Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
31.0 KB
Raw Blame

Upgrades and Auto-Update Notifications

Goal

Users get notified of new GBrain features conversationally, and the agent walks them through upgrading with post-upgrade migrations that make the new version actually work.

What the User Gets

Without this: GBrain ships updates but nobody knows. The user stays on an old version with stale skills and missing features. Or worse, someone runs gbrain upgrade but skips the post-upgrade steps, leaving new code with old agent behavior.

With this: the agent checks for updates daily, sells the upgrade with punchy benefit-focused bullets, waits for explicit permission, then runs the full upgrade flow including re-reading skills, running migrations, and syncing schema. The user gets new capabilities automatically.

Self-upgrade modes

gbrain stays current the way gstack does: it rides invocation frequency. A throttled, cache-read-only check runs at the start of every gbrain invocation (CLI and MCP) and emits an UPGRADE_AVAILABLE <old> <new> marker on stderr, once per 30 minutes per version under one GBRAIN_HOME (a newer version shows at once). It is one line: when stderr is an interactive TTY a human sees only the plain gbrain X -> Y available sentence instead of the machine token; set GBRAIN_FORCE_UPGRADE_MARKER=1 if an agent harness parses the token but runs under a PTY. <old> is always the RUNNING binary's version, so a stale or foreign-written cache never nags about an upgrade this binary already has. No host cron required — every agent kind (Claude Code, Codex, OpenClaw, Hermes, the gbrain serve host behind a Perplexity thin client) converges to current by construction. The behavior is governed by one file-plane config key, self_upgrade.mode:

Mode Behavior Who it's for
notify (default) Emit the marker + a 4-option prompt; never apply without confirmation. Interactive installs / anyone with a human in the loop.
auto (opt-in) Apply silently, but ONLY during quiet hours, ONLY when the brain is idle, doctor-gated, never re-trying a known-bad version, and never installing a release the host's Bun cannot start (Bun floor). Headless / always-on installs (autopilot daemon, the gbrain serve host).
off Never check. Air-gapped / pinned installs.

Enable hands-off upgrades on an always-on install with one line:

gbrain config set self_upgrade.mode auto

auto is deliberately NOT a default anywhere — it's an explicit autonomy grant, because applying code from GitHub unattended is, by design, remote code execution. The trust model is TLS + GitHub (same as gbrain upgrade); signature verification is a tracked follow-up. Apply manually any time with gbrain self-upgrade.

The auto quiet-hours window is configured via the self_upgrade.quiet_hours config key (gbrain config set self_upgrade.quiet_hours '{"start":23,"end":8,"tz":"US/Pacific"}'). The quiet-hours pattern itself — gating any notification or background action on the user's local sleep window — is owned by quiet-hours.md; this doc only covers the self-upgrade hook into it.

Bun floor

Say to your agent: "gbrain says the upgrade needs a newer Bun. Upgrade Bun and finish the gbrain upgrade." The agent runs bun upgrade, then gbrain upgrade, then gbrain doctor to confirm self_upgrade_health is ok.

A source (bun-link) or package install runs the new release on the host's Bun, and a release refuses to start on a Bun below its engines.bun floor. So before swapping, gbrain upgrade, gbrain self-upgrade and the autopilot channel read the target's floor and compare it with the lower of the bun on PATH and the Bun the running process uses:

  • bun-link clone: git fetch, then package.json at the fetched upstream commit; the swap fast-forwards to exactly that commit (git merge --ff-only <sha>), so a floor raised upstream after the check is not installed unchecked.
  • global package (bun): package.json at the ref the install's github: spec names (the default branch when unpinned) on raw.githubusercontent.com.
  • compiled binary: exempt; it carries its own Bun.

When the floor is above the host's Bun, a manual upgrade refuses and changes nothing (exit 78):

gbrain <target> requires Bun >=<floor>; bun on PATH (<path>) is <found>. Fix: bun upgrade, then gbrain upgrade. Docs: docs/guides/upgrades-auto-update.md#bun-floor

Run bun upgrade, then gbrain upgrade again. When the floor cannot be read (offline, no upstream branch, or a floor not of the form >=X.Y.Z), the refusal names the read that failed; after checking the target yourself, gbrain upgrade --no-bun-floor-check (or gbrain self-upgrade --no-bun-floor-check) upgrades without the check. The override is for manual upgrades only; autopilot never passes it.

