts it as removed_after_restore and says so (GST-2128).

Federating a source names the member pages it exposes

gbrain sources federate now names the pages it makes readable to everyone that a member's credential wrote. A token or OAuth client bound to a member and granted its own source (the documented --sources private-<member>,default shape) writes there without readable_by, which is right while that source is not federated: only credentials granted it read it. Federating the source adds it to what every connection with no source grant reads, so each such page became readable by every unbound token and agent connection, and nothing said so. sources federate <id> still federates exactly as before, with the same exit code, and then prints on stderr how many pages in the source a remote token or OAuth client wrote that name no readers and no folder rule, up to five slugs with the member each credential is bound to now, gbrain access who-can-read <slug> --source <id> to check one, and gbrain access rule add --prefix <folder/> --readers <member> --source <id> (which previews before --yes) to keep a folder to its member; every command is shell-quoted. No record keeps which member a credential was bound to when it wrote, so pages from a credential unbound, rebound or deleted since then are listed too. A page whose remote author was followed only by a tag change (which moves the revision attribution without a version snapshot) is not listed. sources federate --json prints one object (id, federated, and pages_exposed when there are such pages) and moves the human lines to stderr. sources restore, which re-federates by default, prints the same list, and a managed restore returns it as pages_exposed. The pages are read from the database's write attribution after the change commits, so a repo-less brain reports them the same way. Writes made through an attested actor (x-qm-actor) carry the attesting system's credential and are listed under it.

Itemized changes

  • src/core/access/federation-exposure.ts: remotePagesFederationExposes reads the live pages of a source with no readers and no folder rule whose live revision or any earlier version a token or OAuth client wrote; remotePagesFederationExposesOrWarn runs it only for a change that federated the source and reports a failed read as a warning without changing the federation.
  • src/commands/sources.ts: sources federate (and --json) and the unmanaged sources restore print the list on stderr after the change commits.
  • src/core/persistence/administration.ts: a managed source_lifecycle restore that re-federates adds pages_exposed to its result.
  • Docs: docs/guides/multi-source-brains.md.
  • Tests: test/sources-federate-member-pages.test.ts (repo-less PGLite: a member-bound token's remote put_page into an unfederated source, then sources federate names that page and its member and not a page with readers or an operator page, and an unbound token then reads it; --json; a credential unbound or deleted since it wrote and a slug holding $(...), quoted in the commands; unmanaged restore and --no-federate; the managed restore result). It fails on master with an empty stderr.

[0.60.648.0] - 2026-11-10

A managed sync no longer imports a file that Git stores through a filter, such as a git-crypt encrypted note, as if its stored bytes were the page. The file alone is held content_filter_owned with a message that says gbrain does not run a repository's filter programs during sync (and names git-crypt when the file starts with its header), and the rest of the source imports and moves on. Reported by @JustinTSmith (#9024) and first fixed by @Masashi-Ono0611 (#9026).

Itemized changes

  • src/core/persistence/sync-blobs.ts: readFilterDrivers reads the filter attribute of many paths at the synced commit in one git check-attr -z --stdin --source=<commit> (git 2.40 or newer), or from a temporary index of that commit with check-attr --cached on an older git. Every Git call it makes runs with the hardened arguments (fsmonitor, hooks and every transport off) and GIT_NO_LAZY_FETCH=1, so reading the attribute runs no repository program and fetches nothing. The module's other reads also switch off core.fsmonitor and hooks.
  • src/core/persistence/sync-entry-fault.ts: filterOwnedRefusal refuses a blob that starts with git-crypt's header, or a committed blob whose path names a filter driver other than Git LFS, whether or not this checkout configures it (LFS pointers keep their own handling). For a plain-text file under an attribute nobody uses, a held file's message says to remove the attribute and run gbrain sources retry-held. missingAttributesRefusal holds a file git_object_missing when a partial clone lacks a .gitattributes that applies to it.
  • src/core/persistence/sync-screen.ts, sync-run.ts: the freeze screen and the sync.holds=fail path check committed imports with it before the NUL check, so a git-crypt file gets the clear hold instead of content_null_bytes. Working-tree reads are unaffected. A company-brain source checks its approved revision and stops with content_filter_owned.
  • New hold and error code content_filter_owned (error registry, fault class, public category, docs/guides/write-refusals.md, the sync unblock runbook).
  • Tests: test/sync-filter-owned-9024.test.ts and test/sync-filter-owned-company-9024.test.ts on PGLite and Postgres, with a clean/smudge filter, an fsmonitor and an upload-pack marker in the test repository, and a check that none of them ran.

[0.60.647.0] - 2026-11-10

A team member's remember can now share a fact with the team, and every remember reply says where the fact went. A member credential still saves privately by default, on its own page memory/members/<member>, so only that member recalls it. Passing share: true saves the fact on its subject's page instead, where everyone who reads the source recalls it, as every other caller's remember already does. The tool description tells agents to share facts about the team's work and keep facts about a person private. Before this, a member's visibility: "world" was quietly saved as private; it is now refused with member_share_required and nothing is saved, so no one believes they shared a fact the team cannot see. For the same reason, share: true onto a page whose facts are kept private (a page with readers named, or a brain whose fact default is private) is refused with share_target_restricted. A shared fact belongs to the team: only the brain host can withdraw it. Nothing already saved moves.

Itemized changes

  • src/core/verbs.ts: remember takes a share boolean on every MCP surface, starter included. A non-boolean share, share: true with visibility: "private", and a member's visibility: "world" (member_share_required) are refused with invalid_params before anything is saved, so one such item refuses a whole batch.
  • src/core/persistence/memory-mutations.ts: a member's share: true saves on the entity page with world visibility; when that page's facts would be private (whatever visibility the call passed), the save is refused (share_target_restricted), and preparation refuses it again if the page turned restricted after admission. In a batch, items saved before a refused item stay saved, as with other save-time refusals. New and replayed replies carry page and visibility.
  • src/core/facts/writeback-instructions.ts, writeback-config.ts, src/mcp/instructions.ts, the HTTP serves, the stdio serve bound through mcp.stdio_member and memory_conversation_mode: with ambient writeback on, a member caller is told to omit visibility and to pass share: true only for facts about the team's work, instead of the brain-wide "pass visibility: world" line its saves would now be refused for. A credential that may name a member per call (x-qm-actor on remember, as QM's service does) is told both: the member rule on calls made for a member, visibility: "world" on the rest. The member_share_required refusal names the exact retry.
  • src/core/remember-batch.ts: share is accepted per item and as a shared top-level key, and each item's receipt carries page and visibility.
  • test/mcp-schema-budget.test.ts: the starter tool-list ceilings and the remember budget grow by what share costs (+144 characters, 200 headroom kept).
  • QM's memory_write and its per-person contract (docs/contracts/qm-memory-provider/v1.json) are unchanged.
  • src/core/behavior-change-notice.ts: one row for the changed remember behavior.
  • Docs: docs/tutorials/company-brain.md, docs/mcp/ADMIN.md, docs/TOOL_CATALOG.md (regenerated).
  • Tests: test/remember-share-gst2243.test.ts (a shared fact is recalled by another member and an unbound reader, an unshared one is not; all three refusals; replayed and non-member replies; batch share; memory_write still saves on the member page); test/access-member-facts.test.ts drops the visibility: "world" an attested member save no longer accepts.

[0.60.646.0] - 2026-11-10

A migration that adds a column under a running server no longer fails writes on direct Postgres. On a direct Postgres connection the driver reuses each connection's prepared statements. When a migration adds a column to a table that a SELECT * reads, the next run of that statement inside a transaction fails with SQLSTATE 0A000, "cached plan must not change result type", once per pooled connection. Outside a transaction the driver quietly prepares it again; inside one, the transaction is already aborted and the error reaches the write. Two places lost writes this way. A legacy-token write whose authority reads access_tokens failed with storage_error and committed nothing; a retry worked. Worse, after a migration added a column to persistence_requests, an accepted write could end failed/storage_error at publication and was never retried, so a caller that left on write_pending lost it unless it polled the receipt; other submissions were refused at admission with internal_error. Four of the last 50 migrations (v254, v261, v292, v294) changed that table. Now the two access_tokens reads name their columns, so a new column leaves their result unchanged; the stale-plan error counts as a transient interruption of the write (cause stale_plan), so the request is claimed again after the usual backoff and stays bounded by persistence.max_transient_retries; and admission re-runs on it like a serialization failure. Only the stale-plan error matches (the server's plan-revalidation routine, in any locale, or that English message): every other 0A000 stays a refusal. Both reads also keep working on a schema behind the build (a brain whose migrations could not be applied): an accepted legacy-token write still replays, a new one is refused migrations_pending, and shared-skill membership still works. A transaction-mode pooler (prepare=false) and PGLite were never affected. Writes that already failed this way stay failed; gbrain doctor's terminal-failures check lists them, and each needs a resubmit with a new request_id.

Itemized changes

  • src/core/retry-matcher.ts: isStalePlanError matches SQLSTATE 0A000 with the server routine RevalidateCachedQuery, or, when no routine is reported, the "cached plan must not change result type" message.
  • src/core/persistence/transient-retry.ts: transientInterruption returns { cause: 'stale_plan', sqlstate: '0A000' } for it, so finishUnpublishedFailure, the recovery route and grouped publication charge it against persistence.max_transient_retries and release the claim with backoff; past the limit the write ends write_interrupted. src/core/persistence/coordinator.ts words that receipt for the new cause.
  • src/core/persistence/admission-retry.ts: retryWriteAdmission re-runs an admission that hit it, within its existing time budget.
  • src/core/persistence/coordinator.ts: transientDatabaseFailure also accepts the stale-plan error, so with persistence.preparation_deadlines false it is released uncounted as database contention, as a 57014 is; the effect worker, the consumer's release paths and the recovery resolver share that predicate.
  • src/core/grants/model.ts: TOKEN_GRANT_ROW_COLUMNS lists the columns grantFromTokenRow reads, and the async tokenGrantRowSelect(engine) selects the ones the table has; src/core/persistence/authority.ts (publication authority, FOR SHARE) and src/core/shared-skills/membership-authority.ts use it instead of *, at the cost of one catalog read per legacy-token call.
  • Docs: docs/guides/write-refusals.md and docs/guides/error-codes.md (write_interrupted), docs/architecture/key-files/persistence-stalls.md.
  • Tests: test/persistence-stale-plan-retry.test.ts adds a persistence_requests column under a pool of 3 and requires all 9 later put_page calls to commit with no request past the retry limit (on the previous code 3 to 9 of the 9 failed; without the admission retry or without the publication classification it still fails). test/access-token-grant-columns.test.ts runs both access_tokens reads on every pooled connection before and after an added column, plus a legacy-token put_page (all three failed on the previous code). Its worktree-backed case does the same on a git checkout and requires each page's file once with one Git commit (on the previous code 5 and 6 writes failed). test/access-token-grant-columns.test.ts also drops private_pages to stand in for an older schema: both reads still work, an accepted request replays and a new one is refused migrations_pending (42703 or unavailable on the previous code). test/stale-plan-classification-gst2314.test.ts covers the routine match, a non-English message and that other 0A000 errors stay refusals. test/publication-failure-owner-exception.test.ts now uses a different 0A000 as its sample owner exception, since the stale-plan one retries. test/access-ungranted-member-writes.test.ts resets its connections after the column-rename case, which left stale plans for the next test; no assertion changed.

[0.60.645.0] - 2026-11-10

A GitHub (or other API) connector source no longer stops for good when one of its items is a page stored in segments. A long issue (a body plus hundreds of comments) grows past 1 MiB and is stored in segments. Before, that item could never import, and once such a page existed, every later sweep of a bound source stopped on page_body_segmented with no item named, so nothing after it imported and the cursor never moved. Now an item up to 24 MiB imports and re-renders like any other, and one the sweep cannot read whole (over 24 MiB) fails on its own from the first sweep, and is held after three (gbrain sources status), while the rest of the run imports and the cursor moves; a new comment on it re-tries only that item. Reported and first diagnosed in #9062.

Itemized changes

  • src/core/persistence/connector-sync.ts: the connector publication reads a segmented page back as strings (up to 24 MiB) before it carries its timeline and fences into the re-render, seals a segmented body through its import (never the inline text seal, which read the whole timeline), and runs the import's after-commit step, so a segmented page left unsealed is chunked. The freshness read (page()) reads a segmented page back the same way.
  • src/core/github-source.ts: a stored page the sweep cannot read fails that item (a content hold after three runs), not the run.
  • src/core/connectors/connector-hold-session.ts: on a bound source, page_body_segmented raised before any write request is counted against the item instead of aborting the run, as invalid connector text already was.
  • Tests: test/connector-segmented-page-9062.test.ts (PGLite) and its PostgreSQL arm test/e2e/connector-segmented-page-9062.test.ts.

[0.60.644.0] - 2026-11-10

gbrain doctor now ends on its --budget-ms deadline when git hangs in an OpenClaw workspace. The bootstrap entry checks push health by running git status, git branch and git rev-list on the install receipt's workspace, and on a workspace whose last push failed. Those calls ran synchronously with a 10-second limit each. While one hung, nothing else in the process could run, so the run overran its budget and every later check, the cheap incident checks included, started with no time left. They now run asynchronously, like the home-dir probes since #8979. Each call is killed with SIGKILL, together with its process group, at 10 seconds or at the entry's deadline, whichever comes first, and a call that fails or is cut short reads as unverified, as before. Every local probe's output is also capped at 1 MiB again: a probe that prints more than that is killed and reads as absent, which is what execFile's buffer limit did before #8979.

The Postgres crash robot no longer fails a full run when a race makes one fewer slot than its count run saw. The robot first runs each schedule to count the seams it reaches, then crashes one run at a counted slot. In sync_and_connector_race_direct_write, the connector republish races a direct write, and the race legally settles two ways: one makes 5 publications, the other 6, with the git effects, preparations and commits following. A crash planned at the 6th publication was never reached by a run that took the 5-publication outcome, and the job failed with fault_not_reached (#8078). A run that misses its fault now reports how many slots of that seam it made. If it was exactly one short of the count run, it runs once more at its last slot, and the manifest records retargeted on the run and lists it under faults_retargeted. A run that found violations, made no slot, fell two or more short, or misses again still fails as fault_not_reached.

Itemized changes

  • src/commands/doctor/local-probe.ts: localProbe moved here from checks/local-audits.ts. It takes ownTimeoutMs (default 5 s) and windowsShell (git runs without cmd.exe on Windows), and it kills a probe and returns null once its stdout passes LOCAL_PROBE_MAX_OUTPUT_BYTES (1 MiB).
  • src/commands/doctor/bootstrap-checks.ts: bootstrapDoctorChecks(engine, entryDeadlineAt) runs every git call through localProbe with a 10 s bound. commitsAheadOfOrigin, treeMatchesOrigin and recordedPushFailureCheck are async.
  • Tests: test/doctor-bootstrap-git-deadline-8990.test.ts. A git stand-in that ignores SIGTERM and waits on a 6 s child is killed within a 1000 ms budget, on both the workspace probe and the recorded-failure probe, and neither it nor its child survives (each fails on the previous code, at 18 s and 12 s). A probe that floods stdout is killed within 1.5 s and reads as absent, one that prints 2 MiB and exits reads as absent, and one under the cap returns its whole output (the first two fail without the cap).
  • scripts/persistence/crash-robot.ts: the run role's fault hook reports reached, the slots of its seam it made. scripts/persistence/robot-driver.ts: a planned crash carries counted (the count run's slots), and retargetable reruns a run once when it found no violations, counted is set and 0 < reached, counted - 1 <= reached < nth.
  • Tests: test/persistence-crash-robot.test.ts. A run one slot short crashes at its last slot and passes; a short run with violations, a run two short, a run short again after the retarget, a run with no slot and a run with no counted plan still fail. The failed #9071 manifest entry, given the counted: 6 its count run saw, failed 4 of 4 on Postgres before the fix and passed 3 of 3 after it, crashing at the 5th publication.

[0.60.643.0] - 2026-11-10

Retrying an interrupted gbrain sources remove or purge can no longer delete a re-added source's pages. A managed remove or purge deletes a large source's pages in batches, and a run that stops partway can be finished by rerunning it with the same --request-id. Until now nothing tied that request id to the source it was started against: if the source was removed and added again under the same ID in between, the rerun deleted the new source's pages. The first batch now records which incarnation of the source the request was admitted against, in the same transaction as its first deletion. A rerun against the same incarnation still resumes and finishes; a rerun after the source was replaced refuses with source_retire_superseded and changes nothing. Restoring, archiving or otherwise changing the source after a run stopped abandons that run the same way, so a purge stopped before a restore can't hold up deactivate or graduation for good, and both ask the user before a stopped run is finished, since finishing deletes the rest of the source's pages. A stopped automatic expired-archive purge is left to the next automatic run instead of being handed to a person, and gbrain sources writer status --source <id> --json now lists a source's stopped and superseded remove and purge requests.

Itemized changes

  • src/core/persistence/topology-receipts.ts: admitTopologyRetire records a retire admission (a persistence_topology_changes row in state recovering with no recovery record, bound to the source incarnation, holding one lifetime ID and a full receipt reservation after the same quota check assertTopologyChangeCapacity made); recordTopologyChange settles a request's own admission committed and trues up its receipt bytes; supersedeTopologyRetires settles other requests' admissions failed when a remove or purge deletes their incarnation, or when any other lifecycle operation (archive, restore, rebind, claim) commits a change to it; it skips only the caller's own (principal, request id) pair, so a retry under a replaced CLI registration settles the old registration's admission too.
  • src/core/persistence/source-lifecycle.ts: a request id whose admission names another incarnation than the source registered now refuses source_retire_superseded before any lock or deletion (a dry run too) and settles the admission failed, so a later retry refuses the same way; the batches and the final transaction resume past their own admission (#8984).
  • src/core/persistence/source-lifecycle.ts: a remove or purge retry also refuses source_retire_superseded when another CLI registration (one replaced since) admitted the same request id against a different incarnation of the source, and settles that admission failed; idempotency is per registration, so before this the retry ran as a new request and deleted the re-added source's pages. An admission bound to the current incarnation still resumes under the new registration (#8984).
  • src/core/error-registry.ts: new caller code source_retire_superseded, with a read-only writer-status fix; sources writer status --json gains topology_requests (recovering admissions and superseded requests).
  • src/core/persistence/deactivation.ts, graduation-drain.ts: an unfinished retire is listed as a blocker that names its request id and asks the user whether to finish it (rerun the same command with its request id) or keep what is left (another lifecycle change abandons it); graduation marks it needsUser. A stopped automatic expired-archive purge (its admission records expired_only) instead says the next automatic run finishes it, with needsUser: false.
  • Tests: test/source-retire-incarnation.test.ts (PGLite) and test/e2e/source-retire-incarnation-postgres.test.ts: a retry after the final transaction failed, and a retry after committed batches already deleted part of the source, both refuse once the source was replaced and keep the new source's pages (both resolved committed on master); a same-incarnation retry resumes and settles; a restore after an interrupted purge abandons it, so its retry refuses and no blocker stays. test/source-retire-batched.test.ts and its Postgres twin now expect the interrupted run's admission row.

[0.60.642.0] - 2026-11-10

A purged page's facts no longer come back. When the 72-hour retention purge removed a soft-deleted page's row, it left the page's facts active, because facts have no foreign key to pages. Once the tombstone was gone, those facts showed up again for every reader, local and remote, and a new page written at the same slug picked them up. The purge now expires those facts in the same transaction as the page delete: the facts filed on the page, and the facts about it that no page holds. The rows are kept, expired, never deleted. A fact another page holds about the slug stays active, even when that page is soft-deleted, so restoring it brings the fact back. An ontology observation about a live entity that was sourced from the purged page also stays active.

Itemized changes

  • src/core/facts/gone-page-facts.ts: expireGonePageFacts, the facts a purged page takes with it, as two updates that each use their slug index.
  • src/core/engine-sql/pages.ts (purgeDeletedPages, the bulk purge) and src/core/persistence/purge-deleted.ts (expireRecordedTombstones, the belief-ledger purge of a recorded source) call it in the purge transaction. The managed purge already sweeps the facts filed on the page.
  • Tests: test/purge-expires-page-facts.test.ts (PGLite and Postgres), which fails before the fix on the bulk and recorded purges.

[0.60.641.0] - 2026-11-10

Facts, timeline entries, tags and ontology observations now follow the page they are about, so a remote caller who cannot read a page learns nothing about it from fact reads, and a deleted page's facts, timeline and tags stay hidden from everyone until it is restored. Before this, a fact saved on an open page about a page with readable_by or visibility: private still reached remote callers through recall, ontology_get, volunteer_chronicle, delta, find_trajectory, hot memory and the entity card, and a soft-deleted page's facts, timeline rows and tags kept answering every reader. Recorded-time reads (get_page and get_versions with known_as_of), plain get_versions with include_body, and delta's belief changes now also keep a quarantined page's bodies and claims from remote callers, including a page quarantined only by a later write and a version history that spans several sources. Nothing is deleted or expired, so restore_page brings every row back as it was. A new page written at a deleted page's slug still shows the deleted page's facts, and a purged page's facts still come back, as before; both are follow-up work.

Itemized changes

  • src/core/search/private-visibility.ts: privateFactFilterForCaller and privateEntityFilterForCaller add the entity-page read gate to the provenance-page gate; recall, the entity card, contested claims, attribution, saved-fact matching, entity facts and the proactive brief use it.
  • src/core/eligibility/sql.ts: deletedPageProjectionHiddenSql hides facts on or about a soft-deleted page, and that page's takes and timeline rows, in every projection read; fact list reads, ontology reads and the unattributed-facts review list apply it even without a read policy.
  • src/core/ops/tags.ts: get_tags reads live pages only, for local callers too.
  • src/core/ops/get-page-projection.ts, admin.ts, known-as-of.ts: historyQuarantine gives each stored version the quarantine view of its own page (quarantined now or in that version), so a remote caller gets no body for it.
  • src/core/belief/history.ts, changes.ts: a quarantined page's fact and take events and fence claims stay out of a remote caller's belief history and delta.
  • src/core/persistence/page-prepare.ts: a put over a soft-deleted page takes its tags from its own content and does not write the deleted page's database-only timeline rows into the new page's first version.
  • Tests: test/fact-reads-page-access.test.ts (new), test/recorded-read-gate.test.ts (new cases 4 to 7), test/get-tags-remote-privacy.test.ts, on PGLite and Postgres.

[0.60.640.0] - 2026-11-10

The skill resolver's hook lane is faster and safer at its 50 ms deadline, offers skills again after a Codex compaction, supports Pi and Cursor, and takes per-harness settings. The hook starts its worker first and skips the CLI dispatcher, so the worker bundle drops from 172 KB to 62 KB and the lane's p90 drops from 58.9 ms to 38.1 ms on a 2-CPU machine (0 of 40 turns over 50 ms, was 7). A turn the deadline ends mid-commit now removes its session temp file and releases the session lease. The resolver stays off by default, lexical-only and fail-open.

Itemized changes

  • Lane speed: the hook fast path (src/core/skills/resolver/hook-lane.ts, hook-worker.ts, lane-worker.ts) starts the worker before reading settings and never loads the CLI dispatcher. That hook process no longer prints the update heartbeat or the backup nag; GBRAIN_HOOKS=0 and the CLI preflight still apply. A worker that exits before answering is logged as skill_resolver_internal at once instead of waiting out the deadline.
  • Deadline: the session commit holds a release handle, so a deadline exit removes the temp file and the .json.lock lease it took.
  • Codex compaction: gbrain setup codex --skill-loading resolve also writes a SessionStart hook (with its trust entry), and a SessionStart with source: "compact" resets the session's offered skills. An existing install gets the hook on its next setup run.
  • Harnesses: pi and cursor join the path harnesses (each suggestion carries its SKILL.md path). The turn-input and build-receipt schemas list them.
  • Resolve: min_match_strength (--min-match) drops weaker suggestions before the cut to top_k; exclude lists take up to 1,024 ids (EXCLUDE_MAX); skill_catalog_mismatch carries catalog_hash; a skill the turn names keeps its line after the budget is spent; with no catalog word in the message, previous_user_message is ranked instead. RESOLVER_FEATURES and MATCH_STRENGTHS are exported from gbrain/skill-resolver.
  • Config: skills.resolver.top_k, skills.resolver.min_match, skills.resolver.pinned and skills.resolver.per_harness.<harness>.{top_k,min_match}, validated by gbrain config set (which refuses per_harness.constructor). With nothing set, Claude Code and Codex keep 5 suggestions and no floor.
  • Logs: skill-index/logs/lane-<UTC day>.jsonl and resolver-<UTC day>.log, kept 14 days and capped at 2 MB a day; the old undated file is removed and doctor reads today and yesterday.
  • Library bundle: gbrain-skill-resolver.mjs is built with whitespace minification, so the CLI build and the release build have one sha256 whatever the working directory.
  • Bench: scripts/bench-skill-resolver.ts runs a discarded warm-up round, then judges the median of nine rounds of 100 cold turns; the bars stay 10 ms at 1,000 skills and 40 ms at 7,500.
  • Tests: test/skill-resolver-hook-lane.test.ts (deadline stalls during the lease and after the temp write, Codex compaction, fast path), test/skill-resolver.test.ts, test/skill-resolver-contract.test.ts, test/setup-skill-resolver.test.ts, test/skill-resolver-packaging.test.ts (two-directory sha256), test/skill-resolver-cli.test.ts and test/cli-fast-path.test.ts.

[0.60.639.0] - 2026-11-10

Access hardening: job results, write receipts, the status snapshot and graph edges stay inside the caller's grant; a forgotten fact stays forgotten; transcript redaction, .env handling and test databases are tighter. An agent job, its result and its progress now answer only a caller acting for the same audience the job was submitted with, and write receipts follow the same rule. get_status_snapshot lists only the sources a remote caller's grant reads. Link reads, graph walks, entity cards and context packs leave out edges whose authoring page is deleted or, for a source-scoped caller, lives outside the caller's sources. A forgotten fact stays forgotten when the default visibility narrows or extraction runs again, while an explicit remember of the same claim still saves it. Transcript redaction also catches credentials assigned to names that end in a key word (STRIPE_KEY=), a .env file in the working directory can no longer set GBRAIN_SURFACE, the stale-editor-save check works on any Postgres encoding, and the throwaway test databases publish on loopback only.

Itemized changes

  • Jobs: jobActorSql (src/core/minions/types.ts) matches the job's stored audience as well as its attested actor, so get_agent_job, get_job, wait_job, get_job_progress, list_jobs (both paging modes) and cancel_job answer only the audience the job was submitted for; a job with no audience answers only callers that name none.
  • Write receipts: sameAttestedActor (src/core/access/member-facts.ts) compares a stored audience the same way, so get_write_request, list_write_requests, cancel_write_request and request_id replay stay with the audience the write was made for.
  • get_status_snapshot (src/core/ops/skills-catalog.ts) confines its sources to the sourceScopeOpts ladder sources_list and get_health use; the local CLI keeps every source.
  • Links (src/core/engine-sql/links.ts, src/core/verbs/entity-card.ts): every arm of getLinks/getBacklinks, traversal, listLinkSources and the entity card drop an edge whose origin page is soft-deleted, and, under a source grant, one whose origin page is outside it. For remote callers, find_orphans counts only links whose other endpoint and origin page are inside the caller's sources; the local operator's count stays brain-wide.
  • Facts (src/core/facts/withdrawal.ts, backstop.ts, fence-write.ts, src/core/persistence/facts-prepare.ts, facts-maintenance.ts): every automatic admission path (extraction, the refused-window retry, the withdrawal retry and the moved-facts follow-up) checks the withdrawal ledger under the lock at any visibility, so narrowing facts.default_visibility or extracting again never brings a forgotten claim back. remember with content-origin user_said still saves it.
  • Redaction (src/core/secret-scan-assignment.ts): the high_entropy_assignment key also accepts an identifier whose last word is key (SUPABASE_SERVICE_ROLE_KEY=, deploy-key:); MONKEY=, a bare KEY= and short or low-entropy values stay as written.
  • .env (src/core/env-trust.ts): GBRAIN_SURFACE joins CWD_DOTENV_PROTECTED_KEYS. A shell export still sets the stdio surface; a repository's .env cannot widen it.
  • Local-edit adoption (src/core/persistence/local-edit-adoption.ts): the stale-editor-save comparison trims and compares revisions in TypeScript, so it works on SQL_ASCII and LATIN1 databases. New doctor check database_encoding warns when a Postgres brain's server_encoding is not UTF8.
  • Test databases: docker-compose.ci.yml, docker-compose.test.yml and every documented docker run -p publish on 127.0.0.1 only.
  • Skills: skills/conventions/untrusted-content.md and the ingest skills describe the controls that ship (remote-read quoting in transcript-quote.ts, TURN_CONTEXT_ENVELOPE) and drop flag and fence names no code reads; text an agent ingests itself is stored at owner tier, unquoted.
  • Tests: test/access-on-behalf-of.test.ts, test/status-snapshot-source-scope.test.ts, test/link-origin-scope.test.ts, test/transcript-redaction-key-names.test.ts, test/env-trust.test.ts, test/test-db-loopback-ports.test.ts, test/adopt-local-edit.test.ts, test/doctor-database-encoding.test.ts, test/managed-sweep-refused-window-retry.test.ts, test/managed-facts-withdrawal-recovery-6506.test.ts, test/managed-facts-withdrawn-skip-6506.test.ts, test/facts-entity-moved-7602.test.ts, test/untrusted-content-convention.test.ts (PGLite and Postgres where an engine is involved).

[0.60.638.0] - 2026-11-10

Atoms extracted from a world page never carry a fence row the page shows only to its owner, and gbrain repair atom-fence-rows finds and narrows atoms stored before. A world page's facts table can hold rows marked visibility: private, and its takes table is never part of what a remote caller reads. Atom extraction, take proposals and the patterns phase read a page as a world reader sees it, so what they produce holds only world rows. Atoms an earlier version stored may still hold such a row: doctor atom_fence_rows counts them and names the read-only preview, and only gbrain repair atom-fence-rows --apply --expect <hash> makes the previewed atoms private. Nothing runs on upgrade.

Itemized changes

  • src/core/cycle/extract-atoms-input.ts: a page's extraction input is its publicBody (facts rows other than world and the takes fence removed); the cap, the watermark and quote verification apply to that text, on the cycle phase and the managed path alike.
  • src/core/cycle/propose-takes.ts: the extractor receives the page's publicBody, and a proposal that restates a take held by someone other than world is dropped before it is filed; the content hash, fence skip and dedup rows still read the stored body.
  • src/core/cycle/patterns.ts: reflection excerpts in the subagent prompt and the quote-grounding sources are the reflection's publicBody.
  • src/core/repair/atom-fence-rows.ts (new, registered in repair/core.ts, repair/registry.ts, persistence/preview-approval.ts): explicit-only, preview-bound kind; the preview lists each world atom with its source page and the row kinds and numbers it carries (never the row text) and changes nothing; the apply sets visibility: private on exactly the previewed set through a revision-bound put_page.
  • src/commands/doctor/checks/atom-fence-rows.ts (new, a host-only WAVE_CHECKS entry, BRAIN category): atom_fence_rows warns with the count per source and the preview command.
  • Docs: docs/guides/repair.md#atom-fence-rows (what is listed, the apply, cost), skills/RESOLVER.md, key-files rows in core-cycle.md, hindsight.md, core-cycle-patterns.md and memory-purge.md.
  • Tests (PGLite and Postgres): test/extract-atoms-world-fence-rows.test.ts (cycle phase and managed brain: prompt, atoms and a remote read of each atom hold only world rows), test/cycle-world-fence-rows.test.ts (propose_takes input and proposals, patterns prompt and quote grounding), test/repair-atom-fence-rows.test.ts (doctor finding, preview changes nothing, apply without --expect refuses, apply narrows exactly the previewed atoms, unmanaged and managed).

[0.60.637.0] - 2026-11-10

A file Git puts back after sync deleted its page brings the page back, and the retention purge never deletes a committed file. On a managed brain, a commit that removed a Markdown file made sync soft-delete its page. When the same file came back (a git revert, switching to a branch and back, a collaborator's delete-then-restore, a file missing from disk for one working-tree sync), sync settled the path as "deleted, file unchanged" and left the page out of get, search and recall while gbrain sync status said synced. gbrain pages purge-deleted --yes then deleted the restored file from the checkout. Sync now tells a tombstone it made from a missing file apart from a person's delete and imports the file over it, so the same page comes back; a brain an older version left in that state heals on its next sync. The purge leaves a file committed at the source's last synced commit or at HEAD alone, unless a person's delete of the page never removed the file (for example because another page shared it). A page purge also stops leaving timeline text in the write journal.

Itemized changes

  • src/core/persistence/sync-pass.ts, sync-classify.ts: the level read marks a tombstone a managed sync deletion made (its open sync-delete-tombstone record or uncompacted managed_sync_delete receipt); once its file is back with clean committed bytes it classifies import (the import revives the same page row), and a file_drift hold an older version raised over it re-runs. A person's tombstone keeps deleted_same / deleted_edited.
  • src/core/persistence/sync-pass-run.ts: the selection fingerprint carries sync_deleted_restore, so the first sync after the upgrade reads the whole source once and brings back pages left deleted beside their file.
  • src/core/persistence/page-prepare.ts, sync-markers.ts: a retention purge of a tombstone whose file is committed at the source's last_commit or HEAD removes no file: a sync-made tombstone defers (revision_conflict) to the sync that restores it; any other is purged database-only (write_through.skipped: committed_file), except a person's delete that made the tombstone and never removed its file (write_through.written: false, for example because another live page shared it), which still has the purge finish that removal. Publication re-checks, so a file committed while the purge was prepared is kept too. A sync-made tombstone defers only while its file is on disk; off disk (working-tree mode) it is purged database-only without a deletion marker, so the file's later return imports it. An unreadable checkout counts as committed; a checkout with no commit yet does not. Repo-less sources are unchanged.
  • src/core/persistence/page-purge.ts: a page purge redacts the entry summary and source and the bullet text of the page's add_timeline_entry receipts, compacted ones included (up to 10,000 per page; past that journal_sweep.complete is false).
  • Tests: test/sync-restored-file-purge.test.ts (PGLite and Postgres), test/sync-classify-6712.test.ts, test/purge-eval-fixes.test.ts.

[0.60.636.0] - 2026-11-10

A fact or take no longer disappears when gbrain re-creates a ## Facts or takes section that a page rewrite dropped. When an agent rewrote a page without its facts or takes section, gbrain correctly kept those rows live until the user decided. But the next remember, takes add, fact extraction, gbrain facts relink or dream-cycle merge of a duplicate stub page into that page built a new section holding only its own row, and the write after that (a tag, a timeline entry, an append, or a get and put round trip) read the new section as the user's answer: the remembered facts expired and the takes were deleted, on repo-less and Git-backed brains alike. Those writers now render the held rows back into the section they create, so nothing is lost; a fact the user forgot, a take they removed, and rows retired by an explicit empty section stay withdrawn. gbrain repair lost-fence-rows previews the rows an older gbrain already lost this way and, after --apply --expect <hash> --yes, puts exactly those rows back from page history.

Itemized changes

  • src/core/facts/unfenced-facts.ts: heldFenceFactRows lists the live rows a facts fence of the page held; withHeldFactsFence renders them into a new facts section when the page has none. Used by remember (fence-fact-append.ts), managed fact extraction (persistence/facts-prepare.ts, which also serves Google loops and the put backstop), facts relink (relink-publish.ts), unmanaged writeFactsToFence (fence-write.ts), cycle fact adoption (planFence) and the dream cycle's phantom redirect (cycle/phantom-redirect.ts, cycle/phantom-redirect-managed.ts).
  • src/core/persistence/takes-prepare.ts: heldTakeRows reads the stored takes while the take fence is held fence_absent; a takes write that creates the fence renders them first (covers takes add, takes propose --accept and consolidation).
  • src/core/repair/lost-fence-rows.ts: new explicit-only, preview-bound repair kind lost-fence-rows, registered in repair/core.ts and repair/registry.ts; persistence/preview-approval.ts accepts its approvals.
  • Docs: docs/guides/repair.md#lost-fence-rows.
  • Tests: test/fence-absent-append-keeps-rows.test.ts (remember, takes add, managed extraction and relink, each followed by a tag, timeline entry, append or round trip; forget, takes remove and an empty section stay withdrawn; a managed phantom redirect into the page keeps the held fact; repo-less and Git), test/repair-lost-fence-rows.test.ts (preview writes nothing, apply restores exactly the listed rows, deliberate removals and empty sections are not listed, a changed page is skipped).

[0.60.635.0] - 2026-11-10

A connected agent can no longer change, or learn about, a page it is not allowed to read. A remote write token that could not read a visibility: private page could still overwrite that page's stored raw data, add links to or from it and remove its links, and extract_entities would link new entity stubs to a source page the caller could not read, or add a timeline line to an existing person page it could not read. Each of those answers also differed from the answer for a page that does not exist, so a caller could tell which hidden pages exist. They now answer exactly as for a missing page and change nothing. ontology_propose refuses an entity page its caller cannot read and stores an observation no wider than that page. get_ingest_log no longer names pages its remote caller cannot read or the host's checkout path, and get_health no longer lists such a page among the most connected. An MCP read that names a source or filter parameter with a slip (source for source_id, type for types) now gets an error naming the right parameter instead of an answer drawn from every source the token can read. gbrain serve --http now applies the /mcp rate limits and body cap when GBRAIN_HTTP_RATE_LIMIT_IP, GBRAIN_HTTP_RATE_LIMIT_TOKEN or GBRAIN_HTTP_MAX_BODY_BYTES is set; before, only the legacy transport read them. Trusted local CLI calls are unchanged.

Itemized changes

  • src/core/persistence/page-visibility.ts: remoteTargetHidden tests the caller's whole read policy (member access, then privacy with its grant and hatch), the policy get_page applies, instead of the member-access part alone; remotePageAccess and remoteReadPolicySql give the same test to a batch of slugs. requireWritablePage (so put_raw_data, add_link on either endpoint and the managed add_timeline_entry preflight) and remove_link use it; a delegated subagent write keeps its own privacy refusal.
  • src/core/ops/extraction.ts, src/core/enrichment-service.ts, src/core/trust/taint.ts: extract_entities treats a source page its remote caller cannot read as missing (no backlink, no trust read from it), skips an entity whose page the caller cannot read (skipped_reason: page_not_found, under the slug its name maps to, nothing written) and carries on with the rest of the batch, follows a retired entity page's redirect only through pages the caller can read, and counts mentions only on pages the caller can read.
  • src/core/ops/chronicle.ts: ontology_propose from a remote caller requires a readable entity page when one exists (a hidden one answers like put_page's create on it) and tightens the observation's visibility to that page, as remember and extract_facts do.
  • src/core/ops/ingest-log.ts: for remote callers, get_ingest_log shows a row only when every page it names (in pages_updated, or in source_ref for rows that name one page there) is a live page the caller can read, and shows source_ref without host paths (a sync row keeps only its commit, an import row reads <local path>).
  • src/core/ops/admin.ts: get_health's most_connected lists, for a remote caller, only pages its get_page would show.
  • src/mcp/validate-params.ts, src/mcp/dispatch.ts: narrowingUnknownParams; a read rejects an unknown scope or filter parameter in warn mode too (mcp.strict_params keeps its default), and source, sourceId, sources and sourceIds suggest source_id / source_ids.
  • src/commands/serve-http-mcp.ts: POST /mcp runs the IP bucket before the token lookup and the token bucket after it, each only while its env var is set, and passes GBRAIN_HTTP_MAX_BODY_BYTES (default the SDK's 4 MiB) to the transport; a value that is not a positive whole number is ignored with a warning. SECURITY.md, docs/guides/memory-boundaries.md, docs/guides/remote-mcp.md state the real defaults.
  • Tests: test/access-hidden-page-writes.test.ts (PGLite and Postgres; repo-less and managed), test/serve-http-mcp-limits.test.ts (a live server), test/validate-params.test.ts.
  • Also in this release (#9060): the chronicle long-conversation tests no longer fail when they run across midnight UTC, and the architecture file index describes the chronicle judged-input digest and the duplicate-events doctor check.

