Starter kits

clodex-adversarial-review — the two-agent review loop

A real maker–checker loop across two agents: Claude implements and fixes, an adversarial reviewer (Codex, or a subagent stand-in) reviews against a…

A real maker–checker loop across two agents: Claude implements and fixes, an adversarial reviewer (Codex, or a subagent stand-in) reviews against a severity threshold, up to 5 iterations. Never describes an errored or exhausted run as approved.

Category: B · PR & review · Heartbeat: event (per PR, up to 5 iterations) · Level: L1 (report-only, drafts-not-pushes) · Cost: Medium

Source: Loop #9, "The Clodex adversarial-review loop," from Forward Future's Loop Library, original prompt by Lukas Kucinski.

Quickstart

1. Read LOOP.md — the six parts, the prompt, the limits, the three stops,
   and the source's verbatim prompt.
2. clodex-adversarial-review-state.md is your spine — already committed
   with an empty per-PR review-cycle register.
3. Claude Code: wire a PR event trigger at
   .claude/skills/loop-task/SKILL.md — the loop-verifier agent stands in
   for Codex if the real `codex` CLI isn't installed. OpenCode: rename
   opencode.json.example → opencode.json, add the AGENTS.md line, wire
   skills/loop-task.md into the Actions event.
4. Rehearse once, by hand, on a real PR before the trigger ever fires:
   "Run the clodex-adversarial-review skill now on PR #N, exactly as the
   event would."
5. Grade the rehearsal — the loop-verifier checks that a stalled/exhausted
   cycle was never reported as "approved." PASS → arm the trigger. FAIL →
   fix the skill, never the verdict.

What's in this kit

File	Role
`LOOP.md`	the definition — six parts, prompt, limits, ownership, three stops
`clodex-adversarial-review-state.md`	the spine — per-PR iteration count, findings history, verdict (resumable)
`loop-budget.md`	caps: 5 iterations/PR (the source's own), 400k tokens/PR
`loop-constraints.md`	the constitution — never calls an exhausted run approved, reviewer must be genuinely separate
`loop-run-log.md`	append-only, one line per beat
`.claude/skills/loop-task/SKILL.md`	the procedure (Claude Code)
`.claude/agents/loop-verifier.md`	the read-only reviewer — doubles as the adversarial-review stand-in when `codex` isn't available
`opencode.json.example` · `skills/loop-task.md`	the OpenCode twin — same skill text, event heartbeat

Loop Ready checklist

Provable success condition — the source's own spec: reviewer approves, or only accepted findings remain
Run limit — 5 iterations/PR (the source's own cap)
Spine written first, committed empty — resumable per the source's requirement
Report-only (L1) start — drafts fixes in a throwaway worktree, never pushes to the real PR
Human gate placed — you read the verdict and findings and push the fix yourself
One log line per beat, no silent runs
Kill switch — loop-pause-all in loop-budget.md

How this relates to the course

This kit is a real, working instance of Step 11's maker–checker split: Claude is the maker, the adversarial reviewer is a genuinely separate pass, never the same agent grading its own work. The source's own iteration cap (5) and "progress stalls" stop are exactly the maker–checker standoff bound — "a rejection cap... stops the pair and escalates the disagreement to the human gate." The source's sharpest rule — "never describe an errored or exhausted run as approved" — is this loop's own version of CLAUDE.md rule 12, "green ≠ done."

Full attribution: resources/sources.md.