OpenAI Codex

Topic file for running pr-reviewer from inside OpenAI Codex (CLI / app). Not yet exercised end to end in staged mode; the dispatch shape below matches the collaboration.spawn_agent schema observed in one Codex session (2026-09). Read your Codex's actual tool schema and model list before dispatching.

Fetch tooling

Codex has shell access in its sandboxed workspace and is competent with gh and glab. Network requirement: Steps 2 and 4 need outbound network. If the sandbox is offline, the user must check out the PR head beforehand and the skill can only run the "current branch" path, verifying HEAD == headRefOid against metadata the user pastes in. Codex must not invent metadata it could not fetch. The default workspace-write sandbox may not allow $HOME/.cache; Step 3's temp-dir fallback covers that.

Models

Same aliases as the-mister's Codex file.

Codex takes effort as a dispatch parameter, so "high effort" is reasoning_effort: "high" rather than prompt text. Record the model and effort you passed.

Fan-out primitive

spawn_agent takes task_name, message, fork_turns, model, reasoning_effort. Three calls for the reviewers, then wait_agent repeatedly until all three have returned (it returns on the first completion); only then spawn the assessor:

{
  "task_name": "pr_reviewer_a_correctness",
  "message": "<reviewer prompt from references/passes.md, fully instantiated>",
  "fork_turns": "none",
  "model": "gpt-5.6-terra",
  "reasoning_effort": "medium"
}

fork_turns must be "none": a model override requires a non-full-history fork, and a clean slate is what keeps one reviewer's context out of another. Reviewers must not be able to see each other's results; do not fork from a turn that already contains one.

If the sandbox has no spawn_agent, fall back to single mode and record mode: single (fallback: <reason>) in findings.md so the report says so.

Looping primitive

Codex CLI has no /loop primitive at the time of writing (codex#15679; broader scheduling at codex#25466). Use one of:

• Iteration prompt: "Run /pr-reviewer <num> --rounds=3 and keep invoking it until findings.md records converged: true for the current revision."
• External driver (CLI): a shell while loop that breaks on the convergence marker (see the awk check in generic.md). Always pass --rounds=N.
• Codex app Automations for coarse scheduled re-reviews.

Posting

Step 11 of the skill. If the sandbox is offline, print the report and let the user post it; do not simulate posting.