codex-pr-closeout-review

Codex PR Closeout Review

Run the bundled helper as a closeout gate, not as permission to broaden the task.

Contract

• Freeze scope before review: original request, target base, intended behavior, owner boundary, changed files, and rough non-test LOC.
• Review only the selected diff bundle.
• Treat review output as advisory; verify every finding in the real code before patching.
• Accept only findings introduced or exposed by the current diff and fixable inside the original task boundary.
• Classify broader design, cleanup, hardening, or adjacent-surface issues as follow-up.
• Stop and report scope breakage when the best fix requires a new public API, migration, protocol, storage contract, release-process change, or owner-boundary move.
• Prefer end-to-end proof over isolated unit proof when getting code ready for PR or reviewing a PR. Unit tests are useful for tight logic, but closeout confidence should come from the smallest realistic workflow that exercises the changed behavior.
• If review-triggered edits change code, rerun focused proof and rerun this helper.
• Stop after one clean structured review. Do not rerun just for nicer wording.

Commands

Use the skill-local helper:

skills/codex-pr-closeout-review/scripts/codex-pr-review --mode branch --base origin/main

For dirty local work:

skills/codex-pr-closeout-review/scripts/codex-pr-review --mode local

For an already-committed change:

skills/codex-pr-closeout-review/scripts/codex-pr-review --mode commit --commit HEAD

Run focused proof in parallel only after formatting is stable:

skills/codex-pr-closeout-review/scripts/codex-pr-review \
  --mode branch --base origin/main \
  --parallel-tests "python3 scripts/validate-release.py"

Optional context must be repo-relative:

skills/codex-pr-closeout-review/scripts/codex-pr-review \
  --mode branch --base origin/main \
  --prompt-file review-notes.md \
  --evidence proof/summary.json

Helper Behavior

• Builds one sanitized git bundle with safe git diff flags.
• Runs one Codex review in read-only mode with structured JSON output.
• Keeps web search off unless the caller passes --web-search.
• Rejects evidence paths that are absolute, escape the repo, point through symlinks, look sensitive, or contain secret-like values.
• Filters findings outside the reviewed changed files.
• Writes only requested output files plus normal stdout/stderr.
• Fails the Codex engine path cleanly if the model call exceeds --timeout-seconds.
• Exits 0 for clean review, 1 for actionable findings or incorrect verdict, 2 for parallel proof failure, 3 for unsafe input or target setup failure, and 4 for Codex/JSON/schema failure.

Final Report

Include the review command, proof command, accepted findings and fixes, rejected or follow-up findings with rationale, and the final clean helper result or the conscious reason a remaining finding was not fixed.