crew

Dispatch as the LAST step of a ticket, after crew:mr-review clears and before the orchestrator finalizes. Harvests the ADVISORY, non-blocking findings that crew:reviewer and crew:mr-review left on the MR (MINOR, advisory MAJOR, explicitly out-of-scope-of-this-MR) and files them as GitHub issues — one per distinct actionable finding — labeled review-followup (read from config; NEVER agent-ready, so the loop never auto-picks them) and BLOCKED BY the source ticket — the issue this MR Closes — via a GitHub blocked-by dependency (plus the board's blocked status), so GitHub auto-unblocks the follow-up when the MR merges and closes that issue, deduped against existing open review-followup issues, each with a backlink to the source MR + comment, file refs, and severity, and assigned to the configured findings-assignee (if set). Posts a short summary comment listing the filed issues. Changes no code. Project conventions (the review-followup label, board statuses) are read from CLAUDE.md at runtime.

Dispatch to build a ticket. Normal mode (first dispatch for an issue): the GitHub issue IS the spec — read it + CLAUDE.md + the codebase, implement step-by-step, write unit/integration tests, run the project checks, then create the branch, open the DRAFT MR (Closes #<issue>), and push. Fix mode (re-dispatched by run after a reviewer FAIL): read the reviewer's findings from the MR comment and fix ONLY what was flagged on the same branch — do not re-implement. Output is always an MR comment, never an on-disk report. Project conventions are read from CLAUDE.md at runtime.

Independent, blind code-smell review of an MR diff. Dispatch as the LAST gate, after crew:reviewer returns PASS, before the orchestrator flips the MR to ready-for-review. Reads ONLY the diff and the surrounding code — never the other agents' MR comments, the reviewer's verdict, or the progress_log — so its judgment is genuinely independent. Focuses on maintainability concerns a correctness review misses: duplication, dead code, leaky abstractions, naming, complexity, weak tests. Posts findings as an MR comment. A CRITICAL smell may bounce the MR back to implementation once.

Dispatch after crew:implementation has opened the draft MR and pushed its first commits. Verifies the implementation against the issue's acceptance criteria: routes each criterion to the right venue (Gherkin scenario / lint rule / unit test / impl check-result), extends the ONE whole-app e2e/gherkin suite so the ticket's behavior is proven inside existing journeys, runs the suite, commits the test code to the MR branch, and posts a coverage map + pass/fail-per-criterion MR comment. Re-dispatched after a crew:implementation fix round to re-verify. Owns the e2e tree; never opens the MR and never fixes implementation bugs.

Adversarial correctness reviewer. Dispatch after crew:qa, on the same per-ticket worktree, once an implementation exists on the MR branch and qa has posted its coverage. Distrusts every prior phase by design: reads the GitHub issue (the spec), the actual diff, the implementation, and the qa results, then renders a binary PASS/FAIL verdict with specific, severity-tagged, actionable issues. On FAIL the orchestrator routes back to crew:implementation in fix mode; the reviewer itself changes no code. Re-dispatched for each review round of a ticket.

Onboards a project for the crew loop — scans the structure, detects and validates the test / lint / build / e2e and app-start commands, detects the GitHub remote and optional Projects board, offers a one-time gated bare-clone worktree migration, advises on setup gaps, and writes a Workflow Config block into CLAUDE.md (commands, ticket-source label, board columns incl. a needs-human/blocked and a done column, the priority field, the review-followup + merge-approval labels and merge method, branch convention, worktree layout, and the stack-run config: start command / readiness check / per-ticket isolation, and an optional GitHub App crew-identity — installing + testing its token helper) that every other component reads at runtime. Use when the user invokes /crew:adjust.

Merge-queue auto-approver. Human-invoked, NOT part of the autonomous loop. Sweeps the open ready-for-review MRs that carry no merge-approval label, judges each on its actual diff (origin-agnostic), and adds the merge-approval label (default `approved`) only to the provably-low-risk ones — small, green-CI, no open reviewer/mr-review MAJOR/CRITICAL, confined to low-impact surfaces (docs, comments, tests, i18n strings, config, dead-code/type-only cleanups) and never user journeys (routes, components, API handlers, DB migrations, auth/payments, business logic). Conservative by construction: anything large, red, conflicting, journey-touching, or uncertain is LEFT for the human. It only labels (with an audit comment) and verifies the write landed — `/crew:merge` still does the merging. Reads CLAUDE.md ## Workflow Config, keeps the sandbox on. Use when the user invokes /crew:approve.

Senior-advisor codebase audit that feeds the crew backlog. Surveys the whole project read-only across nine dimensions (correctness, security, performance, tests, tech debt, dependencies, DX, docs, direction), vets and prioritises findings by leverage, then — by default — files the ones you select as GitHub backlog issues in the crew ticket contract (Context / Out of scope / Acceptance criteria), labeled agent-review (NEVER agent-ready) for a human to promote during planning. Never modifies source, never opens MRs, never executes — /crew:run owns execution. Project conventions and labels are read from CLAUDE.md at runtime. Use when the user invokes /crew:improve.

Team-lead merge-queue skill. Sweeps the open MRs a human has green-lit — a GitHub Approval OR the merge-approval label (default `approved`) — oldest-first: squash-merges every green-lit MR that is CI-green and conflict-free (auto-closing its issue via Closes #N, moving the board card to Done, deleting the branch), and resolves EVERY conflicting MR back to mergeable regardless of label (dispatching crew:implementation for real conflicts) so the queue stays clean — merging the green-lit ones, leaving the rest conflict-free for a human to green-light. No one-blocker-per-invocation cap; each fix is bounded by the shared 3-round cap. Stays thin (dispatches subagents, never writes code itself), reads CLAUDE.md ## Workflow Config, keeps the sandbox on, honors the §4.13 ownership claim, and verifies every GitHub write landed. When the project opts in (`auto-merge: low-risk` config or the `--auto-approve` flag), it also auto-merges the provably-low-risk slice of unlabeled, fully-reviewed MRs — judged on the diff by the same conservative bar as /crew:approve (small, green-CI, no open MAJOR/CRITICAL, low-impact surfaces only; default DEFER) — so trivial follow-ups (docs, tests, dead-code) land without a human label; off by default. Use when the user invokes /crew:merge.

Autonomous orchestrator loop. Pulls the next agent-ready GitHub issue, processes it end-to-end in a per-ticket git worktree by dispatching crew:implementation → crew:qa → crew:reviewer (with a capped fix loop) → crew:mr-review → crew:findings (files leftover advisory findings as review-followup-labeled, MR-blocked follow-up tickets), then flips the draft MR to ready-for-review and moves to the next issue. GitHub is the source of truth: each agent commits and comments on the MR; the loop never does domain work and never waits for a human merge. Project conventions are read from CLAUDE.md ## Workflow Config at runtime. Use when the user invokes /crew:run.