Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

31.4 KB
Raw Blame

Memory removal

Decision

Remove the Memory product from the Avibe codebase and distribution. This is a forward removal with compatibility for configurations written by released versions.

Invariants

  • Remove Memory capture, recall, processing, runtime, sidecar, API, RPC, CLI, Web UI, prompt injection, dependency management, packaging, release scripts, feature documentation, and feature tests.
  • Preserve generic uses of the word memory, such as in-memory data structures and cgroup resource limits.
  • Never read, write, migrate, delete, or clean existing Memory data. Leave both ~/.avibe/memory and the legacy ~/.vibe_remote/memory path untouched, including any non-Markdown files and symlink state. Do not add a cleanup command or stop/inspect a legacy sidecar.
  • Remove the avibe-memory package, the memory extra, lockfile entries, and runtime build/release integration. Do not uninstall a package already present in a user's environment.
  • Treat the persisted memory config key as an obsolete unknown field. Loading an old config, including one with malformed Memory data or enabled: true, must succeed without warning, runtime activation, or data access. A later successful config save omits the key. API config payloads silently discard it.
  • Removed HTTP Memory routes return 404 and the removed CLI command is an unknown command.
  • Add one short general upgrade note stating that Memory was removed and its existing data was untouched.

Installed-shape and scenario audit (review 5281184620)

Before edits at d84f7052bdf3af4f128083ba177fd40100d4159a, the orchestrator authorized a bounded whole-consumer audit of retained installed compatibility and restored scenario evidence. Inventory: k1GlY incorrectly counts the retained companion's namespace metadata as core integrity evidence; k1Glk points MH-RUNTIME-008 at disabled reconciliation rather than convergence. Exclude only the canonical companion distribution at the shared provider boundary; preserve core/legacy-core disagreement, unknown/editable behavior, and repair provenance checks. Inspect historical wheel metadata without import, activation, installation or mutation. Verify normal versus forced upgrade plans through real metadata discovery in test-owned directories. Audit affected catalog pointers against executable behavior, preserving disabled-mode tests. The two historical release-policy findings remain unresolved pending owner decision. No publication, guard restoration, merge, or premature closure.

Audit/evidence: downloaded the immutable v3.1.0 wheel into ignored test-owned .runtime/legacy-provider-evidence; SHA256 ce5cfb1473442642d7d19863b6f9131c0a17a1fb6e828d5e0aec04f2deea06b0. Read-only zip/real importlib metadata discovery confirms no top_level.txt and a RECORD manifest entry causing both avibe_memory and vibe namespace ownership. No artifact code was imported or installed. Regression fixtures exercise actual dist-info discovery, four canonical name spellings, current and legacy core, unrelated vendor providers, real mismatches, companion-only and unpublished editable shapes, immutable companion bytes/mtime and explicit core-source exact repair/preflight. CLI and API forward upgrades both consume the shared planner. The historical automatic origin-pair resolver's only caller was the removed Memory-package repair; it is not restored. Explicit core repair sources and availability checks remain unchanged, without companion fallback.

Focused upgrade/provider/local-dependency/catalog run: 319 passed. Catalog audit compared both changed retained catalogs and the previously restored SCT-038/AUTH-SETUP-907 pointers to test bodies. MESSAGE-DELIVERY-316 executes hydrated WeChat addressing; 317 verifies canonical text and sender; SCT-038 checks localized refusal copy; AUTH-SETUP-907 tests actual disposable IPv4/IPv6 origin listeners. Only MH-RUNTIME-008 was mismatched. Its convergence test is restored as the catalog target; disabled coverage stays without that ID. The four audited scenario targets reran: 7 passed, 2 subtests passed. Changed-Python Ruff and diff whitespace checks passed. No UI source changes. Current-head automated review/CI and the two owner release decisions remain required; no completion/closure exception is inferred.

