Repository navigation
fix: unblock structured and multi-group Codex reviews - #50
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
CodexBot could reject the first Responses request, demand redundant shell inspection in later review turns, or lose required image inputs after native Codex compaction. This fixes those three execution failures without changing inference budgets, credentials, authorization, quotas or approval/merge policy.
Real ChatGPT acceptance on the exact final source 2a5f0e8:
All native events were COMMENT because the unchanged TopkAPI policy disables automated approval and requested changes. All results bind to the exact fixture heads. The staging webhook deliveries were operator-signed fixtures derived from fenced real comments, followed by actual App dispatch, Actions OIDC, protected Environment credentials, isolated Codex, official Responses inference and native publication. This is not evidence of direct GitHub webhook delivery to the separate test backend. No API-key fallback or KMU review was used.
Deterministic/native regressions reproduce the schema rejection, per-turn inspection failure and actual pinned-Codex image-dropping compaction before their corrections. The compaction case retains all four distinct images on every upstream request through six groups, continuation and synthesis. Validation passed: npm run check (246 tests, typecheck, lint, build); clean checkout with npm ci --ignore-scripts; workerd/D1; browser tests with synthetic providers; catalog; npm audit (zero vulnerabilities); Actionlint; Wrangler dry-run; Docker and all 19 pinned-Codex synthetic-provider inspection scenarios. Both full GitHub CI runs on this head passed, including image scanning and sandbox checks.
Cleanup completed: the official CLI confirmed revocation of the temporary renewable ChatGPT session; TopkAPI's original vault secret was restored securely through stdin. Cleanup PR29 restored the exact original workflow bytes and entire original main Git tree. Fixture PRs23–25 were closed without merging. The temporary Worker, D1 and local secret copy were deleted; resource absence and unchanged production deployment/health were independently checked. Production/KMU configuration and the production trusted workflow pin were not changed. This source PR remains unmerged.
Limitations: the historical production HTTP400 did not retain a usable provider rejection code, so its sole rejected keyword is not retrospectively proven. Successful live requests now establish the corrected path works for these fixtures, not arbitrary PRs or guaranteed model accuracy. An earlier small test's Unicode-escape false positive was disproved with the original bytes and isolated assertions; that review remains visible. The equivalent literal-input fixture passed, and the false positive is not concealed or counted as a clean success. Four small distinct PNGs establish image transport/provenance, not semantic accuracy on complex screenshots. All existing hard resource and completeness limits remain enforced.