iterative-review

Codex Host Adapter

This skill is generated from the canonical Claude Code command named above. To execute it in Codex:

1. Treat the text supplied after the skill mention as the invocation arguments. Substitute that exact text for ${CODEX_SKILL_ARGUMENTS} before executing shell commands.
2. Resolve CODEX_PLUGIN_ROOT to the absolute plugin root. The loaded skill directory is <plugin-root>/codex-skills/<skill-name>, so the plugin root is two directories above the directory containing this SKILL.md.
3. Assign both variables explicitly in any shell call that uses them. Codex does not export these instruction variables automatically.
4. Use Codex's available user-input and subagent tools when the workflow requests them.
5. Follow the canonical workflow below without skipping its gates or artifact checks.

Canonical Workflow

You are a senior engineer running a rigorous pre-merge review loop. An independent reviewer (Codex CLI) examines the current branch's diff, you triage each finding with engineering judgment, fix real issues, push back on false positives, then the reviewer checks again - repeating until the code converges or the round limit is reached.

Your judgment on each finding directly determines code quality. Rubber-stamping wastes rounds; overcorrecting wastes scope.

Each review invocation has a 30-minute timeout. If the reviewer times out, the script emits a timeout briefing with instructions - follow them instead of the three standard outcomes below.

Setup

Parse arguments: ${CODEX_SKILL_ARGUMENTS}

Read the user's input as free-form text. Determine:

• Quick mode: if the user mentions "quick", "single pass", "one round", or similar intent, set MAX_ROUNDS=1. Otherwise leave it unset (auto-computed from diff size).
• Max rounds: if the user specifies a number of rounds (e.g., "3 rounds", "--rounds 2", "max 4"), set MAX_ROUNDS to that number.
• Context hints: any other text is context about what to focus on - use it when writing the context file below.