Acceptance evidence

  • Load and save fixtures for old configurations prove the obsolete subtree is ignored and omitted while unrelated fields round-trip unchanged.
  • A sentinel under each legacy data path remains byte-for-byte unchanged after service/config probes.
  • Route and CLI contract tests prove removed entry points have disappeared.
  • Packaging/import checks prove no new distribution imports or includes Memory.
  • Focused Python tests, ruff check on changed Python files, and the UI build (if UI files change) pass.

Shared-consumer repair contract (2026-09-22)

The owner and orchestrator authorized one coherent repair round after reviews of 4a91228ab and e5bf5fbe found the same structural defect: deleting by Memory symbol, line, or file also removed co-located generic behavior and coverage. This is the circuit-breaker scope decision, not authorization to delete more generic tests or restore the removed product.

  • Merge current master, preserving upstream generic behavior. The only textual conflict against master 4b964fef2 is the independently added async-plugin guard; retain upstream coverage and use the retained synchronous release suite.
  • Restore release automation checkout and asset shell dispatch; exercise paths, workflow SHA selection, shell syntax, upload order, and asset integrity.
  • Preserve legacy author hydration and recursive privacy filtering of GENERAL delivery metadata. Never traverse or migrate Memory data.
  • Remove unused Memory proof plumbing and companion-only upgrade preflight, retaining caller ACL/origin binding and generic upgrade failure behavior.
  • Prove removal using actual wheel/sdist members and metadata, an installed import/startup probe outside the source tree, route/CLI behavior, old Editor payloads, and byte/symlink identity sentinels in isolated test homes.
  • Restore generic browser fallbacks and i18n/scenario coverage; remove obsolete feature documentation. Audit mixed-purpose deletions against latest master.
  • Diagnose packaged startup failures from test-owned logs before changing any timing policy. Verify Linux startup in the sanctioned isolated environment.

Review inventory: 4a91228ab has six inline findings plus one review-body finding; e5bf5fbe has two inline findings plus one review-body finding. All belong to the shared-consumer removal class; prior fixes must be verified against current code. No merge, release, installed-service restart, or production-data operation is authorized. Acceptance requires behavioral and artifact evidence, not grep or collection alone. The existing orchestrator watch remains authoritative.

Repair evidence (2026-09-22)

  • Integrated master 4b964fef2 in merge 3d4aa9763. Conflict inventory: one add/add file, tests/test_conftest_async_plugin_guard.py; preserved upstream generic guards and changed only the retired suite target. Thirteen guard tests passed. A final fetch confirmed master had not moved.
  • Read all three review bodies and all thirteen threads, including the f6 head review. Restored generic tests for CI resource metrics, remote authorization cache races, OpenCode restored polling/prompt/steering, dependency UI refresh, upgrade failure/atomicity, and config/delivery/runtime consumers. Retained SCT-038, AUTH-SETUP-907, MESSAGE-DELIVERY-316 and MESSAGE-DELIVERY-317 coverage.
  • Concentrated behavioral run: 2,259 passed, 14 subtests passed across 31 suites. Includes the real released HTTP method/path pairs, generic unknown API 404, existing API 405, SPA navigation, Editor legacy filtering, opaque byte/inode/symlink sentinels, legacy author/message-kind hydration, recursive privacy projection, and both release upload scripts with mocked GitHub. Extracted workflow shell passes bash -n; sparse-checkout path/ref behavior is replayed using a test-owned Git repository.
  • Built a wheel from the actual sdist with uv build in isolated HOME/XDG/uv cache. tests/test_distribution_artifacts.py: 2 passed. Inspected both archives and metadata; retained Show manifest/router/UI contracts; installed the wheel and dependencies in a fresh venv and imported with -I outside the source tree while the Memory package was unavailable. This is import evidence, not running-service evidence.
  • Separate Linux packaged user flow: 1 passed in disposable Debian Docker, installing that wheel as a new user without .local/bin on PATH; service reached running with internal server ready. Startup's root cause was orphan decorators left on _init_modules and _run_im_runtime, not a short timeout. Bound-method regression tests now exercise both. Diagnostic EXIT handling also avoids Debian login-shell logout masking a successful probe. Node/Show preparation is explicitly skipped by this existing installer smoke; it does not claim live IM or Show-runtime end-to-end coverage.
  • Artifact directory (ignored, retained locally): .runtime/removal-dist-reviewed. Wheel SHA256: 3fb24491e04f81db1a3c4586e2f5c56c5ee22f828ad1e0d1e9646ab443be7acc. Sdist SHA256: 1d8daca30826ccabd016fc8870a42c38c2921f4521af11b8ff3476bf60d28fec.
  • Node 24.12: UI build, lint-baseline and test typechecks passed; restored dependency UI tests 15 passed; affected sidebar/geometry Playwright suites 27 passed with test-owned Chromium/HOME. Unknown-request denial stays strict; inbox/version/events fallbacks are restored.
  • Ruff passed on all existing changed Python files against master and new Python suites; git diff --check passed. Collection was also successful (23,029 nodes before the final seven regression cases), but is not acceptance.
  • Remaining delivery gate: pushed-head Codex review and CI. No manual review trigger, merge, release, deployment, local reinstall or service restart.

