Releases are cut by hand and every non-prerelease moves next and stable together, so a bad release reaches every host on stable at once.
Wanted
A manual release workflow (version or bump level) moves [Unreleased] in CHANGELOG.md to the new version, runs task news, commits, tags and pushes atomically. It refuses an empty [Unreleased]. The subject-dump generator (tools/release.sh, release:preview, release:prepare) is deleted; notes are written with each change.
A release publishes v<version> and next only.
The canary is the production host on next (decided). The deploy run points it at next and installs through its dashboard API.
A scheduled evaluator compares the canary's 24h after the release was issued with its 24h before (decided: one production host, so the control is the canary's own earlier window):
hard fail at any check: the installed version is not the release (the watchdog rolled it back), readiness is degraded, or a new action alert is open;
relative fail at 24h: error-level log lines over 2 × baseline + 10; task park+dead rate over baseline + 10 points, needing at least 5 finished tasks in the window, otherwise that signal is skipped.
Promote (automatic, decided): stable is retagged to the release's signed manifest, refused if that would move stable to an older issued_at. No re-signing: the signature covers the statement.
Roll back: stable stays put; the canary is pinned to the stable version and reinstalled. If the release migrated the database, the binaries cannot go back without restoring the pre-update snapshot (losing writes since), so the evaluator holds the canary and reports instead.
RELEASING.md describes the new flow.
Done when
Dispatching the workflow with an [Unreleased] section produces the tag, the signed artifact on v<version> and next, and the release-notes page with no local steps.
The evaluator returns promote or roll back from those signals over that window; stable moves only on promote.
task check passes on each commit; a local registry:2 rehearsal of the pipeline is recorded here.
Releases are cut by hand and every non-prerelease moves `next` and `stable` together, so a bad release reaches every host on `stable` at once.
## Wanted
- A manual `release` workflow (version or bump level) moves `[Unreleased]` in CHANGELOG.md to the new version, runs `task news`, commits, tags and pushes atomically. It refuses an empty `[Unreleased]`. The subject-dump generator (`tools/release.sh`, `release:preview`, `release:prepare`) is deleted; notes are written with each change.
- A release publishes `v<version>` and `next` only.
- The canary is the production host on `next` (decided). The deploy run points it at `next` and installs through its dashboard API.
- A scheduled evaluator compares the canary's 24h after the release was issued with its 24h before (decided: one production host, so the control is the canary's own earlier window):
- hard fail at any check: the installed version is not the release (the watchdog rolled it back), readiness is degraded, or a new `action` alert is open;
- relative fail at 24h: error-level log lines over `2 × baseline + 10`; task park+dead rate over `baseline + 10` points, needing at least 5 finished tasks in the window, otherwise that signal is skipped.
- Promote (automatic, decided): `stable` is retagged to the release's signed manifest, refused if that would move `stable` to an older `issued_at`. No re-signing: the signature covers the statement.
- Roll back: `stable` stays put; the canary is pinned to the `stable` version and reinstalled. If the release migrated the database, the binaries cannot go back without restoring the pre-update snapshot (losing writes since), so the evaluator holds the canary and reports instead.
- RELEASING.md describes the new flow.
## Done when
- Dispatching the workflow with an `[Unreleased]` section produces the tag, the signed artifact on `v<version>` and `next`, and the release-notes page with no local steps.
- The evaluator returns promote or roll back from those signals over that window; `stable` moves only on promote.
- `task check` passes on each commit; a local `registry:2` rehearsal of the pipeline is recorded here.
Built in a7f6e517a (release workflow, task release, generator deleted) and a4a7c96da (publish to next only, archied release canary start|check, archied release promote, canary.yml, RELEASING.md).
Rehearsal
Local registry:2 on loopback, a throwaway signing key, a stub canary dashboard serving the same endpoints (/api/chat/update, /health/detailed, /api/alerts, /api/tasks, /api/logs, update-settings replace, install), and a rehearsal-only archied built with go build -overlay to backdate issued_at (nothing in the tree).
task release RELEASE=minor in a clone: refused with an empty [Unreleased]; with an entry it wrote ## [1.57.0] - 2026-10-31 under a new [Unreleased], ran newsgen, committed and tagged v1.57.0.
1.0.0 issued 25h ago, first release: canary start set the channel to next (source and key kept) and installed it; canary check → promote; stable resolves to 1.0.0.
1.1.0 issued 2h ago, readiness degraded, no migration: → roll back; canary pinned to 1.0.0 and reinstalled; stable unchanged; exit 1. The next check: "held on exact-pin 1.0.0", no action.
1.2.0 issued 1h ago, migrations changed, degraded: → roll back; canary held on 1.2.0; stable unchanged.
1.3.0-rc.1 on next: nothing to judge. release promote refuses it (prerelease) and refuses 1.0.0 once stable is 1.1.0 (issued before).
task check passed on both commits. TestPromote and TestJudgeCanary cover the promotion refusals and the verdict table; reverting the issued-before or prerelease check fails TestPromote.
Not yet exercised
The workflows on the live runner, and the dashboard API on the real canary. They need the CANARY_URL variable and the CANARY_TOKEN secret (an instance-admin API token on the canary), and the runner must reach that URL.
Built in a7f6e517a (release workflow, `task release`, generator deleted) and a4a7c96da (publish to `next` only, `archied release canary start|check`, `archied release promote`, `canary.yml`, RELEASING.md).
## Rehearsal
Local `registry:2` on loopback, a throwaway signing key, a stub canary dashboard serving the same endpoints (`/api/chat/update`, `/health/detailed`, `/api/alerts`, `/api/tasks`, `/api/logs`, update-settings replace, install), and a rehearsal-only archied built with `go build -overlay` to backdate `issued_at` (nothing in the tree).
