claude-codex-loops

Run ONE bounded ping-pong review round with the Codex peer. Pull the peer's latest findings, verify each against the real code, fix what is real TDD-first (delegating the maker/checker legs to the dev-wave workflow), then push a redacted acknowledgement plus a fresh review-request packet and print the exit state. Exactly one round per session (turnBudgetThisSession = 1); raising the budget requires an explicit --turns. All brakes are inherited from the pack state machine — a terminal thread is a no-op that reports its reason. Trigger: /pingpong.

Drive ANY multi-iteration task — build, refactor, migration, or research — as a bounded, self-paced orchestration loop. Each iteration you do a unit of work (the maker), run an INDEPENDENT checker, record the outcome to a named loop's disk state, then let the pack's bounded state machine decide whether to continue or exit. Brakes are inherited: a round cap (maxRounds), a no-progress brake, and the kill switch all force a sticky terminal so the loop cannot spin. This is BUILD-TIME only — a self-paced session loop that always exits, never a runtime daemon. The helper manages state under <stateDir>/loops/<name>.json and prints the next step; it spawns no models — YOU are the maker and checker. Trigger: /ccl-loop.

Install the claude-codex-loops pack into the current git repo, or run the preflight check with the `check` argument. The check shells to the pack doctor (bin/doctor.mjs), which reports PASS/FAIL for the CLIs and state the pack needs — git, the gh/claude/codex logins, node, a writable state dir, merged hooks, no duplicate scheduler label — and warns if a provider API key is set (the pack neither needs nor accepts one). If the doctor is not yet installed the check reports that gracefully and never throws. The pack ships for macOS/launchd in v0.1.0; the installer refuses other platforms with a clear message. Trigger: /pingpong-install.

Triage the outstanding review-bot findings on a pull request through the bounded pr-review-triage workflow. Optionally pass a PR number; otherwise the workflow resolves the current PR from the branch. It collects the bot comments, classifies each finding evaluator-first, fixes what is real, and records a terminal verdict per finding in the review ledger — bounded by the per-fingerprint two-pass brake and the whole-sweep triage.maxRounds cap, so it always terminates or escalates to blocked-needs-human. Runs inside the Claude Code Workflow harness; it adds no new logic beyond that workflow. Trigger: /pingpong-pr.

Run a bounded, self-paced RESEARCH loop that compounds only VERIFIED, cited findings. Each round you do one research pass with YOUR OWN session tools (WebSearch / fetch / read — never a provider API), then verify every claim against a real source: a claim with a source is appended to a per-topic findings ledger (deduped by fingerprint); a claim with NO source is DROPPED and logged. Adversarial verification is built in — unsourced claims never reach the ledger. The loop is bounded: it exits on a QUIET round (no new verified findings) or on maxRounds, and the kill switch halts it first. This is BUILD-TIME only, a self-paced session loop that always exits, never a daemon. The helper manages state and the ledger under <stateDir>/research/; it spawns no models — YOU do the searching and verifying. Trigger: /research-loop.