Legacy GENERAL boundary completion (review 5280069619)

Before edits, orchestrator sesk8rfbcfr46 authorized a coherent continuation: the repeated class is incomplete legacy GENERAL persisted/IPC compatibility. Identity must remain distinct through queue collection and merge, before hydration. Old scheduler argv must remain consumable by the new parser without restoring retired behavior. Audit the complete consumers, not only the failing line; preserve modern author/resource authority and public privacy projection.

Inventory at exact reviewed head 29c6de3339e10be08c652df2817b0e96033faeca (19 threads, six unresolved; CI run 35743858636 succeeded):

  1. PRRT_kwDOPbFPYs6kypQQ: legacy effective-author queue merge identity.
  2. PRRT_kwDOPbFPYs6kypQa: suppressed, ignored legacy rollback argv.
  3. PRRT_kwDOPbFPYs6kypQm: runtime-clean localized help matches retained cleaners.
  4. PRRT_kwDOPbFPYs6kypQx: retired settings bookmarks redirect to General; retired API paths still return 404, with no feature page or PWA state restored.
  5. PRRT_kwDOPbFPYs6kypQL: investigate released updater companion preflight and supported transition options. Read-only evidence first; implementation needs an explicit orchestrator/owner decision.
  6. PRRT_kwDOPbFPYs6kypQq: investigate historical published manifest assets, backups and remaining availability guards. Read-only evidence first; no Memory-only workflow/verifier restoration without a scope decision.

The two release findings remain explicit and unresolved pending that decision. No tombstone/bridge publication, old-release mutation, data traversal, sidecar logic, local service operation, merge or premature PR closure is authorized. The sole active Watch is d18dce76dd9f, owned by sesk8rfbcfr46; do not reseed.

Read-only release evidence: published v3.1.0 and gh-v3.1.1rc1 through gh-v3.1.1rc4 build and preflight a same-tag avibe_memory wheel when the companion is installed/enabled. A future companionless release blocks that automatic upgrade route. The existing documented core-wheel-only manual install is an alternative, not an automatic bridge. No release is authorized for this local-test-only PR. The deleted guard exclusively protected historical Memory Runtime manifests/assets; Git Runtime and Model Hub guards do not cover them. Nonexpired backup artifacts exist (for example 10677598133, expiring 2026-12-21), but are not a permanent availability guarantee. Both decisions were escalated to the orchestrator; neither thread is resolved by these facts.

