Codex Loop (fix until clean)

Critic→Refiner loop with the Codex CLI as external critic: review → fix approved findings → verify → re-review, until codex is clean or rounds run out. Codex reviews; Claude fixes. Complements phx-review (Claude panel) — run either or both before a PR.

Usage

phx-codex-loop                    # diff vs default branch, interactive
phx-codex-loop --uncommitted      # staged + unstaged + untracked
phx-codex-loop --base develop     # explicit base branch
phx-codex-loop --auto             # auto-approve P0/P1/P2, skip P3
phx-codex-loop --max-rounds 2     # default 3

Iron Laws

1. BOUNDED rounds — never exceed --max-rounds (default 3) — each round costs codex quota; report remaining findings instead of looping on
2. Verify BEFORE re-review — mix compile --warnings-as-errors + tests must pass before burning a codex round on broken code
3. NEVER commit or push — leave git to the user
4. Codex findings get Iron Law scrutiny — decline suggestions that violate an Iron Law, with explanation in the round report
5. Skipped findings are reported, never dropped — every finding ends as fixed, declined (with reason), or deferred-by-user
6. Missing CLI stops the skill with an install hint — suggest phx-review as the codex-free alternative; never crash

Workflow

Step 1: Preflight

Run command -v codex. If missing: STOP. Show the install hint (brew install codex or npm i -g @openai/codex, then codex login; codex doctor to diagnose) and suggest phx-review.

Then two cheap checks that save review rounds:

• Dirty tree (git status --short, only when using --base): codex inspects working-tree state too and WILL flag local dirt (stray edits, dirty submodules) as findings. Surface the dirt and ask: clean/stash first, or proceed knowing round 1 may spend findings on it.
• Missing rubric: if AGENTS.md has no ## Review guidelines section, note once that phx-init installs the Elixir rubric that steers codex priorities — then proceed (it works without, on defaults).

Step 2: Detect Diff Mode

• Flag passed (--base/--uncommitted) → use as given
• Commits ahead of the default branch → --base {default}
• No commits yet → --uncommitted

Step 3: Review Round (max --max-rounds, default 3)

1. Review — ONE foreground Bash call, explicit timeout: 600000 (large diffs run 10+ min). On timeout with the process alive: ONE until [ -f {out} ]; do sleep 5; done wait — never poll-spam or pkill a running review (quota is spent either way). ALWAYS silence the streams — only the -o file matters (10k+ lines otherwise):

   codex exec review --base {branch} --ephemeral \
     -o /tmp/codex-round-{n}.md > /tmp/codex-round-{n}.log 2>&1
   

   NEVER pass custom instructions with a diff-mode flag — the CLI rejects the combination; the rubric comes from AGENTS.md ## Review guidelines (phx-init). Parse the output file, NOT the exit code (0 even with findings). Recipes: references/codex-cli.md.

   Retired model (exit 1, no -o file, log: model is not supported / Model not found): resolve the account default — MODEL=$(codex debug models | jq -r '.models | map(select(.visibility == "list")) | sort_by(.priority) | .[0].slug') — retry once with -m "$MODEL" -c review_model="$MODEL" (both: a pinned review_model beats -m), reuse it every round, never hardcode a slug. Tell the user once to set model/review_model = "$MODEL" in ~/.codex/config.toml.

2. Clean check — parse the output file; no - [P{n}] bullets means CLEAN → go to Step 5.

Step 4: Triage, Fix, Verify

1. Triage — emit the findings table (# | P | file:line | title | proposed action) as visible response text BEFORE any AskUserQuestion call — a table composed only in thinking never renders. Then ask approve-or-skip per finding (AskUserQuestion). With --auto: approve P0/P1/P2, skip P3 (list skipped in the round report).

2. Fix — apply approved fixes with user-visible diffs. Check each against Iron Laws first (Law 4).

3. Verify — scoped, before the next round:

   mix format {changed_files} && \
   mix compile --warnings-as-errors && mix test {affected_tests}
   

   If verification fails 3 times, STOP with a BLOCKER report — do not burn another codex round on broken code.

4. Next round (n+1) — back to Step 3.

Step 5: Final Report

## Codex Loop Report — {CLEAN | MAX ROUNDS REACHED | BLOCKED}
Rounds: {n}/{max} | Fixed: {n} | Declined (Iron Law): {n} | Deferred: {n}
{per-round: findings → outcome}
{remaining findings if not CLEAN}

On CLEAN: suggest phx-compound for non-obvious fixes, then commit/PR. On MAX ROUNDS: list remaining findings; offer phx-plan to convert them into a follow-up plan.

Integration

implement → phx-codex-loop (YOU ARE HERE) → clean → commit/PR → phx-watch-pr --codex
     ↑ or arrive from phx-review --codex verdict REQUIRES CHANGES

References

• references/codex-cli.md — invocation recipes, parse patterns, verified gotchas (CLI 0.142.5; model fallback 0.162.1)