The autopilot channel holds instead of refusing: the tick records an unsupported_runtime audit row, gbrain doctor reports Auto-upgrade to <target> held: … under self_upgrade_health with the same fix line, and the target is not marked known-bad, so the first quiet-hours tick after bun upgrade applies it (an unreadable floor is retried on the next tick). A swap that gbrain upgrade itself refuses for the floor (for example, upstream raised it after the channel's check) is a hold too.

After the swap, gbrain upgrade runs gbrain --help (a command that passes the startup Bun check, unlike --version). If the new release does not start, the upgrade exits non-zero and prints recovery: bun upgrade, then gbrain post-upgrade; or return to the previous release with git -C <clone> checkout <previous sha> && bun install (bun-link) or git ls-remote --exit-code --tags https://github.com/garrytan/gbrain refs/tags/v<previous> && bun remove -g gbrain && bun add -g --backend=copyfile github:garrytan/gbrain#v<previous> (package; every global reinstall copies files, because OpenClaw refuses a hardlinked plugin manifest; an in-place tag swap fails or silently corrupts the global lock on current Bun, #5034). Nothing is rolled back automatically, and the autopilot channel records that version as failed.

Tag- or commit-pinned package installs (github:garrytan/gbrain#v<x>, #<sha>) stay pinned through bun update, so gbrain upgrade replaces them itself when a newer release tag is published (bun remove -g gbrain, then bun add -g the tag, restoring the previous release if the add fails), and otherwise prints a reinstall. Every printed reinstall checks that the release tag exists before bun remove -g gbrain runs, and ends with the line that restores the previous version if the add fails. The target is the newest published release tag: master's VERSION is stamped before the release checks pass and the tag is created, so a version can be announced before its tag exists. When the requested version's tag is missing, upgrade says v<x> is not published yet, removes nothing, and offers the newest published release if it is newer than yours (#6662). Windows PowerShell 5.1 has no &&, so on Windows the steps print one per line; run them in order and stop if the tag check fails.

Which version counts as available

gbrain has two release channels, and every update surface (the startup marker, gbrain check-update, gbrain self-upgrade, gbrain upgrade, doctor's self_upgrade_health and the auto mode) follows the one in effect:

Channel What it offers Read from
stable (default) Releases promoted, by the daily stable check or a maintainer, after the soak passed on that release's exact git tree. The evidence record stable.json on the stable-record branch, checked against the stable tag.
alpha Every published release. VERSION at the latest-stable tag, which release.yml moves on every publish.

latest is accepted everywhere as an alias for alpha. The channel comes from the first of these that is set: --channel on the command, the GBRAIN_RELEASE_CHANNEL environment variable, self_upgrade.channel in the brain home's config, then the default. An invalid value at the first one set is an error (invalid_params), never a fall-through to the next. Channel state lives under GBRAIN_HOME, so a repo-less brain and a repo brain behave the same. When a channel is set, gbrain doctor prints channel=<channel> (<source>). New installs start on stable: the documented install (github:garrytan/gbrain#stable, or git clone --branch stable) installs the stable release, and a fresh gbrain init (new content database, no prior config file) writes self_upgrade.channel: stable, so it never sees the notice below.

gbrain config set self_upgrade.channel alpha    # follow every release on this brain home
gbrain config set self_upgrade.channel stable   # back to the default
gbrain upgrade --channel alpha                  # one run on alpha, config unchanged
gbrain upgrade --dry-run --json                 # what an upgrade would install, installing nothing

An install with no channel set prints one notice, once per GBRAIN_HOME, on stderr (never under --json, serve or mcp): [gbrain notice release_channel] This install now follows the stable channel .... It is not an UPGRADE_AVAILABLE marker. Setting any channel silences it.

The evidence record. Promote stable (RELEASING.md) moves the stable tag and commits stable.json to the stable-record branch in one atomic push. The record names the version, commit, tree, the soak run that passed on that tree and a known_bad list of withdrawn releases. The update check reads only that file (no API call, no git), so a missing record or a record without the evidence its generation requires is never trusted. Generation 1 records carry the soak run; from the matrix cutover on, matrix_run is required too. gbrain upgrade also checks that stable and v<version> peel to the record's commit, then installs exactly that release: a bun install is reinstalled pinned to #v<version> (remove, then add; it restores the previous release if the add fails), a bun-link clone fast-forwards to the release commit, and the binary install downloads that release's asset.

