oduffy-delphi/codex-review-gate

Codex Review Gate

Internal skill invoked by /bug-sweep and /workday-complete to get an independent-model code review via the Codex plugin. Codex (GPT-5.4) provides a different model family's perspective, catching issues that intra-family reviewers may share blind spots on.

When This Skill Is Called

The calling command provides context:

• scope: "bug-sweep-fixes" or "workday-diff" — informational, for reporting
• base: git ref for diff base (default: origin/main)
• required: true or false — if false, failure is non-blocking (the normal case)

Execution

Step 1: Check diff exists

git diff --shortstat {base}...HEAD

If empty (no changes against base), report and return:

"Codex review gate: no diff against {base} — skipped."

Step 2: Run the Codex review

Invoke /codex:rescue with the review task:

Review the diff between {base} and HEAD for code quality issues, bugs, security vulnerabilities, and adherence to project conventions. Focus on P0/P1 findings only. Return structured findings.

This routes through the rescue subagent, which calls the Codex companion CLI directly. The /codex:review command has disable-model-invocation and cannot be invoked via the Skill tool.

Step 3: Assess result

Use exit code only for success/failure discrimination:

• Exit code 0 = success. The output contains the review findings.
• Non-zero exit code = failure. The output or stderr contains the reason.

Do not use string matching on output headers — that's fragile against plugin format changes.

Step 4: Report

On success:

• Summarize the Codex review: verdict, finding count, and severity breakdown
• Present the full Codex output for the calling command to include in its report
• If findings exist, categorize by severity (P0/P1 = flag to PM, P2 = note for next session)

On failure (graceful fallback):

• Extract the reason from output/stderr. Common reasons:
  • "not installed" — Codex CLI not installed. Suggest: /codex:setup
  • "not authenticated" — Codex CLI not logged in. Suggest: !codex login
  • Connection/broker errors — Codex infrastructure issue
  • Any other error — report first meaningful line
• Report: "Codex review gate skipped: {reason}."
• If required: false (the default): continue without blocking. The calling command's existing reviewer output stands alone.
• If required: true: halt and report to PM for decision.

Design Notes

• Codex is additive, never blocking in the default configuration. Users on limited ChatGPT plans benefit most from bounded token spend (one review per end-of-day, not continuous).
• The existing persona reviewers (Patrik, Sid, Camelia, Pali) from /code-health provide the primary review. Codex is a second opinion from a different model family — complementary, not a replacement.
• Blind spots may be correlated within a model family — if Claude misses a pattern, its reviewer personas (which share the same model family) are more likely to miss it too. Codex mitigates this by providing an independent sample.