Repository navigation
CodexBot v0.1.1
CodexBot v0.1.1 fixes the review-execution failures uncovered by real isolated acceptance and ships bounded multi-group review orchestration with strict completion checks.
Fixed
- Project group, continuation and synthesis output schemas onto the documented Responses Structured Outputs subset. Uniqueness and string-length constraints remain mandatory in trusted-host validation.
- Stop requiring a redundant shell command in every later turn when immutable, hash-verified unit evidence is already supplied. An execution without successful inspection still fails.
- Restore the original verified image inputs and their labels after Codex context compaction removes image history. Partial loss, substitution, duplicate replacement and over-budget requests remain rejected.
- Preserve validated findings from incomplete analysis as non-approving COMMENT reviews while the analysis job remains failed. Missing essential context, unfinished groups or synthesis never authorize approval or merging.
- Include the preceding context-completeness and sanitized primary/cleanup diagnostic corrections recorded in CHANGELOG.md.
Verified acceptance
Real ChatGPT reviews completed against the exact final implementation in isolated TopkAPI acceptance:
| Fixture | Actual result |
|---|---|
| 2 files / 2 commits | Complete clean review; native review published |
| 33 files | All three groups and synthesis completed; expected cross-file authorization defect published inline |
| 93 files / 89 text patches / four distinct PNGs / 2 commits | All six groups and synthesis completed; complete clean review published |
All three used the same final implementation SHA. Native events were COMMENT under the unchanged advisory-only test repository policy. Test webhook inputs were operator-signed staging fixtures derived from fenced real comments; the downstream GitHub App dispatch, Actions OIDC, protected Environment credentials, isolated Codex, official Responses inference and native publishing were real. This does not claim direct GitHub webhook delivery to the separate test backend.
Validation passed: 246 automated tests, strict typing, lint, build, clean checkout, real workerd/D1, synthetic-provider browser tests, Docker isolation, 19 native pinned-Codex inspection scenarios including actual compaction with four distinct images, Actionlint, npm audit and full GitHub CI including the image scan. A fresh production credential-only OIDC diagnostic passed without inference or publishing.
Deployment and upgrading
The reviewed versioned reusable-workflow source is bf0510ed98ad2bb7a3798ad9fd3d1ae1286954a2. Production Worker trust and the KMU/TopkAPI callers are coordinated to this exact immutable SHA. This release tag targets the reviewed final Worker/configuration commit; the later maintenance commits change only trust configuration and its test, then restore the merge gate.
Self-hosted operators must keep their own Worker trust and caller pins aligned and follow the documented maintenance procedure. No database migration, credential reconnection or repository-policy change is required by this patch. Quotas, isolation, authorization, exact-head checks and GitHub protections are preserved.
CLI-based ChatGPT onboarding remains the supported release path. Browser ChatGPT onboarding and hosted website inference remain disabled. Deployment-level merge authorization is restored after verified cutover; repository-specific opt-in, explicit-request policy where configured, required checks and GitHub protections still apply.
Limitations and cleanup
AI review quality is not guaranteed: an earlier small-fixture Unicode-escape false positive remains visible in the acceptance history. These tests establish successful bounded execution for the tested sizes and image transport/provenance, not universal PR completeness or semantic accuracy on arbitrary screenshots. Existing request, byte and time limits remain enforced. The original production HTTP400 did not retain a usable provider rejection envelope, so its sole rejected keyword is not retrospectively proven.
The temporary ChatGPT session was revoked, its Worker and D1 were deleted, and TopkAPI was restored before the reviewed production pin-only upgrade. Fixture PRs were closed without merging their code. No KMU business code, policy, feature PR, secrets or branch protections were changed; no additional KMU AI review was triggered during rollout.