Never downgrade. A build at or above the stable release is offered nothing, so an install that ran ahead on alpha waits until stable passes it. The one exception is a build on the record's known_bad list: it is offered the current stable release even when that is older. An install whose method cannot be detected is refused with the reinstall command.

The update cache and snooze are per channel. Each channel keeps its own answer (update-check.<channel>.json, which also records the evidence policy) and its own snooze (update-snoozed.<channel>.json). Switching channel reads that channel's answer; a cached answer from another policy, or the unpartitioned cache of older builds, is never read. A snooze holds for its whole window (24 h, then 48 h, then 7 days) however many releases ship meanwhile; a stable rollback below the snoozed version clears it.

When stable has no usable answer. The background check only records and reports these; it never offers alpha instead. An explicit gbrain upgrade or gbrain self-upgrade on stable exits non-zero with the code and installs nothing.

Code Meaning What to do
stable_missing No release has been promoted yet (no record or no stable tag). Wait for the first promotion, or gbrain config set self_upgrade.channel alpha.
stable_invalid The record is malformed, lacks a required evidence key, lists its own version as known bad, or its version's tag is not the stable commit. Wait for a maintainer to repair or re-promote stable, or switch to alpha.
stable_unreadable The record or the tags did not answer (network, or a promotion still propagating). Retry: gbrain self-upgrade --check-only --json. The last verified answer still shows in the notice, never as an install target.

When the channel came from --channel or GBRAIN_RELEASE_CHANNEL, the fix names that override instead of the config command. Builds before this release reported stable_channel_unresolved (retired).

A swap is a swap only when the version changed. When gbrain upgrade finds the same version running after the install step (bun update on a tag whose head is already installed), it prints No newer version was installed, keeps last_upgrade, writes no JUST_UPGRADED breadcrumb and keeps the pending marker. The auto channel then stays up instead of exiting for a relaunch, clears its breadcrumb, records no known-bad version, audits the attempt as skipped (nothing newer was installed) and tries again after 24 hours (self_upgrade.last_check_ts). The channel confirms a swap from a receipt gbrain upgrade writes, or from gbrain --version when there is none; a swap that exits 0 with no newer version confirmed is held the same way, never relaunched. It also checks the version the install would actually get (a bun-link clone follows master, a moving tag can pass the checked version) against the known-bad list before swapping. A version is recorded known-bad only when it was installed and then failed (the post-swap gbrain --help smoke, or a relaunch that did not come up on it), under the version actually installed.

A known-bad version that was never installed

Releases before this fix could record a version known-bad that was never installed: the swap installed nothing, reported swapped, the supervisor relaunched the same version, and the relaunch recorded the attempted version in self_upgrade.failed_versions, so auto skipped that release forever. doctor's self_upgrade_health warns about each such entry and names the preview as its fix:

gbrain self-upgrade --clear-never-installed            # preview: lists each entry and the swap that recorded it; writes nothing
gbrain self-upgrade --clear-never-installed --apply    # removes exactly the listed entries
gbrain doctor --only self_upgrade_health --json        # verify

An entry is listed only when it is proven never installed. The audit rows alone are not proof, because a release that installed, crashed before its first boot finished, and was rolled back by hand leaves the same rows. The proof is the record the no-op swap wrote itself: last_upgrade in upgrade-state.json with the same version as from and to, timestamped inside that entry's swap (between its apply row and its swapped row), plus exactly one recorded failure for the entry, a crash-on-launch row on the version that ran the swap, and no audit row that ever ran the entry's version. Each swap overwrites last_upgrade, so only the most recent no-op can be proven; an older entry stays listed as kept, and you clear it by hand as below once you have confirmed it. Anything else stays known-bad. Nothing is rewritten during an upgrade. gbrain config set self_upgrade.failed_versions '["<version>"]' puts an entry back. On a release that does not have this command, read the list with gbrain config get self_upgrade.failed_versions, confirm the entry was never installed (no v<version> tag existed when it was recorded, or the audit shows the swapped row and the relaunch on the same version), then set the list without it.

Recovery when autopilot is holding an upgrade:

gbrain doctor                 # self_upgrade_health: Auto-upgrade to <target> held: ...
bun upgrade                   # raise the host's Bun to the floor or above
gbrain upgrade                # apply now instead of waiting for quiet hours
gbrain doctor                 # self_upgrade_health ok

Restart gbrain serve and autopilot afterwards if a service manager does not restart them for you, so they run on the new Bun. Hosts still running a release from before this check do not run it, so a Bun floor raise reaches them only after they have upgraded past this release.