[0.60.634.0] - 2026-11-09

Privacy, trust and access checks written in SQL now give the same answer as their TypeScript versions, so remote callers can't see a page through one path that another path hides. Several rules are evaluated twice: once in TypeScript on a page already read, and once in SQL for list, search, backlink and scan queries. Postgres btrim(x) strips only spaces, while JavaScript's trim() strips every Unicode edge whitespace character, and Postgres lower() follows the database collation. A page marked private through the access or confidentiality field, with other whitespace around the value (a YAML block scalar keeps a trailing newline), was hidden from remote get_page but not from remote list and search. The fast page filters also skipped the full check for type: question pages and pages with a pinned_question key, which the full check treats as private.

Itemized changes

  • src/core/edge-whitespace.ts (new): EDGE_WHITESPACE, exactly what JavaScript's trim() strips, edgeTrimSql(), and an ASCII-only lowercase for TypeScript and SQL. The generated SQL is plain ASCII, so it also runs on a Postgres database whose encoding is not UTF8; there the SQL side trims ASCII whitespace only.
  • src/core/search/private-visibility.ts: privacyAliasSql trims and lowercases like the row-side check; privatePagesFilterFragmentFast and the caller fast path send question and pinned_question pages to the full check.
  • src/core/persistence/git-declared.ts: privacyClass runs the lineage check for question pages too, so a question page built from a private input stays out of Git, as remote reads already hide it.
  • src/core/trust/channel.ts, src/core/trust/backfill.ts: the frontmatter trust caps and the backfill's extractor marker trim the same whitespace as the TypeScript caps; the transcript-harness cap lowercases ASCII on both sides.
  • src/core/access/caller.ts (local stdio member binding), src/core/persistence/unbound-pages.ts (database-only pages), src/core/persistence/effect-journal.ts (the auto_link / mcp.remote_auto_links gate), src/core/repair/private-labels.ts and src/core/persistence/local-edit-adoption.ts use the shared set.
  • Behaviour: a page whose alias value carries such whitespace, or a question or pinned-question page without visibility, is now hidden from remote list, search and backlink reads as well as from get_page. Doctor and repair counts that use these checks may move to match. Nothing rewrites stored pages.
  • Tests: test/edge-whitespace-twins.test.ts (PGLite and Postgres) holds each SQL check equal to its TypeScript side on generated inputs and checks remote list and search; test/trust-claim.test.ts, test/trust-backfill.test.ts and test/remote-graph-freshness.test.ts gain edge-whitespace cases.

[0.60.633.0] - 2026-11-09

Added

  • gbrain skills index build|status|remove and gbrain skills resolve: an engine-free skill resolver that ranks a harness's skill catalog against one user turn and prints a short <relevant_skills> block (at most 5 suggestions, at most 400 tokens), so an agent with hundreds of skills sees the few that fit. Off by default; nothing changes until you build a catalog and call it. Works with no brain, on thin clients and while gbrain serve holds the PGLite lock. Word matching only (the BM25F ranker frozen by the Step 1 offline eval); nothing leaves the machine. Sessions never repeat a suggestion and share a 2,000-token budget (skills.resolver.session_token_budget). JSON contract in docs/schemas/skill-resolver.v1.*.json; library as gbrain/skill-resolver and, for Node, one dependency-free module built from the same source: gbrain skills library writes it to GBRAIN_HOME/lib/gbrain-skill-resolver.mjs and prints its sha256, and each release attaches gbrain-skill-resolver.mjs with a .sha256. Each suggestion carries match_strength (named, name, triggers, routing, description or body) so an adapter can drop weak matches. Guide: docs/guides/skill-resolver.md.
  • gbrain setup claude-code|codex --skill-loading resolve: a per-turn hook lane (GBRAIN_HOOK_LANE=skill-resolver) that adds the resolver's block to every prompt, with the harness's own load line. Setup builds the user catalog first and writes the skills.resolver.* keys and the hook only when it found at least one skill (skill_resolver_setup_incomplete otherwise); --skill-loading off and setup --remove undo all of it. The lane runs in the hook process with no brain and no serve, exits within 50 ms with nothing printed when it cannot deliver, drops a block over skills.resolver.lane_payload_chars whole, and lets exactly one wiring deliver. Claude Code project skills get their own catalog, built in the background. gbrain doctor reports it as skill_resolver. Off by default; with it off, hook output and setup's files are unchanged. A config.json that holds only skills.resolver settings no longer reads as a Postgres brain. Release binaries are now built with the lane's worker as a second entrypoint.

[0.60.632.0] - 2026-11-09

A page write that git could not commit because the file is flagged assume-unchanged or skip-worktree now stays in the write-through pending ledger. gbrain's own page writes and gbrain git commit-push --pending asked git status which pending files were still uncommitted, and status does not show a flagged entry, so the write was left uncommitted, dropped from the ledger (doctor dirty_age never reported it and no later write retried it), and --pending answered nothing_to_commit with exit 0. A flagged file whose bytes differ from HEAD now counts as uncommitted.

Situation Before After
A page write to a flagged file not committed, not in the ledger not committed, kept in the ledger (the write itself fails commit_paths_skipped, as in 0.60.630.0); doctor reports it as dirty_age after 30 minutes
A later page write while a flagged file is in the ledger the entry was settled away the entry is committed beside the new write when git picks it up, and otherwise stays
gbrain git commit-push --pending with a flagged pending file nothing_to_commit, exit 0 commit_paths_skipped naming the file, exit 1; the ledger keeps it
A page write or gbrain git commit-push naming a modified skip-worktree file that is on disk git add refused it: commit_paths_invalid, nothing committed, and a page write dropped the file from the ledger the other paths are committed; the file is reported as skipped (commit_paths_skipped, reason skip_worktree, exit 1 from commit-push) and a page write keeps it in the ledger

Itemized changes

  • src/core/git/commit-push.ts: the pending-ledger check (dirtyPaths, used by the write-through commit, --pending, ledger settling and doctor dirty_age) also counts flagged entries whose bytes differ from HEAD, using the same check as the GST-2116 skipped-path detection (flaggedChanges), where an entry git cannot hash counts as changed instead of failing the check and a flagged gitlink is left out; commitPaths no longer hands git add a named skip-worktree file that is on disk (git refuses it and failed every named path), and reports it as skipped; the write-through commit notes its own flagged files that git left out in the ledger, and ledger settling keeps every entry when the check fails.
  • Docs: docs/guides/shared-checkout-git.md (pending write-through section).
  • Tests: test/pending-ledger-flagged-gst2118.test.ts (real git and the real CLI: a flagged write-through file stays in the ledger and doctor's dirty check sees it, a later write commits an older flagged entry instead of settling it away, --pending on a flagged file fails and keeps the entry, and a flagged entry git cannot hash is kept while a later write still commits, a write to a named skip-worktree file beside another page commits the other page and keeps the file for --pending to report, and an ordinary pending file as the control). Five of its six cases fail without the fix; the control passes.

[0.60.631.0] - 2026-11-09

A long conversation that keeps growing is no longer re-extracted, and re-paid for, on every new message. Automatic event extraction reads a page's title, date, attendees and the first 12,000 characters of its body. On a conversation longer than that, each new message recorded a new revision, and each revision paid for another extraction call on exactly the same text. The model words each event a little differently every time, event pages are named after that wording, and an extraction of a page longer than 12,000 characters never retired the previous one, so every run added a fresh copy of every event. One page could collect dozens of paid calls and hundreds of duplicate events. Now a revision that changes nothing extraction reads (text, title, date, attendees, privacy) makes no call and writes nothing (it records judge_input_unchanged); a window the judge keeps failing on (its answer cut off, unreadable, malformed or over the per-page spend cap) is paid for at most the page's attempt limit across all revisions, retries included, then records judge_input_failed until the text changes, a release changes the extractor, or a backfill asks again (a run a newer revision replaced, or a missing provider or price, never counts toward that limit); and a re-extraction that re-read all the text the page's live events came from (for example after a title change, or the first time a conversation grows past 12,000 characters) replaces those events instead of adding to them. An extraction that dropped proposals over chronicle.max_events_per_page is asked again when that setting or the timezone changes, and one that dropped proposals dated after the page's last day is asked again, once, when that day moves. A page whose opening text was edited still keeps its earlier events, as before. Text past the first 12,000 characters is still not extracted; incremental coverage of the tail is tracked in #9013. Reported by @andreineacsu.

Doctor counts the duplicates earlier releases left. The new chronicle_duplicate_events check (info) reports how many pages hold live events from earlier extraction runs beside their newest run's, and how many such events. It changes nothing; a preview-first repair that retires them is tracked in #9012.

Itemized changes

  • src/core/chronicle/execute.ts: before the daily reservation, a row whose judge input digest equals that of the page's latest decided extracted generation (no ineligible revision or chronicle retirement since) finishes skipped / judge_input_unchanged with no call; an automatic row whose input already used the attempt limit without publishing finishes failed / judge_input_failed (a refused input stays judge_refused), with no call, summing judge_input_failures over every row of the page with that digest, the current revision's retries included (only a judge_truncated, judge_parse_failed, malformed_proposal, budget_exhausted or judge_refused attempt increments it); a truncated generation replaces the previous one only when every live chronicle event of the page is recorded by a ledger row whose judged window the new window starts with. Ended-invite projections and held results never count as judged, and each finish that judged writes the judged input outright, so a held retry clears what a failed attempt recorded (a row a newer revision replaced before its retry keeps it); a generation that dropped future-dated proposals records cutoffInputHash (the digest keyed to the screen's cutoff day).
  • src/core/chronicle/extract-events.ts: chronicleJudgedInput (digest of the judge input with the body cut to JUDGE_BODY_CHARS, the extractor version, the depth page's visibility and source incarnation, the timezone and chronicle.max_events_per_page; the window and its sha256) and cutoffInputHash.
  • Migration v297 chronicle_judge_window: chronicle_page_state.judge_input_hash, judge_window_chars, judge_window_hash (nullable) and judge_input_failures (NOT NULL DEFAULT 0); no backfill, so existing rows start at no judged input and zero counted failures. A row recorded before it has no window, so that page's next revision is judged once more and stays additive.
  • src/core/chronicle/reasons.ts, src/core/error-registry.ts: reason codes judge_input_unchanged and judge_input_failed.
  • src/core/chronicle/ledger-stats.ts, src/commands/doctor/checks/auto-chronicle.ts: readChronicleDuplicateEvents and the chronicle_duplicate_events doctor check.
  • Docs: docs/guides/life-chronicle.md, docs/guides/error-codes.md.
  • Tests: test/chronicle-long-conversation-8972.test.ts with its Postgres arm test/e2e/chronicle-long-conversation-8972-postgres.test.ts; schema, migration and doctor goldens regenerated for the new columns and check.

[0.60.630.0] - 2026-11-09

gbrain git commit-push and scripts/brain-commit-push.sh no longer report success when git left a named path uncommitted. git add stages nothing and still exits 0 for a modified file flagged assume-unchanged or skip-worktree and for a file inside a nested Git repository, so the helper answered "nothing to commit" (or pushed the other paths) with exit 0, and the write sat uncommitted with nothing in any log. Now the other named paths are still committed (and, on a classic checkout, pushed), and the run exits 1 naming each skipped path and its reason: writer_coordinator_required for a page in a managed canonical worktree, which belongs to the persistence owner (save it through put_page), and the new commit_paths_skipped for anything else. The helper's own managed-checkout refusal now prints the operator contract (code, why, fix, next) instead of one line.

Situation Before After
A named file is flagged assume-unchanged (or skip-worktree under a named directory) and modified nothing to commit, exit 0; the edit stays uncommitted commit_paths_skipped (reason assume_unchanged / skip_worktree), exit 1; fix (ask the user): git update-index --no-assume-unchanged -- <paths>, and a separate git update-index --no-skip-worktree -- <paths> (given both flags, git clears only the first)
A named file sits inside an untracked nested repository exit 0 with nothing committed; beside other paths, git_unavailable and nothing committed the other paths are committed, then commit_paths_skipped (reason not_staged), exit 1
The same on a managed canonical worktree, for a page nothing to commit; not pushed, exit 0 writer_coordinator_required, exit 1, naming each page; fix (ask the user): save it through put_page with the managed source (gbrain put --source-id <source> <slug>, source_id over MCP)
The push fails after a path was skipped the push error alone commit_paths_skipped still names each skipped path, with the push error beside it, exit 1 (the CLI; the helper without gbrain exits 4)
A gbrain write-through commit (commitPathsLocked) that git leaves a named path out of the commit returned as saved commit_paths_skipped after the rest is committed
scripts/brain-commit-push.sh without gbrain on the machine nothing to commit, push, exit 0 the push still runs, then Error [commit_paths_skipped] with each path, exit 1
scripts/brain-commit-push.sh on a checkout marked managed one line, writer_coordinator_required: submit managed changes through persistence, exit 1 Error [writer_coordinator_required] with Why:, Fix: (put_page; gbrain git commit-push for files that are not pages) and Next: ask_user, exit 1

A skip-worktree file absent from disk (a sparse checkout), a flagged gitlink (its commit lives in the nested repository) and ignored files under a named directory are not reported; a named ignored file was already refused by git add (commit_paths_invalid). Existing brains get the new helper on the next gbrain sources harden.

Itemized changes

  • src/core/git/commit-push.ts: after the commit step, commitPaths reads which named paths HEAD still lacks (flagged entries whose bytes on disk differ from HEAD, named files git did not index, and status entries with unstaged changes) and leaves an unindexed file out of commit --only, so it no longer fails the other paths. The check runs after the commit's index-restore block, so a failure there never stages a reversal of a commit that already landed. The managed put_page fix carries the managed source in its CLI argv and MCP arguments, so following it cannot write a same-slug page in another source. commitAndPush (classic, managed and remoteless) throws commit_paths_skipped or, for a page in a managed worktree, writer_coordinator_required, after committing and publishing the rest; detail carries the skipped paths, the commit and the pushed head.
  • src/core/brain-repo-durability.ts: the rendered scripts/brain-commit-push.sh prints the operator contract on its managed refusal, and its fallback (no gbrain) runs the same check after its commit (NUL-separated listings, so a non-ASCII or quoted name is read exactly; a symlink is compared by its link text), leaves a file git did not index out of commit --only, and exits 1 with commit_paths_skipped after the push.
  • src/commands/git.ts: when the run fails after the other paths were pushed, the push log still records that push.
  • src/core/error-registry.ts: new code commit_paths_skipped (reasons assume_unchanged, skip_worktree, not_staged).
  • Docs: docs/guides/shared-checkout-git.md (the commit_paths_skipped row, the writer_coordinator_required row, the fallback paragraph), docs/guides/error-codes.md (regenerated).
  • Tests: test/git-commit-push-skipped-paths-gst2116.test.ts (the real CLI and the rendered helper against temp repositories, a bare remote and a managed-root registry: an assume-unchanged page, a nested-repository page beside a committable one, a skip-worktree entry under a named directory, a managed page and a managed non-page file, a remoteless checkout, a named directory inside a nested repository, a run with both flags, the helper's fallback (a non-ASCII path, a nested-repository file beside a committable page, a clean flagged symlink), the helper handing over to gbrain on a managed root, its managed refusal; a named directory beside a sparse entry and an ignored draft still commits with exit 0). Also: the write-through door, a push that fails after a skipped path (CLI and helper), a named assume-unchanged gitlink beside a committed page (the commit stands and no reversal is staged), and the managed fix's source. Sixteen of its eighteen cases fail on master; the two that pass there are the sparse-entry control and the gitlink case, which guards this change's own post-commit path.

[0.60.629.0] - 2026-11-09

gbrain repair file-drift now agrees with sync on a held origin-twin page. A legacy page left as an origin twin with no recorded file (no import record) could be held file_drift by sync while repair file-drift --slug <slug> reported no_drift. The hold's own suggested fix couldn't clear it, and scripts that use the preview as a drift check passed falsely. The preview now reads such a twin against its slug-path file, the same way sync imports it, and lists it with an apply command.

Itemized changes

  • src/core/persistence/drift-conflicts.ts: liveFileDrift reads an adoptable origin twin with no recorded file against its slug-path file. A twin in a read-only mirror source, whose slug path no file can import as, whose slug-path file sync doesn't admit (outside its scope or under sync.exclude), or whose slug-path file another page records as its source path keeps database_only. A slug path several files import as keeps the earlier reading instead of reporting slug_path_ambiguous. A twin under a declared db_only path, which sync holds but no repair can adopt, reads unknown (db_only_slug_path) instead of passing as no_drift, and so does a slug-path file another page names only by file URI (shared_origin), since sync still imports it into the twin.
  • src/core/persistence/reconcile-state.ts: the slug-path adoption check moves into one exported adoptableSlugOrigin, which reconcile and the drift preview now share. Reconcile's behaviour is unchanged.
  • Applying either side to such a twin binds the slug-path file as its source_path and clears origin_twin. Maintenance writes to a twin whose slug-path file differs are now refused as drift, like any drifted page.
  • Tests: test/repair-file-drift-origin-twin-gbra108.test.ts (PGLite, managed source, real git): a legacy page with no import record held by sync, previewed and applied for both --prefer database and --prefer file, after which a re-sync clears the hold; also mirror-read-only, db_only, non-importable slug path, missing file, shared-origin, excluded slug path, URI-only claimant and ambiguous slug path cases.

[0.60.628.0] - 2026-11-09

A page a member-bound credential creates in a federated source is now readable only by that member. A token or OAuth client bound to a member (gbrain auth rescope --member <id>) and granted --sources <id> for a federated source wrote its pages there with no readers. Every credential without a source grant reads default plus every federated source, so another member's no-grant token and an unbound token read those pages over get_page, search, query, recall and list_pages. Such a page is now written with readable_by: [member:<id>], as a page the credential writes to default already was, and POST /ingest refuses the credential. A write source that is not federated keeps the audience its grants give it. Pages written before the upgrade keep their readers, and gbrain access who-can-read <slug> shows who reads one.

Itemized changes

  • src/core/source-resolver.ts: noGrantFloorReads says whether a caller with no source grant reads a source unqualified: default, or a source in the federated set noGrantFederatedScope threads for that floor. It throws when the sources cannot be read.
  • src/core/persistence/page-visibility.ts: defaultSourceMemberWriter keys on noGrantFloorReads instead of the literal default, so a member-bound write to a federated source is stamped too. A source table it cannot read refuses the write instead of publishing the page shared.
  • src/commands/serve-http-webhooks.ts: the POST /ingest refusal names the credential's write source.
  • Docs: docs/guides/multi-source-brains.md, docs/tutorials/company-brain.md (Model C), docs/mcp/ADMIN.md.
  • Tests: test/access-federated-member-writes.test.ts (PGLite, and Postgres with DATABASE_URL): a member-bound token or OAuth client granted a federated source creates a page that another member and an unbound token, reading the federated set as the HTTP transports do, cannot read; with default isolated the floor is default alone and nothing is stamped; a streamed upload over the 24 MiB ceiling is kept to its member too; a write whose source config cannot be read is refused, not published unstamped; a non-federated source stays unstamped and out of unbound reads; a member-bound write cannot leave its granted source; POST /ingest refuses with nothing queued.

[0.60.627.0] - 2026-11-09

A webhook capture can no longer drop a page's readers (GST-2112). POST /ingest wrote through a trusted local path whose only check was that the sender could read the page, so a member-bound token granted a source could replace a page whose readable_by named it with a body that named no readers, and every caller could then read the page. The route and the ingest_capture job now apply the remote put_page readers rule: a body that drops readable_by from a page that names readers, or names readers that leave out the sender, is refused readable_by_invalid (403 at the route, nothing queued; at write time for a job already queued). The job binds its write to the revision it checked, so readers added between the check and the write refuse the write with revision_conflict and the retry refuses readable_by_invalid. A capture already queued from a member-bound credential whose write source is default is refused permission_denied at write time, as the route refuses it before queueing. Bodies that keep the readers, and pages without readers, import as before.

Itemized changes

  • src/commands/serve-http-webhooks.ts: the route applies the remote readers rule to the capture body before anything is queued.
  • src/core/minions/handlers/ingest-capture.ts, src/core/import-file.ts: the job reads the page snapshot before the visibility check, applies the same rule at write time, binds its write to the checked revision (revision_conflict when it moved; a page that did not exist stays a create), and refuses a member-bound default writer permission_denied.
  • src/core/behavior-change-notice.ts: one row for the changed capture behavior.
  • Docs: docs/guides/write-refusals.md.
  • Tests: test/ingest-capture-keeps-readers.test.ts (PGLite and Postgres).

[0.60.626.0] - 2026-11-09

gbrain sources record-origin --remote now locks down the existing checkout too, and on a managed brain it can no longer report success while a reclone still clones with git-lfs on. On a managed brain the record took only the source row lock, so a reclone that had already read the source, or held its worktree locks, could clone with git-lfs downloads and hooks on after the command said the source was recorded. The record now takes the worktree locks and rows a reclone takes (refusing the id source_busy while one runs), and the reclone reads the record under its own locks. --remote --apply now also turns git-lfs downloads and hooks off in the existing checkout's .git/config, because a reclone only happens when a checkout is lost or replaced, and until then doctor read green while the checkout still fetched git-lfs objects and ran hooks. The git-lfs check now reads a committed .gitattributes under a directory whose name holds a tab (it was cut at the tab and skipped), ignores commented-out filter=lfs lines, says "declares a git-lfs filter" when a rule only declares one, and records a clone of an empty repository instead of telling the operator to restore an intact checkout. A failed write is reported for its own id instead of ending the run without saying which ids were recorded.

Itemized changes

  • src/core/url-source-origin.ts: on a managed brain each apply runs inside withTopologyLocks and lockTopologyRows (a held lock or a running refresh refuses source_busy); --remote --apply writes the clone-time lock-down (lfs.fetchexclude=*, core.hooksPath=/dev/null, lfs.locksverify=false) into the checkout's .git/config with git config --file before the row, into the repository's shared config (--git-common-dir, so a linked worktree is covered), then reads back the values Git uses there and refuses write_failed, naming the key, when a worktree-scoped or included setting overrides one; a source already recorded as remote whose checkout lacks the lock-down gets it (recorded, with no lfs_tracked refusal even when its files now use git-lfs), and one that has it stays already_recorded; each key's existing values are removed before its one value is written, so one that holds several values gets the one clone-time value, and like Git, it takes config.lock beside the config's real path, copies the config into it (mode kept) and renames it over the config in one step, so a failure part way leaves it as it was; a throw is reported per id as write_failed, saying whether the checkout already got the lock-down, and the batch continues. The committed-attributes read splits each ls-tree -z record at its first tab, reads only blobs, and ignores comment lines; a declared-only filter refuses lfs_tracked with its own wording. An unborn HEAD with an empty index counts as no committed attributes. The top-level comparison realpath-normalizes both sides. lfs_unknown reasons name the cause from Git's own error: checkout missing, not its own repository's top level (an enclosing repository, or none), owned by another user (safe.directory), or a Git read failed (for example a corrupt .git).
  • src/core/hardened-git.ts: rev-parse --git-common-dir joins the read-only allow-list, so the lock-down finds a linked worktree's shared config.
  • src/core/persistence/topology-clone.ts: a managed clone or reclone takes the source config it clones with from its locked admission read, not from the read before it took its locks.
  • src/core/behavior-change-notice.ts: one row for the changed record-origin behavior.
  • Docs: docs/guides/repair.md#url-source-origin (checkout lock-down and why, source_busy, write_failed, lfs_tracked and lfs_unknown causes), the key-files row in docs/architecture/key-files/files-and-sync-2.md, docs/architecture/canonical-writers.tsv.
  • Tests: test/url-source-origin.test.ts (PGLite and Postgres): a record attempted while a reclone holds its locks is refused and writes nothing; a reclone that read the source before a record landed clones with the recorded settings; the existing checkout gets the lock-down on apply and not on preview; a tab-named directory's .gitattributes is read; commented rules are ignored and a declared-only rule is worded as declared; an empty-repository clone records; each lfs_unknown cause is named; a mid-batch write failure is reported for its id while the others record; a record during an active refresh is refused source_busy; a lock-down that fails part way is reported as a partial change; a worktree-scoped override is refused write_failed naming the key; a linked worktree gets the lock-down where Git reads it; a source recorded as remote before the lock-down existed gets it on the next apply; a corrupt .git and a safe.directory refusal each get their own lfs_unknown reason; a source recorded as remote whose files now use git-lfs still gets the lock-down, and a key that already holds several values gets the one clone-time value.

[0.60.625.0] - 2026-11-09

gbrain upgrade no longer consults the ClawHub registry; a large Google address book, and a large source's removal, can now finish. Advisory: an unofficial package on the ClawHub registry shares the name gbrain. gbrain upgrade, and autopilot's unattended self-upgrade, treated any install it could not otherwise classify (for example a release binary run under its published file name, gbrain-linux-x64) as a ClawHub install whenever the clawhub CLI was on the host, and ran clawhub update gbrain, which installs whatever the registry resolves for that name from any publisher. Upgrade now never runs it: a compiled binary is recognised whatever its file name, and an install it cannot classify prints the manual steps. When Google's People sync token expires, the contacts re-list now checkpoints after every page and resumes there, so a large address book no longer restarts from page one in every sync job and Calendar and Gmail sync again. A managed gbrain sources remove or purge of a large source now deletes its pages in bounded batches under the 5-second statement bound instead of one cascading statement that could never finish, and a rerun resumes. The secret scanner also catches keys right after terminal color codes, \xNN and \UXXXXXXXX escapes and quoted-printable =3D, and redacts a bearer token containing an escaped slash in full.

Itemized changes

  • src/commands/upgrade.ts: detectInstallMethod drops the clawhub method and its clawhub --version probe; a process whose argv[1] is Bun's compiled entrypoint (/$bunfs/, B:\~BUN\) is binary; the clawhub upgrade branch and hint are gone. src/core/self-upgrade.ts: canSelfUpdate no longer accepts clawhub. src/commands/check-update.ts: no clawhub update hint (#8464).
  • src/core/google/people-calendar-sweep.ts, google-clients.ts: a full contacts re-list goes page by page (listConnectionsPage), saves contacts_relist (next page token and field mask) after each page lands through the same state store as Gmail's checkpoints (no git, so repo-less brains work), resumes there, restarts once from page 1 when Google rejects a saved page token (400/410), and lists at most 10 pages per run so Calendar and Gmail still run (#8906).
  • src/core/persistence/source-lifecycle.ts: managed remove/purge deletes pages 500 at a time, each batch its own bounded transaction that first rechecks the final transaction's preconditions (including the CLI writer registration and the receipt quotas, assertTopologyChangeCapacity) and halves its size when it hits the statement bound; a refusal after committed batches says how many pages that run deleted. topology-transaction.ts: the write_pending hint says a retry resumes (#7702, #8905).
  • src/core/secret-scan.ts: the escape-boundary twins also fire after ESC[…m (raw or written \u001b[, \x1b[, \033[), \xNN, \UXXXXXXXX and =3D; the bearer value accepts \/ (#8266).
  • Tests: test/upgrade.serial.test.ts, test/self-upgrade.test.ts (a renamed binary or an unclassified install with a working clawhub CLI never runs it); test/contacts-relist-checkpoint-8906.slow.test.ts, test/contacts-relist-checkpoint-managed-8906.test.ts; test/source-retire-batched.test.ts, test/e2e/source-retire-batched-postgres.test.ts (60,000 pages); test/secret-scan-encoded-boundary-8266.test.ts.

Reported and first fixed by @javieraldape (#8910, #8764) and @Masashi-Ono0611 (#7704, #8907, #8908).

[0.60.624.0] - 2026-11-09

gbrain doctor now ends on its --budget-ms deadline when a local probe hangs, and the cheap incident checks keep their share of the budget. The home_dir_in_worktree entry, which runs before the connection checks, ran npm root -g, which -a gbrain and crontab -l synchronously, with 5-second limits on the first and last and none on which. While one of them hung, nothing else in the process could run, the deadline included. So a 2500 ms run took about 5100 ms, and every later check, links_extraction_lag and persistence_capacity included, came back unknown with elapsed_ms: 0. Those probes now run asynchronously and are killed with SIGKILL at the entry's deadline (5 s at most, less the 250 ms report reserve), so none of them outlives the run. The entries before the connection checks may use at most half of the time left when the check run starts, so the ops sections and the cheap incident checks keep the other half. --only, --fast and a run with no database connection still give those entries the whole budget. This was the flaky #6803 doctor deadlines (postgres) E2E case (#8979). It failed on every backend-matrix job that ran on the 20261026.1 runner image, where a probe ran into its 5 s limit.

Itemized changes

  • src/commands/doctor/checks/local-audits.ts: localProbe runs a local command asynchronously, bounded by scanDeadline (5 s, clipped to the entry deadline read when the entry starts). At that bound it kills the probe with SIGKILL, on POSIX its whole process group so a wrapper's children die too. A probe the entry deadline cuts short marks the entry expired, so its checks report unknown instead of absent. npm_squat (npm root -g, which -a gbrain) and cron_gbrain_home (crontab -l) use it, and a probe that fails or times out reads as absent, as before.
  • src/commands/doctor/checks/cron-gbrain-home.ts: the synchronous readCrontab is gone.
  • src/commands/doctor/registry.ts: with an engine and without --fast or --only, an entry before the connection lane runs to the overall deadline less half the time left when the check run starts.
  • Tests: test/doctor-entry-deadline-6803.test.ts. One test puts stand-ins for npm, crontab and which on PATH. They ignore SIGTERM and wait on a 6 s child of their own, yet a 1000 ms run finishes within its budget, npm_squat reports unknown, and neither the wrappers nor their children are still running afterwards. Another checks that a probe that answers in time still feeds cron_gbrain_home. In the second (PGLite and Postgres), an entry before the connection lane that outlives a 2000 ms budget still leaves links_extraction_lag and persistence_capacity reporting. A third checks that under --fast such an entry keeps the whole budget. The first two fail on the previous code.

[0.60.623.0] - 2026-11-09

The brain owner's own agents now read the whole personal brain, and pages stamped access: private by default can be cleared in reviewed batches. On a single-user brain, writer templates stamped access: private on ordinary research and ops pages, and every remote caller got not_found for them, the owner's own agent included. gbrain access owner set <member> now names the owner (config access.owner_member, database state, so repo-less brains work the same). Every token, OAuth client or stdio server bound to that member reads visibility: private pages, privacy-alias pages, derived and pinned-question pages and private facts through get_page, search, query, recall and the other page reads, within its source grant. readable_by and folder rules still bind it, a shared-room audience gets no relaxation, and other members, unbound and public callers read exactly as before. gbrain repair private-labels previews the existing labels with counts per category, keeps real restrictions, and clears the over-stamped ones in reviewed, recorded, reversible batches. The bulk-ingestion page template no longer writes an access: line, and docs/guides/access-designations.md defines every designation.

Itemized changes

  • src/core/access/owner.ts (new), src/commands/access.ts: gbrain access owner [set <member>|clear]; the member must be enabled; gbrain access check --token <name> reports owner.
  • src/core/access/caller.ts, src/core/search/private-visibility.ts: CallerAccess.owner, set when the credential's live member is the owner and no audience other than the owner is asserted; the owner reads with privacy relaxed (member access stays mandatory); the request's config snapshot carries the owner key; _meta.access.relaxation gains owner.
  • src/core/ops/facts.ts: remote recall returns private facts to the owner.
  • src/cli/as-preview.ts: --as <owner> previews the owner's view.
  • src/core/repair/private-labels.ts (new), src/commands/repair-private-labels.ts (new): gbrain repair private-labels, explicit-only and preview-bound; classes alias, alias-and-visibility, git-publish (named only) and restore; a ledger under the gbrain home records each change.
  • Skills and docs: skills/conventions/page-privacy.md (new), skills/bulk-ingestion (template), skills/brain-ops, skills/_AGENT_README.md, skills/frontmatter-guard; docs/guides/access-designations.md (new), linked from multi-source brains, company brain, page anatomy and brains-and-sources; docs/protocol/MCP_META_CHANNELS.md.
  • src/core/persistence/page-visibility.ts, src/core/persistence/page-prepare.ts, src/core/persistence/upload-streamed.ts: a new default page the enabled owner's credential creates through put_page or a streamed upload is not kept to the owner (no readable_by: [member:<id>]); a rewrite keeps its stored readers, and POST /ingest still refuses the owner.
  • src/core/persistence/page-prepare.ts: on a git-canonical source, a page that was private from its first write (no file ever recorded) writes its file for the first time when it stops being private, so --class git-publish and an ordinary edit no longer refuse it as a removed canonical file.
  • src/core/repair/private-labels.ts: restore plans from each page's latest clear and its current labels, so a restore that was accepted but never landed stays restorable; a ledger line that fails to save after a write reports the values to undo by hand; a git-publish clear is called one-way, since Git history keeps the page.
  • Tests: test/access-owner-reads.test.ts, test/access-owner-default-writes.test.ts, test/repair-private-labels.test.ts.

[0.60.622.0] - 2026-11-09

think keeps an entity's trajectory when the question names a month. A question such as "What was the MRR of acme-example in April 2026?" bounds think's evidence to April (PR #5086), and that window used to cost the trajectory block twice over: a metric whose readings fall before and after April (March and June) injected nothing, although the March reading was the value in force all April; and an entity page last edited outside April was dropped from the gathered pages, so when the question's noun phrase ("the acme-example team") did not resolve on its own, no entity was found. Each metric's latest reading before the window now joins the readings inside it, unless its valid_until ended by the time the window opens, and pages the window drops still name the entities whose trajectory is read; they stay out of the evidence. BrainBench Cat 25's hermetic gate goes from 4 of 6 probes (FAIL) to 6 of 6 (PASS).

A purge no longer waits behind a refresh. With persistence.refresh_barrier on, a write that arrives during a gbrain sources refresh is held as deferred. When that write would publish a staged body holding the claim being purged, gbrain forget <id> --purge refused purge_blocked_pending_recovery ("is publishing") until the refresh released or the barrier expired, up to 15 minutes, so leaked content stayed in the brain that long. A deferred write has not been claimed, so the purge now cancels it with purged_content, as it already did a queued one; a running or recovering write still refuses.

Itemized changes

  • src/core/think/temporal-window.ts: trajectoryPointsInWindow keeps the points inside the window plus each metric's latest point before its start when that point is still valid at the start (events are not carried); filterPagesToWindow also returns the dropped pages.
  • src/core/think/gather.ts: ThinkGatherResult.outsideWindowSlugs lists the pages the window dropped.
  • src/core/engine-sql/facts.ts, src/core/engine.ts: findTrajectory points carry valid_until (the MCP find_trajectory wire shape is unchanged).
  • src/core/think/index.ts: the trajectory step filters through trajectoryPointsInWindow, and its candidate entities come from the gathered pages plus the pages the window dropped.
  • src/core/persistence/intent-redaction.ts: cancelRequestsHoldingStagings treats a deferred request like a queued one (GST-1602).
  • Tests: test/refresh-barrier-purge-gst1602.test.ts (PGLite and Postgres: a put_page deferred behind a real paused refresh, publishing a staged body that holds the claim; the purge commits, the write ends cancelled/purged_content, the staging is aborted and its parts deleted, and after the refresh releases the write stays cancelled with its counters released; it fails purge_blocked_pending_recovery with the fix reverted); test/think-trajectory-window.test.ts (a month with no reading of its own gets the March reading and not January's or June's; an entity page dated outside the month still routes its trajectory). Each test fails with its own fix reverted.

[0.60.621.0] - 2026-11-09

A page a member-bound credential creates in the default source is now readable only by that member. A token bound with gbrain auth rescope --member <id> and no --sources writes to default, the source every credential without a source grant reads, and so does a token granted --sources default and an OAuth client left at its registration default. Its put_page (and every page write through the same path) used to create an ordinary page there, which another member and an unbound token read through get_page, search, query, recall and list_pages; its remember facts were already kept on the member's own page. A page such a credential creates is now written with readable_by: [member:<id>], in the database row on a repo-less brain and in the file on a repo brain; rewriting it without readable_by keeps the stored readers, and the member widens it by naming readers. An edit to an existing page that names no readers leaves it shared, and POST /ingest, which cannot keep a page to a member, refuses such a credential with permission_denied. Member-bound credentials whose write source is another source (the documented --sources private-<member>,default shape) and unbound credentials write exactly as before, apart from slugs a member now creates. Pages derived from such a page (extracted atoms, chronicle events, derived_from pages, and a synthesized concept that uses it) follow the rule for pages with readers and are readable by no remote caller, the writer included. A page that teammates should read, such as a shared people/ or companies/ entity page or a page in a folder a rule already scopes to a group, names its readers when it is created; otherwise the first member to create a slug keeps it, and every other remote credential writing that slug (another member, an unbound token, a shared agent) gets page_not_found. The operator widens such a page from the brain host by adding readable_by to it (the local CLI is not member-scoped). A member-bound put_page over a page deleted within the last 72 hours is a create, so it is kept to the member too. Pages written before this release keep whatever readers they have: gbrain access who-can-read <slug> shows who reads one.

Itemized changes

  • src/core/persistence/page-visibility.ts: defaultSourceMemberWriter reads, live from the stored principal, whether a remote write to default comes from a token or OAuth client bound to an enabled member (never for a local, unbound, attested or audience write); defaultSourceWriterReaders gives a page such a writer sends without readable_by the stored page's readers, or the writer alone when the write creates the page.
  • src/core/persistence/page-prepare.ts: page preparation applies it to the frontmatter before the content hash, the rendered file and the stored row are made (restores and reverts excluded).
  • src/core/persistence/upload-streamed.ts: a page uploaded over the single-string ceiling gets the same readers.
  • The binding is read straight from the token or client row, so a lookup that fails refuses the write instead of publishing the page shared, and a write bound to a disabled member is refused (readable_by_invalid).
  • src/commands/serve-http-webhooks.ts: POST /ingest refuses a member-bound credential whose write source is default before anything is queued.
  • Docs: docs/guides/multi-source-brains.md, docs/tutorials/company-brain.md (Model C), docs/mcp/ADMIN.md.
  • Tests: test/access-ungranted-member-writes.test.ts (repo-less PGLite and Postgres, plus a repo brain: tokens with no grant and with --sources default and an OAuth client at its registration default; another member and an unbound token read nothing of a new page over get_page, search, query, recall and list_pages; the writer reads, rewrites and widens it; a shared page stays shared; remember unchanged; the documented shape unchanged and unstamped; POST /ingest refused with nothing queued).

[0.60.620.0] - 2026-11-09

A diverged managed checkout now publishes its local commits on refresh by default when Git durability is on, after a secret scan of those commits. Commits a script made with plain git commit in a managed checkout used to strand there whenever the remote moved: every gbrain sources refresh refused refresh_diverged and asked the user, unless sources.refresh_publish_diverged was set to true. With the setting unset, a worktree whose Git durability is on (gbrain sources writer git-durability <source> --enable, or the legacy hook) now publishes them itself: the local commits merged with the upstream in memory, never a rebase or force-push, then the fast-forward. The publish child scans, under its publish lock and for the exact commits it pushes, what they add: commit messages, added lines (including a value a later commit removed) and merge-commit conflict resolutions. A value that looks like a secret, added text too large to scan, or a commit that changes .gbrain-scan-allow or a .gitattributes refuses the new publish_secret_found naming the file, line, rule and commit, never the value, with nothing pushed; commits the scan cannot read in full refuse publish_scan_unreadable instead. The rewritten-upstream refusal and the --expect-remote bound are unchanged. false turns it off; true still publishes with durability off.