- `task release RELEASE=minor` in a clone: refused with an empty `[Unreleased]`; with an entry it wrote `## [1.57.0] - 2026-10-31` under a new `[Unreleased]`, ran newsgen, committed and tagged `v1.57.0`.
- 1.0.0 issued 25h ago, first release: `canary start` set the channel to next (source and key kept) and installed it; `canary check` → promote; `stable` resolves to 1.0.0.
- 1.1.0 issued 2h ago, readiness degraded, no migration: → roll back; canary pinned to 1.0.0 and reinstalled; `stable` unchanged; exit 1. The next check: "held on exact-pin 1.0.0", no action.
- 1.2.0 issued 1h ago, migrations changed, degraded: → roll back; canary held on 1.2.0; `stable` unchanged.
- 1.3.0-rc.1 on next: nothing to judge. `release promote` refuses it (prerelease) and refuses 1.0.0 once `stable` is 1.1.0 (issued before).
- `task check` passed on both commits. `TestPromote` and `TestJudgeCanary` cover the promotion refusals and the verdict table; reverting the issued-before or prerelease check fails `TestPromote`.
## Not yet exercised
The workflows on the live runner, and the dashboard API on the real canary. They need the `CANARY_URL` variable and the `CANARY_TOKEN` secret (an instance-admin API token on the canary), and the runner must reach that URL.
2181b5fd0: release images are published as archied:v<version> and archie-agent:v<version>; promotion (canary check -latest, release promote -latest) moves each image's :latest to its :v<version> before stable, refusing an agent image whose digest is not the one the statement pins.
Rehearsed on a local registry:2: promoting 1.0.0 moved both :latest tags to the :v1.0.0 digest; promoting 1.1.0 whose archie-agent:v1.1.0 differed from the pinned digest was refused and stable stayed on 1.0.0. task check passed.
2181b5fd0: release images are published as `archied:v<version>` and `archie-agent:v<version>`; promotion (`canary check -latest`, `release promote -latest`) moves each image's `:latest` to its `:v<version>` before `stable`, refusing an agent image whose digest is not the one the statement pins.
Rehearsed on a local `registry:2`: promoting 1.0.0 moved both `:latest` tags to the `:v1.0.0` digest; promoting 1.1.0 whose `archie-agent:v1.1.0` differed from the pinned digest was refused and `stable` stayed on 1.0.0. `task check` passed.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Releases are cut by hand and every non-prerelease moves
nextandstabletogether, so a bad release reaches every host onstableat once.Wanted
releaseworkflow (version or bump level) moves[Unreleased]in CHANGELOG.md to the new version, runstask news, commits, tags and pushes atomically. It refuses an empty[Unreleased]. The subject-dump generator (tools/release.sh,release:preview,release:prepare) is deleted; notes are written with each change.v<version>andnextonly.next(decided). The deploy run points it atnextand installs through its dashboard API.actionalert is open;2 × baseline + 10; task park+dead rate overbaseline + 10points, needing at least 5 finished tasks in the window, otherwise that signal is skipped.stableis retagged to the release's signed manifest, refused if that would movestableto an olderissued_at. No re-signing: the signature covers the statement.stablestays put; the canary is pinned to thestableversion and reinstalled. If the release migrated the database, the binaries cannot go back without restoring the pre-update snapshot (losing writes since), so the evaluator holds the canary and reports instead.Done when
[Unreleased]section produces the tag, the signed artifact onv<version>andnext, and the release-notes page with no local steps.stablemoves only on promote.task checkpasses on each commit; a localregistry:2rehearsal of the pipeline is recorded here.Built in
a7f6e517a(release workflow,task release, generator deleted) anda4a7c96da(publish tonextonly,archied release canary start|check,archied release promote,canary.yml, RELEASING.md).Rehearsal
Local
registry:2on loopback, a throwaway signing key, a stub canary dashboard serving the same endpoints (/api/chat/update,/health/detailed,/api/alerts,/api/tasks,/api/logs, update-settings replace, install), and a rehearsal-only archied built withgo build -overlayto backdateissued_at(nothing in the tree).task release RELEASE=minorin a clone: refused with an empty[Unreleased]; with an entry it wrote## [1.57.0] - 2026-10-31under a new[Unreleased], ran newsgen, committed and taggedv1.57.0.canary startset the channel to next (source and key kept) and installed it;canary check→ promote;stableresolves to 1.0.0.stableunchanged; exit 1. The next check: "held on exact-pin 1.0.0", no action.stableunchanged.release promoterefuses it (prerelease) and refuses 1.0.0 oncestableis 1.1.0 (issued before).task checkpassed on both commits.TestPromoteandTestJudgeCanarycover the promotion refusals and the verdict table; reverting the issued-before or prerelease check failsTestPromote.Not yet exercised
The workflows on the live runner, and the dashboard API on the real canary. They need the
CANARY_URLvariable and theCANARY_TOKENsecret (an instance-admin API token on the canary), and the runner must reach that URL.2181b5fd0: release images are published asarchied:v<version>andarchie-agent:v<version>; promotion (canary check -latest,release promote -latest) moves each image's:latestto its:v<version>beforestable, refusing an agent image whose digest is not the one the statement pins.Rehearsed on a local
registry:2: promoting 1.0.0 moved both:latesttags to the:v1.0.0digest; promoting 1.1.0 whosearchie-agent:v1.1.0differed from the pinned digest was refused andstablestayed on 1.0.0.task checkpassed.