Restart: one build on the host

A long-running gbrain process keeps the code it started with, so an upgrade that restarts nothing leaves serve, the jobs supervisor and its worker, autopilot and every MCP server on the old build while one-shot commands and job children run the new one. gbrain upgrade therefore restarts every gbrain process this brain runs on the host, or refuses before it installs anything:

  1. Before the swap it lists the processes (ps, on Linux only processes whose GBRAIN_HOME is this brain's, with the build each one loaded from the consumer heartbeat rows on Postgres, the supervisor PID file and the worker registry; a record whose pid ps does not show as gbrain is stale and ignored) and checks that each one can be restarted: serve and autopilot through their installed services when the service's own process runs them, a process that starts from a separate Bun global copy (for example one the serve service points at, pinned to a commit) has that copy replaced first (bun remove -g then bun add -g the release tag under its BUN_INSTALL), and one that starts from another kind of install refuses with its commands; the supervisor with gbrain jobs supervisor stop then start --detach with the queue, concurrency and PID file it was launched with (its worker, under tini or not, restarts with it). An autopilot that no installed service runs is handed to its launcher when one is known to start it again: this brain's own gbrain cron line (what gbrain autopilot --install writes on Linux without a user systemd bus), or a cron entry or wrapper of your own once you declare it with gbrain config set self_upgrade.autopilot_launcher external (only when it starts autopilot again within 5 minutes of it exiting). The restart stops that autopilot (only the daemon this brain's autopilot lock names, and only while ps still shows the same process; SIGTERM drains it) and waits for the launcher. A serve or worker started by hand, an autopilot started any other way, a running gbrain sync, or an MCP server an agent harness started refuses the upgrade with restart_refused (exit 3), naming the exact step for each process; nothing is installed. For MCP servers the fix asks the user, then reruns with --stop-mcp: the upgrade stops them (each pid only while ps still shows the same MCP server) and each harness starts the new build when gbrain is reconnected (Claude Code: /mcp, then reconnect; other harnesses: a new session).
  2. After the swap and migrations the new build runs gbrain upgrade --restart-only: serve first, then autopilot, then the supervisor, then an autopilot handed to its launcher, then the MCP servers --stop-mcp named. It then polls for up to three minutes (six when an autopilot was handed to its launcher) until no old pid is alive, every process reports the installed build (a process that records no build must have started after the restart from an install that holds it; a compiled binary's build is inferred from its start time and listed as such), each restarted role and the supervisor's worker came back, and, when serve had a warm CLI socket, the socket accepts this build (a handshake-only probe). Anything short of that is restart_unverified with the processes still on another build.
gbrain upgrade --restart-only --dry-run --json   # read-only: every gbrain process, its build and how it restarts
gbrain upgrade --restart-only                    # restart onto the installed build and verify (doctor's runtime_build_skew fix)
gbrain upgrade --no-restart                      # install only; doctor fails runtime_build_skew until the restart runs

gbrain doctor runs the same pre-swap plan read-only: self_upgrade_health warns while the next gbrain upgrade would refuse, naming each process and its step (for an autopilot started outside gbrain's service, gbrain autopilot --install, after retiring the launcher that starts it now, or the declared hand-off above), and names a hand-off or a second Bun global install the upgrade replaces itself. An MCP server a harness started does not make it warn (gbrain upgrade --stop-mcp stops it), and a held or failed auto-upgrade is reported first.

On an always-on host, start with gbrain autopilot pause --reason "upgrade" and wait until gbrain jobs list --status active --json returns no jobs, so no job runs half on each build; resume with gbrain autopilot resume once the checks below pass (the order).

The autopilot --swap-only channel exits for its own supervisor to relaunch. Once the relaunched autopilot runs the new build, it checks the serve that the installed serve service runs (com.gbrain.serve under launchd, gbrain-serve.service under systemd). If that serve is on an older build, the autopilot restarts it through the service and verifies that it came back on the new build. It tries once per installed build (serve-restart-attempt.json remembers the attempt across autopilot relaunches), under the swap's own auto-mode, quiet-hours and idle gates, and records it in the self-upgrade audit as restart_serve (applied, or skipped with manual or unverified). It leaves alone a serve that the service does not run (one started by hand or from another install), a serve service installed for another brain on the same OS user (its wrapper is under another GBRAIN_HOME), the jobs supervisor and MCP servers. For those, doctor's runtime_build_skew fails while two builds run, and gbrain upgrade --restart-only is its fix.