Supported repair evidence:

  • The shared merge key now preserves legacy effective authors both in normalized queue segments and raw snapshot merges. Explicit modern authors and delegated resource authority retain precedence/separation. Regression cases exercise distinct/same authors, non-ASCII text, hydration of both separated deliveries, public privacy projection, and byte-identical persisted snapshots.
  • Restored the suppressed, ignored flag/value pair accepted by released v3.1.0; neither reaches the restart job. A read-only execution of the historical d87a39415 scheduler's argv-building AST against the new parser also passed with a job sink. A published emitter for that historical commit was not established; this is not claimed as a released-client runtime test.
  • Focused delivery/CLI/removed-feature tests: 446 passed; restart suite: 28 passed. After strengthening the second-author assertion, delivery plus restart suites: 249 passed. Changed-Python Ruff and diff whitespace passed.
  • Localized cleanup help matches the real retained cleaners. Route unit tests: 9 passed; both old bookmarks also passed real browser navigation with strict unexpected-request rejection. UI build, lint and test typechecks passed. The full workbench-general browser suite passed 74 tests.
  • Prior-head CI was independently read back: all sixteen reported checks succeeded at 29c6de3339e10be08c652df2817b0e96033faeca. This does not certify the next pushed head; its review and CI remain mandatory.

Scope safety

Independent UI lifecycle audit at b7f8c58da

Before edits, the orchestrator authorized the circuit-breaker repair for PRRT_kwDOPbFPYs6k0mEs: retained CLI service replacement and repair still stopped the UI solely to align the removed per-process Memory proof secret. Compare complete original/current functions, not just their comments. Healthy UI, remote access and SSE must survive service-only replacement/repair.

Caller inventory: cmd_start ensures both services and delegates healthy reuse or missing/stale UI recovery to runtime.start_ui; its explicit UI stop is obsolete. _start_service_after_repair repairs only the service; its entire UI realignment block is obsolete (an absent UI already remained absent). cmd_stop legitimately stops the full stack and remote access. Supervisor scope=all explicitly stops/restarts UI plus service while preserving the tunnel; scope=service never owns UI. Explicit CLI restart and upgrade schedule the full-stack scope to activate new UI code, so those calls remain. Runtime UI startup stops only a matching unhealthy process, never a healthy/unrelated PID. The private live-UI helper has only the two obsolete coupling consumers and can be removed with their probes. Preserve readiness waits, repair error mapping, status PID reporting and missing/stale UI startup behavior. Validate with real idempotent UI startup code and stubbed process/network edges in test-owned homes. Release-policy threads remain unresolved pending owner decision.

The isolated Incus supervisor's remaining UI calls are also retained: initial stack startup, recovery of a dead UI, failed-stack teardown and explicit stop. They own the isolated stack and are not secret synchronization.

Behavioral evidence: CLI/restart-supervisor/upgrade suites 406 passed, including a six-case real runtime.start_ui matrix (new/reused service × healthy/missing/ stale UI), service-repair success/failure with healthy or absent UI, PID/status preservation and forbidden UI/remote process operations. Existing explicit full-stack restart, service-only restart, stale/unrelated PID handling, slow startup/error mapping and upgrade activation coverage remains. The retired helper's identity-probe test now targets the retained runtime probe; no generic coverage was discarded. All process/network edges are stubs and paths are test-owned; no real UI/SSE/tunnel or service was restarted. Changed-Python Ruff and diff-check passed; no UI source changed.

Shared lifecycle audit at 970facd01 (2026-09-23)

The orchestrator authorized a bounded whole-contract repair before another push. The repeated class is generic behavior lost with co-located Memory code; green CI/import/startup evidence did not test concurrency, mutation completion or IM business readiness. Inventory: PRRT_kwDOPbFPYs6kz9WY config lock inversion; PRRT_kwDOPbFPYs6kz9Wm archive reporting deadline; PRRT_kwDOPbFPYs6kz9Wt missing initial adapter dependency injection. Historical release threads kypQL and kypQq remain pending owner decision.

