codex-review

Files (2)

• SKILL.md
• codex_review.py · scripts/

SKILL.md (first lines)

# Codex Review

Use Codex as an independent reviewer for Shannon issue designs, code diffs,
experiment results, and architectural decisions.

The review is advisory. The calling agent remains responsible for deciding
whether Codex's feedback is correct, acting on valid findings, and explaining
any rejected findings.

Do not rely on Codex session memory alone. Session IDs are useful for grouping
related reviews and may help continuity, but every important review prompt must
include the essential issue, diff, log, and question context explicitly.

This skill is the Codex counterpart of `claude-review`. Use either or both when
you want a second model's read on risky work.

## When This Skill Applies

Use this skill when:

- the user asks for "codex review", "ask Codex", or similar;
- an experiment is large, risky, or touches the Nushell/Reedline fork surface;
- an experiment has already failed once and the next design needs review;
- a code change affects architecture, shell state synchronization, process
  handling, signal handling, config loading, terminal behavior, or persistent
  state;
- before closing an issue after a complex series of experiments.

## Review Timeout

Every Codex review has a hard timeout of 15 minutes.

Use the helper script in this skill instead of calling `codex` directly. The
script enforces the timeout with Python's subprocess API and terminates Codex's
process group on timeout because macOS does not provide a portable `timeout`
command by default.

## Files and Logs

All review state and output lives under:

```text
logs/codex-review/
```

The helper script maintains:

| Path                                            | Purpose                         |
| ----------------------------------------------- | ------------------------------- |
| `logs/codex-review/current-session-id`          | reusable Codex session UUID     |
| `logs/codex-review/<timestamp>-prompt.md`       | exact prompt sent to Codex      |
| `logs/codex-review/<timestamp>-events.jsonl`    | raw Codex `--json` event stream |
| `logs/codex-review/<timestamp>-last-message.md` | Codex's final review message    |
| `logs/codex-review/<timestamp>-stderr.log`      | Codex stderr / timeout details  |

These logs are intentionally under `logs/`, which is gitignored. They may still
contain sensitive diffs or local paths. Keep context narrow and do not write
secrets into review prompts.

## Process

1. **Prepare the review context.** Include only the context Codex needs. Prefer
   concrete artifacts:
   - the user's current request;
   - the issue file and experiment section;
   - relevant `git diff` or `git diff --staged`;
   - test commands and observed outputs;
   - screenshots/log paths when those are part of the evidence.

2. **Ask specific questions.** The prompt should tell Codex what kind of review
   to perform. Examples:
   - "Audit this experiment design for contradictions, missing verification, and
     architectural risks."
   - "Review this diff for correctness, regressions, and missing tests."
   - "Given these logs, identify what failed and what experiment should come
     next."

3. **Run the helper script.**

   ```bash
   python3 skills/codex-review/scripts/codex_review.py \
     --context issues/0789-electron-style-pdf-viewer/README.md \
     "Audit Experiment 6. Focus on whether the scope is right and whether the verification would prove the result."
   ```

   For a staged diff review:

   ```bash
   git diff --staged | python3 skills/codex-review/scripts/codex_review.py \
     "Review this staged diff for bugs, regressions, and missing tests."
   ```

   For a fresh thread:

   ```bash
   python3 skills/codex-review/scripts/codex_review.py --new-session \
     "Start a new review thread. Reply with the review session status."
   ```

4. **Read the result.** The helper prints:
   - the session ID;
   - the prompt path;
   - the raw event-stream path;
   - the final-message path;
   - the extracted Codex text when a final message was captured.

5. **Act on valid findings.** If Codex finds a real issue and the user asked for
   implementation, fix it. If Codex is wrong or speculative, do not make the
   change; briefly document why.

6. **Report high-signal results.** Summarize Codex's actionable findings to the
   user. Do not paste long raw output unless the user asks for it.

## Session Handling

Codex differs from Claude here. `claude -p` lets you choose a session id up
front with `--session-id <uuid>`. Codex does not: `codex exec` generates the id
itself. The helper therefore captures the id from the `--json` event stream
after the first run and stores it for follow-ups.

From shannonshell/shannon, MIT.

More from shannonshell/shannon