¶ AutoCoder implementation-handoff production backport — 2026-09-23
¶ Scope and provenance
- Authorized request: deploy the planning-to-implementation handoff repair to production.
- Scoped backport PR: https://gitea.app.lan/AutoCoder/AutoCoder-Engine/pulls/6666, merged into main as
4eaee90621ef06b2524fe5a4753fe75331197c0c at 22:04 UTC. The full dev branch was not promoted.
- Worker-host branch CI for backport head
419b1abd1e9636335b91892902992a999a4b6106: branch tests (including isolated database), native release, and WASM release all passed. Independent static review found no concrete blocker.
- Repository production workflow run 3625 classified backend and frontend for deployment; manifest deployment was skipped. Backend identity validation, both image builds, both rollouts, documentation sync, and cronjob image update passed.
- Backend Deployment: 1/1 ready, image tag
sha-4eaee90621ef, pod image digest sha256:924b07e4ea39eada84ed28482c87bf5967f991ca0b77af84f4bb3db059f9a3a7. In-pod /healthz returned HTTP 200 and ok.
- Frontend Deployment: 1/1 ready, image tag
build-8706-sha-4eaee90621ef, pod image digest sha256:bae042123348229ddcac7263f342c580177d0b578e5880444390e09a10a0c60d. The pod served /pkg/build-revision.txt containing the exact merge revision.
AUTOCODER_BACKGROUND_AUTOMATION_ENABLED=false and AUTOCODER_IMPLEMENTATION_WORKERS_ENABLED=false remained unchanged. No existing queue was released.- The supervised two-task semantic pilot did not run: development has no enabled model credentials, and the operator directed us not to provision them. Therefore rollout and synthetic/database checks are verified, but live model behavior is not.
¶ Rollback and follow-up
- Prior production revision was
68bbc3bd6a548671007963125a93ed187b271208; prior backend tag sha-68bbc3bd6a54 and frontend tag build-8638-sha-68bbc3bd6a54.
- On regression, keep automation disabled, preserve current pod/deployment evidence, and use the repository's reviewed revert/redeployment workflow to restore the prior revision or images. Do not release queued AI work during recovery.
- When a separately authorized dev credential path exists, run the exact supervised two-task pilot before claiming live semantic acceptance.