codex-review — "Claude checks Codex"

A thin broker for the review pair. Codex writes the code; Claude reviews its diff independently. (The old note "codex exec headless hangs on Windows" is obsolete: on codex-cli 0.141+ headless works natively — verified 2026-07-07 on the laptop and 2026-07-14 on the hub.) File handoff, no shared credentials, subscription (not a paid API).

When to run it

The owner made a change through Codex (or there is simply a diff/edit in a repository) and wants an independent "second pair of eyes" check. Triggers: see description.

What I do

1. Identify the target. Ask for / take the path to the repository where Codex made the change (or the path to a patch file). Default — the working repository the owner named.
2. Run the engine (deterministic, 0 of my tokens for the run itself — the review is done by headless claude -p):

   python "%USERPROFILE%\.claude\scripts\cc-review\cc_review.py" --repo "<path to repo>" --model sonnet

   • model: Sonnet by default (the free subscription bucket; in the test it caught both bugs in 25s); Opus for hard/safety-critical changes (--model opus) — the quality gate.
   • other modes: --range "HEAD~1 HEAD" (a specific commit), --diff "<patch.diff>" (a ready patch), --task "<task.md>" (what Codex was asked to do — gives the reviewer context).
3. Report to the owner: the verdict (APPROVE / REQUEST_CHANGES), the list of findings (file:line · severity · what is wrong · the fix), and the path to review-<ts>.md. If REQUEST_CHANGES — propose fixing the root cause (not the symptom).

Boundaries

• The engine forcibly unsets ANTHROPIC_API_KEY → the review runs on the subscription, not on a paid key.
• A giant diff is truncated to 120k characters (with a note) — if you hit that, review it in parts / per file.
• This is a REVIEW, not an auto-fix: fixes are proposed/made separately per AK-47 (easy and reversible — do it yourself; a fork — ask).
• The full auto-pair (Codex headless as well) can be built later under WSL2/Docker — a separate step (not needed now).

2026-06-28 (hub): the engine cc_review.py was lost (not in git / .stversions / on any disk) → rebuilt as a simple Windows-native "normal mode" (the call above). Run live (/tt): bugs → REQUEST_CHANGES, clean → APPROVE, empty/not-a-repo → graceful. Artifacts live outside the vault (reviews/), counter reviews/_log.jsonl. ⚠️ Canon update in [[decision-hermes-multivendor-arbitrage-rejected]] (addendum 06-28).

⭐ The two-way hetero pair (both sides run on the hub)

Full symmetry: both sides review each other, and the final check is always done by the OTHER vendor.

• Claude → reviews Codex (forward) — WORKS: python "%USERPROFILE%\.claude\scripts\cc-review\cc_review.py" --repo "<repo>" --model sonnet (see above).
• Codex → reviews Claude (reverse, the mirror) — WORKS ON THE HUB TOO: python "%USERPROFILE%\.claude\scripts\cc-review\codex_review.py" --repo "<repo>" — the engine was /tt-verified 2026-07-07 (laptop, codex-cli 0.141.0) and 2026-07-14 (hub, codex-cli 0.144.4: npm i -g @openai/codex, the ChatGPT-subscription login was already in ~/.codex/auth.json; smoke: planted bugs → REQUEST_CHANGES with exact file:line in 16s · clean diff → APPROVE · empty diff → graceful without calling Codex). Native codex exec -s read-only on Windows without WSL2. Flags: --range "HEAD~1 HEAD", --diff <patch>, --task <task.md>, --timeout 300. Read-only sandbox: Codex only reasons over the diff. Output: review-codex-<ts>.md + VERDICT: APPROVE|REQUEST_CHANGES.

Full auto mode under WSL2 (archived path — NOT deployed on the hub)

⚠️ The orchestrator codex_pair.py is missing from the hub's live tree (it survived only in the snapshot _config-backup\snapshot-2026-07-04), and WSL2 is not installed there → the mode does not start. To bring it up: restore codex_pair.py from the snapshot + install WSL2/Codex, then /tt. Historical context (the machine where it was built, 2026-06-22): WSL2 Ubuntu-24.04 + Node22 + Codex 0.142.3 (logged into ChatGPT), codex exec --full-auto headless. What it did: Codex implements the task headless inside WSL → takes the git diff → Claude reviews it (cc_review.py) → verdict + codex-change-*.diff + review-*.md. The repo lived in the WSL Linux filesystem (/root/...); a path normalizer fixed the Git-Bash /root/... mangling.

Extension — the reverse side is READY (the hub has been fully two-way since 2026-07-14)

The reverse side (Codex reviews Claude) is built = codex_review.py (see the "two-way hetero pair" block), native codex exec on Windows. The Codex CLI was installed on the hub 2026-07-14 (npm i -g @openai/codex, v0.144.4, login via the ChatGPT subscription) and the smoke test passed — the hetero pair works in both directions. The same day it was also deployed to the anchor VPS (headless: auth.json is copied over with scp, no browser OAuth needed; both engines + smoke ✅) — the anchor is an extra node, and live duo tests still run from the hub in the review chat.

Like this skill? It is one of 100 in second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.

Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).

Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_

Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.