Restart and verify: a deploy ends with data movement

An upgrade is not finished when the new version starts: a managed brain whose processes stay alive while no page commits reads healthy on /health, on the pid and on sync_running: true. Both upgrade and post-upgrade end by printing the supervisor's step, bare, so it inherits the default window:

# on the brain host, after the swap (gbrain upgrade has restarted every process; with --no-restart run:)
gbrain upgrade --restart-only            # every resident gbrain process onto the installed build, verified
gbrain sources writer movement           # waits one window (max(300 s, preparation budget + 60 s)); exit 0 only when pending work moved or nothing is pending
gbrain doctor --json                     # managed_sync_not_moving, two_consumers_on_host, consumers_without_heartbeat, host_identity_mismatch all ok

writer movement [<source>] [--wait <dur>] [--warn-only] [--json] snapshots each active managed source (newest committed managed_sync_* receipt, the head's step, the cursor and its holds), waits the window and judges: moved or current exit 0; held exits 0 with the holds and their route (gbrain repair fences --source <id> for fence holds, gbrain sources writer status --source <id> --json then gbrain sources retry-held <id> for preparation_stalled); within_allowance (a multi-wave group still preparing with its step advancing) exits 0 with retry_after_ms; not_moving exits 1 with managed_sync_not_moving (reason: movement_check) and the writer-status command as its fix; unknown carries the read failure's own code. --warn-only prints the same envelopes and exits 0 for pipelines that cannot fail; --wait <dur> takes sync's duration syntax. Restart every resident process, not only serve: an older jobs worker or autopilot left running writes no persistence_consumers heartbeat row and never defers to the resident consumer (doctor consumers_without_heartbeat), and a worker started under a different HOME/GBRAIN_HOME reads the binding as another host's (doctor host_identity_mismatch). Runbook for a check that fails: the catch-up is parked.

gbrain config set self_upgrade.<key> writes ~/.gbrain/config.json (the file plane every self-upgrade reader uses) and refuses a value the readers would ignore: an unknown key or mode, a quiet-hours window with an hour outside 0-23, equal bounds, or an unknown timezone. Versions are stored in the four-segment form (0.57.1 becomes 0.57.1.0). The keys are machine-local, so they work on a thin client too. Older versions wrote these keys to the database, where nothing reads them; the next config set or config unset of the key removes that row.

Implementation

The Check (cron-initiated)

check_for_update():
  result = run("gbrain check-update --json")

  if not result.update_available:
    exit_silently()  // do NOT message the user

  // Sell the upgrade — lead with what they can DO, not what changed
  message = compose_upgrade_message(
    current: result.current_version,
    latest: result.latest_version,
    changelog: result.changelog
  )
  send_to_user(message, respect_quiet_hours=true)

The Upgrade Message

Sell the upgrade. The user should feel "hell yeah, I want that." Lead with what they can DO now that they couldn't before, not what files changed.

> **GBrain vX.Y.Z is available** (you're on vX.Y.W)
>
> What's new:
> - Your brain never falls behind. Live sync keeps the vector DB current
>   automatically, so edits show up in search within minutes
> - New verification runbook catches silent failures before they bite you
> - New installs set up live sync automatically. No more manual setup step
>
> Want me to upgrade? I'll update everything and refresh my playbook.
>
> (Reply **yes** to upgrade, **not now** to skip, **weekly** to check
> less often, or **stop** to turn off update checks)

Handling Responses

User says Action
yes / y / sure / ok / do it / upgrade Run the full upgrade flow (below)
not now / later / skip / snooze Acknowledge, check again next cycle
weekly Store preference, switch cron to weekly
daily Store preference, switch cron back to daily
stop / unsubscribe / no more Disable the cron. Tell user how to resume

In notify mode (the default), never auto-upgrade — always wait for explicit confirmation. The auto mode (opt-in, see "Self-upgrade modes" above) is the only path that applies without a prompt, and only under its conservative gates (quiet hours + idle + doctor-gate). This per-cron-prompt flow is the notify experience.