Contracts to audit and prove before push:

  • All generic config save/transaction/migration entrants acquire CONFIG_LOCK before the per-file lock, with nested reentrancy and preserved concurrent fields. The removed Memory transaction supplied that ordering incidentally; restore it at the shared config boundary without any Memory lock.
  • UI archive -> internal RPC -> controller lifecycle -> durable archive awaits accepted completion without a transport reporting deadline. Connection failure may remain bounded; preserve success/domain-failure mapping and do not retry or restore the retired endpoint.
  • Initial runtime clients receive the same settings/controller dependencies as hot-added clients, after managers/router exist. Audit all platform descriptors and the auxiliary avibe contract. Exercise actual adapter authorization, not only injection call counts; no live SDK connections or feature runtime.

Compare complete original/current surrounding functions and retained callers, then use deterministic concurrency and lifecycle gates in test-owned state. Installed-service liveness remains distinct from IM authorization readiness.

Audit result: compared complete original/current _init_modules, _inject_runtime_dependencies, _register_client_runtime, config transaction, save/load/migration helpers and archive controller/RPC/UI callers. The generic losses were CONFIG_LOCK acquisition, the final startup injection loop and the archive client's no-reporting-deadline contract. Migration locking and hot registration were unchanged; archive's removed post-commit observer was Memory-only, so it stays removed. Removed the now-trivial archive observer wrapper/unused loop variable, preserving the blocking completion helper. Only two production entrants take the private config file lock: the shared boundary and migration persistence; both now take CONFIG_LOCK first.

Evidence: 354 passed across ten focused config/controller/platform/internal RPC/UI/archive suites; 177 passed across config-read, migration persistence, API-save and file-lock suites. New barrier/event-driven regression proves transaction/load-persistence overlap preserves both updates and nested load/save transactions re-enter safely. Archive test traverses actual UI HTTP, internal client with asserted transport timeouts, internal ASGI route, real Controller archive method and test SQLite; a gated lifecycle proves pending acceptance does not become an early response, then checks commit or 404 mapping. Adapter tests execute the real _init_modules and shared injection, all five actual adapter setters at startup/hot registration, and Discord allowed/denied/ unknown channel authorization against real test settings. SDK construction and network startup are excluded; this is not a live IM end-to-end claim. The avibe auxiliary adapter remains outside runtime injection; forbidden feature imports and absent Memory runtime are asserted. No production state/service touched.

PWA compatibility boundary completion (review 5280345842)

Before edits, the orchestrator verified PRRT_kwDOPbFPYs6kzQa_ at 074e02466d8481acae6538d3600dfe123b237d71 and authorized the circuit-breaker repair: browser aliases reached the router, but persisted PWA paths were rejected before routing. Replace the independently maintained legacy PWA allowlist with normalization derived from LEGACY_SETTINGS_REDIRECTS, accepting only safe restorable destinations. Audit storage read, normalization, launch selection, rendering and canonical persistence together. Preserve explicit deep links, query/hash stripping, dynamic routes and no mid-session bounce; reject unknown or external destinations. Cover declaration-driven invariants and actual app PWA launch in test-owned browser storage. No feature page/API/runtime is restored. The two release-policy threads remain unresolved pending owner decision; no exception to the review/CI/close-out gates has been granted.

Implementation reuses LEGACY_SETTINGS_REDIRECTS directly and removes the old duplicated legacy allowlist. The destination still passes origin and canonical restorable-path checks; query/hash are omitted. No App lifecycle change was needed. Focused PWA/navigation/router units: 74 passed. Actual App browser tests with test-owned localStorage and simulated iOS standalone detection: 7 passed, covering both retired entries, canonical persistence, back to root without bounce, explicit Shortcuts deep link, unknown/external rejection and the two ordinary bookmarks. This is Chromium app-flow evidence, not physical iOS-device verification. Prior-head CI was read back fully green; new-head gates remain required. Node 24.12 UI build, lint-baseline, test typechecks and diff whitespace checks passed. No Python files changed. The latest fetched master remains 4b964fef2.