Itemized changes

  • src/core/git/range-secret-scan.ts (new): reads the range with plumbing only (rev-list, git cat-file commit per message, one git diff-tree --stdin -p --root --cc streamed to a private temporary file and read in chunks), so no log, diff or encoding setting, no replace ref and no system attributes file changes what is scanned, and memory stays bounded however large the binaries in the range are. It scans each commit's message (also as raw bytes when it declares an encoding other than UTF-8), each file's added lines and each merge's lines no parent has with scanText and the workspace allowlist (.gbrain-scan-allow). Binary is git's own rule (a NUL in the first 8000 bytes) with attributes read from the empty tree; a file git still calls binary is checked by its own bytes when its old side was not empty or an attribute could apply, and scanned whole when it is text. Added text over the scan cap is reported unscanned, and output it does not recognise, a git failure or its 300 s budget make the result unreadable.
  • src/core/persistence/worktree-refresh.ts: an unset sources.refresh_publish_diverged publishes only when gitDurabilityPolicy says durable (otherwise refresh_diverged names the durability and the --enable command); the publish child runs with --scan-secrets.
  • src/core/git/commit-push.ts, src/commands/git.ts: scanSecrets / --scan-secrets (push-only) scans trackingRef..target inside the publish lock, right before the push, and refuses publish_secret_found (new error code) on a finding or an over-cap file, and publish_scan_unreadable (new error code: no secret was reported, check the checkout with git fsck) when the scan could not read the range in full; the secret refusal's fix lists the commits and asks the user to remove the value or allowlist a false positive, never to push anyway.
  • Docs: docs/guides/shared-checkout-git.md (publish_secret_found row), docs/guides/write-refusals.md (refresh_diverged), docs/guides/live-sync.md, docs/guides/shared-checkout-git.md, docs/guides/troubleshooting.md, docs/architecture/canonical-writers.md.
  • Tests: test/guarded-publish.test.ts (a secret added then removed, one in a commit message and one in a merge resolution each refuse with nothing pushed and no value printed; an allowlisted fingerprint and a binary file are skipped; a clean range publishes; a NUL past a text file's first 8000 bytes, a repository log encoding log.showRoot=false with a merged root commit and a message declaring UTF-16 never hide a secret; a range the scan cannot read, or a message encoding it cannot decode, refuses publish_scan_unreadable; every scan git call runs with the system attributes file off); test/refresh-publish-diverged-7686.test.ts (unset with durability off refuses naming --enable, unset with durability on publishes, a secret in the local commits refuses publish_secret_found with nothing pushed or fenced).

[0.60.619.0] - 2026-11-08

gbrain git commit-push on a managed checkout no longer reports commits the remote lacks as already pushed. On a managed canonical worktree, commit-push commits only the named paths and leaves publishing to the persistence owner. When the commit already existed (a script's bare commit, or one an earlier call made), it printed nothing to commit; origin/main at -, exited 0 and logged [push] ok-already-on-remote, which the push-log reader behind put_page's lastPushStatus and doctor counts as a push that landed, while the branch sat ahead of (or diverged from) its upstream. It now says nothing was pushed, reports how many commits the upstream lacks as last fetched (unpublished in --json), names the command that publishes them, and writes no push-log line. Nothing is pushed or moved; exit codes are unchanged.

Itemized changes

  • src/core/git/commit-push.ts: commitManagedPaths marks its result managed: true, counts unpublished (@{upstream}..HEAD; null when no fetched upstream can be read) and names the branch's configured upstream (null when it names none); both are in --json.
  • src/commands/git.ts: a managed result reports committed <sha> or nothing to commit, then not pushed: with the unpublished count against the configured upstream (or "is on as last fetched", "may not be on " when its fetched ref can't be read, or "has no upstream"), and returns before the push log is written.
  • Docs: docs/guides/shared-checkout-git.md (managed worktree paragraph and the nothing_to_commit / committed_local rows).
  • Tests: test/git-commit-push-managed-unpublished.test.ts (the real CLI against a bare remote, a sibling clone and a managed-root registry: an existing commit with a moved remote, a fresh commit, a checkout on its upstream, a branch with no upstream, an upstream whose fetched ref is missing, an upstream with a different branch name; no push-log line in any case).

[0.60.618.0] - 2026-11-08

A refresh no longer loses its tick when gbrain's own writes land while it drains. gbrain sources refresh waits for writes admitted before its fence, and finishing them commits on the checkout and pushes, so HEAD moved past the commit the refresh started from and the run refused refresh_source_changed, twice when its same-run retry met the next writes. When every commit since that start is in gbrain's commit ledger and the upstream branch the refresh started with holds them, the refresh now takes HEAD as its start under the fence and fast-forwards to that upstream; any other HEAD move, one commit another agent made among them or an upstream changed under the fence included, still refuses as before.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: replanOnOwnCommits, under the commit lock in fenceAndMerge: HEAD descends from old_head, every sha in old_head..HEAD is in the commit ledger (ownCommits), and @{upstream} is still the upstream_ref the precheck recorded and holds HEAD and the precheck's target; one UPDATE of the fenced row with no merge started sets old_head to HEAD and target_head to the upstream. Nothing is fetched under the lock.
  • spec/tla/refresh-barrier/RefreshBarrier.tla: RefMerge models the re-plan (OwnSince for the ledger check, foreign commits as a set fgn), new invariant ReplanOnlyOwnCommits in every safety config, guard mutant replan-over-foreign-commit (spec and code arms).
  • Docs: docs/designs/REFRESH_WRITE_BARRIER.md, the refresh_source_changed row in docs/guides/write-refusals.md, spec/tla/README.md.
  • Tests: test/refresh-own-commits-replan-8828.test.ts on PGLite and Postgres (own published commits fast-forward; a ledgered commit not pushed yet still refuses; one foreign commit among gbrain's own still refuses and merges nothing; own published commits under an upstream changed under the fence still refuse and merge nothing).

[0.60.617.0] - 2026-11-08

Older URL sources can get the same reclone settings as new ones remote callers add. URL sources added before gbrain recorded who added them carry no record, and nothing stored tells a remote caller's add from the local operator's, so their reclones used the clone settings of a local add. The new doctor check url_source_origin lists them and asks the user who added each one. gbrain sources record-origin <id>... --remote|--local records that answer, as a preview until --apply. A URL source the local operator adds from now on records that it was a local add, so doctor never asks about it. With --remote, every later reclone of the source clones with git-lfs downloads and hooks off, like a source a remote caller adds today. A checkout that tracks git-lfs files is refused, because changing its settings would make the next reclone fail to match its canonical content. Nothing changes on upgrade.

gbrain frontmatter install-hook reads a clone's hooks-off setting from that clone's own config only, and its skip message names only that setting. The check that leaves a clone with core.hooksPath=/dev/null alone now runs with GIT_DIR, GIT_WORK_TREE and the GIT_CONFIG* variables scrubbed, so a variable inherited from the caller's shell cannot point the read at another repository or config file. The skipped — … line says the clone's own config turns hooks off, without naming a cause it did not check.

Itemized changes

  • src/commands/frontmatter-install-hook.ts: hooksOffInClone runs git config --local --get core.hooksPath with gitChildEnv() minus GIT_CONFIG_GLOBAL, GIT_CONFIG_SYSTEM and GIT_CONFIG_NOSYSTEM. hooksPathBlocker returns hooks are off in this clone's own config (core.hooksPath=/dev/null), and gbrain leaves them off. Other git reads in the file keep the caller's environment.
  • src/core/url-source-origin.ts, src/commands/sources-record-origin.ts, src/commands/doctor/checks/url-source-origin.ts, src/commands/doctor/wave-checks.ts: doctor url_source_origin (host-only, read-only) lists active URL sources in the default clone directory with no added_by_remote_caller record, including a clone there from before the ownership marker; a clone in another directory is a local add and is not listed. Only an explicit true reads as a remote caller's add. gbrain sources record-origin is trusted local CLI only (trusted_local_only for any other caller) and writes the record in one transaction that re-reads the row. --remote refuses lfs_tracked (a tracked path with filter=lfs, or a committed .gitattributes naming filter=lfs, so clone-local attributes cannot hide git-lfs files) and inspects the checkout again under the row lock before it writes, and refuses lfs_unknown (a checkout that is missing or Git cannot list); --local never replaces a remote record. Without git-lfs files the tracked bytes of a reclone are unchanged, so the stored writer manifest is not rewritten.
  • src/core/sources-ops.ts, src/core/persistence/managed-sources.ts: a URL add records added_by_remote_caller: false when the local operator adds it, as both add paths already recorded true for a remote caller.
  • Tests: test/frontmatter-install-hook.test.ts; test/sources-remote-lfs.test.ts; test/url-source-origin.test.ts on PGLite and Postgres; the doctor goldens gain the url_source_origin row.

[0.60.616.0] - 2026-11-08

Opt-in local-edit adoption (persistence.adopt_local_edits, off by default) no longer leaves an adoption commit unrecorded when the write it was made for ends early, names that commit on a failed adoption's receipt, and refuses during a bisect or an uncommitted squash merge. An adoption whose gbrain: adopt local edit commit landed before its publication failed used to end write_claim_lost (or be cancelled) once its caller's claim ended, so the commit was never recorded and later writes to the page refused file_database_drift until a sync. It is now published on its replay with its receipt naming the landed commit. An adoption that ends without committing after its commit landed names that commit too (outcome.persistence.git_commit, landed_before_row: true), so gbrain write-request and gbrain sources writer failures tie the commit to the adoption. Adoption now also keeps its refusal while a git bisect or an uncommitted git merge --squash is in progress, and when Git names the checkout's git dir with anything but one absolute, existing directory in valid UTF-8.

Itemized changes

  • src/core/persistence/local-edit-adoption.ts, file-repair.ts: landedAdoptionCommit finds this adoption's commit at HEAD (subject, gbrain-adopted: trailer, commit-ledger line, the pre-adoption blob in its parent and the adopted blob at HEAD). The ledger's commit line is written after git commit and can be missing (a busy ledger lock, or a crash in between), so the intent line written before the commit also proves it when its head_before is the commit's parent and its message digest is the commit's. Preparation and the publication's recheck run the caller check only until that commit lands, and settleAdoption leaves a requeued adoption whose commit landed to the consumer instead of cancelling it. Fixes #8824.
  • src/core/persistence/coordinator.ts: recoverPublication stamps a landed adoption commit as outcome.persistence = {git_state: committed, git_commit, landed_before_row: true} on the requeued or failed receipt, the shape a strict source's landed door commit already uses; completeWrite keeps it on any later end without committing. Fixes #8767.
  • src/core/persistence/pending-import.ts: operationMarkerPaths adds BISECT_LOG and SQUASH_MSG, reads git rev-parse --absolute-git-dir as bytes and answers null on invalid UTF-8, or unless the one line (only its trailing newline stripped) is an absolute path that exists as a directory. GIT_OPERATION_MARKERS and assertNoOperationInProgress (src/core/git/commit-push.ts) are unchanged. A git cherry-pick -n leaves no marker; the adoption and its caller commit only their own path (git commit --only), so the picked path stays staged. Fixes #8826.
  • Tests: test/adopt-local-edit.test.ts on PGLite and Postgres (a replay after the caller was cancelled before recovery requeued it; a terminal failure right after the adoption commit; a replay whose commit line never reached the ledger, and intent lines that prove nothing without their commit, on another parent or with another message; a real cherry-pick -n of another path; a real bisect and a squash merge stopped on a conflict in another path), and test/operation-marker-paths.test.ts (a non-UTF-8 git dir reached through a clean symlink mid-cherry-pick, real squash merge and bisect, and a shim git answering with CRLF, a trailing space, a relative path, a missing directory, a file or no newline).

[0.60.615.0] - 2026-11-08

A refresh killed after git refused its fast-forward no longer leaves the checkout fenced until a person acts. When another agent commits on the start commit with plain git right before a refresh merges, git refuses the fast-forward and the refresh aborts, keeping that commit. If the refresh process died between the refused merge and that abort, recovery used to mark the refresh recovery_required, so writes to the checkout stayed fenced (or deferred, with the refresh write barrier on) until someone reset the checkout. Recovery (the owner's scan or gbrain sources refresh --resume) now aborts it the same way, with the fence lifted and the commit kept, and the next refresh completes. It tells this case from a merge that ran and was then reset by reading the reflog of HEAD and its branch since the merge started; without a reflog, or for a refresh an earlier version started, it still marks recovery_required.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: the update that marks outcome.merge_started also records merge_started_at (epoch ms, the refresher's clock). recoverOne aborts a fenced row refresh_interrupted when its merge started, HEAD descends from the start commit without the target, no member synced off HEAD, and the reflogs of HEAD and its branch (reachedTargetSince) show no move to the target since merge_started_at less 5 s. It reads at most the last 1 MiB of each reflog file and needs an unbroken old-to-new chain back past that time or to the log's first entry; a missing, empty or unreadable reflog, a gap in the chain or a tail that ends inside the window keeps recovery_required, and a reflog read git cannot finish leaves the row fenced for the next scan. Refresh git calls no longer inherit GIT_COMMITTER_DATE. --abandon lifts such a row the same way instead of marking it recovery_required.
  • spec/tla/refresh-barrier: RefusedAbortable models the new recovery rule, FfMerge aborts a refused merge as the code has since #8765, and RecoveryOnlyByPolicy now allows recovery_required only for a merge that ran. The guard mutant resume-marks-refused-merge-recovery reverts the rule.
  • Tests: test/persistence-worktree-refresh-restart.test.ts SIGKILLs a real refresh process right after git refused its merge, on PGLite and Postgres: recovery aborts it, a write deferred behind it commits and the next refresh completes; with the reflog deleted, with no merge_started_at, after a merge that reached the target and was reset (also with those reflog entries expired), or with a reflog tail that holds nothing from before the merge, it stays recovery_required; a 32 MiB reflog still aborts, and a reflog read git cannot finish leaves the row fenced.

[0.60.614.0] - 2026-11-08

With the refresh write barrier on, a write held behind a refresh that ends without completing is now decided in the run that ends it, instead of waiting for a later refresh. A refresh's fence keeps managed sync from importing another writer's commit while it runs, and only the refresh's own sync imports in its place. When the refresh was aborted, abandoned or rolled back after its process died, nothing imported that commit, so a write to that file stayed pending until some later refresh or sync, indefinitely while the checkout stayed diverged (up to the 15-minute wait limit). Now, while a checkout's last refresh ended aborted, every gbrain sources refresh run on the owner host ends by running the managed --no-pull sync of each source behind the checkout HEAD, including a run that refuses refresh_diverged. The waiting write is then prepared again against the checkout as it is. A put_page with expected_revision ends revision_conflict, a blind one lands and names what it replaced in supersedes, and a remember or add_timeline_entry applies on top. The barrier is still off by default (persistence.refresh_barrier); with it off, the only change is that such a run also syncs sources behind HEAD.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: refreshWorktree ends every run (returned or refused, not --dry-run) with syncAfterUnfinishedRefresh. When the worktree's newest refresh row is aborted, it runs the bounded member sync for each member whose last_commit is not HEAD and that the run did not already sync, after the run's locks are released. The run's synced lists those syncs once each; a member left behind HEAD is named with its gbrain sync --source <id> --no-pull command in next (and sync_blocked when blocked) or in a refusal's message. A refusal of the run's own arguments, refresh_in_progress and an unexpected failure skip it.
  • spec/tla/refresh-barrier: new RefreshBarrierAbortLiveness.cfg checks liveness with no scheduled sync and no forced expiry (OutsideHelp = FALSE). OwedSyncBegin models the new sync, and the guard mutant abort-skips-owed-sync removes it (DeferredLeaves fails; the code edit fails the new tests). scripts/spec/mutants.ts names the temporal property that fails by checking each one on its own.
  • Tests: test/refresh-barrier-held-after-abort.test.ts (PGLite, and Postgres when DATABASE_URL is set). It covers a refresh aborted by another writer's commit followed by a diverged retry, a refresh killed at fenced and then abandoned by --resume, and an owner-scan abort followed by a diverged tick, plus a run that already synced the members (no second sync, each member listed once), a member left behind HEAD named with its resume command, and an argument refusal that runs no owed sync.

[0.60.613.0] - 2026-11-08

A POST /ingest capture that runs again after its write landed no longer reverts a newer edit to the page. A worker that died or lost its lock after the write committed, but before the job was marked complete, handed the job to another attempt, which wrote the captured body over any edit made in between. A client replaying the same POST after gbrain jobs prune removed the finished job got a fresh job that did the same. Each capture is now recorded under its job's idempotency key in the same transaction as the page write, and any later run of that key writes nothing and returns the first run's result. The record lasts 90 days, even after the job row is pruned. The upgrade records the captures whose completed jobs are still in the queue, so replaying one of those after its job row is pruned writes nothing either. Separate captures to one page still land in order, because each has its own key.

Itemized changes

  • src/core/minions/handlers/ingest-capture.ts + src/core/ingestion/applied-captures.ts: a keyed capture checks ingest_applied_captures (new migration v296, which also records the completed ingest_capture jobs already in minion_jobs) before writing. On an unmanaged brain the row is inserted inside the page write's transaction, and a run that finds the row already there rolls its write back. On a managed brain each run sends the door a new request_id under a prefix derived from the key, the row is inserted once the door reports the write committed, and a committed request the handler sent as the local writer, found without its row, counts as applied and fills the row in; a remote writer's request under the same prefix never counts. Because the id is new on every run, a stalled run or a gbrain jobs retry (which resets the attempt count) never replays an earlier write of the capture that failed. Tombstones and jobs queued without a key behave as before.
  • src/core/minions/types.ts: handlers see the job's idempotency_key on their context; jobs already waiting when you upgrade carry it from their row.
  • src/core/cycle.ts: the dream cycle's purge phase drops applied-capture records older than 90 days.
  • Tests: test/ingest-capture-applied-once.test.ts (PGLite) and test/e2e/ingest-capture-applied-once-postgres.test.ts (Postgres): a re-run after a newer edit, a replay after the job row is pruned (through the real route), two ordered captures to one slug, a replay of a capture completed before the upgrade, the re-run on a managed brain, a retried managed capture whose attempt id already names a refused or failed write, and a remote writer's request forged under the capture's prefix.

[0.60.612.0] - 2026-11-08

A page write or a managed extract_facts run no longer retires facts that the page's ## Facts fence never held, and gbrain repair unwithdrawn-facts restores the ones earlier releases retired. A fact a put_page writer attached to a page with a row number, but that was never in the page's fence, used to be expired by any later write that projected the fence. That included a write that changed nothing in the fence, often the page's first facts-fence generation, which gbrain repair generations cannot undo because it has no earlier generation. Now a row retires only when a fence held it (a facts-fence generation recorded it, the page as the writer read it carried it, or a stored version of the page did) or when the incoming fence states its claim. Every other row stays active with its id. Removing a row from the fence still retires it, as before. Doctor unwithdrawn_facts counts the facts that were already retired and points to the new repair, which previews first and restores the original fact ids and lifecycle from the generation ledger.

The repair restores only what a later write would keep. It leaves a fact retired when you forgot it, including on your own memory page, and when the page's fence states the same claim. For copies of one claim it restores one per speaker. It picks the narrowest visibility first, and a copy that is still valid before one whose validity had lapsed. On a brain without managed persistence the extract_facts reconcile still retires rows its fence lacks, so the repair lists those facts as excluded:unmanaged_reconcile and restores none of them.

Itemized changes

  • src/core/generations/fence-evidence.ts (new): factsFenceHeld decides whether a fence ever held a fact row, from the facts-fence ledger, the replaced body and the page's stored versions. moveKeptRows moves a kept row whose number an incoming fence row claims above the page's rows.
  • src/core/generations/adapters/fact-rows.ts, src/core/persistence/canonical-projections.ts, src/core/trust/fence-guard.ts: the page projection's fact generation retires only held rows and rows whose claim the incoming fence restates. Conversation-extractor rows still give up a number a fence row claims. Its expire statement and the trust guard leave kept rows alone.
  • src/core/cycle/extract-facts.ts: on a managed brain the reconcile keeps stale rows no fence held, unless the fence restates their claim, and moves one whose number a new fence row needs. An unmanaged brain's reconcile is unchanged.
  • src/core/repair/unwithdrawn-facts.ts and src/commands/doctor/checks/unwithdrawn-facts.ts (new): gbrain repair unwithdrawn-facts [--include-ambiguous] [--apply --expect <hash>], bound to each page's revision at preview. A managed brain applies it through a database-only managed_maintenance_restore_unwithdrawn_facts request. Each fact is listed with its class: evidenced, ambiguous, or excluded with a reason (fence_held, withdrawn, fence_states_claim, active_duplicate, duplicate_candidate, unmanaged_reconcile and others). A page written since the preview restores nothing. Doctor check unwithdrawn_facts.
  • Tests: test/fence-first-generation-8819.test.ts and test/repair-unwithdrawn-facts-8819.test.ts, on PGLite and Postgres.

[0.60.611.0] - 2026-11-08

Doctor foreign_writers no longer reads an unreadable ledger or a skipped repository as clean, the owner-path commit guard fails closed, and the opt-in local-edit adoption (persistence.adopt_local_edits, off by default) gets the hardening it needed before any source turns it on. foreign_writers now marks a scan partial when a root's Git probe fails, when roots pass the 50-root cap, or when the commit ledger exists but cannot be read. It names a foreign commit that carries a gbrain- trailer as possibly an older gbrain without the ledger, and still fails it. gbrain git commit-push refuses a repository that only encloses a registered checkout, even when an older gbrain left a stale marker in it. Adoption keeps its refusal for a conflicted path or a merge in progress, a stale editor save, a segmented or oversized page, and a caller whose claim is gone, and checks the claim, the conflict and the merge again in the transaction that publishes its commit. Its quiet wait is capped at the foreground budget, and an adoption whose publication failed after its commit is retried instead of left orphaned.

Itemized changes

  • src/commands/doctor/checks/foreign-writers.ts, src/core/git/commit-ledger.ts: partial_reasons lists skipped roots. A probe that fails inside a checkout, or roots past MAX_ROOTS, mark the scan partial, and a scan that read no repository but skipped one warns instead of reporting not_applicable. readLedger treats only ENOENT as empty. A failed or timed-out git status names its checkout, and its timeout follows doctor's per-entry deadline (5 to 30 s). Commits with a gbrain- trailer carry gbrain_trailer.
  • src/commands/git.ts, src/core/persistence/root-registry.ts: checkout_not_a_source requires a registered root to contain the checkout, or to be a directory of it that its HEAD tracks. A stale gbrain-managed.json in an enclosing repository's git dir no longer admits it. A registry directory that cannot be listed refuses with a typed error.
  • src/core/git/commit-guard.ts: the owner-path hook exits 1 when git rev-parse --absolute-git-dir fails.
  • src/core/persistence/local-edit-adoption.ts, file-repair.ts, coordinator.ts, pending-import.ts: adoption refuses unmerged paths and in-progress merges, cherry-picks and rebases. It also refuses a file whose body equals one of the last 50 revisions or that drops rows the current revision added, and it stats the file before reading it. It caps the quiet wait at 30 s and records the caller's claim, so an orphaned adoption publishes nothing. It waits on an adoption another task is already publishing, retries a requeued publication against a HEAD that already holds the adopted blob, and never reports released for a lost claim. The publication transaction locks the caller's row and checks again that its claim holds, that the path has no unmerged stages and that no merge, cherry-pick, revert or rebase is in progress, so a cancel or a conflict that arrives after admission stops the commit. A caller that names no claim token ends write_claim_lost instead of a database error, and the stale editor save check strips the same edge whitespace in SQL as in JavaScript, so a body ending in a non-breaking space still matches its earlier revision. An empty persistence.adopt_quiet_ms reads as the 5000 default, and an admission failure logs a phase=adoption line.
  • src/commands/doctor/checks/local-edit-adoptions.ts: counts adoptions from outcome.adopted_local_edit, which receipt compaction keeps.
  • src/core/persistence/refresh-barrier.ts: with persistence.refresh_barrier on, a write waiting behind a refresh or an import now counts a later write to its page as a conflict unless managed sync's own import wrote it. Before, a caller could hide its write by passing a managed_sync_ kind, and a compacted sync receipt counted as a later write. A connector import still counts as a later write after its receipt is compacted.
  • src/core/persistence/page-mutations.ts: a page write that passes a kind starting with managed_sync_ is refused with invalid_params.
  • Tests: test/doctor-foreign-writers.test.ts, test/commit-ledger.test.ts, test/sources-guard.test.ts, test/git-commit-push-not-a-source-8662.test.ts, test/adopt-local-edit.test.ts, test/refresh-barrier-gst1623.test.ts and test/refresh-write-barrier.test.ts on PGLite and Postgres (#8755).

[0.60.610.0] - 2026-11-08

A client that retries a POST /ingest capture accepted before v0.60.609.0 no longer queues a second job that writes the old body over the page's newer edit. v0.60.609.0 added the target slug to the capture's queue key, and nothing looked up the old key form, so a retry of a capture that had already landed was taken as a new one. The route now also looks up a capture's pre-v0.60.609.0 key and answers 202 with that job when it was queued for the same target slug, so the retry is the same request again and nothing is queued.

Itemized changes

  • src/commands/serve-http-webhooks.ts (POST /ingest): before queueing, the route looks up the capture's legacy key (client, write source and content hash, no slug part). A row queued for the same normalized target slug (or with no slug, for a slug-less capture) answers the request; a row for another slug never does, so one body sent to two slugs still queues two captures. A dead or cancelled legacy row frees its key, as under the current key. A legacy row still waiting or running with no recorded submitter (left by an upgrade across v0.50) answers 409 legacy_job_authority with its recovery hint instead of queueing a second job. The 429 at the waiting cap, the 400 for slugs over 1024 bytes and the webhook HMAC are unchanged.
  • Tests: test/serve-http-ingest-legacy-key-retry.test.ts (PGLite) and test/e2e/serve-http-ingest-legacy-key-retry-postgres.test.ts (Postgres) drive the real route: a retry of a completed legacy capture gets its job back, queues nothing and leaves the newer edit in place, and a legacy row with no recorded submitter gets its job back when completed or a 409 while still waiting.

[0.60.609.0] - 2026-11-07

POST /ingest no longer answers 202 for a capture it drops, and an unsigned GitHub webhook can no longer tell from response time which repositories a brain syncs. The same body sent to two slugs now queues two captures, where the second caller used to get the first job back and nothing written to its slug. A capture over the 50-capture waiting cap now gets 429 ingest_backlog_full with Retry-After: 30 and nothing queued, where it used to get another capture's job id. A slug over 1024 bytes gets 400 invalid_slug before anything is queued, well short of the roughly 2.7 KB where no page can be stored under it. Every webhook request now computes the HMAC over its body, whether the repository is unknown, has no secret or is covered by a source, so a 401 takes the same work either way.

Itemized changes

  • src/commands/serve-http-webhooks.ts (POST /ingest): the queue idempotency key carries the normalized target slug as its sha256 (empty when the job picks the slug), so retries to the same slug in any letter case still collapse onto one job. When the waiting cap hands back another job, the route answers 429 instead of 202. An X-Gbrain-Slug over 1024 UTF-8 bytes is refused with 400 invalid_slug.
  • src/commands/serve-http-webhooks.ts (POST /webhooks/github): push and item events compute the HMAC with a throwaway key when there is no source or no secret. For item events, every github-kind source with a secret is checked, and legacy github_repo sources share exactly one check, given to the lowest-id one with a secret, so the work does not depend on how many sources name the repository. A second legacy source on the same repository with a different secret is not tried.
  • Tests: test/serve-http-ingest-write-loss.test.ts (slug-keyed dedup, the 429 at the cap, the 1024-byte slug limit at both edges), and HMAC-count checks in test/serve-http-github-webhook.test.ts and test/serve-http-github-push-webhook.test.ts.

[0.60.608.0] - 2026-11-07

With the refresh write barrier on, a write that waited behind a refresh or an import can no longer land on top of a later write to the same page, and most of the barrier's remaining review follow-ups (GST-1623) are closed. The barrier is still off by default (persistence.refresh_barrier); with it off, only the receipt health, the recovery-time replay order, the refresh merge check, the drain sleep and the new index change. A consumer from an older release that claims a later write first now makes the waiting write end revision_conflict instead of overwriting it. A write held again and again behind imports now expires at a ceiling counted from its admission. Waiting writes count toward their source's fair share of the local CLI cap. A refresh whose record recovery moved before the merge merges nothing. With the barrier on or off, a plain-git commit that lands on the upstream commit right after a refresh's merge no longer leaves the checkout recovery_required with writes refused until someone runs --resume: the refresh records the merge and syncs the new HEAD. One that lands on the start commit right before the merge no longer does either. Git refuses the fast-forward, and the refresh ends refresh_source_changed with the fence lifted, nothing merged and the commit kept, so the next refresh can run.

Itemized changes

  • src/core/persistence/refresh-barrier.ts: a deferred or import-held write ends revision_conflict when a later non-sync write to its page has already committed (appends excepted), and refreshed paths compare case-folded when core.ignorecase is true. A held write's limit is the wait limit since its hold, or four wait limits (at least four hours) since its admission. An expired remember says that only a fact naming its entity gets the automatic replay, and that gbrain repair failed-writes lists the rest.
  • src/core/persistence/import-deferral.ts: a write deferred behind a refresh that meets pending_import takes the refresh-change check on that file first (a refresh that changed it ends the write revision_conflict), and its hold clears the released deferred_refresh_id, so a pre-v294 consumer never claims it (#8759 included: a capture whose captured bytes are the imported file names no supersedes).
  • src/core/persistence/page-prepare.ts: an earlier held write refuses only a journaled request admitted after it, never a Git or withdrawal-mirror effect, lint or a drift check reading the file.
  • src/core/persistence/journal.ts: the local CLI fair share counts deferred rows; managed sync's capacity discount counts only writes still held for its own source (#8759); receipt health treats a released deferred write as an earlier write and reports recovery_required for every write on a worktree whose refresh awaits recovery; a replay of an accepted request id reconciles before any deferral.
  • src/core/persistence/worktree-refresh.ts: the ff-only merge runs only when recording merge_started matched the fenced row; otherwise the refresh ends refresh_source_changed with nothing merged. When the merge succeeded but HEAD already descends from the upstream commit (a plain-git commit landed in between) and no member synced a commit HEAD lacks, the refresh records the merge and continues to syncing. When git refuses the fast-forward because HEAD gained commits on the start commit only (a plain-git commit landed after the HEAD check), the refresh aborts with refresh_source_changed and retries once in the run (GST-1674). A reset to the start commit plus an unrelated commit still ends recovery_required.
  • src/core/persistence/sync-drain.ts: a worktree_refreshing retry hint never sleeps past the drain's refresh budget.
  • Migration v295 adds persistence_worktree_refreshes_completed on (worktree_id, completed_at DESC) WHERE state='completed' for the retry-hint history read.
  • gbrain upgrade --check-downgrade and the guide say a released deferred write publishes only when a process runs the write consumer on the owner host; polling gbrain write-request only reads the receipt.
  • Tests on PGLite and Postgres: test/refresh-barrier-gst1623.test.ts (GST-1623 items 1, 2, 3, 5, 13–22, #8759, and a commit landing right after or right before the refresh merge with the barrier off and on).

[0.60.607.0] - 2026-11-07

A long gbrain sync no longer prints a stall right after the writer was busy finishing a claimed request.

While a sync drains its write queue, gbrain prints a live line such as stalled 45s on queued when the request at the head of the queue stops moving. If the head was claimed for a while (for example while the process published the writes a killed server left behind) and a new queued request then reached the head, that line counted the new request's wait from before the claimed stretch. It reported a stall of a minute or more right after the writer had been working the whole time. Now every read of the head counts as progress, claimed or not, so the next queued head's age starts at the read before it. A claimed head still prints its own step age.

What you'd see

Head reads, ten seconds apart Before Now
a queued request, then a claimed one for 60 seconds, then a new queued request stalled 70s on queued stalled 10s on queued
a claimed request that is publishing the claim's step age the claim's step age (unchanged)

Things to watch

Only the printed age changes. When a sync stops, escalates or reports blocked is unchanged, and nothing stored is touched.

Itemized changes

Fixed

  • The in-pass stall line (inPassHeadAge in src/core/persistence/sync-drain.ts) records claimed head reads, so a queued head that follows a claimed one counts its age from the read before it.

For contributors

  • test/sync-drain-claims.test.ts covers a claimed head between two queued heads. The convergence nightly's 70 s owner-kill stall bound now measures real stalls, not time the writer spent publishing a killed server's writes.

[0.60.606.0] - 2026-11-07

Updates follow a stable channel by default; every release is one setting away on alpha.

gbrain publishes a release for nearly every merged change, more than a hundred a week. Every install followed all of them, so an always-on brain upgraded several times a week onto releases no one had run on a production-shaped brain. Installs now follow the stable channel, which gets about one release a week, each promoted by a maintainer after the soak passed on that release's exact tree. The alpha channel (the old latest, which still works as an alias) keeps every release.

What changed

Situation Before Now
A new or existing install with no channel set offered every published release offered the newest stable release; one notice on stderr says so, once per brain home
Following every release the default gbrain config set self_upgrade.channel alpha (or --channel alpha, GBRAIN_RELEASE_CHANNEL=alpha)
Which stable release is offered VERSION at the stable tag the published evidence record (stable.json on stable-record): version, commit, tree, soak run, withdrawn versions; a record without the evidence is never trusted
Stable has no usable answer stable_channel_unresolved on upgrade stable_missing, stable_invalid or stable_unreadable; the background check only reports it, an explicit upgrade exits non-zero, neither falls back to alpha
An install ahead of stable (it ran alpha) — offered nothing until stable passes it; never downgraded unless its version was withdrawn
Update cache and snooze one file for every channel; a snooze ended when a newer release shipped one per channel, tagged with the evidence policy; a snooze holds its whole window (24 h, 48 h, 7 days); a stable rollback below it clears it
gbrain upgrade --dry-run refused without --restart-only reports channel, its source, current and target version and the action, installing nothing
doctor self_upgrade_health — names the channel and its source, warns on an invalid channel or a failed stable check
Promote stable workflow moved the stable tag also commits the evidence record in the same atomic push; new withdraw input adds versions to known_bad; promoting the current stable again writes its record from the stable tag's message, moving no tag (the one-time backfill for releases promoted before the record existed; the daily stable check does it on its first scheduled run)

Install commands now name stable (bun install -g github:garrytan/gbrain#stable), so a new install starts on the channel it follows. Every install surface goes through one constant (src/core/install-ref.ts), checked by bun run check:install-ref; latest-stable still installs the newest release for anyone who pins it.

[0.60.605.0] - 2026-11-07

Stable now moves forward by itself: a daily check promotes an important fix within a day and anything else within a week, and only after the soak passed on that exact release. .github/workflows/stable-daily.yml runs at 05:17 and 13:17 UTC. It acts only when a merged PR labelled stable:important is released but not yet in stable, or stable is more than 7 days old. It promotes the newest release whose exact tree passed the production soak with every integrity gate clean, and otherwise dispatches one soak so a later run can promote. It never moves stable backward, pauses while stable is a rollback, and stops entirely with the repository variable STABLE_AUTO_PROMOTE=off.

Itemized changes

  • New scripts/stable-daily.ts and .github/workflows/stable-daily.yml: the reason (a released stable:important fix stable lacks, or the 7-day weekly floor), the candidate (newest release on master newer than stable with a successful full soak of its exact tree, both default cells for at least 2 hours on the requested 600,000-page fixture with privacy probes run and none erroring, and privacy_leaks, read_errors, version_skew, residents_alive and session_refusals passing in every cell; on the important-only path it must contain the fix), one soak dispatch when nothing qualifies (never while a soak runs, never after a soak on a newer release failed to qualify, with a warning once stable is past the weekly floor), the rollback pause and the kill switch. A manual dispatch defaults to a dry run, and dry_run=false needs an admin or maintainer before it writes anything.
  • scripts/promote-stable.ts: scheduleAuthorized() lets only a schedule event of stable-daily.yml on the default branch promote without the admin or maintain role; every other check is unchanged, and the record's promoted_by names the schedule. Manual dispatches keep the maintainer check.
  • docs/RELEASING.md#daily-stable-check documents the label, the schedule, the kill switch and the rollback.
  • test/scripts/stable-daily.test.ts covers the authorization, the integrity gates, each reason and no-op path, the soak dispatch rules, the dry run and the workflow shape.

[0.60.604.0] - 2026-11-07

New installs now start on the stable channel: the soaked release, not every release. The documented install (bun install -g github:garrytan/gbrain#stable) and the clone fallback (git clone --branch stable) follow the stable tag, which moves only to a published release whose exact tree passed the production soak, at least weekly. A fresh gbrain init (and so gbrain setup claude-code or codex) writes self_upgrade.channel: stable, so the update check and gbrain upgrade offer only stable releases. #latest-stable keeps working as the alpha line with every release. An existing brain is not changed, because a re-init never sets or alters its channel.

Itemized changes

  • Install pins move from #latest-stable to #stable in README, AGENTS.md, INSTALL_FOR_AGENTS.md, BOOTSTRAP_FOR_AGENTS.md, docs/INSTALL.md, the MCP, headless, in-agent and tutorial guides, the setup skills, the Codex plugin manifests and launcher, doctor's npm-squat fix, the bootstrap hints and the agent-workspace template. The bootstrap runbook fetch and the in-agent setup helper read the stable ref, and the in-agent helper pins the commit refs/tags/stable peels to. Until a release containing this change is promoted to stable, the runbook and helper served from that ref are the previous stable release's copies.
  • scripts/check-bootstrap-tag.sh sanctions stable as the one distribution ref and now refuses #latest-stable in the entry docs.
  • src/commands/init.ts: a fresh install (new content database, no prior config file) persists self_upgrade.channel: stable on PGLite and Postgres. A re-init keeps whatever the config says, including no channel. test/init-stable-channel.test.ts covers the fresh, re-init and explicit-channel cases, and test/e2e/init-stable-channel-postgres.test.ts the fresh and re-init cases on Postgres.
  • docs/guides/upgrades-auto-update.md#stable-channel describes stable as the default for new installs and latest as the alpha line.

[0.60.603.0] - 2026-11-07

A source that a remote caller adds by URL keeps its clone settings on a managed brain, both when it is first cloned and when it is recloned, and gbrain frontmatter install-hook leaves that clone's hooks off. v0.60.601.0 wrote git-lfs downloads off and hooks off into such a clone's own git config, but on a managed brain the add path did not record the source as remote-added and gbrain sources reclone cloned it again without the settings. install-hook read the clone's hooks-off setting as a hooks directory someone else chose, wrote a hook it would never run, and told the operator how to turn hooks back on.

Itemized changes

  • src/core/persistence/managed-sources.ts, src/core/persistence/topology-filesystem.ts, src/core/persistence/topology-clone.ts: a managed add by a remote caller records added_by_remote_caller on the source, and the managed clone of such a source, on add and on reclone, uses the same git clone --config settings as cloneRepo's skipLfs (shared as LFS_OFF_CLONE_CONFIG in src/core/git-remote.ts).
  • src/commands/frontmatter-install-hook.ts: a checkout whose own (--local) core.hooksPath is /dev/null is skipped (skipped_hooks_off) with nothing written and no wiring advice, --force included, and doctor's frontmatter_hook check no longer asks to refresh a hook there. A machine-wide core.hooksPath=/dev/null keeps the earlier behaviour, so the hook is written unwired with the usual advice.
  • Tests: test/persistence-source-lifecycle.test.ts (PGLite and Postgres), test/sources-remote-lfs.test.ts, test/frontmatter-install-hook.test.ts.

[0.60.602.0] - 2026-11-07

A TLA+ model of the refresh write barrier and its deferred-write queue now runs in CI. It checks that a refresh killed at any phase recovers without stranding the writes deferred behind it, that held writes commit in admission order, and that no lock order deadlocks. Nothing changes at run time: this release adds a model, its CI job and tests.

Itemized changes

  • spec/tla/refresh-barrier/: RefreshBarrier.tla models write admission and deferral, the drain, the fence and the merge, a crash at each phase followed by recovery, ledger commits that move HEAD during the drain, writes held for an import, deferral behind recovery_required, and local-edit adoption. Four configs check every safety property and one checks liveness; two larger configs run nightly. The model reproduces the fenced-orphan recovery bug fixed in 0.60.596.0 and passes on the fix.
  • spec/tla/refresh-barrier/mutants.json: 19 mutants. Each removes one guard from the model, and TLC must report the named property. 15 make the same edit to the code, and a named PGLite or Postgres test must then fail on an assertion. Run them with bun scripts/spec/mutants.ts refresh-barrier.
  • .github/workflows/tla-spec.yml: TLC and the mutants run for both TLA+ specs, nightly and on pull requests and pushes that touch a spec, its runner, the code it models or a test a mutant names. Each spec uploads its own log artifact, tla-spec-logs-<spec>.
  • Tests: test/spec-refresh-barrier.test.ts keeps the configs, the workflow's path filter and the README's mutant table in step with the catalogue. New tests cover an import expiry overlapping a fence (Postgres), a claimed write that holds the drain, a token revoked after its write was claimed, and the owner scan aborting a refresh killed at the fence after an outside commit. The two Postgres race tests now fail if the operation they start never waits on the lock.

