Imported from lifeodyssey/animichi (.claude/skills/use-codex/SKILL.md). Install upstream with npx skills add lifeodyssey/animichi --skill use-codex. Copyright stays with the author.

Using Codex from Claude Code (current policy)

Codex's role in this repo

Code-writing is Policy-C'd to opencode (see use-opencode). Codex has exactly two jobs:

1. Spec adversarial review — the dual-seat seat B: /codex:adversarial-review with gpt-5.6-sol at effort xhigh. Seat A is Fable. Findings → planner revises → re-review → owner sign-off. A spec without both seats signed off is not executable.
2. Visual-capable inspection fallback — inspection work that needs eyes (screenshot judgment) which text-only executors cannot do.

Raw code-writing via the CLI is still forbidden: codex exec --sandbox workspace-write is blocked by the block-codex-exec-codewrite hook, and guard-codex.sh blocks concurrent or looped invocations.

The one rule that has never changed

Never run the raw codex CLI concurrently or in a loop. Two raw processes fight over the same Codex websocket → 403/429 connection contention (not a resettable rate limit), each one still burns quota, and retrying makes it worse. Route everything through the managed plugin — it serialises over one app-server connection.

Dispatch recipe

• Review: /codex:adversarial-review (pass the spec, model + effort).
• Manage: /codex:status, /codex:result, /codex:cancel.
• The forwarder returns before Codex finishes — arm a Monitor keyed on the job log going quiet; the report lives at ~/.claude/plugins/data/codex-openai-codex/state/*/jobs/<task-id>.log. Ask for findings incrementally; never trust the wrapper's return value.

Sandbox facts

• No network: git fetch/push, gh, pnpm install all hang or fail. Build the environment and verify every gate yourself BEFORE dispatching.
• gh is unavailable → export PR/thread context to a file in the working directory.
• Brief must include: "if anything is missing, STOP and report — do not install, do not reach the network."

Commits

Codex's sandbox refuses most .git writes (Operation not permitted); occasionally it commits anyway. When the job stops: git -C <target> log --oneline -3 and git -C <target> status --short; commit anything uncommitted yourself immediately. Then run every gate yourself — a "gates pass" claim is not evidence.

The brief must contain

• Working directory, plus pwd / git rev-parse --abbrev-ref HEAD / git rev-parse --short HEAD echoed back in the report. Give the full SHA, never a truncated commit message — a truncated message is a trap you set for yourself.
• Every gate command in full, with pnpm run prefixes, each one run by you first.
• Baseline numbers measured in this directory, never copied from a sibling clone.
• The house rules: no suppressions, 1-10-50, no any. State explicitly that a suggestion which can only be satisfied by silencing a rule is not to be actioned.
• What "done" looks like as a file when the deliverable is analysis — the report can be lost, so the findings must land on disk.

Do not

• codex exec for code (hook-blocked).
• Dispatch into a directory another agent is editing — two writers in one tree produce silent conflicts.
• Dispatch into a long-lived worktree. Delete and recreate, or clone fresh — stale brokers are the ones that fail to start.