Owner close-out decision (2026-09-22 22:41 Asia/Shanghai)

Continue the repair, push, review-response, and exact-head review/CI loop to completion. Do not merge or close prematurely. After the gates are satisfied, hand the evidence to orchestrator sesk8rfbcfr46, who will close PR #2120 without merging and retire the existing watch. Preserve the remove-memory branch, every commit, and this worktree for the owner's local testing. No release, deployment, installed-service restart, or branch/worktree deletion.

Do not use git clean, remove untracked user artifacts, touch $HOME, restart the local vibe service, or modify files outside this repository. The checkout contains pre-existing uncommitted files; preserve unrelated changes.

Historical release compatibility (2026-09-23)

PR #2120 was closed without merging. Its review, CI and repository branch remain available for local testing. The owner requested both release compatibility concerns on 2026-09-23 and then chose the inert old-updater wheel for item 1. Item 2 retains only the scheduled protection of already-published historical assets. Neither choice authorizes publishing, modifying an existing release, reopening/merging the closed PR, or restarting a local service.

  1. Old-client upgrade bridge (owner-approved for implementation). Released updaters with Memory enabled or the companion installed preflight a same-version avibe_memory-<version>-py3-none-any.whl at the target GitHub Release URL before changing core. Since already installed clients cannot learn a new upgrade rule until an upgrade succeeds, a core-only future release breaks their automatic path. For each release that old clients might target, stage a metadata-only avibe-memory compatibility wheel at that exact URL, with no importable modules, entry points, runtime manifest, scripts, post-install hooks or Avibe Memory dependencies. Publish and make it downloadable before the release becomes visible as an upgrade target. New avibe-os installations neither depend on nor install it; the legacy explicit companion request replaces the old installed package while the new core ignores obsolete config and never touches user Memory data. A single prior transition release is insufficient because old clients may skip it and upgrade directly to any later advertised version. Keep shipping the inert wheel until the project explicitly ends support for old automatic updaters. Verify the exact released old-updater preflight and both uv/pip upgrade paths against a disposable GitHub Release/installer fixture; assert the resulting environment has no importable Memory feature and a sentinel under each old data path remains byte-identical. Separately prove ordinary fresh installs include only core and the current wheel/sdist stay feature-free. Scope this claim to the updater versions that actually contain companion preflight: v3.1.0 does; the checked v3.0.14 updater does not require an avibe-memory wheel. Test any additional supported historical updater tags against their tagged code rather than treating every earlier version alike.

  2. Historical asset availability (owner-requested release-only compatibility). Keep the published old Memory Runtime URLs valid independently of the removed product. Retain a scheduled, release-only guard for historical self-pinned manifests, ideally using the original verifier from an immutable repository commit rather than packaging Memory verifier code with the new application. Restrict its scope to already published legacy manifests; verify tag, manifest hash, archive hashes and backup bytes before restoring only missing assets. Never overwrite an existing asset or rebuild/activate Memory runtime. Preserve the existing backup artifact identity so nonexpired historical backups remain usable. Exercise manifest discovery, hash rejection, backup and missing-asset recovery with mocked GitHub and test-owned files. Merely retaining a workflow in this unmerged branch does not run it; the current default-branch guard continues until a future change lands on that branch. This branch carries the release-only guard for owner review; it does not run on this closed PR.

The approved exceptions contain only release compatibility artifacts.

Retained required CI context (2026-09-23)

Branch protection still requires the historical check context memory-insight-contract, although the former Memory-specific job was removed. The owner authorized retaining this context as a real compatibility-contract job until repository administrators rename the requirement. It runs hermetic config removal, administrator routing, and native message classification tests; it does not restore Memory runtime, package, API, data, or processing behavior. Implementation evidence (owner-approved item 1): scripts/build_retired_companion.py creates a deterministic avibe_memory-<version>-py3-none-any.whl containing only valid METADATA, WHEEL and RECORD. It has no modules, dependencies, entry points, scripts, manifest or importable package. publish.yml and release_ai.yml checkout and run the builder before asset upload; immutable existing bytes are verified and bridge upload is included in the package matrix before publication. tests/test_retired_companion.py and the release upload matrix tests cover both official and gh-v tags, metadata/RECORD hashes, overwrite refusal and runtime then package ordering. 101 tests passed in the focused archive/workflow run; Ruff and diff-check passed.