[0.60.601.0] - 2026-11-07

Security wave from a /cso audit: four remote-surface holes closed on gbrain serve --http and gbrain mcp expose. A write-scoped token could replace, through POST /ingest, pages that every MCP write refuses it: another member's readable_by page, an owner visibility: private page, another member's fact page. A read-scoped caller could make search_by_image read a file on the brain host by passing a file:// URI or an absolute path as image_url, which echoed the file's first bytes, sent server images to the embedding provider, or hung or exhausted the server. And an unsigned GitHub push webhook answered whether a repository was wired to the brain, its tracked branch and its source id before checking the signature. A remote caller adding a source by URL could also make the brain host contact an address the cloned repository names for itself, which the URL check never sees. A default keyless install (stdio MCP, no HTTP surface) was not exposed to any of them.

Itemized changes

  • src/commands/serve-http-webhooks.ts, src/core/minions/handlers/ingest-capture.ts: POST /ingest stores the caller's token or OAuth client on the job. The route answers 404 page_not_found for an X-Gbrain-Slug the caller cannot read and 403 permission_denied for a member fact page or the scope page, and queues nothing. An X-Gbrain-Slug that is not a valid slug gets 400 invalid_slug instead of a queued job that fails later (#8763). The job checks the same writer again right before it writes, so a page restricted after the 202 is still refused, and a queued webhook job with no stored writer is checked as a guest. Fresh slugs and pages the caller can read import as before.
  • src/core/ops/image.ts, src/core/search/image-loader.ts: search_by_image accepts only an http(s) image_url from any caller that is not the trusted local CLI, and refuses image_path unless the context says the caller is local. The loader reads local paths only when its caller opts in, and checks for a regular file and the size cap before it reads.
  • src/commands/serve-http-webhooks.ts: a push event is verified against the source's secret before anything about the repository is answered. An unknown repository, a source with no secret and a bad signature all get the same 401 signature_mismatch; the missing-secret case is logged on the server with the command that sets it.
  • src/core/git-remote.ts, src/core/sources-ops.ts, src/core/ops/sources.ts: a URL source a remote caller adds is cloned with git-lfs downloads off, written into the clone's own git config (lfs.fetchexclude=*, which outranks the repository's .lfsconfig, with hooks off so git-lfs's pre-push hook never runs), so every later pull, managed refresh and commit-push in that checkout keeps large files as pointer files and contacts no LFS endpoint. The source records added_by_remote_caller: true so a re-clone keeps it off. A URL source a remote caller added before this release carries no such record, so remove it and add it again (from the remote caller or with gbrain sources add --url, which keeps git-lfs) to choose its setting. A remote caller whose clone fails gets a fixed message; git's own error goes to the brain host's log. Sources added from the local CLI keep git-lfs.
  • Tests: test/serve-http-ingest-member-access.test.ts (real token verifier and Express route, two members), test/search-by-image-op.test.ts, test/cross-modal-phase2.test.ts, test/serve-http-github-push-webhook.test.ts and test/sources-remote-lfs.test.ts (real git-lfs against a local LFS endpoint).

[0.60.600.0] - 2026-11-07

With the refresh write barrier on, writes waiting behind an import no longer use up the capacity the import itself needs, so a full cap no longer stalls managed sync until the waits expire. The barrier is still off by default (persistence.refresh_barrier), and with it off nothing changes. Before, every held write kept its outstanding-request slot and intent bytes, so enough of them could leave a managed sync job, or a bulk import group, refused queue_capacity, and the writes it would have released waited out the limit and expired. A held capture that replaces an imported page now names what it replaced, as a blind put_page does, and the expiry pass no longer expires a hold that an import and a new deferral replaced while it ran.

Itemized changes

  • src/core/persistence/journal.ts: a managed sync job, and a bulk sync group in admitWriteGroupInTransaction whose writes are all managed sync jobs, is admitted as if the writes held for an import or a refresh took no outstanding-request slot and no intent bytes. The held writes are counted only when the plain check would refuse, so an admission that fits reads nothing more. The importer can pass a cap by at most what the held writes hold. Other writes still count them.
  • src/core/persistence/import-deferral.ts: a held capture that lands over imported content records supersedes on its receipt, the gbrain-supersedes: commit trailer and the commit-ledger line, as a blind put_page does.
  • src/core/persistence/refresh-barrier.ts: expireDeferredWrites rechecks, under the lock, that the hold it scanned is still waiting on the same import and is still older than the wait limit, so a hold that an import and a new deferral replaced between the scan and the lock is left alone.
  • Tests on PGLite and Postgres: test/import-write-barrier.test.ts (a held capture's supersedes, the capacity a single import and a bulk import group need, the expiry racing a new hold) and test/refresh-write-barrier.test.ts (the capacity a refresh's own sync needs). Follow-ups are in #8759.

[0.60.599.0] - 2026-11-07

A page write refused source_changed now says why in its receipt, and a source can opt in to having a remember, capture or add_timeline_entry adopt an uncommitted local edit instead of being refused. Before, only pending_import, sync_held and canonical_file_missing refusals carried a cause, so a write lost to a hand edit, an occupied path or a file that moved mid-write looked the same in gbrain write-request and doctor. Every source_changed refusal now records error_detail.origin. With persistence.adopt_local_edits.source.<id> on (off by default), a one-sided local edit (Git HEAD still holds the database's copy, or any edit on a repo-less brain) is committed as its own write first, then the caller's write lands on top. Keep the flag off until the follow-ups in #8755 are done.

Itemized changes

  • src/core/persistence/publication-failure.ts, coordinator.ts, page-prepare.ts, git-door.ts: a source_changed refusal's error_detail is { origin }, one of file_database_drift, canonical_path_occupied, raw_file_changed, pinned_git_worktree_conflict, unsafe_file_target, source_binding_changed or unbound_source. Classification, retries and replay are unchanged. gbrain sources writer status leaves these routine refusals out of its recent database refusals.
  • src/core/persistence/local-edit-adoption.ts (new), file-repair.ts, effect-git.ts, commit-ledger.ts: the adoption is a managed_file_repair admitted under the installation's CLI writer with a request id derived from the caller's, so a write adopts at most once. It waits until the file has been unchanged for persistence.adopt_quiet_ms (default 5000, at most three windows), imports the exact bytes, keeps the replaced revision with change reason adopted_local_edit and, on a Git-durable checkout, commits only that file (gbrain: adopt local edit to <path>, trailers gbrain-adopted: and gbrain-author: unknown, a commit-ledger adopted line). The caller's receipt names adopted: { path, blob, replaced_revision }.
  • The adoption keeps the refusal for an edit to the title, type or a protected frontmatter key, a file that does not import, an edit on both sides, bytes that are not valid UTF-8, a HEAD that moved before publication or that Git cannot read, a cancelled caller, put_page, batches and every other verb. A file changed again during the adoption ends raw_file_changed with nothing written. An adoption still waiting on the worktree lock when the caller gives up is cancelled.
  • src/commands/doctor/checks/local-edit-adoptions.ts (new): doctor local_edit_adoptions lists the last 24 hours of adoptions per source.
  • Tests on PGLite and Postgres: test/adopt-local-edit.test.ts (each verb, every kept refusal, the quiet window, the lock and cancel races, a moved or unreadable HEAD, the doctor check, a repo-less brain) and test/remember-pending-import.test.ts.

[0.60.598.0] - 2026-11-07

A segmented page that sync replaces on Postgres now publishes on the first try instead of being refused body_staging_expired ("does not cover its N-byte body") and blocking its source until the next sync. The import ran its version write, which for a segmented page prunes old versions and sweeps their segment rows, at the same time as the flip that publishes the new body. On Postgres the two overlap on one connection, so the sweep could delete the freshly published tail row between the flip clearing its staging and bumping the page's generation. PGLite runs them in order, so only Postgres was affected. The refusal now records which staging was short, a retry inside the run really prepares the page again, and gbrain sources writer failures answers within its bound on a source full of large sync writes.

Itemized changes

  • src/core/import-file.ts (#8729): a segmented import writes its version before the page write, as unpinned and demoting imports already did, so the version prune and segment sweep never run beside the flip.
  • src/core/persistence/sync-prepare.ts: a waiver screen aborts the segmented staging its import opened as soon as the import hands it over, on the unbounded engine so a screen past its deadline still aborts it. Before, every screened replace left a second staging live until the idle sweep.
  • src/core/persistence/sync-pass-batch.ts, sync-run.ts: an in-place re-freeze in a pass takes passRequestId(…, failedRequestId) instead of the failed request's own id, so the retry prepares the entry again instead of reading the failed receipt back. The first freeze keeps the v1 id; STOPS_PER_ENTRY still bounds retries. A chain restarted in the same pass can still re-derive an earlier chain's ids (#8750).
  • src/core/persistence/publication-failure.ts, coordinator.ts: a body_staging_expired refusal's error_detail is { origin: 'body_staging', staging: { staging_id, state, reason, rows, segment_count, covered_bytes, body_bytes } }, read field by field from the engine's detail. Remote callers still see only origin.
  • src/core/persistence/terminal-classify.ts, terminal-failures.ts: the supersession probe tests the failed row's intent kind before it reads any later request's intent, and readTerminalFailures selects its page of rows before the lateral join, so the drill-down and the summary no longer detoast every later sync write's frozen file per failed row. On 200 failed and 2,000 committed large-intent writes the drill-down went from unknown at the 5 s bound to about 2 s.
  • Tests on PGLite and Postgres: test/large-pages/segmented-sync-replace-8729.test.ts and test/e2e/segmented-sync-replace-8729-postgres.test.ts (tail-only replaces while every version prunes, no staging left live, a forced refusal recorded and prepared again), test/e2e/terminal-failures-large-intents-8729-postgres.test.ts, test/publication-failure-staging-8729.test.ts, test/terminal-supersession-kind-8729.test.ts, test/sync-pass-retry-id-8729.test.ts.

[0.60.597.0] - 2026-11-06

gbrain upgrade now restarts a jobs supervisor started from a source checkout as bun run src/cli.ts jobs supervisor instead of refusing the upgrade with restart_refused. The restart planner recognised a gbrain process only when its command line held a path containing "gbrain", so a supervisor launched with a relative src/cli.ts was invisible, its worker looked unsupervised, and the upgrade stopped. Four smaller fixes ride along with it, covering a flaky spend-guard SIGTERM test, the PGLite snapshot cache keys, the stress gate's pooled-owner match, and the ownership stamp on a checkout that became a Git repository after it was stamped.

Itemized changes

  • src/core/upgrade-restart.ts: gbrainEntry resolves a relative entry point against the process's working directory (/proc/<pid>/cwd on Linux), skips interpreter flags that take a value (--preload x, --cwd x, -r x; after the value of any other flag, such as --dns-result-order ipv4first, the entry is the next path) and treats src/cli.ts as gbrain when the package two levels up is named gbrain; an unreadable package falls back to the old path match. bun run <file> runs the file in its own process and is classified; bunx, bun x and bun run <bin> are wrappers whose child is the gbrain process, so they are no longer counted twice. A worker whose parent is not recognised as gbrain is refused with that parent's pid and executable name; its command line, which can carry credentials, is never printed or put in --json output. On macOS, where the working directory is not read, a relative entry still matches a gbrain path; otherwise its worker is refused with the parent's pid and executable name (ps -p <pid> -o command= shows the rest).
  • scripts/eval-spend-guard.sh (#8717): the lease's start time is read before the signal traps are armed, so no command substitution is parsed between arming them and wait. A TERM that landed there ran its trap while bash 5.2 still expected the lease line's ), the trap failed with "unexpected EOF while looking for matching )'", the signal was lost and the guard waited out its command and exited 0. Under CPU load the original guard did that in 3 of 3000 forced TERMs and the fixed guard in 0 of 3000, interleaved. test/eval-spend-guard.test.ts` pins that no command substitution sits in that window; the exit-143 assertion is unchanged.
  • .github/workflows/test.yml, e2e.yml, stress.yml (#8700): the 15 pglite-snapshot cache keys hash src/core/**/*.ts (plus the snapshot builder and its test helper) instead of src/core/*.ts and a hand list of subdirectory files, so a schema input in a new src/core folder needs no workflow edit. A change anywhere under src/core now rebuilds the snapshot cache.
  • scripts/stress/run.ts (#8702): a pooler-gated test is exempt from the stress gate only when the workflow job that names it, or the script lines beside it, set GBRAIN_CI_REQUIRE_PGBOUNCER; a flag elsewhere in the same file no longer counts. test/e2e/graduation-cli-topologies.test.ts is no longer exempt, and a script owner is reported as an "owning pooled runner".
  • src/core/persistence/physical-root-record.ts (#8713): verifying a physical root inside a Git checkout writes the ownership-marker exclude once per process, so a repo-less brain that later ran git init stops offering its .gbrain-owner.json to git add -A. The conversion refusal for a root with no Git history now says to add the exclude before the first git add -A. Doctor's existing tracked_ownership_marker check (#5186) still names a stamp already in the index with the git rm --cached fix, which keeps the file and history.

[0.60.596.0] - 2026-11-06

With the refresh write barrier on, a page write no longer bounces because a sync is running, and a gbrain sources refresh killed at any point recovers on its own or stops for an operator, never losing work. The barrier is still off by default (persistence.refresh_barrier), and with it off writes behave as before. With it on, a put_page, capture, remember or add_timeline_entry whose page file holds a committed change managed sync has not imported yet is answered pending instead of refused pending_import, then prepared again once the import lands. A write during a refresh stuck in recovery_required waits too, and the per-refresh deferral cap is gone. Wherever worktree_refreshing still refuses, retry_after_ms is estimated from the refresh instead of a fixed second.

What changed

Situation (barrier on) Before Now
A page write onto a file with a committed change sync has not imported refused pending_import pending; lands after the import (a put_page with expected_revision ends revision_conflict when the import moved the page)
A blind put_page that replaces imported content n/a lands and names the revision and commit it replaced in supersedes, a gbrain-supersedes: commit trailer and the commit ledger
The import never lands within the wait limit of the hold n/a pending_import_expired, nothing applied; a remember ends as its pending_import refusal did and is replayed after the import
A write during a refresh in recovery_required refused refresh_recovery_required waits behind the refresh and lands once it is resumed or abandoned
Writes past persistence.refresh_barrier_max_deferred refused deferred; the setting is no longer read
retry_after_ms on worktree_refreshing always 1000 the refresh's typical remaining time, at least 15 s while syncing and 2 s otherwise

Itemized changes

  • Import holds (src/core/persistence/import-deferral.ts, journal claim, refresh-barrier.ts): the consumer moves its own claim back to deferred naming the last imported commit, and the row is claimable again once sources.last_commit moves. A later write to the same page never passes a held one, and expiry rechecks under the source lock that the import still has not landed. An idle owner counts an overdue hold as work, so expiry runs with nothing else queued. A hold happens only when every live consumer on the owner host advertises refresh_barrier_import_1; otherwise the write keeps the refusal. Schema migration v294 adds persistence_requests.deferred_import_from and widens the deferred-row guard.
  • Refresh crash recovery (src/core/persistence/worktree-refresh.ts): a refresh records merge_started in a committed transaction before git merge --ff-only. A refresh killed at fenced before its merge, whose HEAD then moved on its own line (gbrain commits from its drain, or anyone else's), aborts as refresh_interrupted instead of being recorded as merged. One whose HEAD reached the target records the merge. A started merge followed by a reset and an unrelated commit still stops at recovery_required for an operator.
  • Retry hint (src/core/persistence/refresh-retry-hint.ts): the median duration of the worktree's last completed refreshes minus how long this one has run, leaving out refreshes that waited in recovery_required. The message says the write was not saved and when to retry, and the CLI's human output ends with one line saying the same.
  • Tests on PGLite and Postgres: test/import-write-barrier.test.ts (every held operation, supersedes, expiry, the remember replay, the hold clock, the capability gate, the idle-owner expiry, recovery_required and same-page ordering), test/refresh-retry-hint-gst1582.test.ts, and the crash-recovery cases in test/refresh-write-barrier.test.ts. Design note: docs/designs/REFRESH_WRITE_BARRIER.md, "Writes that wait for an import" and "Crash recovery".

[0.60.595.0] - 2026-11-06

A memory_recall that names sources reads as that scope, never as the attested person, so a shared turn can no longer recall one member's private facts into a shared channel. memory_recall gains source_id and source_ids. Named, the read drops the person's own recent facts and every member-page fact, keeps restricted pages hidden and never relaxes visibility: private. The contract for it is the new docs/contracts/qm-memory-provider/v2.json; v1.json stays byte-identical and a test holds it there. Three smaller provider fixes ride along.

What changed

Situation Before Now
memory_recall with source_id / source_ids the argument was ignored (unknown_param) and the call read as the person, own facts first reads the named sources as their audience: no own facts, no member-page facts; a source outside the grant is permission_denied, __all__ is invalid_params
memory_recall without sources the person's read unchanged
acting_user with no x-qm-actor, from a credential bound to a member argument_without_attestation its own member id (or a linked canonical id) is dropped and the call runs as the credential; any other value is argument_mismatch
recall with max_chars unknown argument 200-12000: whole facts, then results, that fit; truncated says if some were cut
A read that ignores an undeclared argument a [gbrain notice unknown_param] text block in the result only _meta.warnings, so a harness that stores tool output never stores it; writes still refuse

QM clients that read v1 keep working unchanged. A shared-scope turn should name its sources and read the v2 contract.

QM can now save a shared scope's memory (a project, channel or team) as well as each person's, and nothing else can read that scope. memory_write gains source_id: named, the facts go to that scope source's memory/scope page, world-readable inside the source, still attested to the person who said them. The operator lists the scope sources on the QM credential with gbrain auth rescope --attest-sources, which marks each one a scope source for good. A scope source is never federated and never joins another credential's read grant, so its memory is read only by naming it from the scope credential, and a person read never sees it.

Situation Before Now
memory_write with source_id refused (unknown_param) saves on memory/scope in that source when the credential lists it in --attest-sources; otherwise permission_denied reason source_not_attested; the credential's own write source is invalid_params reason source_is_member_source
gbrain auth rescope --attest-sources a,b|none no such flag lists the scope sources; refuses default, a federated source, the own write source, a source another credential reads, ops beyond the memory tools and the private-pages grant
A grant, token or registration that reads a scope source n/a refused (QM scope source), also after the scope credential is revoked
sources federate / sources restore on a scope source federates refused / restored unfederated; a federated flag on a scope source is ignored
Remote write to memory/scope (put_page, remember, delete_page, tags, links, timeline) allowed like any page refused unless it is a scope memory_write
memory_recall without sources from a credential that reads a scope source n/a never reads the scope source

A long recall question that opens with an instruction preamble finds the facts its last message asks about on a keyword-only brain, and refusals say what they need. Recall used to score a question on its first 12 content words, so a provider preamble filled all 12 and nothing cleared the admission floor.

Situation Before Now
recall question over 12 content words scored on its first 12 searched on its last 24, scored on the 12 of those the fewest candidate facts hold; shorter questions unchanged
Candidates found, none above the admission floor empty facts, nothing else empty facts plus facts_below_floor: <n> (absent when nothing matched)
recall question over 2,000 characters bare invalid_params invalid_params reason question_too_long, detail max_chars=2000 received_chars=<n>; never cut
acting_user refused argument_without_attestation no detail detail missing=x-qm-actor, missing=attest_right or both, also in structuredContent.error

[0.60.594.0] - 2026-11-06

A page write that reaches a checkout while gbrain sources refresh runs can now wait for the refresh instead of being refused, and every commit gbrain makes is recorded, so doctor can name commits made outside it. The write barrier is off by default (persistence.refresh_barrier). With it on, a single page write that arrives while a refresh drains or merges the checkout is journaled and answered pending at once: MCP gets isError: false with status: "pending", and the CLI exits 10 with the command that polls it, on the owner socket and on a thin client alike. The write commits after the refresh. It ends revision_conflict with nothing applied when the refresh changed its file, and refresh_barrier_expired when the refresh outlives persistence.refresh_barrier_max_wait_ms (15 minutes), even on an owner with nothing else to do. The commit ledger and doctor's new foreign_writers check run on every brain with a Git checkout. The owner-path guard installs only when you run gbrain sources guard --install.

Itemized changes

  • Refresh write barrier (src/core/persistence/refresh-barrier.ts, journal admission and claim, the consumer): requests gain a deferred state tied to the refresh they wait behind. Only a claim that declares the barrier can move a deferred row, and only after its refresh is released. A released write replays its change check against the refresh's diff, so a blind write never overwrites a file the refresh changed. Syncs, jobs and batches keep the worktree_refreshing refusal, and a live consumer that cannot drain deferred rows keeps it too. Restore, cancel, token revocation and source lifecycle changes end a deferred write with nothing applied and release its counters. Schema migration v292 adds the columns, the release trigger and an online partial index over the four open request states, so receipt polls stay index-served on a long receipt history.
  • gbrain upgrade --check-downgrade exits 3 downgrade_blocked_deferred_writes while deferred writes wait. gbrain sources writer status lists them.
  • Commit ledger (src/core/git/commit-ledger.ts): every gbrain commit site appends one line to <git common dir>/gbrain/commit-ledger.jsonl, rotated at 4 MiB. A failed append never fails the commit.
  • New doctor check foreign_writers: a commit in a source checkout on this host that the ledger does not name fails the check with the newest such commit and a fix that asks who makes it. Unconfirmed commits and page files uncommitted for over 30 minutes warn. Its git reads take no optional locks, so a doctor run never holds index.lock against a commit. A repo-less brain reports not_applicable.
  • New gbrain sources guard <source> --install|--uninstall|--status [--mode enforce]: the commit guard on a managed source's worktree, from its owner host, in a hooks directory private to that worktree. gbrain's own hooks at the old hooks path, the brain-durability post-commit hook included, are copied along, so installing it never turns Git durability off. --mode enforce also refuses cherry-pick, revert, rebase and sequencer commits.
  • gbrain git commit-push --base <path>=<blob>|absent refuses commit_base_moved, committing nothing, when the file at HEAD is no longer the blob the caller read. It also refuses checkout_not_a_source when a managed brain's command resolved a checkout from the working directory that is none of its sources (#8662).
  • gbrain repair file-drift with --held, --side, --state, --cursor or --provenance and no --prefer refuses with the usage line instead of running the all-drifted survey (#8615).
  • A gbrain sources refresh run refused refresh_source_changed by a moving HEAD or late work retries once from a fresh precheck in the same run (#8617).
  • Tests on PGLite and Postgres: test/refresh-write-barrier.test.ts (every deferred outcome, including edit_page, forget and source lifecycle changes, the idle-owner expiry and the owner-socket exit 10), test/commit-ledger.test.ts, test/doctor-foreign-writers.test.ts, test/sources-guard.test.ts, test/mcp-client.test.ts (a thin client's pending deferred write) and the state audit in test/refresh-barrier-state-audit.test.ts. Design note: docs/designs/REFRESH_WRITE_BARRIER.md.

[0.60.593.0] - 2026-11-06

gbrain repair failed-writes now previews a brain with a large write journal in about a second instead of timing out. On Postgres, the preview of refused writes read the request table once per refused write, and one of those reads walked the table from that write to its end. With hundreds of refused writes and a journal of over a million requests, the preview ran for minutes and the server's statement timeout ended it with a bare internal_error, so the operator had no way to see or recover the lost writes. The preview now classes refused writes together in set reads, each sized so it stays short even on a page with hundreds of refused writes and a long history. On a synthetic journal of 1.5 million requests with over 800 refused writes it took about a second (it took 265 seconds before), with the same classes and the same preview hash. When a preview read does still run past the timeout, it refuses repair_preview_timeout with a fix instead of internal_error, and nothing is saved.

Itemized changes

  • src/core/repair/failed-writes.ts: recovered fact requests are read once for all refused writes instead of inside the candidate read; replay attempts are read for every refused write together (request_id = ANY); the later requests of each refused write's page are read for groups of writes sized by how long their pages' later history is, so one page with hundreds of refused writes and a long history cannot make one read outlast the statement timeout; the file check runs once per page. The listing, classes, residuals and preview hash are unchanged.
  • New refusal repair_preview_timeout. When the request indexes the preview reads are missing or INVALID, it names them and its fix is the read-only gbrain repair request-indexes --json preview, whose apply command builds them; when they are valid, it tells the brain host to run the same preview once with a longer statement timeout.
  • Schema migration v291 builds two request-table indexes online (persistence_requests_replay_id on the request id, persistence_requests_slug_sequence on source, slug and sequence). They join the indexes gbrain repair request-indexes and doctor's persistence_request_indexes check report. Without them the preview is still bounded, only slower. PGLite and repo-less brains get them unchanged.
  • The request-index states doctor and gbrain repair request-indexes print now come back in the order the indexes are declared on Postgres as well as PGLite.
  • test/repair-failed-writes-bounded-8707.test.ts (PGLite and Postgres) seeds a journal with refused writes of every class and checks that the preview's request-table reads stay flat as refused writes grow, that each write gets its class, and that a read ended at the statement timeout refuses repair_preview_timeout and saves nothing, and (Postgres) that 500 refused writes on one page followed by 20,000 later commits preview inside a 1 s statement timeout.

[0.60.592.0] - 2026-11-06

On a page with no import record, converting a source to git-canonical, sources git-gap and repair rows-behind now pick the winner from content evidence, never from a timestamp, and nothing overwrites a page that changed after it was reviewed.

A page whose database copy differs from its committed file and that has no import record used to count as database-ahead. Converting a source to git-canonical then adopted it, so a plain Git edit of that page could be lost. Now its archived versions are compared with the blobs its path held in Git. An older version that equals the committed file means the database moved, so it is adopted. A page that equals an older committed blob means the file moved, so Git wins. A page where both sides moved, or where nothing is shared, waits for the owner's decision. A revert on either side counts as a move, and a history that was capped or pruned gives no verdict. The owner reviews a list, decides each page in a decisions file, or approves the whole plan once. Either way, only pages unchanged since the review apply.

Itemized changes

  • src/core/persistence/git-gap.ts: one direction classifier (adopt_db_moved, git_wins_file_moved, both_moved, no_evidence) reads each page's newest 32 versions and 32 blobs from one scoped history walk run from the Git top level. Renames are followed, version bodies are read under a byte budget, and a segmented page or a version with no comparable content gives no one-sided verdict. sources git-gap reports direction_evidence.
  • gbrain sources canonical <id> --to git: each page is placed by its evidence kind, and the kind is bound into the preview hash. both_moved and no_evidence are new discards (no_record_both_moved, no_record_no_evidence). --list-out <file> writes the review list, --decisions <file> applies the owner's decisions, and --approval-out / --approved apply an approval across time. Only items unchanged since the approval apply, and the rest comes back as git_conversion_delta with counts. These files are written 0600 and refused inside a source checkout or the Git working tree that holds it (conversion_file_in_source_root).
  • The conversion's strict passes may replace a database version only where the plan approved it. A page that changes during the apply is held instead of overwritten, the apply stops with changed_since_preview, and the source stays mixed. A failed apply reports recovery (mode, commits, push, writes and the non-destructive rollback).
  • gbrain repair rows-behind never re-imports a row with no import record whose database moved since (excluded:database_moved). --include-ambiguous also no longer re-imports a row with no import record where both the page and its file moved. Such a row is listed excluded:both_moved, and gbrain repair file-drift decides it.
  • sources git-gap and doctor db_ahead_of_git report each source's live_pages. A scan the time budget stopped reads PARTIAL, <read> of <live> live page(s) read.
  • The hardened git runner accepts a second log shape for the history walk and bare cat-file --batch-check; every other form still refuses.
  • Tests: test/git-gap-direction-evidence-d4.test.ts, test/git-canonical-conversion-evidence-d4.test.ts and test/rows-behind-checkpoint-8071.test.ts run on PGLite and Postgres. Guides: docs/guides/git-canonical.md and docs/guides/repair.md.

[0.60.591.0] - 2026-11-06

gbrain repair file-drift --prefer file now imports a page stored in segments that keeps protected parts, instead of refusing every apply with no next step.

A held page whose body is stored in segments (pages.segmented_storage, a page over 1 MiB) and whose database copy keeps a protected frontmatter key and a timeline row the file lacks previewed cleanly, listing both exceptions and an apply command. That apply refused every time with invalid_params "Unsupported or missing reconciliation fields." and no fix, so the page stayed held and blocked writes to it. The apply now imports the file with the same protections (the file keeps its bytes plus the appended key and row, the page keeps both), the page stays segmented and searchable, and the next sync releases the hold. A preview now refuses a page its own apply would refuse, and every refusal on this path names its next step.

Itemized changes

  • src/core/persistence/reconcile-state.ts: a reconcile preview of a segmented page validates: its snapshot's bodyGeneration and the page's body_storage / body_bytes are accepted (the preimage checks are unchanged). A page with no revision yet (the revision backfill has not reached it) refuses revision_backfill_pending in the preview, with gbrain apply-migrations --force-schema in fix (or gbrain repair orphan-children when the backfill already used every attempt on the row), instead of a stale-preview refusal with none; revision_backfill_pending joins the error registry. Read-only drift classification and the audit read such a page as before.
  • src/core/persistence/reconcile-prepare.ts: the reconcile publication of a segmented page leaves its text seal to the segmented import (as the managed file repair does), runs the import's afterCommit, and reads the published page back materialized.
  • src/core/repair/file-drift.ts: the preview validates each page's reconcile artifact as the apply reads it back, so a page the apply would refuse is listed refused in the preview. A refused preview row carries the refusal's fix and its why; a refused or changed apply outcome carries both, with a read-only preview of that page alone (for a held deleted-side page, its held preview) when the refusal has no fix. gbrain repair prints each row's fix with its consent.
  • src/core/persistence/reconcile-merge.ts: "Unsupported or missing reconciliation fields." names the fields in why.

[0.60.590.0] - 2026-11-06

A managed sync no longer stops for good when a write queues behind another write that waits on the sync, and the convergence nightly's kill loop tests only points a sync actually reaches.

A write queued behind an earlier write that waited on a sync row naming its page used to count as a write that could go first. That held back the very sync row the earlier write waited on, so the three rows waited on each other and nothing committed until the drain gave up, about an hour into the nightly run. The claim now skips such a write, so the sync row commits and both writes follow in order. The live stalled Ns on queued line also stops reporting a stall while the head keeps moving, for example while this process commits the in-flight writes of a serve that was killed.

Itemized changes

  • src/core/persistence/journal.ts: the FIFO claim no longer counts a write that sits behind an earlier queued write waiting on a sync row naming its page. That write can't go first anyway, because a write never passes an earlier one.
  • src/core/persistence/sync-drain.ts: the in-pass stall line counts an unclaimed head's age from when that head became the head, and never from before the drain's own last commit.
  • src/core/persistence/sync-run.ts and fault-points.ts: a new sync:passed_over fault point after the cursor step that passes over an entry whose file changed after admission. It does nothing without a test hook.
  • test/sync-convergence.serial.test.ts: the kill loop drops sync:prescreened and sync:retry_deferred, which only company sources reach since #6712 PR 4b, and kills at sync:passed_over and sync:batch_entries instead. A primed edit that the round's sync never froze is now reported by the "every planned fault fired" check, where it used to crash the harness on an empty commit.
  • test/claim-fifo-foreground-waiting.test.ts covers the chained write on PGLite and Postgres, and test/sync-drain-claims.test.ts covers a head that keeps moving.

[0.60.589.0] - 2026-11-06

gbrain can now wake an agent's conversation when a job it started finishes, through Wakeshift. It's off until the brain owner turns it on.

An agent that submits a long job no longer has to poll. With reporting back on, it calls watch_job with the job id and ends its turn. When the job completes, fails, dies or is cancelled, Wakeshift wakes that conversation with one line naming the job and the get_job call that reads the result. The wake is a pointer (the job id, kind, status and times), never page text, titles or slugs, so it can't widen what anyone reads. The owner turns it on with gbrain wakeshift enable on the brain host, using a key from their own Wakeshift app, and gbrain wakeshift disable turns it off at once. Until then nothing changes: the tools aren't listed, no rows are written and nothing is sent. This release reports back only on brains that serve their owner alone. wait_job and gbrain jobs wait <id> remain the way to wait with nothing leaving the machine. Guide: docs/guides/wakeshift.md.

Itemized changes

  • New MCP tools watch_job, list_watches and unwatch, listed only while reporting back is on and a usable key is on the host. watch_job runs the same owner check as get_job: a job the caller doesn't own reads like a missing one and nothing is sent. A watch belongs to the credential that made it, and a revoked credential's watch ends without a wake.
  • New gbrain wakeshift enable | disable | status | test. enable asks consent (it sends data off the machine and stores a credential), refuses a brain with live access members, checks the key with Wakeshift and publishes gbrain's gbrain.job.finished event. The key comes from WAKESHIFT_API_KEY or a 0600 ~/.gbrain/wakeshift.key, and is never stored in the database or printed (status shows its fingerprint).
  • Delivery runs inside gbrain serve (stdio and HTTP) and the job worker every 5 seconds, and once more at the end of gbrain jobs work. A watch on a job that already finished wakes at once. A failed delivery is retried at most 5 times over about 40 minutes, then the watch ends and the conversation is told. Jobs are never slowed by Wakeshift.
  • New doctor check wakeshift: a key file other users can read, reporting back enabled without a usable key, and watches that died or were refused as id_reused in the last 7 days.
  • Schema migration v290 adds wake_subscriptions (empty unless reporting back is on). Repo-less brains get it unchanged.
  • --log-full-params on gbrain serve --http now redacts a tool call's session_key argument.

[0.60.588.0] - 2026-11-06

Engines that make their first write at the same moment no longer fail one of the writes registering the same writer. Two or more engines opened from the same settings on one Postgres database (the same writer id, credential and lane), or two processes that share one credential file, could both find no registration and both insert it. The loser hit the credential's unique index and failed with duplicate key value violates unique constraint "persistence_local_writers_credential_hash_key"; four engines on a fresh database lost three of their four first writes. Registration is now idempotent under that race, and every refusal still holds. PGLite was never affected, because its single connection runs the read and the insert one after the other. Contributed by @bradgessler.

Itemized changes

  • src/core/persistence/identity.ts: the registration insert names no conflict target, so a conflict on either unique index (the id or the credential hash) does nothing. When nothing was inserted, the row under the id is read again and checked exactly as an existing row is. The same credential and lane pass, while a revoked row or another credential or lane under the id refuses permission_denied. No row under the id means the credential belongs to another registration, which now refuses permission_denied with a fix where it used to raise a raw 23505.
  • test/e2e/engine-open-postgres.test.ts: four engines opened from one writer value make their first writes at once, twelve rounds on a scratch database (master: ['23505','resolved','resolved','resolved']). A second case forces the lost-insert branch and checks each refusal on Postgres.
  • test/engine-open.test.ts: the same lost-insert cases on PGLite, plus four concurrent first writes on one engine landing under one registration.

[0.60.587.0] - 2026-11-06

An HTTP MCP caller can now say which AI made a write, and attribution reads show it. Send X-Gbrain-Agent-Label: Claude Code (or Codex, or a job name) on a request, and the writes that request makes itself show agent_label in gbrain attribution / get_write_attribution and in get_versions for admin and local readers. This answers "which of my AIs wrote this?" when several share one OAuth client or token. The label is text the caller declares and gbrain does not verify. The principal is still who was authorized, and no grant, access rule or trust tier reads it. A request without the header is recorded exactly as before, so nothing changes until a caller sends it. Contributed by @JamisonMercurio (#8664).

Itemized changes

  • Both HTTP transports (the OAuth serve --http mount and the legacy bearer transport) read the header. Admission stores it on the request's existing authority record as agentLabel, so there's no new column or migration, and repo-less brains record it too. Stdio MCP and the local CLI have no request headers and record no label.
  • A label is 1 to 120 printable ASCII characters after trimming. Anything else (control, bidi, zero-width or non-ASCII characters, or a longer value) is dropped, never refused.
  • Jobs a labelled request queues (submit_job remote jobs and agent jobs) run under their own authority and record no label.
  • Both transports' CORS preflight now allows the header, so a browser client on an origin in GBRAIN_HTTP_CORS_ORIGIN can send it without its whole MCP call being blocked.
  • test/write-agent-label.test.ts covers both transports, a repo-less brain, the parser bounds, a stored authority that is otherwise identical with and without a label, and a read-only token that is still refused. The export-surface golden gains the optional agentLabel and agent_label fields.

[0.60.586.0] - 2026-11-06

A remember refused for its arguments now names the whole valid call, so an agent that never read the schema saves on its next try.

An agent on a harness that lists MCP tools as one line and shows parameters only on request sent remember its host's own memory shape, facts: [{text}]. gbrain answered "did you mean fact?", the agent resent {fact, entity} without provenance, and only the third call saved, so every saved fact cost two refusals. The refusal now points facts at items and text at fact, and quotes a valid call with placeholder values, provenance included. provenance stays required.

Itemized changes

  • src/mcp/validate-params.ts, src/mcp/dispatch.ts: an unknown facts on remember suggests items and an unknown text suggests fact; the refusal then appends the operation's callShape, a valid call as JSON with placeholder values (<one claim>, <subject-slug>, <where you learned it, with the date>) that says provenance is required. callShape is a new optional Operation field; only remember sets it.
  • src/core/verbs.ts: remember's description opens with the call itself, {fact, provenance (required: source), \entity` (subject; recall needs it)} or items[], inside the 117 characters a one-line tool listing keeps. A missing fact` quotes the same placeholder call instead of an example without provenance, and the missing- and empty-provenance refusals quote its provenance placeholder instead of a realistic source an agent could copy.
  • src/core/remember-batch.ts: an item that carries text is told to rename it to fact.

[0.60.585.0] - 2026-11-05

CI shard weights are re-mined from the 2026-11-05 nightlies, so the unit and serial lanes pack their shards on current timings. Nothing changes for users.

Itemized changes

  • scripts/test-weights.json and scripts/serial-weights.json: each test file's weight is its median over the five newest green Test nightlies, 2026-11-01 through 2026-11-05; scripts/*-weights.metadata.json names the runs. The heaviest of the 28 unit shards is predicted at 538 s against the 600 s budget (537 s before). Test files with no weight drop from 61 to 36 of 4547 in the unit lane and from 4 to 1 of 457 in the serial lane.
  • scripts/e2e-weights.json is unchanged: the 2026-11-05 full-corpus E2E nightly failed, and the newest green one already sourced the map.
  • test/pathological/a18-ledger-path-text.test.ts and test/pathological/s-70-classifier-origin.test.ts: the injected import failure picks the import's page read by its arguments instead of by an importOnePath stack frame, so it still fires when an earlier file in the same process opened an embedded engine. The new packing runs test/engine-gateway.test.ts before a18 in one unit shard, where five of its cases failed every time.

[0.60.584.0] - 2026-11-05

gbrain sources refresh --abandon no longer lets the next sync delete pages a member source already synced.

A refresh interrupted after a member source synced the upstream commit, then moved to recovery_required with the checkout back at the commit it started from, could be abandoned: --resume refused and named --abandon, --abandon lifted the fence, and the next sync diffed that member back to the start commit and soft-deleted every page it had synced. --abandon now refuses such a row (refresh_recovery_required, nothing changed) once any member's last_commit is past the start commit (set, and neither that commit nor an ancestor of it; a commit git cannot find counts as past), and --resume no longer names --abandon there. Both name the one way out that keeps the pages, as an ask-the-user fix. It previews with git -C <checkout> status --short, fast-forwards the checkout with git merge --ff-only to the commit the members synced (the upstream commit, or a later one a member already synced), then runs --resume; when git refuses, it names git reset --hard to that commit as the destructive fallback. --resume now also finishes a recovery_required refresh whose HEAD descends from the upstream commit instead of refusing it, and never finishes one at a HEAD that does not contain the commit a member source synced on this refresh's line (or a commit git cannot place). Writer status uses the same member check. A row whose members never moved past the start commit is abandoned exactly as before, and sync's deletion behaviour is unchanged.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: one member check (membersPastStart) feeds --abandon, the --resume refusal at the start commit, the stored recovery_required fix and writer status.
  • Tests: test/persistence-worktree-refresh.test.ts 7r (the three-step chain: crash while syncing, member synced, HEAD reset; pages survive, then reset-to-target and --resume completes) 7s (no member moved: --abandon still works) 7t (a member behind the start commit still abandons; an unknown commit refuses), 7u (a member synced past the target keeps its pages), 7w and 7x (--resume never finishes at a HEAD that lacks a member's synced commit), 7y (a member already on rewritten history before the refresh never leaves crash recovery stuck) and 7v (an unrelated HEAD names no --abandon; a HEAD past the target resumes), on PGLite and on Postgres through test/e2e/worktree-refresh-postgres.test.ts.

