Codex Stage Review Loop is a lightweight Codex Skill for putting a quick gate in front of the next step. It gives Codex a repeatable way to pause, check the work, and avoid moving on just because the task looks finished.
The workflow is intentionally simple: return to the user's original request, verify against real outcomes such as diffs, files, tests, logs, or errors, fix small low-risk issues when they are clearly confirmed, and stop for the user when the next move is uncertain or risky.
It is not meant to guess every hidden bug. Its purpose is to give AI-assisted work a clear acceptance floor: no vague "looks done" handoff, just a scoped check with evidence and explicit stopping conditions.
AI coding agents can complete the main task but still miss the last mile: a skipped check, an unverified assumption, an optional refactor treated as required, or a final answer that sounds more certain than the evidence supports.
This skill provides a small reusable closeout loop for those moments. It is useful after code changes, documentation edits, generated artifacts, local configuration work, or changes to another Codex Skill.
- Locks the review scope to the user's request, touched files, acceptance evidence, and explicit exclusions.
- Checks that each finding, fix, or suggestion still serves the user's original goal instead of creating process-driven extra work.
- Prioritizes inspectable evidence such as diffs, files, tests, build output, logs, rendered UI, or smoke checks before model judgment.
- Runs a fresh-perspective pass to look for missed requirements, secondary bugs, and user-facing gaps.
- Fixes verified blocking issues when the fix is inside scope and low risk.
- Reports optional improvements as suggestions instead of silently expanding the task.
- Stops for user confirmation when the next step is ambiguous, risky, or a product decision.
The default mode for frequent checkpoints. It runs a compact acceptance pass and a fresh-perspective review in the current session, with a small bounded fix loop.
Use it for everyday checkpoints:
Use $stage-review-loop to lightly verify the current work and fix real blocking issues.
A stronger mode for final confidence or higher-risk work. It prefers a separate reviewer, fresh thread, fresh model surface, or other genuinely independent review surface when available.
If no independent surface is available, the skill should report that it only ran a constrained fresh pass inside the current context.
Use $stage-review-loop in independent reviewer mode to check whether this matches the original request.
This skill is not a replacement for Codex's built-in code review or PR review. Use built-in review for focused code-diff findings.
Use this skill when the question is broader than a diff: "Did this task actually meet the request, did we verify it, and are any blocking issues still open?"
They can also be combined: run a code review for diff-level defects, then run this skill as a task-level acceptance pass.
This skill can reduce missed acceptance checks, scope drift, and evidence-free completion claims. It cannot reliably catch problems that require unavailable external facts, missing domain knowledge, absent tests, unclear requirements, or a truly independent reviewer.
The workflow is deliberately lightweight. It should not become a large release process unless the user explicitly asks for a stronger final review.
Copy the skill folder into your Codex skills directory:
stage-review-loop/
SKILL.md
agents/
openai.yaml
Then invoke it by name:
Use $stage-review-loop to verify the current work.
SKILL.md: the skill instructionsagents/openai.yaml: Codex UI metadata
MIT