SuperAPI AI Directory

run-codex-review

Coding & development

View repository

Updated 10-16 04:23

About this skill

run-codex-review

Deprecated compatibility shim for the legacy /codex:review install path. This skill no longer ships its own implementation — it routes the request to the canonical skill that owns the user's actual intent. Do not restore the old runner. Do not invent a new one. Pick the route, then load the target skill.

When to use this skill

Trigger on any of these phrasings, branch shapes, or environment cues:

  • "run codex review on these branches", "converge these branches with codex", "fix-loop this branch list"
  • "resume my codex review run", "the codex review crashed mid-run, pick it up", "rescue the review fleet"
  • "open the PR / hand this off for review" paired with intent to use codex review afterward
  • "the bots reviewed my PR — Copilot + CodeRabbit + Devin — sort it out"
  • "go back over the codex review feedback and tell me what to act on"
  • "installed run-codex-review but the slash command does nothing" (legacy install-path users)
  • A branch list (comma-separated, branches.txt, or ≥2 git refs) plus the words "review", "ship", "merge", "close out"
  • codex --version succeeds AND the user names ≥1 branch and a review keyword

Do NOT use this skill when:

  • The user wants to review someone else's GitHub PR for merge readiness → load review-pr
  • The user wants generic non-codex parallel coding fan-out → load run-codex-2 exec mode directly
  • The work is opening a PR with no codex review involved → load review-self directly
  • The repo is dirty across multiple worktrees with no immediate review intent → load run-repo-cleanup

Routing table — pick the canonical skill, then exit this shim

Match the user's request against one row, surface the route, hand control to the named skill. The shim's job ends after routing.

User intent Canonical skill Entry point
Per-branch codex review fix loop (branch list + review keyword) run-codex-2 review mode node skills/run-codex-2/skills/run-codex-2/scripts/run-codex-2.mjs review --branches <list>
Resume / rescue an interrupted codex review run run-codex-2 rescue mode node skills/run-codex-2/skills/run-codex-2/scripts/run-codex-2.mjs rescue --manifest <path>
Open a PR for the author's own branch (with self-review body) review-self Load skills/review-self/skills/review-self/SKILL.md
Triage received feedback from one or many reviewers (human, Copilot, CodeRabbit, Devin, Bito) review-feedback Load skills/review-feedback/skills/review-feedback/SKILL.md
Reviewer-side merge-risk review of a PR or diff review-pr Load skills/review-pr/skills/review-pr/SKILL.md
Worktrees + branches dirty after a review fleet run-repo-cleanup Load skills/run-repo-cleanup/skills/run-repo-cleanup/SKILL.md

Disambiguation rules (load-bearing)

  1. Codex review ≠ PR review. run-codex-2 review mode runs codex exec review --base <base> per branch and converges via classifier rounds. It does NOT open PRs. PR creation is review-self. Reviewer-side merge calls are review-pr.
  2. Branch list signal wins for fix loops. If the user supplies ≥2 branches AND mentions "review/ship/merge/close out", route to run-codex-2 review — even if the prompt also mentions PRs. PR handoff is a separate downstream step routed to review-self after convergence.
  3. Multi-bot evaluation ≠ codex convergence. When the user names third-party reviewers (Copilot, CodeRabbit, Devin, Greptile, Bito) or says "the bots reviewed my PR", the work is feedback triage — route to review-feedback. This shim does not own bot-wait timing.
  4. Manifest existence forces rescue. If ${CLAUDE_PLUGIN_DATA}/state/<slug>-<hash>/run-codex-2/manifest.json (or the codex-companion fallback under ${TMPDIR:-/tmp}/codex-companion) exists with non-terminal entries AND the user says "resume / continue / pick up / rescue / prior run", route to run-codex-2 rescue. Do not overwrite a live manifest.
  5. No restoration of the old runner. This shim does NOT spawn codex, does NOT write a manifest, does NOT arm a Monitor. Routing is the entire job. The canonical skill owns runtime.
  6. Slash-command boundary. The vendored /codex:review lives inside run-codex-2 under scripts/codex-cc/ and is not used by review mode. Review mode uses native codex exec review. There is no slash command owned by this shim.

Trigger-precision cues for the canonical route

When run-codex-2 is the route (review or rescue mode), confirm these environment signals before handing off so the canonical skill does not bounce the request:

  • codex --version exits 0 (skill assumes 0.130.0+ for codex exec review flags)
  • git rev-parse --is-inside-work-tree succeeds in cwd
  • Branch list resolves (comma-list, branches.txt, or positionals) — at least 2 refs for review mode
  • codex login status is a soft probe; bearer-token / managed-companion / proxy auth is allowed even when status reports "Not logged in" (set USE_CODEX_SKIP_CODEX_AUTH=1 to bypass the bash bootstrap hard-fail)
  • For rescue: a manifest path exists under the resolved state dir and manifest.mode == "review" for review-fleet rescue

If any cue fails, surface the gap in one line and stop — do not auto-improvise into run-codex-2 and let it fail downstream.

Anti-patterns (do not repeat the legacy shim's old mistakes)

  • Never restore the old run-codex-review runner. The old third-party bot-wait timing constants are retired; bot waits route to review-feedback.
  • Never spawn codex from this shim. Routing is the entire job.
  • Never bypass the routing table to "save a step" — wrong-route handoffs lose the user's manifest, branch list, or feedback context.
  • Never collapse "codex converges branches" and "PR is ready for review" into one skill call — they are two routes (run-codex-2 review → review-self).
  • Never invent a /codex:resc or similar slash command. Rescue is a CLI subcommand of run-codex-2, not a slash command of this shim.

Final behavior

Pick the row from the routing table, state the chosen route in one line, and load the canonical skill's SKILL.md. The shim itself produces no other artifacts — no manifest, no Monitor, no codex spawn, no PR. The canonical skill owns every downstream contract.