[0.60.583.0] - 2026-11-05

Sync and source-refresh follow-ups from wave #8123: an interrupted refresh names what holds it and how to finish it, managed-sync preparations stop waiting on table locks, code sources converge on their first full sync, and sync.write_through=false, impossible timeline dates, image names, dry runs and missing source directories behave as documented.

gbrain sources refresh says what is actually holding an interrupted refresh, and its owner recovery stops holding locks it doesn't need (#8123).

Writer status for a recovery_required refresh now reads the checkout HEAD. At the upstream commit it names --resume, at the start commit --abandon (only when no member source has synced past it), and only another HEAD asks the user to reset. An orphan whose HEAD git cannot read names git -C <checkout> status instead of promising that the next refresh finishes it, and the owner scan logs it once every 15 minutes instead of swallowing it. A refresh lock file this host cannot open reads liveness: unknown with the lock path, and a plain refresh refuses refresh_in_progress instead of failing with a raw lock error. A refresh held on another host names that host in next. A plain refresh that rolled back an orphan and then refused keeps that recovery in its refusal message. A delegated refresh (a resident PGLite owner) prints the same structured fix as a direct one.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: the owner scan checks member progress before it takes the refresh lock, caches a syncing orphan it could not complete because of its HEAD, fails one row (not the scan) on a lock error, and prunes its caches each scan (#7712, #7728, #7730, #7731); refreshHolderReport reads HEAD (#7712, #7731); next is host-scoped and shell-quoted (#7731).
  • src/core/persistence/ipc.ts: a refusal's fix, why, reason and contract_version survive the owner socket (#7707, #8130).
  • Tests: the #7679 state matrix gains the --abandon action for every kind (#7718-#7727, #7729), plus red tests for the owner scan's syncing filter and test 7's adoption guard (#7708) and the unreadable-lock branch (#7728).

A table lock on pages no longer pins a managed sync's preparations past their budget, sync and waiver provenance records name a file only when it holds the bytes they record, and the rows-behind doctor check holds at most 64 MiB of committed files per batch (#6318, #8089, #8093, #8094).

A managed sync bounds each preparation by persistence.sync_preparation_ms. Its raw reads already ended on the server at that budget, but the engine's own reads (the page snapshot, the import pipeline's duplicate and content-hash lookups) ran as plain autocommit statements, so under LOCK TABLE pages held elsewhere they waited as long as the lock did, through a pooler that drops the session statement_timeout. In a managed sync those reads now run in a transaction whose statement_timeout is the time the preparation has left, and a read the server ends there is the member's preparation_deadline, held as preparation_stalled like the head member. Nothing changes without a deadline, inside a transaction, or on PGLite.

Itemized changes

  • src/core/read-deadline.ts (new), src/core/postgres-engine.ts withScopedReadTransaction, src/core/persistence/bounded-reads.ts: a call through the managed-sync preparer's bounded engine view carries the clock's deadline, and the Postgres engine's scoped reads issued under it run on a reserved connection as BEGIN READ ONLY; SET LOCAL statement_timeout, the read and COMMIT, pipelined (runDeadlineRead in src/core/postgres-engine/bounded-statement.ts) (#6318). Page and maintenance preparations keep plain engine reads (#8307). Probe: test/e2e/persistence-preparation-engine-reads-lock-postgres.test.ts.
  • src/core/persistence/sync-prepare.ts (publishedSyncNoop, new heldDiskBytes), src/core/persistence/sync-waivers.ts (waivedProvenance): a published no-op re-reads its file at apply and records provenance only when the file still has the raw hash; a waiver records the file it read only when it holds the imported bytes (up to line endings), the rule syncDiskBytes already follows (#8089). Probe: test/sync-provenance-writers-8089.test.ts.
  • src/core/persistence/rows-behind.ts: doctor sync_rows_behind_checkpoint reads and compares a batch's committed files in chunks of at most 64 MiB (ROWS_BEHIND_BATCH_BYTES), drops each chunk's buffers after the comparison (a differing row keeps a copy of its bytes, so it never pins its chunk), checks the deadline before the per-source tree reads and each chunk, and reads its screen config only once a batch starts (#8093 item 1). The docstring names not_eligible (#8094).
  • test/cycle/dream-reverse-write-commit.test.ts: a refused commit in one source's checkout holds only that source's page (#8094).
  • scripts/bench/managed-sync-stall-repro.ts: --sync-preparation-ms and a lock_release result listing sync statements still waiting on a lock 3 s past the budget.

A first gbrain sync --full on a code-strategy source now settles in one pass, and a plain gbrain sync status no longer issues one database statement per abandoned uncommitted edit (#8047, #7339).

A code or image import recorded its provenance without the open --full episode, so every round of the pass saw those files as still owed and imported them again. A fresh code source reported twice its file count, stopped not_converged three passes in a row and then held every file settle_mismatch. Code and image imports now record the episode the way Markdown imports do, and the pass converges in its first round.

sync status re-reads each worktree_dirty hold that is due to escalate, to see whether someone is still editing the file. It read every such hold's row even when the file's bytes still matched the hash the hold already recorded. On a source with thousands of abandoned edits that was thousands of round trips per status call, about minutes over a remote pooler. It now reads a row only when the bytes changed. With 40 such holds, a plain status went from 54 statements to 15, the same count as with one hold; an edit whose bytes changed is still noticed and stops counting as abandoned.

Itemized changes

  • src/core/persistence/sync-prepare.ts: syncFileProvenance (code and image imports) records full_episode (#8047). The same line is in contributor PR #8051; this fix was derived independently from the #8047 reproduction.
  • src/core/persistence/sync-status.ts: observeDirtyEditors skips noteHoldEdit when the working-tree hash equals the hold's recorded working_hash (#7339).
  • New tests: test/sync-full-code-episode-8047.test.ts, test/sync-status-hold-reads-7339.test.ts.

Managed sync and source checks fixes from the sync/refresh wave (#8123 part C): sync.write_through=false keeps sync's frontmatter corrections out of the checkout, a page with an impossible timeline date no longer blocks sync, images with punctuation in their names sync, sync --dry-run previews what it would do, and a missing source directory is no longer reported clean.

  • With sync.write_through off, a managed sync keeps its canonical correction of a file (the frontmatter, title or tags a page keeps when an upstream edit drops them) in the database and never rewrites the checkout. Before, it wrote the rendering into the working tree and left it uncommitted, including into a read-only mirror's checkout reached through the default source's sync.repo_path fallback (#7084, #8118).
  • A timeline bullet or heading dated with an impossible day (2026-02-30, month 13) is no longer extracted as a timeline row. It used to reach the database's date cast and fail the page's publication, which blocked a managed sync and aborted a whole repair file-drift preview batch (#8049). The text stays in the page.
  • Managed image sync now gives an image the importer's identity (path lowercased, punctuation kept), so notes/Photo+A.png imports as notes/photo+a.png instead of being refused, and stays distinct from notes/PhotoA.png (#7107).
  • A managed gbrain sync --dry-run reports and prints how many files the run would add, modify, soft-delete and rename. It used to report zero, or repeat the previous run's counts (#7087).
  • gbrain sources list marks archived sources with ⚠ archived and --json carries archived per source (#7701).
  • A registered source whose directory is missing is reported missing by the frontmatter audit and both frontmatter doctor checks, with the path and the restore or sources set-path step, instead of clean; a connector source's leftover path is not reported (#7889).

[0.60.582.0] - 2026-11-05

A frontmatter-only disagreement between a page's file and its database copy no longer blocks an unrelated remember, a contact row whose last_used moved forward reconciles automatically, and duplicated contact rows can be collapsed with one previewed repair (#8109).

What changed

Situation Before Now
A person page's file differs from the database only in frontmatter (for example one contact row's last_used moved forward), and an agent saves a fact remember refused source_changed (file_database_drift) until someone reconciled the page the fact lands; the file keeps its own frontmatter with the new body, and the frontmatter drift stays for reconcile. Body, timeline, title, type and tag drift still refuse, as does drift in a policy, privacy, lineage or trust key (visibility, readable_by, derived_from, trust_tier, ...) and a write that changes frontmatter
gbrain sources reconcile <source> <slug> --preview --auto-additive on a contact row whose only change is a later last_used review (list_entries_changed_or_reordered) auto (rule contacts_activity_advance); a removed, reordered or otherwise changed row, or a date moved backward or into the future, still needs review. Unquoted YAML dates compare as dates
Pages whose frontmatter contacts: list repeats a (type, value) row no repair gbrain repair contact-rows [--source <id>] [--slug <slug>]... keeps one row per (type, value) with the newest last_used (rows that differ in another field and every distinct value stay; only real date text counts as a date). A page whose contacts is not a list is skipped. It previews (writes nothing), then --apply --expect <hash> --yes rewrites exactly the previewed pages through revision-bound put_page: the file through the coordinated writer on sources with a checkout, the database on repo-less brains. Drifted pages are listed needs_reconcile with the reconcile command

Nothing runs on upgrade.

[0.60.581.0] - 2026-11-05

CI shard weights are re-mined from the 2026-11-04 nightlies, so the unit, serial and full E2E lanes pack their shards on current timings. Nothing changes for users.

Itemized changes

  • scripts/test-weights.json and scripts/serial-weights.json: each test file's weight is its median over the five newest green Test nightlies (41792242983, 41596314589, 41406386795, 41284734181 and 41180589845). The heaviest of the 28 unit shards is predicted at 536 s against the 600 s budget (523 s before). Test files with no weight drop from 65 to 50 of 4536 in the unit lane and from 7 to 4 of 457 in the serial lane.
  • scripts/e2e-weights.json: re-mined from the full-corpus E2E nightly 41784138099. Test files with no weight drop from 19 to 17 of 864.
  • test/pathological/s-classic-file-holds.test.ts: the unknown-error case picks the import's page read by its arguments instead of by a stack frame, so it still injects its failure when an earlier file in the same process opened an embedded engine. The new packing runs test/engine-open.test.ts before it in one unit shard, where it failed every time.

[0.60.580.0] - 2026-11-05

Trunk Dynamic CI on the pull-request tier, in Learning mode. A dynamic-ci job asks Trunk which of the slow, nearly always green pull-request jobs a push still needs; every job still runs until enforcement is switched on in Trunk.

Pull-request CI

  • test.yml gates security-regressions (with its Windows and macOS cells), native-locks, persistence-validation and stress-changed-tests, and scale-tier.yml gates scale. Together they used about 40,000 runner-minutes on pull requests in the week to 2026-11-04.
  • The filter runs only on pull requests and fails open: no token, a Trunk outage or a job without a verdict runs the job. Pushes to master, the merge queue, schedules and the full-ci dispatch skip the filter and run every job as before, so full-suite, e2e-full-suite and ci-reuse are unchanged.
  • cheap-tier accepts a gated lane that skipped only when Trunk's verdict for it was exactly false.
  • New guard check:dynamic-ci (in bun run verify) fails when the filter can run on another event or a gated job loses the fail-open !cancelled() && != 'false' form. See docs/ci-event-parity.md.

[0.60.579.0] - 2026-11-05

The nightly E2E upgrade check passes on keyed runs again. The upgrade-from-previous-release harness checks that a brain and keyless config file written by the previous release resolve the same gbrain config show output under this build. The keyed nightly E2E lane keeps OPENAI_API_KEY and ANTHROPIC_API_KEY set, the harness's CLI inherited them, and config show printed two redacted key lines the golden doesn't list. The harness now runs its CLI keyless, as the install it models does. Nothing changes for users. A keyless install resolves exactly the golden, and a keyed one shows its keys as before.

Itemized changes

  • test/helpers/upgrade-from-previous-release.ts: the harness's CLI child runs with every provider key and base URL in PROVIDER_ENV_KEYS removed. The golden is unchanged. On real Postgres, master failed 3 of 12 with fake keys set (the same two lines as the nightly run) and passed keyless; with this change it passes 12 of 12 both ways (#6040).

[0.60.578.0] - 2026-11-05

Memory train (GBRA-286): forget edits the page files first, recall can weigh how far a fact deserves belief, page versions carry a human reason and can be read by date, and agents can read and switch the saving rule. Every new mode is off by default, and every feature works on a brain with no Git repository.

This release folds seven approved memory PRs into one: #7616, #7791, #7797, #7803, #7810, #7798 + #7804 and #7795. Each was brought up to current master. Where a new path skipped a guard master added since (member-scoped access with readable_by, the trust-tier quote fence for remote readers, the private-request hold on content imports), the guard now applies there too, with a test that failed first.

Forget writes the page files first (#7616)

Git stays the parent. A withdrawal is now a page write, published to Git before the database changes.

Situation Before Now
forget / forget_fact the database expired the fact; the page files caught up later through a mirror every page whose ## Facts table carries the claim gets its row struck (~~claim~~, a valid_until date, forgotten: <reason>), all in one publication and one commit, then the ledger row and expiries in the same transaction
remember without an entity a database-only row a row of the memory/unattributed page's ## Facts table, created on first use (concurrent first saves all land on it), so Git, recall and search see it
Remembering that claim again with an entity two copies the entity row is written and the unattributed row struck in the same publication
A member forgets a fact on their own member page withdrawn for that page unchanged: the withdrawal stays on that page, and other members' copies stay active

forget_fact on a fact that already expired (superseded, consolidated or dropped from its page) still records the withdrawal and strikes the page row, so a later import can't bring it back, then answers fact_already_expired. A forget now waits for the source's writer the way remember does. With the writer unavailable it returns the pending-write receipt, and the fact is still recalled until it publishes. With sync.write_through off, from a sandboxed subagent, or on a source with no canonical owner, a forget of a claim a page file carries is refused (sync_write_through_off, subagent_sandbox, unbound_source) instead of withdrawing it in the database only; a claim only database-only pages carry is still withdrawn. Migration v287 adds the multi-page recovery record (one record restores every file of an interrupted publication); it rewrites no data.

Review lists and review decisions through the page door (#7791, #7797)

  • extraction_pending gains two lists. unattributed (on by default) lists active facts saved without an entity, each with the relink action. possible_duplicates (opt-in through include) flags new entity pages whose name is a typo or acronym of another's. It only flags: nothing is merged and no identity is written.
  • extraction_review promote and reject are now page-door writes (edit_page of the frontmatter, delete_page), so on a source with a repository each decision is a Git commit. A new distinct action records that two flagged pages are different entities.
  • Remote reads of the unattributed list hide facts the caller can't read, and quote fact text below the owner's tier as source material, as recall does.
  • A promote or reject batch keeps going when one slug fails, returns partial with each slug's outcome, and gbrain extraction-review exits 1 for a partial batch.

Recall weighting, entity grouping and the keyless claim lane (#7803, #7810)

recall can weigh how much a fact deserves to be believed, and can group a no-entity question's facts by entity. Both are opt-in.

What changed

Situation Before With the switch on
recall {question}: a year-old fact you confirmed and a fresh agent-written or external fact match equally the newest fact ranked first the confirmed fact ranks first (recall.confidence_weighting on)
The same, between two agent-written facts newest first newest first (age still decays)
A dated historical fact ("founded in 2015") ranked by match ranked by match; it does not decay
recall {question} with no entity answered by two clients' facts one flat list, the clients' figures interleaved each entity's facts together, facts_groups, and an entity_unspecified notice offering the recall per entity (recall.group_by_entity on)
gbrain config set recall.question_admission <typo> stored and silently ignored refused; the new switches take only on or off

Weighting is relevance × (0.5 + 0.5 × weight): facts you confirmed and your own notes keep full weight at any age, other tiers decay from valid_from and carry a tier weight (tool 0.9, agent 0.8, external 0.6). It only reorders facts that already matched; relevance stays the plain score and rank_score and confidence_factor show the weighting. Grouping runs after the limit and budget cuts, so no entity is cut off by it. Both read database rows only, so they work the same on brains with no git repository.

On a keyless 60-question fixture (old confirmed truth against fresh agent and external text, an agent knowledge update, dated history), weighting moved the confirmed truth to first in 20 of 20 questions, from 0 of 20, and left the update and history cases at 20 of 20. Grouping removed entity interleaving in 10 of 10 two-client questions with the same number of entities delivered.

Typed figures you save can now be checked for conflicts without an API key, and you decide each conflict from the trust review.

remember takes an optional typed claim (claim_metric, claim_value, claim_unit, claim_period). Two saved claims about the same entity with the same metric, unit and period in the same as-of bucket (the day, or the week, month, quarter or year the period names) are a conflict. Figures in different buckets are history, so a monthly series, a run of funding rounds or a forecast next to an actual are never flagged. Finding conflicts is off by default: gbrain config set dream.claim_conflicts.mode propose turns on the nightly phase, and gbrain decide sweep --claims runs it once (--dry-run writes nothing). Each conflict becomes one pending proposal that nothing applies until you decide it.

What changed

Situation Before Now
remember a figure text only optional typed claim, stored on the fact and in the entity page's ## Facts table; a bad value is refused with the corrected call
Two saved figures disagree for the same month found only by the paid conflict sweep found by gbrain decide sweep --claims or the claim_conflicts dream phase (propose), with no key and no model call
A typed figure saved for April next to March's could supersede March's if the text was close both kept: a typed write never supersedes a typed fact in another bucket
gbrain trust review proposals, lowered pages, preferences, holds also contradiction items (dp<id>) with supersede, forget, confirm and drop; --accept-all skips them
recall of a fact in a pending decide proposal no label contested with the other fact's id, only when the caller can read that fact

A run files at most dream.claim_conflicts.max_pairs (default 500) proposals per source and leaves the rest for the next run. A new unique index keeps concurrent runs to one proposal per pair (migration v288); it rewrites no data. Everything reads and writes database rows, so it works the same on a brain with no git repository. Guide: the "Old and conflicting facts" section of docs/guides/memory-trust.md.

Page versions carry a human reason, and pages read by date (#7798, #7804)

A page write can say why it was made, and page history shows it.

put_page, edit_page, revert_version and each put_pages item take an optional reason (one line, at most 500 characters; --reason on the CLI). It is stored on the version the write replaced, and get_versions returns it as replaced_because, so the newest version explains the text you read now; gbrain history prints it as changed because: …. When the page's file reaches Git, the commit carries it as a trailer naming the file, gbrain-reason: <path>: <text>, beside any gbrain-repair: trailer. A brain with no Git repository gets the same history from the database. Every write result reports where the reason went in reason_recorded.

What changed

Situation Before Now
A write passes reason refused as an unknown parameter stored with the replaced version and, for a world-visible file, in the commit
A reason over 500 characters, or one that looks like a credential n/a refused before admission: change_reason_rejected (reason_too_long, reason_contains_secret), and no journal row keeps it
A remote caller reads get_versions no reasons replaced_because when the version and the one that replaced it are both world-visible and the replacement was not pruned
A private page's write has a reason n/a the reason stays out of every commit message
A write without a reason the same commit message and history as before
A newer write to the same file commits first (non-strict source) n/a the receipt says git: superseded with a reason_not_committed notice; the reason stays in history

Guide: page history.

Read a page as it stood on a date

A page can now say when its text takes effect, with valid_from and valid_until in its frontmatter, and you can ask what was in force on any date: get_page {slug, as_of: '2026-03-01'}, or gbrain get <slug> --as-of 2026-03-01. Say to your agent: "What did our refund policy say on March 1?"

  • The answer is the edition with the latest valid_from on or before the date, picked across the current page and every version gbrain kept. Its valid_time block says whether that edition was in_force, not_in_force, not_yet_in_force or before_recorded_history, and whether gbrain held enough history to be sure. Only an in_force answer returns the text at the top level, so an agent that skips the status cannot quote expired text as current.
  • When history is missing (pruned versions, a sync that imported several commits at once, a page that came from a file) or the dates look ambiguous (a corrected date and a renewal leave the same trace), the answer says uncertain: true, why, and how to recover: the file's git log, or asking the user for a backup. It never guesses.
  • Works the same on a brain with no Git repository, the default keyless install.
  • Over MCP, an as_of read follows the plain read's rules. A quarantined page's body is omitted, and an edition's introduced_because is shown only when the edition it replaced is one the caller may read.
  • gbrain get --as-of exits 15 when no edition was in force, and 16 (host_upgrade_required) when the brain host is too old to answer by date, printing nothing historical.
  • Writes say when they keep or change an edition's date (valid_time_correction, valid_time_date_only, valid_from_unparsed), and get_versions rows carry each version's valid_from and valid_until.
  • Behavior change: a plain get_page of a page whose own valid_from is in the future, or whose valid_until has passed, now also carries a valid_time status. Pages without those dates read exactly as before.

Guide: page history.

Memory writeback settings over MCP (#7795)

Agents and settings screens can now read the memory-saving rule, an admin can change it remotely, and a user can say "save everything from here" or "stop saving" for one conversation.

Before, the rule for what an agent saves lived in the MCP instructions, and only the brain operator could see or change memory.auto_writeback, from a shell on the host. Three MCP tools change that. Every one keeps its state in the database, so a brain with no git repository gets all of them.

What changed

Tool Who can call it What it does
get_memory_settings any connection the mode agents follow, where it came from, the exact rule text, every mode's rule, background capture state, and why you can or can't change it
set_memory_writeback admin scope, after gbrain config set mcp.publish_memory_settings true (off by default) changes memory.agent_writeback (inherit, off, salient, all) for the whole brain, with a dry run, an audit row in the same transaction, and a caution on team brains
memory_conversation_mode any connection that can call remember returns the rule for the rest of this conversation: all, salient, off (only what the user asks to keep) or brain; stores nothing

The new memory.agent_writeback key changes only what agents are told. The host capture lanes (Stop hook, compaction harvest, SessionEnd transcripts) keep following the operator's memory.auto_writeback, so a remote "off" can't discard banked session text and a remote "all" can't start capture the operator never enabled. An unreadable or unrecognized override makes agents act as if saving is off and get_memory_settings reports degraded. The MCP instructions gain one line telling the agent to use the switch, after the governing rule, wherever the tool is served (starter and full surfaces). gbrain config set|get memory.agent_writeback works from the CLI.

[0.60.577.0] - 2026-11-05

A write a refresh fence refused is never reported as success, and an interactive write waits the fence out. While gbrain sources refresh drains, fast-forwards and syncs a checkout, writes to it are refused worktree_refreshing. That refusal was unregistered (class server, fix: gbrain doctor), remember with items whose every item it refused printed undefined (fact #undefined) and exited 0, and the same batch over MCP returned isError: false. The refusal is now retryable with fix.next: wait and retry_after_ms, an interactive write retries it with backoff for up to 10 s before refusing, and a batch that saved nothing is an error on every surface (#8621).

Itemized changes

  • src/core/error-registry.ts: registers worktree_refreshing (class retryable, actor provider). src/core/persistence/journal.ts: the fence refusal carries retryable: true and a read-only gbrain sources writer status --source <id> --json fix, so the envelope renders next: wait. A retried admission whose request was already accepted (its acknowledgement lost before the fence) reconciles to that receipt instead of being refused as not recorded.
  • src/core/persistence/admission-retry.ts: waitOutRefreshFence re-runs an interactive write's whole admission (250 ms to 2 s backoff, 10 s cap shared by every item of a remember batch, its clock starting at the first refusal; an explicit wait_ms on put_page or put_pages lowers the cap, so wait_ms: 0 is refused at once); nothing is held across a pause and a refused attempt journals nothing. Past the cap the refusal is returned with how long the call waited. Background writes and the managed sync keep their own retries; database-only sources are never fenced. Wired into admitWrite, admitBatch (put-pages) and forget.
  • src/core/persistence/verb-errors.ts: remember and the other memory verbs keep their frozen unavailable code for the fence and add canonical: worktree_refreshing, its detail, fix and retryable.
  • src/core/remember-batch.ts: a batch whose every item failed throws the first item's error (items[<i>]: ... None of the <n> item(s) was saved.); a batch with any failed item (or an item whose outcome is unknown, which keeps its child request_id) adds status: "error" and the first failed code, so the CLI exits 1 and MCP returns isError with every item's receipt (src/mcp/dispatch.ts). src/cli/remember-format.ts prints a batch as saved X of N fact(s) with one line per item; a failed item keeps its canonical code and retryable flag. The batch body survives the resident owner (src/core/persistence/provider.ts) and the thin client (src/core/mcp-client.ts), and a thin-client CLI exits 1 on it too. A batch error leaves the one-shot MCP notices for the next call.
  • src/core/persistence/ipc.ts: an error from the resident owner keeps its canonical code, detail, retryable and (without a receipt) its fix, so a fenced write through a running gbrain serve still reads worktree_refreshing with next: wait.
  • test/write-refusal-refresh-fence-8621.test.ts (new) pauses a real refresh at its drained boundary on PGLite and Postgres and covers MCP, the wait, the cap, background writes, a database-only source and the CLI exit codes.

[0.60.576.0] - 2026-11-05

A database-only source exported by the 0.53.0 migration now syncs. The export installs the shared memory pack (skillpack.json and skills/*/SKILL.md) into the new plain directory, and gbrain sync used to refuse each pack file with skill_bundle_required, so the source stayed blocked_by_failures and doctor's sync_freshness failed for good (#8620).

Situation Before Now
gbrain sync --source <id> on a plain-directory source of a managed brain that holds the pack three skill_bundle_required refusals, blocked_by_failures, no successful sync ever recorded reserved skill paths stay with the shared skill publisher, as in managed Git sync; the source syncs and records last_sync_at
--dry-run on that source counted the pack's files as "would be imported" leaves them out of the count
Shared skills activated before the DB-only export the v0.53.0.0 checklist didn't say the new root needs the activation again the checklist says to rerun the --shared-skills activation after the export binds the root, then resume the export

The importer still refuses skill paths handed to it directly; classic (unmanaged) brains import a plain directory exactly as before.

[0.60.575.0] - 2026-11-04

Train R, seven pull requests in one release. Searches that name a page by an identifier (#412, INV-2026-00417) or a handle (dan-example) now find it. New gbrain find / find_pages finds a page by name, path prefix or glob, and gbrain stats --by source,type counts pages by kind. query --explain-target --json carries its diagnosis. There's an opt-in keyword phrase factor (off by default), an advisory harness check on pull requests, new docs, and guardrails for the embedded core.

Identifiers and handles in search (#7491, #7492, #7497)

A question that names a page by an identifier ("What did the latest weekly digest (#412) say?") now returns the page holding #412, even when the question's other words keep it from matching every word and near-identical template pages tie with it on meaning. A handle such as dan-example now finds the page it names (people/dan-example) in keyword search, quoted or not. The coverage notice also stops blaming indexing for unrelated image pages.

Identifiers in questions

Situation Before Now
No chunk holds every word of the question, and the any-word retry ranks the page holding #412 first the retry's rows are dropped before fusion whenever semantic search returns anything, so on a brain of near-identical template pages an older copy won the keyword arm runs one more statement matching every identifier the question names, with the same scope, visibility and filters as the strict search; its rows vote like strict rows, and a title holding the identifier leads the title arm
Long pages repeat the question's common words ("say", "report") they outranked the short chunk holding the identifier once, so the page fell out of the keyword arm the identifier match comes first, and the any-word retry ranks rare words first for questions of up to 12 words (it was 4)
One unrelated chunk happens to hold every word, the number included that chunk was the only keyword row; the page holding the identifier never entered the arm the identifier rows follow the strict rows, so the page is a keyword candidate again

An identifier is a # number, a run of 6 or more digits, a code such as INV-2026-00417, an email, a URL or a phone number, asked together with other words. Queries without one, a quoted phrase or a - negation run exactly as before, and the any-word rows that don't hold an identifier are still dropped while semantic search is healthy.

Handles in keyword search

Situation Before Now
gbrain search '"dan-example" OR "erin-example"' --match keyword, where the person pages hold the handles only in their slugs the pages that mention the handles came back, never the pages they name: a hyphenated handle only matches text holding the literal handle the pages whose slug ends in each handle (people/dan-example, people/erin-example) are listed and counted first
The same query with --types '["person"]' no results the two person pages
The same query in hybrid search quoted terms skip the any-word retry, so the title arm found nothing either the named pages lead the keyword arm, as rows that count like strict keyword rows
A search that matches nothing while every retrieval stage ran the CLI said "clean miss — no retrieval degradation" and MCP "this is a clean miss", which agents read as "the brain has no page for this" "retrieval healthy", followed by the next things to try (the display name, a synonym, hybrid search) before concluding the brain has nothing; the query skill says the same

A handle is a hyphenated token with a letter in it, bare or quoted on its own (acme-example-2 counts, 2026-10-24 does not), and resolves only to a page whose slug ends in exactly that handle. A query with a - negation or a longer quoted phrase resolves no handle, and --match keyword resolves handles only when every term is one (or OR), so a handle never brings back a page the query excludes, a phrase it doesn't hold or a word it doesn't contain. If the identifier or handle lookup fails, the result is stamped degraded (keyword_candidates_incomplete) rather than reading as a miss. The named pages pass the same filters as any keyword match: --types, the caller's sources, and for remote callers no private or quarantined page.

Coverage notice

Situation Before Now
A pending image page's slug contains the query's digits (412 in frames/t004120) counted as able to match the query a query word counts only as a whole word of the title or slug
Pending image pages and a text-only search counted, although text search never returns image pages counted only when an image search arm ran
The projection_pending part of degraded_recall "recently written pages are not in the search index yet" for every page text pages "are being indexed"; image pages "will not be indexed on their own" and wait for gbrain projections drain or a paid re-import, with doctor's text_projection_readiness check as the fix

Say to your agent: "What did the latest weekly digest (#412) say?"

What to watch for after you upgrade

  • A search naming an identifier costs one more keyword statement, served by the full-text index. Other searches run the same statements as before, except that an any-word retry of 5 to 12 words now also runs the bounded rare-word read and, when a rare word is found, one more retry statement (both already ran for retries of up to 4 words). The rare-word read skips stopwords, and on Postgres it skips words the column statistics already estimate as common, so its cost follows the question's rare words, not its length.
  • A query with a hyphenated token costs one more index lookup on page slugs. That is one more statement in hybrid search, and one more arm of the same statement with --match keyword. keyword_total and the --match keyword count include the named pages.
  • Pages that hold the identifier but not the other words can now rank above pages that match the question's meaning better. The page holding every word still comes first among keyword rows.
  • _meta.retrieval.coverage.could_affect can carry pending_images, and could_affect.pending_projection can drop for queries whose digits or short words used to match unrelated slugs.
  • A script that matched the text clean miss in an empty result should match retrieval healthy instead; _meta.retrieval is unchanged.
  • Nothing is rewritten on upgrade, no setting changed, and there is no schema migration.

Find a page by name, count pages by kind (#7191, #7233)

Agents fell back to the filesystem for two kinds of lookup that gbrain had no direct answer for. One was finding the page called X or under path Y, and the other was counting pages of each kind.

Situation Before Now
Find a page by name or path ls, find or grep on the checkout, or a ranked search gbrain find <name|prefix|glob> [--type T] [--source S] [--limit N] (MCP find_pages). A bare name matches a title or the last slug segment, served by the trigram indexes. A prefix (contains /) or a glob (*, ?) matches the slug or source_path. Rows sort by slug, with total and complete. The default limit is 50, capped at 500 locally and 100 remotely, and limit_clamped_from says when the cap applied. Under --as <member>, find and stats --by show only what that member reads
Count pages by kind or source gbrain stats gave brain-wide totals only gbrain stats --by source,type [--since 30d] [--collapse-aliases] (MCP get_stats, an admin-scope tool, gains by, since and collapse_aliases) reports page counts beside distinct-slug counts, so twins across sources show. --collapse-aliases folds type aliases through the active schema pack
Which call for which lookup spread across the skills one table in skills/conventions/brain-first.md, copied into AGENTS.md, INSTALL_FOR_AGENTS.md, skills/RESOLVER.md and the bootstrap template by bun run regen:all

Remote callers get the shared read policy (private pages, trust floor, quarantine), their source grant, and the ranked-search hard excludes (test/, .raw/), so every remote find hit is also readable through remote get_page. On Postgres these reads run with JIT off. On a 305k-page brain at 40 ms RTT, a cold find costs about as much as gbrain get of one page (about 1.0 to 1.2 s). No migration.

--explain-target in CLI JSON, and an opt-in keyword phrase factor (#7192, #7175, #7198)

  • gbrain query "<q>" --explain-target <source>:<slug> --json now prints { "results": [...], "explain_target": {...} }, with the fix as a command (next, argv, command); without --explain-target the output stays a bare array. New causes: target_not_embedded (the page has chunks but no vectors), the failed arm named in target_not_retrieved, target_collapsed (folded into a kept row), and each arm's rank in target_beyond_limit. The retry argv keeps array and object arguments (--types, --source-ids, --where as JSON), and the target_not_retrieved retry adds --expand, a paid call that sends the query to the model, so it renders next: ask_user. MCP is unchanged: remote callers get the diagnosis in _meta.explain_target without fix, and private targets still answer like nonexistent ones.
  • gbrain config set search.keyword_phrase true doubles the keyword score of a chunk that matches the query as one phrase (phraseto_tsquery), in the keyword arm on both engines and in keyword mode. It is off in every mode bundle, so default ranking and the default knob hash do not change; opting in appends kph=1.

Advisory harness check on pull requests (#7673)

Pull requests now get an advisory harness-check: a keyless run of harness-evals' gbrain check against the pull request's commit that fails only when a member is shown a page outside their grant.

  • New workflow .github/workflows/harness-check.yml (not a required check). It builds harness-evals' gbrain sidecar image for the pull request's head commit, runs scripts/gbrain-check.ts without a key, and gates on the cross-scope page count being zero. Every other failed assertion (tool counts, retrieval, install bundle, installer files) is a warning in the job log and summary.
  • The job reads one secret, HARNESS_EVALS_READ_TOKEN (read-only, harness-evals Contents). Without it the job skips with a notice and succeeds; a pull request from a fork is refused. No model key and no paid trial.
  • Docs: docs/ci-event-parity.md#harness-check (what it runs, the gate, the skip, a sample run and how to bump the harness-evals pin).

Docs (#7609, #7787)

  • New guides: git-canonical sources, anatomy of a page, where your memory lives and entity-scoped facts.
  • docs/architecture/brains-and-sources.md explains git as the parent of a source with a checkout. docs/guides/memory-boundaries.md covers adding to or uploading a large page. docs/mcp/TOOL_REFERENCE.md gains source_ids and pointers to search_exact and append_page. README and llms.txt link the new pages.
  • scripts/check-key-files-current-state.sh no longer caps README.md; the KEY_FILES.md cap stays.

Embedded-core guardrails (#8098)

The embedded core gets guardrails. They are a ratchet on what src/core reads from the process and the machine, a two-brains-in-one-process suite, a hermeticity test, and an upgrade-from-the-previous-release harness. Nothing changes for local users.

The planned embedded core lets one server process run many brains, each opened from its own settings. This release adds the guards that wave and every later one is measured against, with no runtime change. bun run check:core-world-reads (in bun run verify) counts, per file in src/core, loadConfig calls, environment reads, fs imports, module-level collections, spawn imports, process-wide hooks and AsyncLocalStorage stores, and fails when a count rises or a baseline row goes stale. The two-engine suite opens two brains in one process on PGLite and Postgres and shows what they still share today (AI keys and model clients, the query-embedding cache, the usage and telemetry sinks, link semantics, the admission and retrieval-tracking caches, page-lock names and home paths, and on Postgres the connection pool), each case naming the wave that fixes it. The upgrade harness opens a brain and keyless config written by the previous release, as a repository brain, a content-directory brain and a no-root brain, and requires the same page, schema, resolved settings and untouched config file.

Itemized changes

  • New guard scripts/check-core-world-reads.ts with baseline scripts/core-world-reads-baseline.tsv (rows only fall; --prune lowers stale rows).
  • New two-engine isolation suite (test/two-engine-isolation.serial.test.ts, test/e2e/two-engine-isolation.test.ts), hermeticity test (test/embedded-hermeticity.test.ts, expected to fail until W7) and upgrade harness (test/upgrade-from-previous-release.slow.test.ts, test/e2e/upgrade-from-previous-release.test.ts).
  • New example examples/embedded-two-brains.ts (two PGLite brains in one process).
  • An import inside its own engine scope no longer re-enters it, so a process that opened an embedded engine earlier no longer overflows the stack importing through an engine without a stable key.
  • Docs: running gbrain on servers today and the planned embedded core, including the opt-in embedded_core setting later waves add (nothing reads it yet) and which wave moves each piece of per-process state.
  • The 0.60.524.0 entry now names the command that exits 10 correctly: gbrain trust review --accept-all.

What to watch for after you upgrade

  • Nothing is rewritten on upgrade, no default changed, and there is no schema migration.
  • A script that matched clean miss in an empty search result should match retrieval healthy.
  • A script that parsed query --explain-target --json as a bare array should read .results.

[0.60.574.0] - 2026-11-04

A brain opened from settings no longer leaks into the process, or into another brain, when a call forgets its scope. While an embedded engine is open, an embed, chat, expansion or rerank call that entered no scope now refuses engine_scope_required and sends nothing, even if the host configured a process gateway; it decides by scope presence alone. withLocalScope(fn) marks a call that belongs to the process gateway, and the CLI, gbrain serve and gbrain jobs enter it, so they behave exactly as before. hybridSearch, the import paths, minion jobs and POST /mcp enter their engine's scope themselves. Each embedded engine keeps its own chat-usage sink (it survives reconnect()), batch-shrink, no-key and structured-output memory, and alias and remote-privacy policy, and a closed per-call engine is no longer held by the policy sets. A host that reopens a brain per call can pass settings.cache_namespace to keep its query embeddings warm. openEngine now refuses a settings key it does not know. Everything here is behind embedded_core, which is off by default.

Itemized changes

  • src/core/engine-scope-store.ts (new) and src/core/engine-scope.ts: withLocalScope (exported from gbrain/engine-open) and needsEngineScope. gateway.ts and anthropic-key.ts refuse an unscoped call with the new registered code engine_scope_required (its fix names withEngineScope); engine_scope_missing keeps meaning a provider the engine lacks. The hatch that let configureGateway open unscoped calls is gone, and configureGatewayIfUninitialized stays a no-op for them.
  • src/cli.ts, src/commands/serve.ts, serve-http.ts and jobs.ts enter the local scope at their entry points; serve-http-mcp.ts runs each POST /mcp and minions/worker.ts each job in its engine's scope; hybridSearch, hybridSearchCached, importFromContent, importFromFile, importCodeFile and importImageFile re-enter their engine's scope when called outside it, so a local engine's search inside an embedded engine's scope uses the process gateway. withEngineScope enters its scope even before the first embedded engine opens, so a call scoped then keeps it.
  • src/core/engine-construct.ts (new): openEngine builds its engine without the createEngine usage-stack entry and gives it a sink of its own (registerEngineChatUsageSink), so its usage is never the process fallback and a file-backed PGLite reconnect() no longer drops it (#8499).
  • src/core/ai/engine-gateway.ts: the shrink factor, the once-per-model no_key rerank audit row and the structured-output rejection memory live on each gateway state, and an embedded engine's cached query embeddings are dropped at disconnect (#8499 item 2). settings.cache_namespace tags them with the namespace and a hash of the provider ids, base URLs and key fingerprints instead, never the key.
  • src/core/search/private-visibility.ts: visibility.honor_aliases and search.remote_private_pages read on an embedded engine are kept per engine (keyed by engineKey) and decide inside its scope; a read in no scope while an embedded engine is open takes the strict policy; the CLI's process-wide lean-private rule is unchanged. An unscoped rerank refusal skips as no_key and writes no failure row.
  • openEngine refuses an unknown top-level or nested settings key with engine_settings_invalid naming it, never its value; storage is accepted for Wave C. A BEHAVIOR_CHANGES row and docs/ENGINES.md "Upgrading an embedded host" describe the changes for embedded hosts.
  • examples/embedded-two-brains.ts (new): two brains with distinct fake providers on 127.0.0.1, package exports only; test/examples-embedded-two-brains.test.ts runs it. New tests: test/engine-scope-required.test.ts, engine-scope-entry-points, engine-open-usage, engine-open-settings-keys, engine-gateway-state, engine-alias-policy, engine-cache-namespace, and test/e2e/engine-open-b3a-postgres.test.ts (reconnect usage and 1,000 per-call cycles). Fake transports only, no network.
  • Still to come: the two-engine seam suite moves onto openEngine once train R (#8148) lands; drainPersistence does not exist yet (W5a enters scope there); close order and drainFacts are B3b.