The Full Upgrade Flow (after user says yes)

full_upgrade():
  // Step 1: Update the binary/package
  run("gbrain upgrade")

  // Step 2: Re-read all updated skills
  for skill in find("skills/*/SKILL.md"):
    read_and_internalize(skill)  // updated skills = better agent behavior

  // Step 3: Re-read production reference docs
  read("docs/GBRAIN_SKILLPACK.md")
  read("docs/GBRAIN_RECOMMENDED_SCHEMA.md")

  // Step 4: Check for version-specific migration directives
  for version in range(old_version, new_version):
    migration = find(f"skills/migrations/v{version}.md")
    if migration exists:
      read_and_execute(migration)  // in order, don't skip

  // Step 5: Schema sync — suggest new, respect declined
  state = read("~/.gbrain/upgrade-state.json")
  for recommendation in new_schema_recommendations:
    if recommendation not in state.declined:
      suggest_to_user(recommendation)
  update(state, new_choices)

  // Step 6: Report what changed
  summarize_to_user(actions_taken)

Migration Files

Migration files live at skills/migrations/vX.Y.Z.md. They contain agent instructions (not scripts) for post-upgrade actions that make the new version work for existing users. Example: a migration that sets up live sync and runs the verification runbook.

The agent reads migration files in version order and executes them step by step. Without migrations, the agent has new code but the user's environment hasn't changed.

Cron Registration

Name: gbrain-update-check
Default schedule: 0 9 * * * (daily 9 AM)
Weekly schedule: 0 9 * * 1 (Monday 9 AM)
Prompt: "Run gbrain check-update --json. If update_available is true,
  summarize the changelog and message me asking if I'd like to upgrade.
  If false, stay silent."

Frequency Preferences

Default: daily. Store in agent memory as gbrain_update_frequency: daily|weekly|off. Also persist in ~/.gbrain/upgrade-state.json so it survives agent context resets (the runtime's own bookkeeping lives beside it as ~/.gbrain/last-update-check and ~/.gbrain/update-snoozed).

Standalone Skillpack Users

If you loaded this SKILLPACK directly (copied or read from GitHub) without installing gbrain, you can still stay current. Both GBRAIN_SKILLPACK.md and GBRAIN_RECOMMENDED_SCHEMA.md carry a <!-- source: ... --> header pointing at their canonical copies, and GBRAIN_RECOMMENDED_SCHEMA.md also carries a version marker:

curl -s https://raw.githubusercontent.com/garrytan/gbrain/master/docs/GBRAIN_RECOMMENDED_SCHEMA.md | head -1
# Returns: <!-- schema-version: X.Y.Z -->

If the remote version is newer (or the remote SKILLPACK content differs from your local copy), fetch the full file and replace your local copy. Set up a weekly cron to check automatically.

Tricky Spots

  1. In notify mode, never auto-install. The upgrade waits for the user's explicit "yes." Even if the check detects an update and the changelog looks great, the agent messages the user and waits. The auto mode (opt-in) exists for headless/always-on installs where there's no human to prompt — it applies only during quiet hours, only when idle, doctor-gated, never retrying a known-bad version. Don't enable auto on an interactive workstation; the prompt-first notify flow is the right default there.

  2. Migration files are agent instructions, not scripts. They tell the agent what to do step by step in plain language. They are NOT bash scripts to execute blindly. The agent reads them, understands the context, and adapts to the user's specific environment (e.g., skip a step if the user already has live sync configured).

  3. check-update should run on a daily cron. Don't rely on the user remembering to check for updates. The cron runs gbrain check-update --json daily at 9 AM (respecting quiet hours). If there's nothing new, it stays completely silent. The user only hears about updates when there IS something worth upgrading to.

How to Verify

  1. Run check-update and verify detection. Execute gbrain check-update --json. Verify it returns the current version and correctly reports whether an update is available. If update_available is false, verify the version matches the latest release on GitHub.

  2. Verify migration files are readable. List skills/migrations/ and check that each file follows the naming convention vX.Y.Z.md. Open one and verify it contains step-by-step agent instructions, not raw scripts. The agent should be able to read and execute each step.

  3. Test the full upgrade flow end-to-end. If an update is available, say "yes" and watch the agent execute the full flow: upgrade, re-read skills, run migrations, sync schema, report. Verify each step completes and the agent reports what changed.


Part of the GBrain Skillpack.