claude-code-devkit

One review cycle of the PR review pipeline (Agent A). Designed to be invoked by `/loop /pr-review-tick` inside a long-running tmux daemon session, one per repo. Each tick discovers open PRs that have explicitly opted into the pipeline (via the `[auto PR loop]` marker in title or body), dedups against the (pr_number, head_sha) → review_id map in `~/.local/state/claude-pr-pipeline/<repo>/`, and for each opted-in PR with new content runs a three-angle review — (1) codex GPT-5.5 cross-vendor adversarial, (2) `code-reviewer-opus` user-level Opus reviewer (CLAUDE.md mandated), (3) bespoke Opus PR-coherence pass covering cross-commit / scope / test-completeness / CI / arch-fit / breaking-change angles the per-commit hook can't see — then aggregates findings, posts ONE PR comment, and re-arms via ScheduleWakeup. Sibling skill is `pr-orch-tick` (Agent B, fix + verdict). Use this whenever the user says "start the review daemon", "kick off the PR review tick", "Agent A", or `/loop /pr-review-tick`. NEVER fixes code. NEV

Defend a `git commit` blocked by the `second-opinion-commit-gate.sh` Claude Code hook (a non-Claude `codex review` on the staged diff). Triage each `[P0]` finding into worth-fixing / wrong-analysis / not-worth-fixing, fix the real ones with regression tests, retry the commit, and loop until it lands clean or you push back via `[skip-review]` with documented reasoning. Use this whenever the user asks to "run a commit defense loop", "fix the second-opinion findings", "address the P0s and retry the commit", "babysit this commit through the gate", "make the hook happy", or whenever a `git commit` tool call fails with `=== second-opinion BLOCKING ===` in stderr — even if the user doesn't name the skill, this is the workflow that turns hook output into real fixes instead of reflexive `[skip-review]` dismissals. Sibling to `pr-defense-loop` but tighter (single turn, no ScheduleWakeup, no /tmp state) because each iteration costs ~60s of hook latency, not minutes of waiting on external bots.

Three-stage quality pipeline for landing commits cleanly. Runs Claude's own multi-perspective Opus review, then a `simplify` pass for reuse/quality/efficiency, then runs the Codex / gpt-5.5 second-opinion review as the final gate. Operates in two modes — **Mode A (pre-commit)** runs against a staged diff and ends with `git commit`; **Mode B (post-commit audit)** runs against already-landed commits via `codex review --commit <SHA>` and lands fixes as a follow-up commit. Trigger phrases for Mode A include "ship this commit", "do the pre-commit ritual", "make this commit ready"; Mode B triggers on "audit those commits", "did our commits go through review?", "replay the gate on the recent commits", "run quality pipe on the apply work" (i.e. anything implying retrospective check after commits already landed). Do NOT use when the hook has already blocked a fresh commit — go straight to `commit-defense-loop` for that reactive case.

Defend a GitHub PR against automated code-review bots (Codex / chatgpt-codex-connector, Claude Code Review, etc.) by triaging each round of comments into worth-fixing / wrong-analysis / not-worth-fixing, applying fixes for the worth-fixing ones, committing, pushing to retrigger review, and looping until the bots stop raising new actionable comments. Use this whenever the user asks to "run a defense loop", "fix all the bot comments", "loop until Codex is happy", "babysit a PR through review", or anything that implies polling a PR, addressing review feedback, and pushing fixes on a cadence. Self-paces via /loop ScheduleWakeup; survives multi-bot setups; handles Codex's re-anchored false positives without chasing them forever.

One orchestration cycle of the PR review pipeline (Agent B). Designed to be invoked by `/loop /pr-orch-tick` inside a long-running tmux daemon session, sibling to `/pr-review-tick` (Agent A). Each tick discovers PRs that have a new review comment from Agent A, spawns a fix worker sub-agent (in an isolated git worktree at `/tmp/orch-<slug>-pr<N>`) following the `pr-defense-loop` triage discipline, applies fixes with `[auto N/4]` tagged commits, and after the worker reports "no more actionable findings" runs the merge-gate verdict — either posting a "would auto-merge" dry-run comment, executing `gh pr merge`, or @-mentioning the user with a summary. Use this whenever the user says "start the orchestrator daemon", "Agent B", "kick off the orch tick", or `/loop /pr-orch-tick`. NEVER reviews PRs (that's Agent A's job). Caps at `max_rounds=4` per PR; refuses to auto-merge external contributor PRs regardless of dry-run state; respects per-PR `flock` so it never collides with Agent A.