[0.60.573.0] - 2026-11-04

A host can record bytes it stored itself, and every file row names its page. On an engine opened from settings, the new file_register records the files row for an object the host already wrote to settings.storage at the path file_upload would use (<page_slug>/<filename>, or unsorted/<first 8 hash chars>-<filename>), after reading it back and matching its size and SHA-256. It resolves the page in the caller's source under the caller's read policy and records the row's source_id and page_id. Registering the same path again updates that row, and never moves it to another source or another page, including a new page that took a renamed page's slug. file_upload now records its row's source and page on every engine, so purging a page in a source other than default removes its attachments. A brain the CLI or a server opens refuses file_register with embedded_engine_required; it is never an MCP tool. Contributed by @bradgessler (#8457).

Itemized changes

  • src/core/ops/files.ts: file_register accepts only the upload layout path, refuses page_not_found for a page that does not exist or that the caller's read policy hides, and refuses invalid_params for a path whose row belongs to another source or page id; the pre-check and the ON CONFLICT … WHERE guard both compare page_id with IS DISTINCT FROM.
  • file_upload writes source_id and the page's id in that source on each row it inserts (#8513). Rows written before this release keep source_id = 'default' and no page id; nothing rewrites them. The gbrain files upload CLI is unchanged.
  • file_upload refuses unknown_source for a source the brain has not registered, before it writes any bytes. When a row already holds the path, it takes the uploader's source and page id only if the uploader's source has that page. A row attached to a live page in another source stays with that source. Identical bytes return already_exists without moving it, and different bytes refuse invalid_params before any bytes are written.
  • Page purge and its verification match a file by slug only when the row has no page id, so purging one page never removes another page's attachment at the same slug.
  • src/core/storage.ts: prefixedStorage refuses a key with an empty, . or .. segment or a leading / (Path traversal blocked), so on a local store a path such as ../brain-b/x never reaches another brain's prefix.
  • test/file-register.test.ts: real PGLite engines and a real local store.

[0.60.572.0] - 2026-11-04

A source whose automatic sync is off no longer extracts and reconciles every page on every autopilot cycle. A per-source cycle that runs no sync for its source (config.syncEnabled=false, a claimed source awaiting activation, a gbrain-owned content directory, or a source with no checkout) now extracts links and timeline, reconciles facts and recomputes emotional weight only for the source's pages changed since its previous cycle, as connector sources already did, at most 2,000 pages per cycle. Before, extract_facts walked the whole source each time and extract walked its whole checkout, so a sync-disabled source of about 27,000 pages or more hit the 30-minute job timeout on every cycle and never stamped its freshness. Pages an explicit gbrain sync --source imports between cycles land in the next window, and a large import is worked through over several cycles instead of timing out. A source with no recorded cycle starts from its oldest page and walks forward the same way. Reported by @mlahatte (#8573).

Itemized changes

  • src/core/minions/handlers/autopilot-cycle.ts: sourceChangedSince is passed to every per-source cycle whose phases leave out sync or that has no checkout, not only to connector sources.
  • src/core/cycle/source-window.ts: the window lists the source's pages after a cursor in (updated_at, slug) order, sourceWindow.pageLimit (2,000) at a time. A cycle that reaches the end of the window stamps its listing instant (last_source_cycle_window_at); one that stops at the limit stamps its last page (last_source_cycle_window_slug breaks ties), so the next cycle resumes there. The window advances only when extract, extract_facts and recompute_emotional_weight all ran without failing, so a cycle that failed one of them, or ran without them, gets the same window next time. The listing instant is moved back to just before the oldest transaction still open on another connection, because pages.updated_at is the writer's transaction start: a write still open while a cycle listed its pages lands in the next window. That holds for writers whose transactions the cycle's database role can see in pg_stat_activity (the same role, or one with pg_read_all_stats) and that stay open less than 10 minutes. Cursors bind as text, so Postgres keeps their microseconds.
  • src/core/cycle.ts: the extract phase of a cycle with a checkout and a source window extracts the window's pages instead of walking the checkout. A file edited on disk without a sync is no longer picked up by the cycle's extract on such a source, as on any source whose sync runs; lint still reads every file of the checkout, as it does for every source.
  • test/sync-disabled-cycle-window.test.ts (PGLite and Postgres): an idle cycle of a sync-disabled git source walks no page in extract_facts or extract, and a page with a facts fence that an explicit sync imports is reconciled by the next cycle; a first walk and a large re-import are worked through in bounded windows that resume; a cycle without the window phases, or with one failed, keeps the window; a source with no checkout gets the window; a write committed after a cycle's listing lands in the next window.

[0.60.571.0] - 2026-11-04

A sync hold on a page whose segment rows are missing now says to run the repair, and gbrain sync status finds one hold without paging. A page stored in segments whose segment rows do not cover its body refuses every read, so a managed sync holds its file (page_write_failed naming page_body_rows_unreadable). That hold read retry with no person needed and pointed at gbrain sources retry-held, which can never clear it. It now reads needs_human with safe_actions repair, and its fix previews gbrain repair segmented-tiling --source <id> --slug <slug> (read-only; the apply needs the user's agreement and clears the hold). On a source with hundreds of holds the hold also sorted onto a late page of sync status, so a check reading the first page missed it.

Itemized changes

  • src/core/persistence/sync-fault-class.ts, sync-holds.ts (gitHoldFix), sync-entry-fault.ts: the freeze_page_body_rows_unreadable verdict, the hold's fix and its message name the segmented-tiling repair; when doctor segmented_tiling does not list the page, the fix says to grant the reading role access and run retry-held. gbrain sync unblock refuses the hold by name with that fix.
  • gbrain sync status: each page_write_failed hold names its fault_code, holds_summary.by_fault_code counts them, and --slug <slug> / --code <code> (a hold code or a fault code) list one hold; a filtered page's cursor and next_command keep the filter.
  • Doctor: segmented_tiling runs as its own entry with its own budget, so a default run reports it even when the wave checks of timeline_history run out of theirs; an entry's last known result recorded before a check was added no longer hides that check (it reports unknown, or is named in doctor_last_known).
  • gbrain repair segmented-tiling --apply records the republished revision as imported from the HEAD blob when the file imports to the page's stored title, type, frontmatter and tags, so the next sync holds no file_drift for the page it just repaired. A file whose metadata changed since the last import records nothing, and sync still guards the stored version.
  • #8504 items 6 and 8: untiledSegmentedPages (doctor and the repair) skips archived sources; doctor's fix always names a --source and lists the other damaged sources' commands; a repair preview over 1,000 pages reports truncated and says to preview again after the apply.

[0.60.570.0] - 2026-11-04

GPT-style get_page calls that send several window selectors now read a window instead of refusing. GPT-6.1 Sol fills every get_page argument it can, so it sends lines, bytes and often a real around_chunk_id together, beside body: "full" and include_content: true. get_page refused every such call with invalid_params. The refusal carried a runnable fix, but on GBEV-30's SkillOpt re-read Sol followed it 0 times in 305, so 52% of its briefing reads and 84% of its fact-check reads failed. Now get_page reads one window, in a fixed order: cursor, then around_chunk_id, then lines, then bytes. A whole-page flag beside a selector is set aside, so the answer is still a window and never the page. The response's window_selector names what applied and what was ignored. A call with one selector behaves as before.

Itemized changes

  • src/core/ops/get-page-body.ts: parseBodyRequest still validates each selector on its own, then orders them cursor, around_chunk_id, lines, bytes, and the first one applies. body: "full", include_content and content_only beside a selector are set aside, while body: "none" or "outline" beside one still refuses. When a cursor can't read this page (another page's, another view's, or issued before the page changed) or an around_chunk_id's chunk is no longer on the page, and another selector was sent, the read falls back to the next one; on its own, each still refuses as before. With view: "markdown", around_chunk_id beside another selector is set aside. A read that set anything aside returns window_selector {applied, ignored, note}; when that includes include_content or content_only, the note says the window is not the page's content and that edits go through edit_page.
  • get_page's tool description states the order, and the four selector descriptions drop "One window selector only". docs/guides/windowed-reads.md and docs/mcp/TOOL_REFERENCE.md describe the rule.
  • Tests: test/tool-args-empty-optional-sol.test.ts replays Sol's recorded argument shape on every entry point (MCP over HTTP and stdio, gbrain call, the subagent registry and both SkillOpt registries). test/get-page-body-modes.test.ts checks the order, that the read equals the applied selector sent alone (edit_target included), that a whole-page ask beside a selector returns no page body, and that a failed cursor or stale chunk falls through to the next selector.

[0.60.569.0] - 2026-11-04

SkillOpt's iterative loop is deprecated and will be removed on or after 2026-12-03. The loop tied a single rewrite on gbrain-evals' benchmarks at about 1.8 times the cost, and every edit it kept came from its first step, so it is being retired. gbrain skillopt --loop, a --resume of a run the loop started, run_skillopt with optimizer_mode: "reflect" and a queued skillopt job in loop mode now print the skillopt_loop_deprecated coaching notice once (over MCP, once per session) and otherwise run exactly as before. Nothing replaces the loop, so edit the skill directly or run the default one-shot rewrite, which stays. Nothing is removed and no default changes in this release.

Itemized changes

  • src/core/skillopt/loop-deprecation.ts (new): the notice text, the removal date and the one emitter, called by gbrain skillopt (after it resolves the mode, so a resumed loop run counts), the run_skillopt op (ctx.emitNotice, so MCP and gbrain call render it) and the skillopt job handler. skillopt_loop_deprecated is a new coaching code in NOTICE_CODES.
  • gbrain skillopt --help, the run_skillopt description and its optimizer_mode parameter, docs/guides/skillopt.md, the tutorial and the skill-optimizer skill mark the loop deprecated.
  • Ship policy: .gstack/ship-policy.json declares stamp_placeholders: ["TBD"] (and min_gstack_version 1.94.16.0, the first gstack that reads it), so a branch's own since: 'TBD' BEHAVIOR_CHANGES row that the slot restamp fills reads as a stamp instead of a change after /ship's review. It also declares review_exclude_paths: ["docs/**", "test/fixtures/goldens/**"], so once this is on master /ship's reviewers leave docs and generated goldens out of their diffs and a large wave stays under their 2 MB ceiling; the privacy, gitleaks and doc checks still cover those files, and every code, test and config line is still reviewed.
  • test/skillopt/loop-deprecation.serial.test.ts: each entry point that runs the loop emits the notice exactly once and still runs it; the one-shot default, a one-shot job and the cycle phase emit none.

[0.60.568.0] - 2026-11-03

A brain opened from settings keeps its files in its own store. On an engine opened from settings, every file operation (file_upload, file_url, page purge and the new file_verify) uses the engine's own settings.storage: a local path, or s3 with its endpoint and keys, and an optional one-segment key prefix so brains sharing a bucket each keep their own keys. The process storage config is never used there. file_verify lists which rows still have their object or re-hashes one. A brain the CLI or a server opens is unchanged, and there file_verify refuses embedded_engine_required; it is never an MCP tool. file_register, which lets a host record bytes it stored itself, ships separately. Contributed by @bradgessler (#8457).

Itemized changes

  • src/core/storage/s3.ts: file_url on an s3 store with an endpoint builds the URL from the endpoint the SDK resolves, instead of throwing.
  • A store that does not answer file_verify or file_url is the new store_unavailable code, never a file reported missing.
  • On an engine opened from settings, delete_page purge deletes a page's files from settings.storage too. A file it can't find there, or a store that doesn't answer, is reported in blobs_remaining, never as deleted; files an embedded engine's file_upload wrote before this release live in the process store, so delete those remaining paths by hand. Rows an embedded engine's file_upload wrote before this release point at the process store, so file_url and file_verify look for them in settings.storage.
  • openEngine refuses engine_settings_invalid for a storage setting missing a value its backend needs (so no backend falls back to a default path or environment credentials), an unknown field, a backend other than local or s3, or a prefix that is not one segment; an s3 store must name its endpoint, so no AWS_* variable or ~/.aws file picks it; embeddedSettings() reads back hasSecret instead of the secret.
  • src/core/storage/local.ts and s3.ts stream an object's size and SHA-256 (hashObject); other backends hash one download.
  • test/embedded-file-store.test.ts: real PGLite engines and a real local store; the unreachable store is a closed local port.

[0.60.567.0] - 2026-11-03

/ship now carries its unit, serial, slow and E2E lanes from a passed full-ci run instead of running them again. .gstack/pregate.json declares ship_lanes_from (gstack 1.94+): bun run test, bun run test:serial, bun run test:slow and bun run test:e2e carry from the Test (full-suite) and E2E (e2e-full-suite) runs at the exact, clean tree, and print LANE: <label> carried run=<id>/<attempt>. A dirty tree, a newer red or running run, or a run whose tree.json misses a file the lane runs still runs the lane locally. Each lane's file list comes from scripts/pregate-remote-gate.sh --list on the full gate, so a tree that changes the CI reuse files never carries its own runs.

Itemized changes

  • .gstack/pregate.json: ship_lanes_from maps the four lane commands to the test and e2e run profiles. The E2E lane names both, because test.yml's reconciliation job runs two test/e2e/ files. bun run verify and bun run typecheck still run locally: tree.json records test files only.
  • .gstack/ship-policy.json: min_gstack_version is 1.94.4.0. A gstack below 1.94.0.0 rejects ship_lanes_from as an unknown key.
  • docs/ci-event-parity.md#ship-lanes lists the exact lane commands and the carry conditions; AGENTS.md and CONTRIBUTING.md point /ship at it. test/scripts/pregate-remote-gate.test.ts keeps the table and the config in step.
  • Nightly shard weights refreshed (2026-11-03).

[0.60.566.0] - 2026-11-03

Two installs at once no longer drop each other's instruction-ledger entry, and a Codex skills opt-out reaches the file Codex used before. Every install lane reads the brain-first ledger (~/.gbrain/agent-instructions.json) or the skill-loading ledger, edits the instruction file and rewrites the ledger. Two installs at once (say setup claude-code and setup codex) could each read the ledger before the other wrote it, and one entry was lost; a later opt-out then left that block, and its Claude Code allow rule, in place. Each install now holds one shared lock for the whole step, and the second waits for the first. Separately, Codex loads AGENTS.override.md when it is non-empty and AGENTS.md otherwise, so an explicit skills opt-out made after the override appeared left the index in AGENTS.md, where it came back once the override was removed.

Itemized changes

  • The brain-first and skill-loading ledgers share one lock (~/.gbrain/agent-instructions.lock), held across each install step: the ledger read, the file edit, the Claude Code permission edit and the ledger write (#8484). An install that waited reads the recorded choice again under the lock, so an opt-out the other install just recorded stays in force. An install that cannot take it within 10 seconds, or cannot create it, changes nothing and reports why (the skills outcome's reason is ledger_locked). A dry run takes no lock.
  • gbrain setup codex --skill-loading off, --no-instructions and GBRAIN_SKILL_LOADING=off also strip a spring index gbrain recorded in the other global instruction file of the same CODEX_HOME (AGENTS.md or AGENTS.override.md), and report it as a second skills line in the plain output (paired in --json), in --dry-run too (#8485). A plain re-run strips it only when brain-first instructions are declined on record for that Codex home, and no other Codex home, OpenClaw workspace or Claude profile is touched.

[0.60.565.0] - 2026-11-03

The per-file index describes how pages stored in segments are written and held. The key-files entries for write-through.ts and sync-entry-fault.ts contradicted v0.60.560.0: the first said every page is rendered whole and content always holds the bytes written, and the second said a page-body fault at the freeze stops the run. They now describe the streamed write of a segmented page (measured at its snapshot's revision, checked against the scan cap before source_path binds, inlineOnly and segmented_page) and the page_write_failed hold that names page_body_segmented or page_body_rows_unreadable. The entries for full sync, sync status, the derived write-through and export also name reexport_failures, last_error.request_id: null, the freeze_page_body_* routing, the derived writer's segmented_page skip and the export's revision pin. Docs only.

[0.60.564.0] - 2026-11-03

A brain opened from settings brings its own AI providers. settings.ai names the providers an embedded engine may call, with their keys and base URLs as values, and the models it uses for chat, expansion and reranking. Its embed, chat and rerank calls go only there, never to a key in the process environment, ~/.gbrain/config.json or the process-wide gateway, so two brains in one process never spend each other's keys or see each other's cached query vectors. A call to a provider the settings don't name refuses engine_scope_missing and sends nothing. While such an engine is open, an AI call made outside any engine's scope refuses the same way unless the host configured a process gateway. Each engine's chat usage lands in its own database. Keys never show up in errors, logs, doctor output or embeddedSettings(), also for a call still in flight when the engine disconnects. Every other engine, and the CLI, keeps the process gateway exactly as before. Contributed by @bradgessler (#8456, #8085).

Itemized changes

  • src/core/ai/engine-gateway.ts (new) and src/core/engine-scope.ts (new): an embedded engine's gateway config, built from settings.ai and settings.embedding alone and held per engine until disconnect, and withEngineScope(engine, fn) (exported from gbrain/engine-open), which every operation handler, each persistence consumer tick and the subagent handler enter. gateway.ts reads that engine's config and model clients in its scope, refuses engine_scope_missing for a provider the engine lacks before anything is sent, pins native providers to the engine's base URLs (never OPENAI_BASE_URL or ANTHROPIC_BASE_URL from the process), skips OpenAI model discovery and the config-file fallback for it, and tags its query-embedding cache entries per engine; anthropic-key.ts answers the engine's own key. The scope resolves through engineKey, so page transactions, proxies and a recovered degraded engine share it. In a process that never opened an embedded engine no scope is entered at all.
  • Search on an embedded engine reranks with settings.ai.models.reranker unless the brain or the call names one, and a reranker provider the engine lacks is a no_key rerank skip, not a provider error.
  • chat_usage_log rows from an embedded engine's scope go to that engine's database; an unscoped call books only to a non-embedded engine.
  • The in-process facts queue runs each job in the engine scope it was enqueued from, so one brain's extraction never runs on another brain's keys.
  • openEngine refuses engine_settings_invalid for a malformed ai setting, an unknown provider, a provider with neither apiKey nor baseUrl, two providers that would share one environment key slot with different keys, or one that reads a login or endpoint from the environment (claude-cli, Azure OpenAI), naming the setting and never a key. embeddedSettings() reads each provider back as { baseUrl?, hasApiKey }.
  • File storage is not an engine setting yet; it stays process-wide until Wave C.
  • test/engine-gateway.test.ts: fake-provider transports only, no network.

[0.60.563.0] - 2026-11-03

A degraded or partial read now says so on the result itself. Search, query, recall and the other reads already recorded degraded stages (an embedding timeout, a failed arm, a keyword-only brain, a held file) in notices and _meta fields that a harness had to know to look for, and recall and think dropped their page search's stages entirely. Every degraded or partial read result now carries one stamp naming why with stable codes. An MCP result gets _meta.degraded: { reasons: [...] } and one extra text block, [gbrain degraded: <reasons>] ..., after the others; a CLI --json object gets a degraded key, and other CLI output gets that line on stderr. A clean result carries no stamp, and no existing field changed. The code table is in docs/protocol/MCP_META_CHANNELS.md.

A write refused for naming another source says where the connection can write. A remember naming a source_id outside the connection's write grant is still refused scope_denied (a page write: permission_denied), and the refusal now carries writable_source_ids: the sources the connection's own grant lets it write, never another source, identical whether the named source exists or not. On remember the fix repeats the call against that source once the user agrees (next: ask_user, naming both sources), keeps the caller's visibility (a visibility the caller left out stays out, so the brain's default applies), and carries no request id; forget, forget_fact and page writes carry no fix.

Itemized changes

  • src/core/degraded-stamp.ts (new): degradedStampFor builds the stamp from a result's degraded stages, skipping stages that do not affect recall (reranker_skipped); src/mcp/dispatch.ts adds _meta.degraded and the text block, and src/cli/op-notices.ts adds the --json key or the stderr line. It covers search, query, recall, find_experts, context_pack, delta, entity, get_page (file_held), think and synthesize, search_exact, open_loops and rank_entities.
  • recall reports its page search's degraded stages in _meta.retrieval.degraded and its degraded_recall notice; think and synthesize report them as GATHER_DEGRADED:<stage> warnings; a keyword-only search or recall on Postgres reports keyword_candidates_incomplete when the rank bound applies.
  • A traverse_graph walk cut at the row cap (graph_walk_truncated) offers, over MCP, the same walk one level shallower; a depth-1 walk offers no fix. CLI flag validation is unchanged.
  • src/core/persistence/write-grant.ts (new): grantedWriteSource and writeGrantRefusal, used by remember, forget, forget_fact and page writes. The user_message names the caller's source_id only when it is a valid source id. OperationError.writableSourceIds is exported.
  • Docs: docs/protocol/AGENT_OPERATOR_v1.md, MCP_META_CHANNELS.md, MEMORY_VERBS_v1.md and the agent-operator-protocol skill.
  • Tests: test/degraded-stamp.test.ts, test/degraded-stamp-cli.test.ts, test/degraded-stamp-paths.serial.test.ts, test/write-grant-refusal.test.ts, plus golden and fixture updates for the stamp.

[0.60.562.0] - 2026-11-03

A write to a page stored in segments can no longer publish a body nobody can read. When a fact purge or a staging sweep collected a write's staging while the write was preparing, the write's last preparation step moved the staging back to prepared, so the publish passed its staging check, closed the page's old segment rows and published none. Every read of the page then refused page_body_rows_unreadable and a managed sync held its file on every retry. The write now refuses body_staging_expired (send it again) and the page keeps its previous body; the publish also refuses any generation whose rows do not cover the body it records. A new doctor check, segmented_tiling, lists pages an older version left like this, and gbrain repair segmented-tiling --source <id> republishes each from its file at HEAD, preview first, only when applied with the preview's hash.

Itemized changes

  • src/core/engine-sql/page-body.ts: setBodyStagingState moves only a live staging and refuses body_staging_expired for a move on to prepared or committed of an aborted or expired one; flipBodyGeneration refuses, inside its transaction, a generation whose live rows do not cover body_bytes from 0 with no gap (bodyGenerationIncompleteError, detail reason generation_incomplete).
  • src/core/page-body/schema.ts (liveTilingSql, untiledSql), body-tiling.ts (untiledSegmentedPages): the tiling of a page's live generation as SQL, shared by the flip guard, doctor and the repair.
  • Doctor segmented_tiling (BRAIN, host only, read-only): lists live segmented pages whose live generation does not tile, with the repair's preview as the fix.
  • gbrain repair segmented-tiling --source <id> [--slug <slug>]: explicit-only, preview-bound, trusted local CLI on the owner host. Republishes each listed page from the blob HEAD holds as a new generation, after the import screen the next sync would run and the purge and withdrawal overlay (a file another sync hold withholds, one the screen refuses, one that changes who can read the page, a privacy alias or a private-by-default type included, or one the content gate would quarantine, flag or stop embedding is listed held), versions the damaged generation, keeps the page's metadata and trust tier, clears the previewed sync hold and writes a receipt and an audit line. A source with no Git checkout refuses not_a_git_repo. See docs/guides/repair.md#segmented-tiling.
  • Tests: test/helpers/segmented-publish-tiling-cases.ts on PGLite (test/large-pages/segmented-publish-tiling.serial.test.ts) and Postgres (test/e2e/segmented-publish-tiling-postgres.test.ts): a purge abort, a sweep expiry and staged rows deleted under a live staging while the write prepares; each refuses and the page keeps its previous generation. test/large-pages/segmented-tiling-repair.test.ts: doctor lists a damaged managed page read-only, the repair's preview, refusals and apply, the next sync completes, a privacy alias or a content-gate quarantine holds the page, and a repo-less source refuses. The segment-GC fixture now publishes generations that tile, and the seal-bounded allowlist names the flip's tiling guard (it reads manifest rows, never chunk rows). Doctor goldens regenerated for the new check.

[0.60.561.0] - 2026-11-03

An embedding host can open a brain from a settings value (experimental, off by default). openEngine(settings) from gbrain/engine-open creates and connects an engine from one object: its database, its own pool, its direct URL, its persistence host identity, its local writer and its embedding identity. It reads no config file, environment variable or home directory for any of them, so two brains in one process never share a pool, a host id or a credential, and a stray GBRAIN_DIRECT_DATABASE_URL cannot send one brain's DDL somewhere else. It refuses unless the settings say embedded_core: true. The CLI, gbrain serve and mounts don't use it and behave exactly as before. Contributed by @bradgessler (#8455).

Itemized changes

  • src/core/engine-open.ts (new, exported as ./engine-open): openEngine, refusing embedded_core_required, engine_host_missing and engine_settings_invalid. disconnect() closes the engine for good, also during an in-flight reconnect, and a later connect() or reconnect() refuses engine_closed. A connect never migrates; the host runs initSchema(). Documented in docs/ENGINES.md, "Opening an engine from settings".
  • src/core/persistence/identity.ts: hostIdFor(engine) and setEngineHostId, resolved through engineKey. Every host-id read in src/core goes through it or existingHostIdFor(engine) (publication, admission, the consumer, worktree ownership and refresh, topology, effects, repair and the writer stamp), except gbrain migrate's engine graduation; an unmarked engine still gets localHostId(), so the CLI's behavior is unchanged. An embedded engine never reads or mints host.json, and its consumer heartbeat records embedded:<host-id>, which doctor and writer status read as an identity given at open. registerLocalWriterValue registers a writer the host holds as a value, and an embedded engine's writes use it instead of a <brain>.<lane>.json file. It never widens past the settings. The lane is the one given, a new registration gets the default read/write ceiling, and an existing registration keeps its own.
  • initSchema() on an embedded engine seeds the embedding model and width from settings.embedding and refuses engine_embedding_mismatch on a database that records another one. Postgres initSchema() reads the recorded identity again once it holds the schema lock, so two hosts initializing one fresh database can no longer both pass the check.
  • A settings URL must be one single-host URL that names its host, user and database, and an embedded engine's pools pass every connection field to the driver explicitly, so PGHOST, PGPORT, PGUSER, PGPASSWORD and PGDATABASE never choose its server or credentials.
  • PostgresEngine.reconnect() now ends the superseded connection manager's direct pool once its in-flight work finishes (#8478), so a long hold such as CREATE INDEX CONCURRENTLY or the schema lock held on the direct pool completes first. Before, every reconnect on an engine with a direct route left that pool open past disconnect(); this affected every Postgres engine, not only embedded ones.
  • The decide receipt and spend buffers write rows buffered for one brain to that brain when the next row names another, even while a flush is in flight. Before, those rows landed in the next brain's database.
  • The decide spend cache is keyed by engine and held weakly, and a drained feedback queue no longer keeps its engine alive.
  • File storage and AI provider settings are not engine settings yet; they stay process-wide until Wave C.
  • test/engine-open.test.ts, test/engine-open-lifecycle.test.ts, test/engine-open-upgrade.test.ts (a brain built by an earlier release boots identically through the CLI path and openEngine), test/e2e/engine-open-postgres.test.ts.

[0.60.560.0] - 2026-11-03

A page stored in segments no longer stops a source's sync. A managed sync re-importing a Git-tracked page stored in segments (pages.segmented_storage) whose working-tree bytes differ from the pinned commit, as on a live checkout, compared the file against the stored page with a whole-body read. That read refused with page_body_segmented, recorded as a run failure in the freeze phase, and every later sync stopped sync_blocked at the same file, with a retry command that could not help. The freeze now compares a segmented page up to 24 MiB read back whole, as its publication already did, so the page gets the outcome an inline page gets. A page whose stored body can't be read at the freeze (page_body_segmented, or page_body_rows_unreadable) is held page_write_failed, naming its code, and the rest of the source syncs. gbrain sync status classifies such a hold by that code (page_body_rows_unreadable: retry, no person needed), not as an unplaced page_write_failed. A run that does stop before any write request is admitted (discovery, freeze or admission) now shows last_error.request_id: null in gbrain sync status --json instead of leaving the field out: there is no receipt, root request or write_error_detail to look for, and the phase, code and message name the error.

A full sync's DB-only reconcile (#2426) now re-exports a page stored in segments instead of refusing it: the file is measured from the segments of one generation, then streamed through the staging writer the withdrawal mirror and restore_page use, written under an in-progress name and renamed once whole, with the bytes gbrain export writes and the existing file's mode. A file over the pre-push secret-scan cap is skipped before the page's source_path is bound (a body already over the cap is not streamed to measure it) and listed as over_push_scan_cap, with a fix that does not send the agent back into the same full sync. Pages whose re-export fails are listed in the result's reexport_failures, and the run completes. A segmented read no longer refuses a healthy page page_body_rows_unreadable because a new body published between its reads, and a read whose generation a publish retired meanwhile is checked against the size it read with that generation, so the rows the new body still shares never pass as the whole old body. A re-export streams the body of the revision whose frontmatter it writes: a publish in between fails that page's re-export, listed in reexport_failures, instead of writing one revision's frontmatter over the next one's body.

Itemized changes

  • src/core/page-state/import-guard.ts (sameCanonicalImport): a stored body left unread (over 24 MiB) differs from any parsed body, as sameContentAnyKeyOrder already said, instead of reading the throwing getter.
  • src/core/persistence/sync-screen.ts (frozenImportScreenInput): the freeze screen reads the page snapshot (and a rename's base) through materializedSnapshot.
  • src/core/persistence/sync-entry-fault.ts (freezeFaultHold): page_body_segmented and page_body_rows_unreadable thrown while freezing one entry hold that entry page_write_failed (origin freeze, write_failure.code the page-body code) instead of stopping the run. The item breaker still applies.
  • src/core/persistence/sync-fault-class.ts, public-category.ts, sync-status.ts (holdVerdict): freeze_page_body_segmented (upgrade, a person) and freeze_page_body_rows_unreadable (retry, the segment-table grant) rows, applied to a page_write_failed hold whose write_failure.code names the page-body code; a write receipt with either bare code keeps its class (a caller refusal, a gbrain defect); last_error.request_id is null for a failure no request was admitted for.
  • src/core/write-through.ts (writePageThrough): a page stored in segments is measured from its segments (unreadPageFile) and checked against the pre-push scan cap before source_path is bound, then streamed (stageStreamedContent, which refuses bytes that changed since the measure) on the engine it was given, so a caller inside a transaction (an owner repair) works, keeping the existing file's mode, and renamed over the target after the file-changed check; frontmatter overrides on such a page are refused with an error result.
  • src/core/persistence/page-prepare.ts (fileMatchesSnapshot): a segmented snapshot is compared read back whole, so the freeze drift screen no longer reads a format-only difference as drift.
  • src/core/cycle/derived-write-through.ts (writeDerivedFile): a generated page stored in segments is skipped (segmented_page, writePageThrough's inlineOnly), since the batch rollback compares file strings.
  • src/commands/sync/full.ts: the DB-only re-export no longer refuses segmented pages, and each failed re-export is returned in reexport_failures (slug, path, code, message, fix).
  • src/core/page-body/body-tiling.ts (readTiledManifest), ingest.ts, read-range.ts, projection.ts, segment-rewrite.ts: a live read names its generation first and checks that generation's rows against that generation's size; a live read whose generation a publish retired meanwhile reads the new live generation (up to 3 attempts), and is checked against the body_bytes read with the generation (assertGenerationTiles's bodyBytes), never against a missing size.
  • src/core/export-snapshot.ts (segmentedExportMarkdown), page-prepare.ts (unreadPageFile): the body is opened pinned to the snapshot's revision, so a publish after the snapshot refuses the file revision_conflict (the code a changed staging hash gives, so a tag write or restore prepares again or ends conflict) instead of mixing revisions.
  • docs/guides/sync-unblock-runbook.md: the two freeze_* fault-table rows.
  • Tests: test/large-pages/managed-sync-segmented-freeze.test.ts (the live-checkout freeze, with its inline twin, and a page whose rows don't tile held while the rest of the commit imports), test/sync-reconcile-db-only-segmented.test.ts (2 MiB and over-24 MiB re-exports byte-equal to gbrain export, one failing page listed while the run advances, a plain-directory source, a write inside a transaction keeping the file's mode, frontmatter overrides refused, a file over the scan cap left unbound, the derived skip), test/segmented-fault-classes-gbra287.test.ts (hold and receipt classes), test/segmented-read-concurrent-publish.test.ts (a forced publish between a reader's statements, a retired generation leaving only a shared prefix, a publish between a snapshot and its streamed file), test/pathological/d15-sync-reexport-overwrite.test.ts (a segmented DB-only page is re-exported, not held). Every case but the plain-directory one fails on master.

[0.60.559.0] - 2026-11-03

gbrain setup <harness> --skill-loading spring can now give an agent a short index of gbrain's bundled skills instead of the full skill text. Today gbrain either reaches the agent with no skill text at all, or copies every skill into the harness's native skills directory, where Claude Code lists about 87 skills on every turn and usually ignores them. With --skill-loading spring, setup writes a compact index (one line per skill: name and trigger) beside the brain-first block in the harness's instruction file, and the agent loads a skill's full text only when it needs it, with <gbrain> bundled-skill <name>. The option is off by default, and MCP get_skill is unchanged. In an A/B eval on real Claude Code and Codex CLIs with three frontier models (gbrain-evals#192), the index got the agent to load the right skill before acting in 29 of 36 Claude Code trials, against 15 with native skills and 0 with no skill text, and tied native on Codex at 22 of 24, for about a quarter of native's standing tokens per turn.

Itemized changes

  • gbrain bundled-skill [<name>] [--asset <path>] [--json]: prints a bundled skill's text, or a file it links, from inside the binary. It needs no brain or engine, and on a thin-client install it serves only the skills that apply there (host-only skills such as db-repair are refused with skill_not_found).
  • --skill-loading off|spring on gbrain setup (Claude Code, Codex, OpenCode, OpenClaw, Hermes, Cursor), gbrain connect --install and gbrain bootstrap harness, plus GBRAIN_SKILL_LOADING. The choice is recorded in ~/.gbrain/skill-loading.json, so a re-run keeps it. --no-instructions and --remove take the index out with the brain-first block.
  • The index names the absolute launcher the install recorded. Setup refuses to write it when that launcher serves a different skill catalog or fewer skills than the index lists, and on a thin client the index lists only the skills the loader serves there.
  • Claude Code gets one narrow allow rule, Bash(<launcher> bundled-skill:*), so a skill load doesn't stop for a permission prompt. Setup removes the rule on opt-out only when it added it. gbrain setup <harness> --skills-index-only refreshes the index after an upgrade and keeps the rule in the Claude profile the install recorded.
  • gbrain doctor: harness_wiring reports a stale, missing or damaged index and warns when its launcher is gone or the record cannot be read. resolver_health reads past the index block (an unterminated block hides nothing), and skill_currency no longer counts indexed skills as missing from the native directory.
  • scripts/check-skills-embedded.sh (in verify) proves a compiled binary serves every bundled skill file byte-identical to skills/.
  • Two known issues remain on the opt-in path. Concurrent installs can lose a skill-loading ledger entry (#8484, shared with the brain-first ledger), and setup codex --no-instructions leaves an index in AGENTS.md after AGENTS.override.md takes over (#8485).

[0.60.558.0] - 2026-11-03

The QM guide and the admin guide now warn that the attest client belongs only on QM's personal-scope memory route. An attested call reads as the named person, including their own private facts, and gbrain can't tell who else reads the answer. Registering that client as a model-callable MCP connector, or routing it to a shared scope (a channel, group, project, team or the org), could show one person's private facts to a whole room. Shared conversations wait for x-qm-audience (wave 2). Docs only: no behavior changes.

[0.60.557.0] - 2026-11-03

SkillOpt rollouts now see the same tool errors MCP callers get. When a read tool refused or failed during a SkillOpt rollout, the model saw only the error's one-line message. MCP callers get the full error for the same call: its code, the suggestion and a runnable fix. So skills were tuned against a weaker error surface than real callers read. Both SkillOpt registries (the read-only rollout registry and write-capture) now hand the model the error text a read-scope HTTP MCP caller receives for the same call, and they keep code for callers that branch on it. A fix names only tools the rollout offers. MCP output is unchanged. Write ops keep refusing as before.

Itemized changes

  • src/mcp/dispatch.ts (asAgentToolError): turns a failed read into an error whose message is the MCP error text for the same call, rendered for the caller's grant (a read-scope reader when the loop has none) and the loop's own tools. An AI invocation policy stop passes through unchanged, because the tool loop rethrows it.
  • src/core/skillopt/rollout.ts, src/core/skillopt/write-capture.ts: read handlers throw that error, naming the source the read used.
  • test/skillopt-rollout-agent-errors.test.ts: on both registries, a two-selector refusal and a missing-page error read exactly as a read-scope MCP caller gets them, including the suggestion and the runnable fix. A policy stop passes through, and a revoked remote grant reads permission_denied. The four parity cases fail on master.

[0.60.556.0] - 2026-11-02

Mounted brains stop sharing settings and search telemetry. In a process that serves several brains, search telemetry for every brain was written into the first brain's database, and a handful of per-brain settings were read once from whichever brain asked first and then applied to all of them. Each of those caches now belongs to its engine, and each engine gets its own telemetry writer, which stops its timer when the engine disconnects and keeps its buffer for a reconnect. A process with one brain behaves exactly as before. Contributed by @bradgessler (#8454).

Itemized changes

  • src/core/engine-key.ts: engineKey(engine) resolves a page-transaction clone, a read-through view and a statement-deadline wrapper to the engine behind them. Every cache below keys on it.
  • src/core/search/telemetry.ts: one writer per engine (getTelemetryWriter(engine)), held weakly, flushing only into its own engine. The disconnect drain bounds every live writer. A disconnect suspends the writer (no timer, buffer kept), so a PGLite reconnect, which is a disconnect and a connect, still flushes what was buffered; a collected engine retires its writer. setEngine is gone.
  • Keyed by engine: search.track_retrieval (last-retrieved.ts), persistence.publication_ceiling_ms and the idle bound derived from it (publication-deadline.ts), the minion admission policy (admission.ts), the mcp.strict_params last-known-good fallback (validate-params.ts), the write_capacity_low throttle (capacity-snapshot.ts), and search.hnsw_iterative_scan / search.vector_legacy_guard, which are still latched once and now latched per engine.
  • test/engine-keyed-caches.test.ts: two engines per setting, the second read after the first warmed the cache. All six behavior cases fail on master.

[0.60.555.0] - 2026-11-02

gbrain query --limit N without a token budget returns N rows again, a budgeted facts block takes at most N rows, the facts block never outscores the pages, and the CLI names fact rows instead of printing undefined (#8411). Since the facts block went on by default (v0.60.491.0), it also ran on queries that passed no budget, because the shipped no-budget evidence plan carries an implicit 24,000-token budget. Its 22% share let up to 40 fact rows in ahead of the pages, so --limit 3 could print more than a dozen fact rows before the 3 pages. Each block row was scored a hair above the top page, so they all showed the same score and outranked every hit, and the human CLI printed them as [0.84xx] undefined -- … because fact rows have no page slug.

What changed

Situation Before Now
query without token_budget (CLI or MCP) the facts block ran on the implicit budget: up to 40 fact rows on top of limit pages the block does not run; the facts arm fills only free slots, so limit caps the whole result
query with token_budget up to 40 block rows (22% of the budget), whatever the limit up to 40 block rows, 22% of the budget and at most limit rows, still oldest first ahead of the pages
Block row score top page score + n × 1e-6: above every page the top page's score: placed first, never ranked above a page
Human CLI line for a fact row [0.84xx] undefined -- [observed …] and the fact text on a second line [0.84xx] fact:<fact_id> <entity slug> -- [observed …] <fact text> on one line

The contract: limit caps the page rows. Without token_budget, saved facts only fill free rows, so the result never exceeds limit. With token_budget, the facts block adds at most limit rows (and at most 40, within 22% of the budget) ahead of the pages. --json rows already carried fact_id and keep it. The held-out arm that earned the default (token_budget 5,500, limit 25, a mean of 7.5 fact rows) runs the same block. search.query_facts_block stays on by default.

[0.60.554.0] - 2026-11-02

The admin UI loads again on installed builds, and setup, doctor and MCP verify stop failing on a set of install and serve shapes. /admin, including the OAuth consent page, was blank on every installed build because admin/dist/index.html loaded a bundle the embedded manifest didn't carry. It is rebuilt from source, and check:admin-embedded now runs in verify so the two can't drift again. Migration 0.53.0's packaged-skill install no longer fails idempotency_conflict when a later release ships a different payload. gbrain setup refuses the bootstrap harness lane only when that lane wired the same configuration it would write, so one GBRAIN_HOME can serve several HOMEs. Doctor and the turn-rule lane now count a gbrain hook in Claude Code's managed settings. gbrain mcp verify saves its fixture under its own folder, so a nearby page no longer makes remember refuse it. On the MCP side, get_versions and known_as_of stop pointing at a get_page read that doesn't exist, schema inspection uses the caller's own source pack, and open_loops explains an empty answer when none of a scoped caller's sources is a Google source, without revealing anything outside its grant.