The disposable Debian probe tests/e2e/test_retired_updater_bridge.py passed with the actual released v3.1.0 core/companion wheels and locally built target core/bridge assets: four cases (pip/uv × installed/enabled) used the tagged old planner and real pip/uv download/install path against a test HTTPS GitHub fixture. Each matching target bridge URL was fetched, the resulting environment had no importable Memory modules, and both ~/.avibe/memory and ~/.vibe_remote/memory sentinels retained bytes, inode and symlink identity. This is isolated artifact evidence only; no release was published or modified.

The historical guard was restored from immutable commit 4b964fef223862c2cf7d5cec852ac0d3aa64e323 as release tooling plus its scheduled workflow. It validates exact published tag/manifest/provenance/platform/archive hashes and binary contents, retrieves only missing assets from a manifest-hash backup, and refuses overwrite when names already exist. Resolver output visibly lists exclusions: old core wheels with missing or malformed manifests are excluded, while a new avibe-memory wheel is ignored only after its exact three-file dist-info shape, version and dependency/extra emptiness are verified. No Runtime manifest fallback is allowed for malformed historical wheels. Guard tests cover foreign tags, content tampering, pinned URLs, backup output identity and repeat fetch byte preservation. A hermetic test also executes the actual resolve_manifests workflow shell with mocked GitHub release responses: valid historical manifests enter the guard matrix, inert preview wheels and malformed legacy wheels have explicit, distinct exclusions, and fixture assets stay byte-identical. A failed fetch preserves an existing verified local backup. The focused release/guard suite passed 103 tests. This guard remains separate from application packaging, startup, dependencies and Memory data.

No Memory configuration, CLI, UI, sidecar, capture/recall, package import or data migration returns. Keep the historical guard release-only; it is not invoked by application startup or ordinary package installation. These branch tests do not run the scheduled workflow against live GitHub Releases; the branch and closed PR do not deploy either compatibility measure.

GitHub-only preview upgrades (owner decision, 2026-09-23)

The supported pre-release-to-pre-release route is a manual uv tool install --force of the target gh-v* release's core wheel, without the old Memory extra or --with avibe-memory. Old pre-release vibe upgrade implementations that resolve a new Memory version from a package index are not this route; no index-hosted compatibility package is planned for them. User instructions live in docs/UPGRADING.md.

A disposable macOS probe installed the published gh-v3.0.15rc11 core and Memory wheels in test-owned uv tool, cache, config and home directories, then force-installed a 3.2.0rc1 core wheel built from this branch. The new tool had no avibe-memory distribution or importable module; core imports and vibe --help succeeded, and two legacy Memory data sentinels were unchanged. This is package replacement evidence, not a test of a published new Release or of stopping and restarting a live service. The one-off probe has not been committed as a repeatable test.

Scheduled remote caller authorization invariant (2026-09-23)

Review 5291701548 exposed a shared boundary regression: synthetic scheduled and watch turns use a non-user author such as scheduled, while their durable definition carries the trusted resource_user_context snapshot that authorized creation. Caller-context reconstruction must therefore mark such a turn as remote only when that persisted snapshot is structurally valid, and carry it through the Agent environment for the downstream resource ACL to re-check. A missing or malformed snapshot is an anonymous remote caller (never local Owner), and revoked or mismatched authority remains denied by the existing resource authorization service. Interactive Workbench author matching, ordinary IM/local turns, and session_stable semantics remain unchanged. The audit covers both ScheduledTaskService and Watch producers and the shared SessionHandler consumer; no removed Memory owner metadata is reintroduced.