Itemized changes

  • admin/dist, src/admin-embedded.ts: rebuilt from admin/src (the fresh build is byte-identical to the embedded bundle; the stale one goes). The generator no longer stamps the date, and check:admin-embedded runs in verify. test/admin-embedded-manifest.test.ts (#7126).
  • installPackagedSharedSkills: the durable adoption id digests the request id with the skills and initial files, so an exact replay resumes and a changed payload gets its own key. test/shared-skills-packaged-key-7734.test.ts (#7734).
  • gbrain setup: the harness_lane owner conflict applies when a recorded harness target sits in the config directory setup would write (paths compared after resolving symlinks, so another spelling of the same HOME still counts), when the harness wired project-scoped hooks, or when the receipt records no paths for that host. test/setup-harness-lane-scope-7771.test.ts (#7771).
  • inspectTurnHook and turnRuleHookYields read managed-settings.json and its managed-settings.d/*.json drop-ins in the platform system directory. A managed turn-rule entry never makes the turn-rule lane yield to itself, and the install check names a managed hook as one it does not run. test/turn-hook-managed-settings-8057.test.ts (#8057).
  • gbrain mcp verify: the memory-writer fixture lives under gbrain-verification/<fixture>; the entity resolver is unchanged. test/harness-verify-entity-fixture-7699.test.ts (#7699).
  • get_versions / known_as_of: a segmented version's body_ref carries version_id and total_bytes only (#7800).
  • get_active_schema_pack, schema_explain_type, schema_graph, schema_lint: resolve the pack through the caller's own source, the same path capture uses. test/schema-pack-source-inspection-7942.test.ts (#7942).
  • open_loops: a loops_no_google_in_scope notice for a remote caller whose read covered no Google source and returned nothing. It names only the sources the call read, all inside the grant, and stays off when the source read itself failed; the result is byte-identical with and without sources outside the grant. test/loops-remote-scope-notice-7286.test.ts (#7286).
  • test/serve-stdio-lifecycle.test.ts: the watchdog forward-progress probe gave a 20 ms RPC loop too little headroom under a loaded runner, so one scheduler stall failed it. A forced 45 ms stall failed 3 of 3 on the old margin and passes 3 of 3 now (40 RPCs of 20 ms in a 200 ms window); the assertions are unchanged, and removing noteForwardProgress still fails it (#8446).
  • scripts/{test,serial,e2e}-weights.json: shard weights refreshed from the Nov 2 nightly runs.

[0.60.553.0] - 2026-11-02

GPT models that fill every optional tool argument with an empty value can now read pages and walk the graph. GPT-6.1 Sol fills optional arguments with 0, [], {} or false as well as "" and null. The "" and null case was fixed in v0.60.549.0, but read operations still refused the others. get_page refused around_chunk_id: 0, lines: {} and a context sent without grep. traverse_graph refused hops: [] sent beside depth, link_type and direction. Twenty-seven other optional arguments refused the same way (limit: 0 on nine read operations, types: [], only: [], schema_version: 0, get_versions diff: false and more). gbrain now reads these as not provided, but only where the argument's own schema rules the value out: 0 when the valid range starts above zero, [] when the list must not be empty, {} when the object needs a field, false on a flag that only acts when true, and context when there is no grep, since it has no effect without it. Any value inside the valid range is kept, and so is any value on an argument that declares no such rule. So the normalized call is always one the caller could have sent by leaving the argument out, and it never reads more than that call would. This runs on every path that hands model arguments to an operation: MCP over HTTP and stdio, gbrain call, the subagent tool registry and both SkillOpt registries.

get_page with known_as_of still refuses arguments that select a window or a view of the page, because those cut the current page's text and chunk ids and a recorded version has neither. The refusal now names each argument to drop and carries a fix: the same call with only the arguments a recorded-time read takes. Replaying the 651 GPT-6.1 Sol calls recorded in the SkillOpt probe, 30 of 38 traverse_graph calls now succeed. The other 8 send a non-empty hop chain beside depth. The 436 get_page calls that sent context without grep now pass that step and stop at the next rule, because they also send both lines and bytes (and often around_chunk_id), real window selectors that change the answer. Those still refuse, but get_page's description and its selector parameters now say to pass at most one window selector, and the refusal's fix is a runnable single-selector call built from the caller's own arguments: the first selector alone, with body: "window". All 436 fixes read the page in the replay.

Itemized changes

  • src/core/ops/contract.ts (ParamDef): new schema fields minimum, exclusiveMinimum, minItems, actsWhenTrue and appliesWith. Only the normalizer reads them, and they are left out of the advertised tools/list schema.
  • src/mcp/validate-params.ts (normalizeOptionalParams): drops the values those fields rule out, plus {} on a read operation's object whose properties require a member (a write keeps {}, so edit_page within: {} still refuses). It then drops any argument whose appliesWith argument is absent. Required arguments are never touched.
  • Read operations declare the fields where a sweep of every read operation showed the value refuses: minimum: 1 on the swept limits, ids, max_bytes, return_window, schema_version and min_n; exclusiveMinimum: 0 on search_exact max_seconds; minItems: 1 on types, only and hops; actsWhenTrue on get_versions diff; appliesWith: 'grep' on get_page context.
  • src/core/ops/get-page-body.ts, src/core/ops/pages.ts, src/core/ops/get-page-slice.ts: the get_page description and the lines, bytes, around_chunk_id, cursor and context descriptions state the one-selector rule. The two-selector refusal names the selectors to drop and carries a read fix. test/mcp-schema-budget.test.ts: the starter list ceilings and the get_page budget rise by 100 characters for that sentence.
  • src/core/ops/known-as-of.ts: the known_as_of does not combine with … refusal names the arguments to drop for the caller's transport and carries a read fix.
  • test/tool-args-empty-optional-sol.test.ts: a table test over the sweep's 29 cases on every entry point (each one refuses on master), checks that in-range and non-empty values still refuse, and replays the Sol shapes for get_page and traverse_graph.

[0.60.552.0] - 2026-11-02

A master-red flake in the deferred-publish synthesis test is fixed. test/managed-synthesis-postprocess.test.ts "#5854: a postprocess publish still pending after its wait is deferred, then finished by the next ordinary cycle" released the writer lock before disposing the in-process persistence consumer. Whenever the consumer drained the deferred publish in that gap, the "next ordinary cycle" found nothing left to write, and written_slugs came back [] (master run 41476223139, shard 14). The test now disposes the consumer before it releases the lock, so the next cycle is the only thing that can finish the publish. The assertions are unchanged. Forced probe: a 1.5 s pause after the lock release fails the old order with the exact CI diff and passes the new order (3 of 3).

[0.60.551.0] - 2026-11-02

The nightly opencode door now gates its nonce check on a paid Haiku turn, and a free-tier refusal fails in seconds with its reason instead of after eight silent minutes (#6049). OpenCode's anonymous free models now cap usage per IP, and CI runners hit the cap. Past it, opencode run waits out the server's retry-after, which is hours, and prints nothing, so the keyless check timed out twice (about 487 s) with an empty reply and Heavy Tests went red on days the runner's IP was capped. The only trace was a stream error line in opencode's own log file. The paid check also could not have run in CI, because the test preload strips ANTHROPIC_API_KEY unless GBRAIN_TEST_KEEP_PROVIDER_KEYS=1, so a requested paid tier refused itself as keyless.

What changed

Situation Before Now
Free-tier quota refusal during a door turn two 240 s timeouts, then Expected to contain: "kestrel-…" Received: "" the turn ends on the logged refusal; the test fails in about 11 s with provider error(s): AI_APICallError: Free usage limit reached… and makes no second attempt
Nightly Heavy Tests, opencode door keyless lane gates on 5 tests, including the free-tier nonce check (T4) keyless lane gates on T1-T3 (4 tests); the credentialed lane runs every night and its paid nonce check (T5) gates; T4 runs only in the non-gating latest-version canary
Paid nonce check model anthropic/claude-haiku-4-5 anthropic/claude-haiku-5-5, validated against opencode models --refresh before any spend (1.18.18's bundled catalog predates it)
Credentialed lane in CI ANTHROPIC_API_KEY stripped by the test preload, so the paid tier refused to run the step sets GBRAIN_TEST_KEEP_PROVIDER_KEYS=1; one Haiku turn a night

Itemized changes

  • test/helpers/agent-harness.ts: opencodeOneShotTurn reads main-turn stream error lines from <XDG_DATA_HOME>/opencode/log written during the turn, returns them as streamErrors, and ends the turn on a quota refusal (providerRefusal). A provider rate limit, which opencode retries through, does not end it. runOneShotSpawn takes an abortWhen poll, and once it kills the child, a descendant that holds the output pipes open gets 8 s instead of the rest of the timeout.
  • test/helpers/agent-harness.unit.test.ts: stub-binary probes pin the quota fast-fail (it hangs to the timeout without the fix), the fast-fail with a descendant holding the pipes, a transient error and a rate limit that opencode recovers from, and per-turn log offsets.
  • test/e2e/install-real-opencode.serial.test.ts: T4 runs only with GBRAIN_REAL_OPENCODE_FREE_TIER_E2E=1, adds logged stream errors to its provider-error message, and skips its second attempt after a quota refusal. T5 pins anthropic/claude-haiku-5-5, refreshes the model list before its spend gate, and, like T4, reads the turn's JSON events and requires a gbrain_* MCP tool call, since opencode's file tools could find the nonce without one.
  • .github/workflows/heavy-tests.yml: the schedule requests the credentialed lane (GBRAIN_PAID_LANES=0 still turns it off), the keyless lane expects 4 passes and its summary warns that no nonce check gates the run when GBRAIN_PAID_LANES=0, the credentialed step keeps provider keys, and the canary sets GBRAIN_REAL_OPENCODE_FREE_TIER_E2E=1. test/scripts/heavy-door-workflow.test.ts pins all four.
  • docs/mcp/OPENCODE-CLI-PIN.md: the per-IP free cap, where its trace is logged, the new gating split and two triage rows.

A merge slot's full-ci runs can now clear gstack's remote gate on their own, so each tree runs the full suite once. gstack-pregate clear --remote pregate.json/remote_gate --run <Test run> --run <E2E run> --allow-repo-commands verifies both runs on GitHub (same tree, both full jobs green in the newest attempt, no newer red run, under 24 hours old) and checks that the files they executed cover the gate's inventory. Checked against Test run 41342917113 and E2E run 41342918866 on one tree: REUSE: accepted … tests=68132 and PREGATE_VERIFY: ok, with all 5,701 inventory files covered. A tree that changes the full-ci workflows or the scripts that attest their runs still can't vouch for itself. It gets no inventory, so gstack refuses those runs and the gate runs, the same rule remote-gate reuse already follows. That rule now covers seven files: ci-manifest.ts joins the list.

Itemized changes

  • .gstack/pregate.json: run_profiles test (test.yml, full-suite, artifact full-suite-tree) and e2e (e2e.yml, e2e-full-suite, artifact e2e-full-suite-tree), both accepting workflow_dispatch and push, plus remote_gate_runs with bash scripts/pregate-remote-gate.sh --list as the inventory.
  • .gstack/ship-policy.json: min_gstack_version 1.91.377.0, the first gstack that reads these keys. An older gstack now refuses with an upgrade fix instead of an invalid-config error.
  • scripts/pregate-remote-gate.sh: the files whose change stops a tree's own full-ci runs from attesting to it are now seven: test.yml, e2e.yml, ci-reuse.ts, ci-full-suite-tree.sh, ci-executed-counts.ts, lib/gh-issue.ts and, new, ci-manifest.ts, which writes the ci-manifest.json that ci-full-suite-tree.sh trusts for completeness, executed files and counts. A tree that changes ci-manifest.ts therefore no longer reuses its full-ci and runs the gate. --list refuses (no inventory) for a tree that changes any of the seven, or the inventory's own builders pregate-remote-gate.sh and ci-ubicloud.ts, relative to origin/master, when there is no origin/master, and under GBRAIN_PREGATE_REUSE=0. Both paths run one check, and --list adds the two builders. test/scripts/pregate-remote-gate.test.ts pins this.
  • docs/ci-event-parity.md: the profiles and the refusal.

ubi-runner.sh reap now destroys only the caller's own VMs, and pack, usage and reap work on macOS and on an empty project (gstack #3756, #3746). Reap used to destroy any owner's VM whose TTL had ended. That is how one project's up destroyed another lane's VMs (GST-573). Now it destroys only VMs whose owner tag equals UBI_OWNER and logs reap: left N expired VMs of other owners for the rest. The operator reaper, reap-expired.ts, still covers every owner. gbrain keeps UBI_REAP=0 by default, so up still reaps nothing unless asked. On macOS, pack exited 1 under bsdtar after the VM was already created. On a project with no VMs, Ubicloud's bare-newline list read as a VM named -, and reap died on it. A failed size read made usage print a short total, which the quota preflight took as room.

Itemized changes

  • scripts/ubicloud/ubi-runner.sh, ported from gstack 74836e41 and 693ee713:
    • cmd_reap checks the owner tag first and counts other owners' expired VMs. The quota refusal says reap frees only this owner's VMs and names usage and reap-expired.ts for the rest.
    • vm_rows skips a blank row.
    • cmd_usage exits 1 with no total when the list or a size read fails, and skips a VM whose show answers 404.
    • quota_preflight retries an unreadable usage while UBI_QUOTA_WAIT lasts, then logs it and leaves the decision to the API.
    • cmd_pack builds files.tar, appends .git with tar -rf and gzips it, with COPYFILE_DISABLE=1 on every tar call.
  • test/scripts/ubi-runner-run.test.ts and test/scripts/ubi-runner.test.ts pin the owner-scoped reap, the empty project, the unreadable usage and COPYFILE_DISABLE. The six new cases and the two updated reap cases fail on the old runner. The updated cases expected another owner's VM destroyed and now expect it kept.
  • test/scripts/ubi-runner-bsd-userland.test.ts runs pack under bsdtar. It fails on the old runner and is skipped where no bsdtar is installed. test/helpers/mock-ubicloud-api.ts gains listBody, ghosts and a show fault.
  • scripts/ubicloud/README.md: reap's owner scope.

Two test-isolation flakes are fixed (found when this release's new test file moved files between CI shards).

  • test/context-pack-pinned-federation.test.ts set search.remote_private_pages=visible and left its engine in the process-wide privacy cache, so a later test file in the same CI shard saw private pages and test/private-visibility.test.ts failed. Its afterAll now resets that cache.
  • test/db-lock-fencing.test.ts waited a fixed 120 ms for the lock refresher's second retry, which a busy runner could miss. It now waits for eight retries, up to 5 s, so its not-aborted check still covers a run of failed refreshes, with the same assertions.

[0.60.550.0] - 2026-11-02

Security: the stdio MCP server and the remaining hook lanes no longer let the session directory's bunfig.toml or .env run, and the hook's git probes no longer run a checkout's core.fsmonitor (#8426, #8428, #8427, #8435, #8436). Claude Code, Codex and opencode start user-scope stdio MCP servers in the session's directory, just as they run hooks there, so with a script-mode install the gbrain serve entry gbrain setup and gbrain bootstrap hooks registered ran a cloned repository's bunfig.toml preload once per session, with nothing to approve. The user-global opencode entry gbrain bootstrap hooks --harness opencode writes did the same. Those entries now start serve through the same <gbrain config dir>/hook-launcher/gbrain-hook the setup hooks use, and gbrain serve reads .gbrain-source, a mounted brain's persistence socket, relative file_upload paths and the upload root from the session directory the launcher records, without changing its own directory.

  • The hooks from gbrain bootstrap hooks (the workspace lane), gbrain bootstrap harness and the Codex SessionEnd capture both write now start gbrain through the launcher too, and the gbrain plugins' per-turn hook changes into the plugin directory before it starts gbrain. The plugin hook now needs gbrain 0.60.547.0 or newer (it exits silently on an older CLI, which would read the plugin directory as the session's).
  • gbrain post-upgrade moves a user-scope gbrain setup claude-code MCP entry setup still owns onto the launcher, checking it again under setup's lock, and takes the entry's GBRAIN_HOME only from hooks or an MCP entry that install's receipt owns, so another install sharing the Claude settings file never redirects it to a different brain. It rewrites no Codex entry, the turn-rule hook included (a rewrite recomputes the trust of every install's groups in the shared hooks.json): for gbrain setup codex's MCP table and hooks it prints the gbrain setup codex command, and for every gbrain bootstrap entry the bootstrap command that rewrites it, run from the workspace. An entry you edited after setup wrote it is kept and named, and the refresh stops when another install replaced the receipt it planned from.
  • gbrain doctor (harness_wiring, hook_session_cwd) now also covers the stdio MCP entries and the bootstrap, harness and Codex capture lanes, with the command that fixes each. Only a user-scope Claude Code entry setup owns gets an agent-runnable fix. It reruns the install that owns the entry with its recorded --no-hooks, --source and --gbrain-bin, so the rerun turns nothing on and keeps the binary the install chose instead of whatever gbrain is first on PATH. Every gbrain setup codex rerun is now the user's to run, because it rewrites the shared hooks.json and its trust.
  • gbrain hook's git probes run with -c core.fsmonitor=false, and a session outside the workspace the install receipt names runs no git at all, so a .git/config from an extracted archive or shared folder cannot run a command at session start.
  • The readiness fix for an unwired Claude Code or Codex now names gbrain setup <harness> --no-hooks instead of claude mcp add / codex mcp add, so the entry it writes goes through the launcher; it also writes the brain-first block and records hooks as off. When gbrain setup already wired that harness, the fix reruns that install with the decisions its receipt recorded (the Codex rerun is the user's to run), so it never turns an install's hooks off. The install docs now teach gbrain setup claude-code / gbrain setup codex instead of a bare stdio mcp add. A project-scope .mcp.json and committed .claude/settings.json hooks stay PATH-resolved, because they belong to the repository.

[0.60.549.0] - 2026-11-02

GPT models that fill every optional tool argument with "" or null can read pages again in SkillOpt rollouts and gbrain call. GPT-6.1 Sol sends known_as_of: "", cursor: "", section: "" and grep: "" instead of leaving those arguments out. MCP dispatch and the subagent tool registry already treated "" on an optional string argument, and null on any optional argument, as not provided. The SkillOpt rollout and write-capture registries and gbrain call passed the arguments to the handler unchanged. There known_as_of: "" sent get_page into its recorded-time branch, which refused with "known_as_of must be an ISO 8601 instant", so a SkillOpt probe on GPT-6.1 Sol saw every get_page call fail and the model answered from search snippets. The same empty arguments broke about 30 other read operations on those paths (search, get_links, traverse_graph, recall, whois and more). All three paths now run the same step as MCP dispatch. Empty optional arguments are dropped, then the arguments are checked against the operation's schema. A missing required argument or a non-empty invalid value still refuses with invalid_params. The SkillOpt registries now also refuse a wrong-typed or out-of-enum argument before the handler runs, with the same invalid_params error MCP returns. Over remote MCP, memory_recall with acting_user: "" or acting_user: null refused with actor_not_attested, because the attested-actor check read the arguments before they were normalized. It now reads the normalized arguments, so an empty acting_user means the same as leaving it out. A non-empty acting_user is still accepted only on an attested call and only when it matches the x-qm-actor header.

Itemized changes

  • src/core/skillopt/rollout.ts, src/core/skillopt/write-capture.ts: each read tool drops "" and null optional arguments (normalizeOptionalParams), then validates (findInvalidParam, refusing through schemaInvalidParams), before calling the handler. In a rollout, the shared-skill read authorization still runs first.
  • src/mcp/server.ts (handleToolCall, behind gbrain call): drops empty optional arguments before validation, and the handler receives the cleaned arguments.
  • src/mcp/dispatch.ts (dispatchToolCall): the audience and attested-actor checks on a remote call read the normalized arguments.
  • test/tool-args-empty-optional-sol.test.ts: drives get_page with every optional string set to "", and again with every optional argument set to null, through MCP over HTTP (remote) and stdio (local), gbrain call, the subagent registry and both SkillOpt registries. It checks that a missing slug and non-empty bad values still refuse, then sweeps every read operation and checks that empty optional arguments give the same outcome as omitting them. Reverting any one of the three fixes fails it.

[0.60.548.0] - 2026-11-02

Meeting notes now flow into the brain on a schedule, with no custom script.

A new meetings source kind imports meetings from Fireflies (the first provider behind a provider interface) into one page per meeting: title, date, attendees, the provider's summary, action items and keywords, and optionally the transcript. You save the API key once on stdin with gbrain connectors auth fireflies --token - (probed before it is saved, 0600, never in argv, config or the database), add the source with gbrain sources add meetings --kind meetings --provider fireflies, and run the first sync yourself. After that, autopilot keeps it fresh every --sync-interval-min minutes (default 60) and waits out the provider's rate limit. Nothing syncs until you register a source and run its first sync.

It works on a brain with no Git repository, where pages and the sync position are recorded in the database. On a brain with a Git checkout, --dir inside it records each page with a commit before its database row, and a missing checkout holds the meeting instead of writing it to the database alone. The record mode is chosen when the source is added and never re-inferred.

Situation What happens
The key's owner did not attend the meeting the page is private
A meeting the provider is still processing re-checked every sync; published when its summary arrives (or after 48 hours)
A meeting edited upstream the page updates and facts extracted from the old version are retired
A meeting deleted upstream (inside the lookback) the page is deleted and its facts are retired
A key for a different account refused (connector_account_changed) before anything is imported, even after the source's config changes
Summary text that looks like a facts table, citation or timeline row made inert before it lands in the page

Guide: docs/guides/meeting-and-slack-sources.md.

Slack channels can flow into the brain the same way. A new slack source kind reads the channels you name through your own internal Slack app and writes one page per channel per day, in the chat layout the conversation parser reads, with thread replies under their parent. Save the bot token with gbrain connectors auth slack --token -, then gbrain sources add slack --kind slack --channels C0123ABCDEF; the add checks the token and every channel first and refuses with the fix (a missing scope, or "invite the app to the channel"). It works on a brain with no Git repository, re-reads the last 7 days on every sync so edits and deletions converge (retiring the facts of the old page), keeps private channels private and out of any Git checkout, and moves a channel that becomes private out of the checkout. A channel the app loses is held while the others keep syncing.

For both kinds, gbrain sync --source <id> --full re-reads the whole history window, --dry-run fetches and writes nothing (these sources have no preview), and gbrain connectors status fireflies (or slack) reports whether a key is saved without printing it.

Known limits

  • Fact retirement on a changed page still needs a redesign (#8432): on a managed brain a retirement can race the page's own write and is not retried if it fails, every new Slack message retires and re-extracts its whole day's facts, and facts from an unchanged prefix of a long page are not re-asserted.
  • Slack (#7971): a channel window too big for the per-run request cap (s_max_requests) or over 500 history pages does not make progress, and in Git record mode a channel that becomes private leaves HEAD but stays in Git history.
  • A meetings window with more than 20,000 meetings stops with listing_unbounded; use a smaller --history-days. connectors status does not read a source's --token-env.

[0.60.547.0] - 2026-11-02

Security: gbrain's hooks no longer let the session directory's bunfig.toml or .env run (#8424). A script-mode install (bun install -g, bun link) runs gbrain as a Bun script, and Bun applies the working directory's bunfig.toml (a top-level preload runs before any gbrain code) and loads its .env files first. Claude Code and Codex run hooks in the session's directory, so with the hooks gbrain setup claude-code|codex writes, or the per-turn rule hook from --instructions-only, gbrain connect --install and gbrain bootstrap harness, opening a session in a cloned repository ran that repository's code as you on every prompt, with nothing to approve. Each of those hooks now starts gbrain through <gbrain config dir>/hook-launcher/gbrain-hook, a small sh script that records the session directory in GBRAIN_HOOK_CWD, changes into its own directory and starts gbrain there. Its directory holds only the launcher, and it refuses to start gbrain if anything else appears there.

  • gbrain post-upgrade moves existing hooks onto the launcher. That covers the turn-rule hook in Claude Code's user settings and in Codex's hooks.json, and the hooks of each user-scope gbrain setup claude-code install whose MCP entry is unchanged. A Codex turn-rule hook moves only when Codex already runs it, so an upgrade never trusts a hook you did not approve, and a hook you edited after setup wrote it is kept and named instead. gbrain setup codex's hooks are not moved by an upgrade: rewriting them re-trusts them, so post-upgrade prints the gbrain setup codex command that does it, and doctor keeps flagging them until it runs. gbrain setup <harness> and gbrain setup <harness> --instructions-only rewrite them on their next run.
  • gbrain doctor (harness_wiring) warns with the new code hook_session_cwd when a hook still starts gbrain in the session directory, or its launcher is missing, changed or shares its directory with a file gbrain did not write, and names the setup command that fixes it. Project-scope installs get the command to run from their project directory.
  • Not covered yet: the stdio MCP server gbrain setup registers, which the harness also starts in the session's directory (#8426); the capture hooks from gbrain bootstrap's own lanes and the gbrain plugins' hooks (#8428); and script-mode commands you run by hand in an untrusted checkout (SECURITY.md).
  • The launcher rewrite never leaves a temp file in its own directory, and doctor finds a setup hook whose _gbrain marker Claude Code dropped by its receipt hash.
  • gbrain hook resolves the recorded session directory the way the shell's working directory resolves, so project-scope rules still match under a symlinked path, and a .env in the session directory cannot set GBRAIN_HOOK_CWD.

[0.60.546.0] - 2026-11-02

QM can use gbrain as its memory provider. A trusted system that names the person (x-qm-actor) reads and writes as that linked member, each member's saved facts stay on their own page and are recalled by them alone, and memory_recall / memory_write give QM one plain-text read and one batched write (GST-359).

What changed

Situation Before Now
A trusted system that names the person (QM) a client-credentials client granted gbrain auth rescope --client <id> --attest-actor qm [--attest-ops <list>] sends x-qm-actor: <canonical id>; each attested operation runs as the member that id is linked to (gbrain access member link <member> --qm <id>, unlink, links), never reading private pages, and queued writes and jobs re-check the link and the operation list. capture, extract_facts, take writes and list_write_conflicts are not attestable yet. Refusals carry structuredContent.error (actor_not_attested, actor_unlinked). Every MCP server's serverInfo now has a one-line description
A member saves a fact remotely (remember attested as a person, or from a credential bound to one member) from a member-bound credential, went on the entity page as a fact every reader of the source recalled; an attested call could not run remember goes on the member's own page memory/members/<member> (created with readable_by: [member:<id>]), private, with the entity kept as metadata; only that member recalls it, and it never dedups against, supersedes or contests another page's fact, never consolidates into a take and keeps its readers through a fence rebuild and relink. Two people's identical words make two facts. The page must name that member alone, or the fact is refused; remote put_page, edit_page, delete_page and every other remote page write under memory/members/ are refused (permission_denied) unless the caller is that member. A member id a repo brain would store under another slug (a--b, zed-) holds no member facts (invalid_params)
One attest client serving several people hot memory, delta cursors, write receipts and agent jobs were keyed by the client each is keyed by the person too: Alice and Bob on one session id never share a hot-memory payload or a delta cursor, Bob and the bare client get not_found for Alice's receipt (get_write_request, list_write_requests, cancel_write_request) and agent job (get_agent_job, and the client's own get_job, wait_job, get_job_progress, list_jobs and cancel_job), and Alice's request_id never replays for Bob. remember, delta, the receipt ops, submit_agent and get_agent_job are now attestable
QM's memory provider memory_recall {query, max_chars?} answers the attested person's own recent facts first, then the facts they may read ranked by the query, one plain line per fact cut within max_chars (200-12000, default 6000), keyword-only on a keyless brain; memory_write {facts, idempotency_key} saves 1-20 facts on that person's own page through remember's writer, with request ids derived per person from the key so a replay returns the first receipt. Results ride structuredContent, and every refusal carries structuredContent.error; an unlinked person gets an empty recall with code: "actor_unlinked". get_write_request joins the default attested operations. A request carrying x-qm-audience is refused until channel audiences are served (audience_not_supported). forget from an attested person or a member-bound credential reaches only that member's own page. remember and memory_write take valid_from only as an ISO 8601 date or time in years 0001-9999. Contract: docs/contracts/qm-memory-provider/v1.json; guide: docs/guides/qm.md
An attested forget or forget_fact refused (op_not_attested) withdraws only a fact on the person's own member page, and only there: another member's copy of the same claim stays active and rememberable

What to watch for after you upgrade

  • A call that sends x-qm-actor or an acting_user argument from a credential without the attest right is now refused (actor_not_attested) instead of reading as the credential; tokens issued before this upgrade never attest, so a QM client takes a fresh client-credentials token after it is granted.
  • Migration v286 adds access_member_links, the attest columns on access tokens and OAuth clients, and the grant type on OAuth tokens; it is database-only, so a repo-less brain gets it unchanged.
  • Writers never dedup, supersede or contest across member pages: extract_facts, capture and supersession leave member-page facts out of their candidates, so a write never learns that another member saved the same claim.
  • A member's facts go only on memory/members/<member>, so a credential whose bound_slug_prefixes do not cover that page is refused (permission_denied) by remember and memory_write for that member.
  • memory_write binds its idempotency key to the whole batch, so a retry under the same key with more, fewer or different facts refuses (idempotency_conflict), and a batch with an invalid fact saves nothing.
  • The QM contract QM vendors is docs/contracts/qm-memory-provider/v1.json; the guide is docs/guides/qm.md.

[0.60.545.0] - 2026-11-02

The hosted gbrain.io install now tells the agent the brain is there.

Before, the hosted install in README.md, docs/INSTALL.md, INSTALL_FOR_AGENTS.md and AGENTS.md stopped at claude mcp add or codex mcp add, so nothing in the harness told the model to use the new MCP server, and in harness trials a server nobody mentions is almost never called. Each now installs the gbrain CLI with Bun (bun install -g github:garrytan/gbrain#latest-stable) and runs gbrain setup claude-code --instructions-only or gbrain setup codex --instructions-only after the MCP entry is in place. That writes the brain-first block into ~/.claude/CLAUDE.md or Codex's AGENTS.md (or its non-empty AGENTS.override.md) and the per-turn rule hook, without creating a local brain. Each harness line checks for an existing gbrain entry first and keeps it (mcp get output goes to /dev/null, since it can print a stored token), adds the hosted entry otherwise, and skips the setup step when that harness isn't installed. Adding --no-instructions to that gbrain setup command takes them out, and a bare mcp add still works.

[0.60.544.0] - 2026-11-02

The remote gate stops waiting on an artifact GitHub has refused for good, and bench and eval scripts no longer fail pregate for having no test lane (#8370, #7690). When reading a full run's executed-identities artifact failed, fetchArtifactJson marked every failure transient, so a 403 or an expired artifact's 410 kept the gate polling for up to 90 minutes before it fell back to VMs. Only no response, a 5xx, a 429 or a rate-limit 403 is transient now. Any other 4xx refuses reuse at once and the gate goes straight to VMs. This never made a gate pass, because a transient read was always none and never reuse. Separately, scripts/bench/** and the system-one eval runners and generators now declare that they need no test lane, so a PR that touches them no longer fails pregate with PREGATE_NO_LANE. The unit-lane PostgreSQL arms job also gets real headroom back. Its committed weights had drifted about 24% below what the files now take, so by measured times the busiest of master's 13 shards takes 694 s against its 570 s budget.

What changed

unit-postgres-arms (both arms, the CI cell) Before Now
Weights mined from older runs, scaled by 0.763, 6 listed files unweighted (charged the 29 s p75) median of the ARM_FILE times in eight full-ci runs from 2026-11-01 (41296667994, 41307060384, 41310113203, 41324518709, 41342917113, 41347700875, and train M's 41351675051 and 41365617585, covering every suite #8404, #8405, #8407 and M added), every listed file weighted
Shards 13 19
Two-worker makespan, largest shard, against the unchanged 570 s budget 526.4 s (7.7% headroom) by the committed weights; 694.3 s by the fresh ones 476.3 s (16.4% headroom)
Measured job wall time 438-923 s per job in train M's run 41365617585, against the 20-minute step timeout about 475 s per shard, by the same measurements
PostgreSQL-only cell, largest shard 487.2 s 333.8 s

Itemized changes

  • scripts/lib/gh-issue.ts: fetchArtifactJson marks a failed artifact read transient only for no response, a 5xx, a 429 or a 403 whose message names a rate limit (#8370). Both clients' artifact downloads now carry the HTTP status and the response text, so the gh CLI client (used when no GH_TOKEN is set) classifies the same way. A rate-limit 403 stays transient whether its headers (x-ratelimit-remaining: 0, retry-after) or its text (rate limit, abuse detection) say so, and so does a refusal from the artifact storage the download redirects to, since a fresh read mints a new signed URL.
  • test/scripts/ci-reuse.test.ts: a probe checks that 403, 404 and 410 are refusals and that 429, a rate-limit 403, a 502 and no response stay transient, and that a refused receipt ends the gate's wait at once. It fails without the fix.
  • .gstack/pregate.json: scripts/bench/**, docs/eval/system-one/runners/** and docs/eval/system-one/generators/** declare no lane ([]), which applies only where no test maps the file, so the VM-setup scripts keep test/scripts/vm-setup-usage.test.ts (#7690).
  • scripts/postgres-arm-weights.json: re-mined with bun scripts/postgres-unit-arms.ts mine from those eight runs' 98 job logs, unscaled.
  • scripts/postgres-unit-arms.ts: ARM_SHARDS 19; .github/workflows/persistence-validation.yml: the matrix, job name and run argument follow it (pinned by test/scripts/postgres-unit-arms.test.ts, whose budget and assertion are unchanged). Check names become Unit-lane PostgreSQL arms N/19 / Bun …; none is a required check.
  • docs/architecture/key-files/ci-health-and-test-guards.md, scripts/ubicloud/README.md: shard count.

[0.60.543.0] - 2026-11-01

A shared brain can now keep a page to the members it names, inside one source. Put readable_by: [alice-example, group:team-a] in a page's frontmatter, bind each member's token, OAuth client or stdio server to a member, and remote reads of that page answer only for them. Nothing changes until a page names readers or a member is created.

What changed

Situation Before Now
A page names readable_by ignored: every remote caller granted the source read it read remotely only by a credential bound to a listed member (directly or through a group); everyone else gets the canonical not_found on every read path, and a remote write to it answers page_not_found
Escape hatches and --private-pages granted-sources relaxed visibility: private still relax only visibility: private; they never open a page to an unlisted member
A visibility: private page that also names readers hidden remotely still hidden from a listed member unless that member's credential also has --private-pages granted-sources
Atoms, concepts and chronicle events derived from a restricted page followed the origin's visibility readable by no remote caller
Members and groups gbrain access member add|disable|enable|rm|list, gbrain access group add|rm|list, gbrain access who-can-read <slug>, gbrain access check --token|--client <name>; --member and --may-act-for on auth create, auth rescope and auth register-client; config mcp.stdio_member for a stdio server; --as <member> previews one member's view from the local CLI; whoami and grant_diagnosis show member and groups
Doctor access_readers_malformed, access_unknown_readers, access_recent_narrowings, access_page_readers_backfill, access_policy_pending, access_policy_widening_held, access_policy_invalid, access_policy_store_mismatch
Folder rules gbrain access rule add|rm|list|status|history|revert|export|import|migrate, gbrain access apply and gbrain access resume keep whole folders of a source to named members; a page under a rule is read remotely only by a member listed on every rule covering it and on its own readable_by. Rules live in the database on a repo-less source and in .gbrain/access.yml on a git-backed one
Changing folder rules narrowing applies at once; widening is held until gbrain access apply --source <id> previews it and --yes --expect <plan_hash> approves it; an unfinished apply reads as nobody remotely until gbrain access resume --source <id>
gbrain export and gbrain import pages only an export carries each source's applied folder rules under .gbrain/, and import installs them before publishing any page (--access-policy <source> maps a multi-source export)
A shared agent (one bot serving a channel or many DMs) read as its own credential, so it saw one view for everyone its harness names each turn's audience (X-Gbrain-On-Behalf-Of: alice-example, guest or request-level _meta.on_behalf_of); the call reads only what every entry may read, within the agent's source grant, and only a one-member audience keeps the agent's private-page relaxation. The credential lists whom it may act for (gbrain auth rescope --token <name> --may-act-for <list>); every response carries _meta.access, and whoami returns it

What to watch for after you upgrade

  • A page that already carried readable_by before this release reads as nobody remotely until you run gbrain repair access (preview) and then gbrain repair access --apply. Doctor's access_page_readers_backfill lists them.
  • On a brain with any restriction, the escape hatches no longer turn the remote read filter off, and secondary surfaces (proactive brief, retrieval reflex, entity resolution, known-as-of) hide restricted pages from every remote caller, listed members included.
  • A credential bound to a disabled or removed member is refused at verification (access_member_disabled).
  • Works the same on a repo-less brain: members, groups, bindings and folder rules live in the database. A source records which store holds its folder rules when its first rule is created, and only gbrain access rule migrate --source <id> moves them; doctor access_policy_store_mismatch names it when the filesystem changes.
  • The member-access and folder-rule SQL functions end their pinned search_path with pg_temp, like every other gbrain function (#7632), and frontmatter recovery never guesses a duplicated readable_by key: the file is held until a person fixes it.
  • Remote callers can no longer write pages under .gbrain/.
  • A database rebuilt from a git-backed source replays its committed .gbrain/access.yml history, so a removal that was held stays held. The source reads as nobody remotely until the replay finishes, and a shallow clone stays that way until git fetch --unshallow.
  • A page renamed into a ruled folder takes its atoms, concepts, chronicle events and derived_from pages with it, through later renames too and after another page takes its old slug (migration v285 keeps rename history for this; slug resolution does not use it).
  • On Postgres, a database role that neither owns the folder access tables nor bypasses row-level security now needs read policies on them, and an unfiltered policy on slug_rename_history (docs/ENGINES.md, "Folder access tables"). Without them its page writes are refused with writer_coordinator_required (unreadable_state) and its reads show nothing, instead of reading the tables as empty and treating every page as unruled. The owner, BYPASSRLS roles and PGLite are unaffected.
  • A harness that sends an audience its credential's may_act_for does not list gets on_behalf_of_not_permitted with the operator's command; an audience inside tool arguments is always refused (on_behalf_of_in_arguments). may_act_for: * is accepted and doctor warns (access_may_act_for_wildcard).

[0.60.542.0] - 2026-11-01

gbrain skillopt now makes one gated rewrite by default; the iterative edit loop is opt-in with --loop, and benchmark harnesses can drive every SkillOpt rollout.

In gbrain-evals' Cat 31 re-run the full loop and a one-shot rewrite both scored 0.54 on held-out tasks (paired bootstrap p = 1.00), and the loop cost about 1.8x as much. Until a benchmark shows the loop wins, the cheaper method is the default. The default is reversible: --loop on the CLI, or optimizer_mode: "reflect" on the run_skillopt MCP operation and queued jobs, runs the loop as before. The one-shot rewrite now reflects on training tasks only and must beat the baseline on the selection gate before anything is written, so a worse rewrite leaves the skill unchanged. Four bugs that made the loop look worse than it is are fixed too.

Itemized changes

  • One-shot rewrite is the default for the CLI, run_skillopt, --all, queued jobs and the dream-cycle phase. A job queued before this release without optimizer_mode keeps the loop it was queued for, and --resume continues in the mode its checkpoint ran (a changed mode is refused); resuming a one-shot run whose rewrite gave no usable reply retries the rewrite. Receipts carry optimizer_mode.
  • The optimizer's success criteria come from training tasks only, and a rule argument or LLM rubric is quoted only when every training task shares it; a per-task answer is described ("a task-specific required text"), never quoted. Before, every selection and test answer was shown to the optimizer. criteriaText replaces the generated criteria.
  • Held-out results never reach the optimizer. The held-out check runs once, on the final champion, before anything is written, and a refused champion is not filed in the rejected-edit buffer. Legacy held_out_regression rows on disk are skipped when the buffer loads.
  • Only edits that were applied are filed as rejected or recorded as accepted; edits the learning-rate budget clipped stay free to propose again.
  • The success-reflect prompt states the edit schema, so its edits are no longer all dropped (#7257; carried from #8324).
  • Library surface for benchmark harnesses (src/core/skillopt/index.ts): explicit tasks: {train, sel, test?} lists instead of a benchmark file; a rolloutHook that receives each scoring site's whole batch (baseline, training forward pass, selection gate, held-out, test) and whose budget refusal or malformed reply aborts the run without writing (the hook runs outside the run's budget, gets an abort signal at the runtime deadline, and a batch still running or answered after it aborts the run); scoreSkillPerTask; and receipt counts (rollouts by phase, candidates_scored_on_sel, optimizer_calls, cost_scope).
  • A skill can add the reviewed read tools recall and get_timeline to its rollouts by naming them in its frontmatter tools:; rollouts never offer add_timeline_entry or any other write.
  • Tests: test/e2e/skillopt-wave-e1b1.serial.test.ts (17 cases, each red before its fix), test/skillopt/benchmark-hooks.test.ts, and a mode case in test/skillopt/models-four-paths.serial.test.ts.

[0.60.541.0] - 2026-11-01

A saved fact from a quarantined page, or one waiting to be re-derived after a purge, no longer reaches agents through search, query, trajectories, entity cards, standing questions, local gbrain recall or consolidation.

Before, the facts search and query put first when a query names an entity skipped the read-eligibility check that recall and saved_facts apply, so remote MCP callers too got facts saved from a page the content gate quarantined, and facts a purge had hidden until a deriver regenerates them. The same check was missing from a few other fact reads. Now each one hides those facts the way recall does.

Agents connected over MCP also get a quoted page's title on search rows, its tags, and a chronicle event's kind quoted too, and a remote put_page of a whole-markdown read saved with CRLF line endings stores the page instead of its fence.

Before, a remote reader got a page's title raw on search and query rows, its tags raw on get_page, fetch and get_tags, a context card's title outside the card's trust label, and a chronicle row's kind (from an event page an agent may have rewritten) raw, while the rest of each row was quoted. A whole-markdown read whose line endings an editor converted to CRLF was stored with the read's marker and fence in the body and its frontmatter lost. A fenced title pasted into frontmatter next to a naive datetime (date: 2026-10-03 23:30:00) moved that date to UTC, so effective_date could land on another day. Local reads and owner-tier text are unchanged.

Itemized changes

  • search and query entity facts (search/entity-facts.ts) hide quarantined-page and needs_rederive facts, local and remote, and apply the caller's trust floor as saved_facts does.
  • find_trajectory, gbrain think's trajectory read, gbrain eval trajectory and gbrain founder scorecard hide them too.
  • The entity card's active_fact_count counts only the facts the card may show.
  • A standing question's refresh gathers no fact, take or timeline row awaiting re-derivation as evidence.
  • Local gbrain recall (every arm, --supersessions and --since-last-run included) hides them like the recall op.
  • The consolidate phase never promotes such a fact into a take.
  • Tests: test/search-entity-facts-eligibility.test.ts (PGLite and Postgres), red with each read's filter removed.
  • Remote put_page unfences a whole-markdown read whose lines end in CRLF, closing fence kept or replaced, and stores what a plain CRLF write of the unfenced markdown stores.
  • search, query, assemble_evidence and remote recall rows quote a quoted row's title with its text (fenceTranscriptRows).
  • context_pack and the pushed pack render a card below the owner's tier with its title under the card's trust label.
  • Remote get_page (every shape), fetch and get_tags quote a quoted page's tags; remote add_tag and remove_tag take a tag copied out of those reads without the fence.
  • The chronicle reads and volunteer_chronicle quote an event row's kind by its event page's type and tier; --narrative keeps that fence on its own lines.
  • Remote put_page rewrites only the frontmatter keys that held a fence, so every other line (a naive datetime included) keeps the bytes sent; a rewrite that would not parse back to the same values (a flow mapping continued on the next line, an anchor another key aliases) rewrites the whole block instead.
  • A keyword row returned without its text (evidence omitted under a token budget) gets its title quoted for a remote reader like the text would be, and counted in the budget.
  • A remote reader's other_names on search and query leaves out names a quoted row declares.
  • chronicle_since (kind) and list_pages (tag) take a value copied out of a quoted read without its fence.
  • Tests: test/search-transcript-quote-8384b.test.ts (PGLite and Postgres), red on master for each change.

Contributed by @tarush1989 (#7115, rebuilt on current master).

[0.60.540.0] - 2026-11-01

A write to a page with a facts or takes fence no longer re-embeds the whole page: remember, timeline and fence writes now embed only the chunks whose text changed (#6961).

Vector reuse was turned off for every page with a protected fence, so each write to an entity page deleted and re-embedded all of its chunks. One reporter measured 13 to 15 million embedding tokens a day from that alone. Entity pages carry the fences and take the most writes.

Itemized changes

  • Fenced page reuse (#6961). A fenced page now reuses a stored vector when its recorded embedding input equals the one the import would embed, the same check every other page gets. The stored index must still be sealed by the fence-safe chunker, so a reused vector's text holds no private fence row. A chunk whose text changed, such as the fence chunk a new fact lands in, is embedded again. On a fenced page the importer reuses only vectors with a recorded embedding input, so a legacy vector with none is embedded again. A one-row fact append on a five-chunk entity page now makes one embedding call instead of five. Contributed by @javieraldape (#7055, rewritten on master).

[0.60.539.0] - 2026-11-01

A remote put_page of a page an agent read whole now keeps its frontmatter and the code blocks it added, and timeline sources and ontology values are quoted below "your notes" like the rest of their rows.

Before, an agent that read a page with include_content, added a fenced code block, or had a newline prepended and the closing fence replaced, could store the page with the read's fence still in the body (its frontmatter inside it) or with its own code block's opener stripped. A timeline entry's free-text source and the ontology values volunteer_chronicle, ontology_get and ontology_conflicts return reached remote readers raw. Now the write keeps the page as the agent meant it, those fields are quoted by tier, and text copied back from them is stored without the fence. A remote read quoted only because the page's origin is unverified names the command that clears it.

Itemized changes

  • Remote put_page finds a whole-markdown read's closing fence as the first line that is exactly that fence outside a code block the content opens itself, so a backtick code block added inside, in place of or after the read is stored whole; a read with blank lines before its marker and its closing line replaced keeps its frontmatter. A fence naming another page, or one after other text, is stored as written. Content holding many markers that never close is kept as written in linear time (the unquote shares one scan budget per text).
  • get_timeline, the chronicle reads and volunteer_chronicle quote a timeline row's source with its summary and detail; remote add_timeline_entry stores a copied source without the fence.
  • volunteer_chronicle's ontologies, ontology_get and ontology_conflicts quote a value by its fact's stored tier for remote readers (quoted: "transcript"); remote ontology_propose stores a copied value without the fence.
  • A remote get_page or fetch of a page quoted because its trust tier is unknown carries a quoted_unverified_origin info notice, once per session, naming gbrain trust claim-sources (preview with --dry-run --json) and gbrain trust backfill. Quoting is unchanged.
  • Tests: test/search-transcript-quote-8384.test.ts (PGLite and Postgres), red on master for each change.

[0.60.538.0] - 2026-11-01

Agents connected over MCP now see the chronicle reads, takes reads and a quoted page's title and frontmatter quoted as source material, and the context gbrain's hooks inject quotes facts below "your notes" too.

Before, #8367 left chronicle_day, chronicle_since, chronicle_on_this_day, volunteer_chronicle's timeline, takes_list, takes_search and a quoted page's title and frontmatter returning agent-written text raw, and gbrain's per-turn hook and pushed context pack injected agent-written facts raw. Now a remote reader gets that text quoted when its stored tier is below operator_curated; owner-tier text and local reads are unchanged. Harness authors: hook output can now hold quote fences (a marker line naming the fact, then a code fence that closes on its own line).

Itemized changes

  • Chronicle reads quote each timeline row's summary and detail by its own page; takes_list and takes_search quote a claim below the owner's tier and its free-text source and resolved_source.
  • Remote get_page and fetch quote a quoted page's title and every frontmatter string; include_content and fetch text fence the whole serialized markdown, and the put_page round trip still stores the page unchanged. A thin client's gbrain get then gbrain put, or an agent that rebuilds the markdown from title and frontmatter, stores the title and frontmatter without the fences. Titles on search rows and context cards, and a chronicle entry's kind, are still returned as stored.
  • A thin client's paged gbrain get of a quoted page (above the inline limit) prints the page, and its --json content, inside one quote fence after verifying size and sha256.
  • The per-turn push hook and the pushed context pack quote hot and entity facts below the owner's tier; a fact injected before this change is not injected again.
  • context_pack counts the entity facts it withholds when the quote lookup fails (suppressed) instead of dropping them silently.
  • Remote remember stores a fact copied out of a quoted read without the fence, and remote add_timeline_entry stores an entry copied out of get_timeline without it, so the replay key matches and the entry is not duplicated. takes_add, takes_update and takes_supersede do the same for a claim or source copied out of a quoted takes read.
  • The put_page tool-call guard counts only summaries outside the content's own tool-call blocks, and admits a new turn that quotes a summary while every stored call stays; its refusal names edit_page first.
  • Tests: test/search-transcript-quote-followups.test.ts (every item, red on master for the changed reads and writes) and test/trust-keep-maintenance-reimport-8352.test.ts (maintenance re-imports keep an owner note's tier), both on PGLite and Postgres.

[0.60.537.0] - 2026-11-01

One source's stuck writes can no longer starve another source's sync: a write refused for a missing ownership marker ends terminally with its own code and restore step, the shared local principal keeps a fair share for every source, and doctor names the source holding the slots (#8368)

On Postgres, a bulk sync group whose checkout failed the ownership check when its worktree lock was taken threw out of the consumer and stayed running; after each lease lapse the owner claimed it into the same refusal, forever, while the source kept admitting more. A hundred of them filled the local CLI principal's outstanding cap, which every source's sync shares, so another source's sync stopped write_capacity and doctor read ok. Now a deterministic ownership refusal finishes every claimed, unpublished member failed (the single path already did), which releases its slot; transient failures keep retrying as before. A missing .gbrain-owner.json is refused owner_marker_missing instead of the generic identity refusal, and its fix is the verified self-transfer that re-creates it. At the principal cap a source holding less than its fair share (the cap over the sources holding outstanding requests on the local CLI principal) is still admitted, a sync publication group or page batch whole, the brain cap stays hard, and a refusal names the source holding the most slots. persistence_capacity warns when an outstanding-request cap is at 90% or more and its oldest request has waited over ten minutes, and names the source holding the most stalled requests.

Itemized changes

  • src/core/persistence/consumer.ts: failRefusedClaims settles a claimed group refused recovery_required or owner_marker_missing before publication, fenced on the claim token, no recovery record and publication not started; a member whose claim moved is left to its claimant and a connection loss gives the rest back.
  • src/core/persistence/owner-marker.ts, physical-root-record.ts, ownership.ts: owner_marker_missing for an absent stamp under a reservation that identifies the directory exactly, with the self-transfer restore as its fix; a stamp naming another owner keeps recovery_required. Registered in error-registry.ts, WRITE_ERROR_CODES, the sync fault table and the terminal classifier (the receipt's fix is the restore step).
  • src/core/persistence/journal.ts: outstandingRefusal applies the local CLI principal fair share on single and group admission (checked before the admission, so a group larger than the share is admitted whole) and names the top holder.
  • src/commands/doctor/checks/persistence-capacity.ts: details.outstanding[] with holders; a holder's writer status as the fix when it is the only finding.
  • test/write-cap-wedge-8368.test.ts (PGLite and Postgres, with test/e2e/write-cap-wedge-8368-postgres.test.ts).

[0.60.536.0] - 2026-11-01

Agents connected over MCP now see agent-written, ingested and unverified-origin text quoted as source material on search, recall, page, chunk, timeline, version and write-receipt reads, and conversation tool calls come back summarized on the reads that still returned them verbatim.

Before, the quote fence (#8066) covered chat page types and conversation-derived facts only. A note an agent wrote over MCP, a fact saved with remember, a chat message saved into a member source, and every placement read through get_timeline, search_exact, get_chunks, think's extractive answer or a write receipt reached a remote reader as plain text, and a conversation page's tool-call JSON came back verbatim through get_page, fetch, search_exact, get_chunks and think. Now any text whose stored trust tier is below "your notes" (operator_curated) is quoted for a remote reader on those reads: written by an agent, tool data, external, or unverified origin. The chronicle_* reads, takes_search and takes_list, and a quoted page's title and frontmatter still return stored text unquoted; #8377 tracks them.

Remote MCP readers now see unverified-origin text quoted until you run trust backfill. A note saved before gbrain tracked trust has no recorded origin, so it reads as unverified and is quoted for agents connected over MCP. Claim the sources that hold your own notes with gbrain trust claim-sources and run gbrain trust backfill; gbrain doctor (trust_tiers, trust_sources_unclaimed) names how many pages are affected and verifies the fix. Local reads are unchanged, nothing stored changes, and no trust tier is raised or lowered.

Itemized changes

  • src/core/search/transcript-quote.ts: quotesTier / quotesStored (a missing tier reads as unknown, so it fails closed); quoteFacts reads each fact's stored tier from the database for a remote reader.
  • Search, query, recall, assemble_evidence, saved facts, context_pack, delta, _meta.brain_hot_memory (its cache key carries the quoting), get_page (body, timeline, windows), fetch, get_timeline, get_chunks, search_exact, think's extractive fallback, get_versions and known_as_of reads (by the version's tier snapshot; a version without one reads as unverified) and list_write_requests / get_write_request outcome text quote by tier for remote readers.
  • Tool-call blocks in conversation pages are summarized on get_page's reading shape and view: read windows, fetch, search_exact, get_chunks and think; include_content, view: markdown and gbrain's thin client keep the stored bytes inside the fence. A remote put_page whose content holds a tool-call summary that a whole read of the stored conversation page never produces, or keeps a call's summary but fewer of that call than the page holds, refuses with invalid_params and stores nothing.
  • A remote put_page round trip of any quoted page stores the body, never the fence; a forged fence in stored text stays inside the quote.
  • Rendered context_pack/delta lines put a quoted fact's label on the marker line and close its fence on its own line.
  • gbrain doctor: trust_tiers and trust_sources_unclaimed say remote readers see unverified pages quoted and name the claim and backfill commands.
  • BrainBench poisoning: a payload inside a closed quote fence is judged by its fence's marker line, where its label is rendered.
  • Tests: test/search-transcript-quote-tiers.test.ts (every placement and read surface, red on master), test/search-transcript-quote-legacy-claim.test.ts (legacy note quoted, raw again after claim and backfill), both on PGLite and Postgres; existing fixtures that stand for the owner's own notes now write at the owner tier (test/helpers/owner-tier.ts).

[0.60.535.0] - 2026-11-01

Master's fence-census suite passes again after the 2026 DST fallback. The test 'the incremental pass in batches of one reads every page across a DST fallback in the session time zone' wrote fixed updated_at values in the 2026-11-01 New York fallback hour. The census only reads pages updated at or after its watermark (the database clock minus two minutes), so once the clock passed about 05:22Z on 2026-11-01 those pages began falling behind the watermark, and the test has failed on every run since. It now uses the next fallback that is still ahead of the database clock and asserts that the four instants read 01:20-04, 01:40-04, 01:10-05 and 01:30-05 in the session time zone, so the DST coverage can't quietly disappear. It also runs with the clock pinned inside the 2026 and 2027 fallback hours.

Itemized changes

  • test/fence-census.test.ts: the DST-fallback census test derives its instants from the next New York fallback after now(), checks their local text offsets, and runs at the real clock plus two pinned clocks (2026-11-01T05:30Z and 2027-11-07T05:30Z). With the old fixed instants, all three fail.

[0.60.534.0] - 2026-11-01

Merge train S: five sync, refresh and maintenance fixes land in one release. Page writes create a missing source folder and name a missing checkout (#8106). A diverged managed checkout can publish itself on opt-in, and every refresh fetch is logged (#8102). Sync-lane wave 2 repairs quoted fence markers, releases agreeing file-drift holds, syncs plain directories and retires former-binding effects (#8077). The synthesize phase survives a local writer replace (#8111). A door commit during a refresh is waited out instead of failing the refresh as refresh_dirty (#8117).

Missing source folders and checkouts (#8106)

Page writes to a source whose directory is missing on the brain host no longer fail with a misleading storage_error. A source that is a folder inside a checkout now gets its folder created on the first write, and a source whose whole checkout is gone gets a refusal that says so. A checkout restored or copied into place, or one whose private ownership records lost their mode, now refuses with the exact owner step that fixes it.

What changed

Situation Before Now
The source is a folder inside a checkout and the folder is missing (Git keeps no empty folder, so a fresh clone of a repository whose member folder has no files yet lacks it) every put_page, capture and add_timeline_entry to it failed with storage_error ("The canonical file target cannot be accessed as a regular file"), so a token granted only that member's source could save nothing the write creates the folder and the page, and the same token reads it back
The source's whole checkout is missing (a brain copied to another machine or into a container without its source directories) the same storage_error, whose fix told the agent to repair file permissions recovery_required with detail: source_checkout_missing, naming the source; the fix is the read-only gbrain sources writer status --source <id> --json, which shows the registered path. Remote callers are not shown the path. Nothing is written, as before
The checkout was restored, copied or recreated at its recorded path (new inode or birth time) recovery_required whose fix only said to read the receipt detail: physical_root_directory_changed, and the fix is the owner-run relocated-root self-transfer (gbrain sources writer transfer prepare then accept, both with --self-transfer --confirm-relocated-root) with next: tell_user_to_run. The re-stamp stays a deliberate step on the brain host
An ownership record of the checkout became readable by other users (a recursive chmod of a copied brain) recovery_required ("The private physical-checkout identity cannot be verified"), and doctor's foreign_ownership_marker called the file foreign and told the operator to remove it detail: ownership_record_not_private with gbrain doctor --only foreign_ownership_marker as the fix; doctor now reports this brain's own record with its mode and the exact chmod 600 '<path>', never removal, and lists each record once

A token's source grant decides where it writes exactly as before; this changes only how the file path inside the granted source is resolved and how these refusals name their exit.

Publishing a diverged checkout (#8102)

A managed checkout that gets local commits while a sibling checkout pushes the same branch can now stay current without a person in the loop. Turn on sources.refresh_publish_diverged and a diverged gbrain sources refresh publishes the local commits merged with the upstream, then fast-forwards.

Before, such a refresh always refused refresh_diverged, and the database lagged the upstream until someone ran the keep-checkout publish by hand or a scheduled publish caught up. That repeated every interval. The setting is the user's standing consent to push those local commits. It defaults to off, so nothing changes for existing brains.

What changed

Situation Before Now
Diverged checkout, setting on refused refresh_diverged pushes HEAD merged with the upstream in memory (the gbrain git commit-push --push-only --keep-checkout publish), fast-forwards to the merge and reports published (sha, local commits, upstream commits merged)
Diverged checkout, setting off (the default) refused refresh_diverged unchanged; the refusal now also names the setting
A file both sides changed, setting on n/a refuses integrate_conflict naming the file; nothing is pushed, fenced or moved. With gbrain.quarantine on, the remote's version is published instead and the refresh names the ref that keeps this machine's version
An uncommitted edit to a file the upstream changed, setting on n/a refuses refresh_dirty before anything is pushed
An upstream rewritten by a force-push (it no longer holds a commit it held before), or no record of the upstream before the fetch, setting on n/a refuses refresh_diverged and pushes nothing, so the commits the rewrite removed never come back without the user
A force-push that lands after the refresh checked the upstream but before the publish pushes, setting on n/a refuses push_unconfirmed and pushes nothing. The publish is bound to the upstream the refresh checked (--expect-remote), so the removed commits never come back. Run the refresh again; it checks the new history first
A refusal after the publish landed n/a the refusal says the remote already holds the local commits
--dry-run refused refresh_diverged unchanged: a dry run never publishes
gbrain config set sources.refresh_fetch_timeout_ms / sources.refresh_drain_wait_ms refused as an unknown key, although the refresh reads both accepted

Things to watch after you upgrade

  • The publish needs a Git identity in the checkout for the merge commit, the same as the manual keep-checkout publish.
  • It never uses a --pull that moves the checkout. The merge is built in memory and pushed, and the checkout moves only through the refresh's usual fast-forward.
  • The push skips the checkout's pre-push hook, as the manual keep-checkout publish does. If you rely on a pre-push secret scanner, leave the setting off.
  • The setting covers every managed checkout in the brain.
  • If another publish of the same checkout is running, the refresh refuses at once instead of waiting; run it again.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: when sources.refresh_publish_diverged is on and the run is not a dry run, the precheck runs the keep-checkout publish (gbrain git commit-push --push-only --keep-checkout in a bounded child, never on the PGLite owner's event loop) for a diverged checkout (after refusing refresh_dirty first when an uncommitted edit touches a file the upstream changed), re-fetches and continues; it refuses refresh_diverged instead when the checkout holds a commit the upstream held before (reachable from the remote-tracking ref before this fetch, from its reflog, or from a completed refresh's target for the same upstream) but no longer contains, or when no record exists (no reflog entry, no completed refresh, and core.logAllRefUpdates off, so the ref before the fetch may already be past a force-push); every refresh fetch, with the setting on or off and in a dry run, records the move in the reflog, so a refusal's own fetch can never erase what the upstream held; the result carries published (with any quarantined or superseded paths, which the text output names), and a refusal after the publish says the remote already holds the local commits. The refresh_diverged suggestion names the setting.
  • src/commands/sources-refresh.ts: a text line for the publish.
  • src/core/config-runtime-keys.ts: registers sources.refresh_publish_diverged, sources.refresh_fetch_timeout_ms and sources.refresh_drain_wait_ms.
  • Docs: docs/guides/live-sync.md, docs/guides/write-refusals.md, docs/guides/troubleshooting.md, docs/guides/shared-checkout-git.md, docs/architecture/canonical-writers.md and its census.
  • Tests: test/refresh-publish-diverged-7686.test.ts (setting on and off, a conflict, quarantine, a dirty overlap, a refusal after the publish, a publish child that fails, and a dry run, on PGLite and on Postgres through test/e2e/worktree-refresh-postgres.test.ts). Tests 7c and 7d in test/persistence-worktree-refresh.test.ts share one fixture across their cells and run in about half the time, with every assertion kept (#7782).

How to get it

gbrain config set sources.refresh_publish_diverged true

Sync-lane wave 2 (#8077)

Sync-lane wave 2: a transcript that quotes a fence marker is repaired instead of held for a person, held files that already agree with their pages can be released, a plain directory added with --force syncs, and a parked effect left behind by a replaced binding has a guarded way out.

A page that quotes a facts or takes marker in prose was held invalid_fence as repeated_marker for a person to fix. It is now held as quoted_marker, and gbrain repair fences, gbrain repair content, gbrain sync unblock --apply and the maintenance run fix it by wrapping each mention in backticks, with a receipt per file. No reader counts fence lines any differently (#8059). Older file_drift holds on pages with no import record are checked once more against the file's earlier versions (#7199, re-port of #7228). A held file that agrees with its page apart from formatting is released (#7931, #7950). A refused repair file-drift --apply no longer reserves its request id (#7931).

Itemized changes

  • #8059 quoted fence markers. The new src/core/fence-repair/quote-markers.ts wraps prose mentions: an exact marker outside code and table lines, with text after it on its line or on a blockquote line, that doesn't close a fence holding rows and isn't directly above a table. The edit is kept only when every row of every fence keeps its line and kind, and when nothing protectedRegions hid becomes visible. Only the repair makes the edit; write paths just name the reason quoted_marker. Existing invalid_fence holds get one more screen through a hold-only quote_version, and the parser version stays the same. forget now finds the facts fence with the parser's code-aware locator.
  • Quoted-marker screen stays linear. The write-path screen for prose-quoted markers found each mention's line by re-reading the page up to it, so a page with 40,000 mentions took 12 to 18 seconds and a 24 MiB page minutes. It now reads each page once, and the same 40,000 mentions take about 100 ms.
  • markerCarry. A segmented read carries only a marker that a cut splits, not a whole one.
  • #7199 (re-port of #7228). A record-less page that equals the file at the last synced commit, or one of its three earlier committed versions, imports without a hold. Side-unknown hold text now says what was compared. A pass that holds files for a person reports held_needs_human. Git tree listings are memoized by object id (GBRA-135).
  • #7931 parts 2 and 3. A held file whose page agrees apart from formatting is released. The request id of a repair file-drift apply now includes the approved artifact digest, so an apply refused before publishing doesn't block the next preview.
  • #7950. An agreeing held page found through --slug is listed release_hold, bound to the hold's observed_at. An empty preview applies nothing. Contributed by @Masashi-Ono0611 (supersedes #7953).
  • #7974. The stale-mirror drift read no longer throws on an ambiguous slug path, and it compares contents. slugPathFile uses the source's slug-root mode and doesn't follow symlinks. release_hold needs the held path and a file hash. The keep_file_bytes and file-wins docs are corrected.
  • #7973. gbrain sync on a plain directory added with sources add --force runs the import path instead of failing on git rev-parse. The managed Git path refuses with a typed sync_not_applicable.
  • #7988. Activation marks a non-git root in its own directory, never in the metadata of the repository that contains it, such as a dotfiles $HOME.
  • #7985. Parked effects of a former binding get their own group, with settle-effects as the next step. The exact retry names both bindings. settle-effects --retire (proof former_binding_retired) settles one only when its source was replaced or its page moved past the owed revision. The settlement is preview-first, bound to a hash, reversible with --unsettle, and never replays anything.
  • #7982. Repeated transient pass stops build the abandoned-pass streak. An unfenced lease_lost stop never inserts a row. The pass-row and drain-holder readers compare against the database clock.
  • #8009. The repair provenance-stamps preview and apply share one held-page filter, so the preview's count equals what the apply writes.
  • #7998 flake. The page-embedding-lease harness pushes other queued effects out of the way before it drains the one under test.

Synthesize postprocess after a writer replace (#8111)

A dream cycle's synthesize phase no longer fails on every pass after gbrain auth local-writer register cli --replace (#8110). Postprocess settles on the earlier request it already found, checked under the current local CLI writer's grant, instead of looking it up again under the new principal and publishing against a stale revision (revision_conflict forever). A synthesis child output whose own stdio writer was revoked warns once as a permission_denied conflict and leaves the page alone. Receipt lookups stay principal-scoped.

Itemized changes

  • src/core/cycle/synthesize-postprocess.ts: settle on the found row; a revoked child writer is a one-time permission_denied conflict.
  • Test: test/managed-synthesis-postprocess.test.ts (PGLite, and Postgres through its E2E wrapper).

Door commits during a refresh (#8117)

A gbrain git commit-push landing during a gbrain sources refresh no longer makes the refresh fail as refresh_dirty (#8117). The refresh now runs its HEAD check and fast-forward merge under the checkout's commit lock, the same lock the door uses, so it waits for a door commit to finish instead of colliding with it on .git/index.lock.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: fenceAndMerge takes the checkout's commit lock (brain-commit.lock plus the legacy push lock, waiting GBRAIN_COMMIT_LOCK_WAIT_SECONDS, default 60) from the HEAD check through the merge. A lock still held past that wait aborts the fenced refresh as retryable commit_lock_busy. A git process outside the lock that holds .git/index.lock gets three short retries, then a retryable git_index_locked refusal instead of the caller-class refresh_dirty. Either way HEAD is unchanged and the next run starts fresh.
  • Test: test/worktree-refresh-index-lock-8117.test.ts.

[0.60.533.0] - 2026-11-01

A delegated sync no longer reports done while it is still finishing. The serve's sync runner published done before it re-checked the wanted markers and scheduled the deferred embeds. A client polling in that window saw a finished job with no embed backlog queued yet, and a second sync_start could begin while the first run was still settling its markers. Since #7941 the marker re-check also reads the checkout HEAD through git, which widened the window enough to fail test/serve-sync-runner.serial.test.ts intermittently on master (#8354). Both lanes (the shared-secret lane and the verified lane) now set done only after those steps, and a deadline or client abort that arrives after the run has finished no longer marks it aborting.

Itemized changes

  • src/core/serve-sync-runner.ts: job.state = 'done' moves after settleRunMarkers and scheduleDeferredSyncEmbeds in both lanes (#8354); the deadline and sync_abort skip a job whose run already returned its result.
  • test/serve-sync-runner.serial.test.ts: a probe slows the post-run marker read and checks that done is only visible once the markers have settled and the deferred embeds are pending; a second probe aborts during that settle and checks the job still ends done. Both fail without the fix.

[0.60.532.0] - 2026-11-01

CI wave 3: a passed full-ci pair can now clear the remote gate without a second full suite on VMs, the import-rate scale gate no longer fails on one stall, an effect pass that loses its database connection gives its claims back at once, and the Ubicloud runner's teardown retries instead of leaking VMs (#7939, #8103, #8096, #7607, #8119).

The full-suite-tree and e2e-full-suite-tree artifacts now record every test file the run executed and the executed test count, taken from the run's complete ci-manifest.json, so gstack's pregate clear --run can accept a passed full-ci pair on the same tree as run evidence. bash scripts/pregate-remote-gate.sh --list prints the gate's own inventory, and the seven files that never left a receipt on GitHub now do, or leave the gate. The import-rate gate sets aside at most one stall per import (a paused database container, a WAL checkpoint) and never work that recurs; a cost that lands only once in a 10k import is a known gap the 20k nightly still sees. An effect pass whose database connection drops while claiming or flushing git effects gives back every claim it had not finished as soon as the database answers, within 30 s and stopping on shutdown, instead of leaving them to a 2-minute lease. The remote gate's CI-reuse wait keeps waiting for the head's own dispatch, refuses when the run in progress fails, and treats a failed artifact download as retryable. The Ubicloud runner's down retries a timed-out or 5xx destroy with backoff until UBI_DOWN_TIMEOUT, reap bounds each VM's wait, and UBI_REAP refuses spellings it cannot read.

Itemized changes

  • scripts/ci-full-suite-tree.sh, .github/workflows/test.yml, e2e.yml: files and tests in tree.json from a complete, matching manifest; the full jobs need executed-receipts. The reconciliation crash suites run through run-with-receipt.sh (now accepting bun --no-env-file test); the keyed-only test/e2e/skills.test.ts leaves the keyless gate.
  • scripts/pregate-remote-gate.sh, scripts/ci-ubicloud.ts: --list prints the gate inventory and fails on an empty one.
  • scripts/scale/gates.ts: one-stall rule for import_rate; import.<basis>.excused names the time set aside (#7939).
  • src/core/persistence/effects.ts: connection-loss give-back for claimed, set-aside and unflushed git effects, fenced on each claim's token (ported from #7607).
  • scripts/ci-reuse.ts, scripts/lib/gh-issue.ts: the gate waits for a run in progress beside a missing one, refuses a refused receipt only on the head itself, and retries a failed download; an earlier attempt's ledger counts only when its job was carried over (#8096).
  • scripts/ubicloud/ubi-runner.sh: down destroy retries with 1 to 60 s backoff, reap per-VM bound UBI_REAP_DOWN_TIMEOUT, strict UBI_REAP (#8103, #8096).
  • scripts/mine-shard-weights.ts: prunes against tracked files in every mode; shard weights refreshed (#8119).
  • Test fixes ported from #7607: page-embedding lease ordering, git-commit-push-6530 cases, timeline-chunk-reuse handback, segmented-prior-view heap sample, awk portability lint.

[0.60.531.0] - 2026-10-31

Five severe fixes: the secret scan now redacts keys that follow a JSON or percent escape, a connector page that turns private leaves Git HEAD, deleted pages no longer stay searchable under a scoped writer role, a stale freshness marker can no longer starve the job queue, and a slow embedder no longer loses remember writes (#7978, #7971, #7637, #7941, #8135).

The secret scan's rules required a non-word character before a key, but a JSON escape (\n, \t, \u0022) and a percent-escape (%22) end on a word character, so a key at the start of an escaped tool-output string or inside a URL-encoded value was never redacted. Every rule now also matches right after such an escape (an escaped quote between a keyword like password and its value is still a gap, tracked in #8265), and a vendor key that recurs glued to a word (token_<key>) is scrubbed wherever the same key is claimed elsewhere in the document. On a managed connector source bound to a Git checkout, a page that turned private kept its public file at HEAD; it is now removed in the same publication. A page that already turned private before this release keeps its old public file until it is removed by hand (#8269). Under a database role with row-level security that lacks policies on the derived-state tables, deleting a page left its retrieval rows searchable and page versions lost their trust and attribution stamps without an error; those writes now refuse with writer_coordinator_required, and docs/ENGINES.md lists the policies such a role needs. A freshness marker naming a commit the source had already passed kept every sync queuing another top-priority follow-up; a completed run now settles it once the checkout has nothing left to import. A remember whose embed took longer than its 30 s preparation budget ended preparation_stalled with nothing saved; the fact is now saved without a vector and embedded later.

Itemized changes

  • src/core/secret-scan.ts: escape-boundary twin for every rule, run only on lines holding an escape; vendor and JWT values (20 to 512 characters) join the echo dictionary under their own 32-value cap, so planted vendor strings cannot crowd out a bearer token's echo (#7978). The high_entropy_assignment shapes moved to src/core/secret-scan-assignment.ts (module size cap).
  • src/core/persistence/connector-sync.ts: a strict private import of a public page publishes the file removal from preparePrivateTransition (#7971 item 1).
  • src/core/derived-state-maintainers.ts, migration v281: the eight derived-state maintainer triggers stay SECURITY INVOKER and fail closed when the caller's row-level security hides a table they need; docs/ENGINES.md "Derived-state maintainers" recipe (#7637).
  • src/core/notifier/store.ts, handoff.ts, resident.ts: settlement takes the checkout HEAD after the run (#7941).
  • src/core/facts/single-prepare.ts, src/core/persistence/memory-prepare.ts, service.ts: the fact embed is bounded inside the preparation deadline (#8135).

[0.60.530.0] - 2026-10-31

gbrain sources refresh no longer exits with a bare internal_error when another database transaction holds a row its fence needs. A fence that waits past its 5 s lock timeout now refuses as refresh_lock_timeout, and why names the open transactions holding those tables. Nothing moves, and a later run completes once the holder is gone. Any other unexpected error now prints a typed internal_error refusal with its stack in why and the refresh row it left, which the next run resumes.

Before, a stuck transaction holding a source, worktree or brain row (a long-running job, for example) made an upstream-moved refresh fail after about 5 s with canceling statement due to lock timeout. That surfaced as a top-level internal_error with no status, no stack and no holder. If it happened after the drain, it also left a draining row for the next run to clean up.

Itemized changes

  • src/core/persistence/worktree-refresh.ts: Postgres 55P03 in either fence transaction (fenceDraining, fenceAndMerge) is the new refresh_lock_timeout refusal (registry, catalogue, docs/guides/write-refusals.md, docs/guides/error-codes.md). Its why lists the five oldest open transactions in the current database holding locks on persistence_brain, persistence_worktrees, sources or persistence_worktree_refreshes (pid, application, state, seconds open). These are candidates, and the blocker may not be among them, because pg_locks does not show which row each one holds; the fix says to confirm the blocker before ending any transaction. Plain readers, which cannot block the fence, are left out, the holders' SQL text is never copied into why, and a lookup that fails or takes longer than 3 s says so (the statement is cancelled at that bound, and after the drain the lookup runs only once the topology and worktree locks are released). A refusal after the drain aborts the draining row as other fence refusals do. That abort has its own 5 s lock timeout. When the same holder blocks it, the row stays draining for the next run instead of the refresh hanging, and the refusal says so.
  • src/core/persistence/worktree-refresh.ts: refreshWorktree turns any other non-OperationError throw into internal_error with the error name, error code (a SQLSTATE or a Node code) and message (credentials, connection details and the home directory redacted), the stack in why, and a fix naming the active refresh row and its state, or saying the row could not be read (also when that lookup takes more than 3 s; its statements are cancelled at that bound). A deadlock (40P01), serialization failure (40001), cancelled statement (57014) or lock timeout outside the fence (55P03) is marked retryable. The fix repeats --dry-run, --abandon, --resume, --wait-drain and --fetch-timeout-ms when given. Home-directory redaction replaces only the home path itself, not a sibling path that shares its prefix. The row is not changed, and a host identity file that cannot be read still yields that refusal. A database access error the classifier recognizes (connection refused, a missing table) is rethrown unchanged, so it keeps its gbrain db-repair or gbrain apply-migrations fix. The already-current path no longer throws a TypeError of its own when a member sync rejects with a non-object.
  • src/commands/sources-refresh.ts: refusals print why (a why key under --json, a Why: line otherwise).
  • Test: test/worktree-refresh-unexpected-throw.test.ts (PGLite, and Postgres under DATABASE_URL; the lock-timeout cases hold a source row from a second connection, before the fence and after the drain, and check that a plain reader is not named and no SQL text is echoed; an unreadable host identity file still gives a typed refusal, an access error is not wrapped, credentials and the home directory are redacted without touching a sibling path, --abandon and --dry-run survive into the fix, a lookup that never answers still gives the refusal and is cancelled (row lookup and lock-holder lookup), transient SQLSTATEs are retryable, the holder list is worded as candidates, and an abort blocked by the same holder still returns and says the row stays draining).

[0.60.529.0] - 2026-10-31

A pull request that adds a file to the PGLite snapshot inputs no longer gets stopped at gstack's pregate secrets check by its own workflow cache keys (#8107).

Pregate read every key: pglite-snapshot-${{ runner.os }}-${{ hashFiles(...) }} line in the workflow files as a possible secret (env.kv), and a branch can't allowlist findings on its own lines. .gstack/pregate.json now allowlists that one finding on those cache-key lines, and only when every hashFiles argument is a lowercase path or glob under src/, test/ or scripts/. An argument with digits, capitals or a ./ prefix still gets flagged, and so does any recognized credential on the line. A new test checks that every cache-key line in the workflows is covered and that secret-shaped values are not. Nothing changes for users of the gbrain CLI.

[0.60.528.0] - 2026-10-31

The test-isolation scan is about 25 times faster on Linux and 12 times faster on macOS, so bun run verify and the macOS validation job no longer come close to their per-check caps (#8053, #8099).

check:test-isolation started several processes per test file, about 2,400 in all. It now reads every file in one awk pass with byte-identical output under mawk, gawk and macOS awk. Linux: 123 s to 1.3 s alone (about 5 s inside verify). macOS 26 validation (bash 3.2, a 240 s cap): 134-146 s to 10 s. Neither cap changed.

This release also brings the Ubicloud runner up to date with gstack main, refreshes the shard weights from the median of ten recent runs, fixes CI gate edge cases and removes several flaky tests at their root cause. Nothing changes for users of the gbrain CLI.

Itemized changes

  • scripts/check-test-isolation.sh: one awk pass over all files instead of per-file greps (#8053, #8099).
  • scripts/ubicloud/ubi-runner.sh, .gstack/ubicloud.json, scripts/ubicloud/reap-expired.ts: runner synced from gstack main, owner-tagged VM names with an expiry (3 h default), a per-command idle watchdog, a committed remote verify, and UBI_REAP=0 by default until legacy VMs without a TTL are gone. scripts/ubicloud/ci-item.sh now exits with the worst item's exit code, so the committed verify fails when its item fails.
  • scripts/ci-ubicloud.ts, scripts/pregate-remote-gate.sh: a gate that exits before requesting a VM reports left=0 vms=0, so its failure stays a gate result instead of an infrastructure failure.
  • scripts/ci-reuse.ts, scripts/ci-executed-counts.ts: gate reuse and executed-count edge cases (#8001, #7927). scripts/ci-mcp-tap.sh, scripts/ci-openclaw-mcp.sh: Tier 2 counts a tool call only when its request was answered (#8011). The crash robot stops a lane after a worker failure (#8062).
  • scripts/mine-shard-weights.ts, scripts/test-weights.json, scripts/serial-weights.json, scripts/e2e-weights.json: weights are the per-file median of several runs; a complete re-mine drops files only older runs measured.
  • Tests: flaky tests fixed at the cause (#7922, #7933, #7957, #7964, #8025, trust-release fence); the #7679 refresh matrices 7c and 7d moved to their own files to fit the per-file E2E cap (#8025).
  • docs/TESTING.md, scripts/ubicloud/README.md, docs/eval/: runner TTL and reap defaults; long repro recipes pass --ttl-hours.

[0.60.527.0] - 2026-10-31

**A shared brain can now let each member's agent read that member's own private pages, and nobo