multicrew

Install a battle-tested multi-agent dev-squad configuration onto a new Multica project (installer, NOT a generator). Use when the user wants to set up, provision, bootstrap, or install a Multica dev squad on a repo — drives phases P0-P7 with hard gates.

Use when reviewing a diff for cross-file coherence, coupling, or abstraction-boundary regressions — the architecture lens. Supplies precise vocabulary (depth/seam/locality), the deletion test, the adapter-count seam rule, and a code-graph-first caller-impact recipe with a graph-free fallback. Do not use for writing code, style policing, or other reviewers' lenses.

Use when {{QA_SEAT_NAME}} must drive a real browser against the target repo's web surface ({{UI_SURFACE}}) with the playwright MCP — supplies hybrid DOM+screenshot grounding, the tiered-locator order, a self-heal loop that ESCALATES instead of weakening assertions, a step cap, and the headless no-clarifying-question contract. Do NOT use for native-channel rendering (that is un-automatable), for API/contract testing (qa.md HTTP oracles), or for scenario-shape / non-determinism metrics (your eager QA scenario contract + eval-harness own those).

Use while implementing a sub-issue at a judgment point — a repeatedly failing build/test, a needed dependency, a wrong-looking AC (HALT vs push through), or metabolizing the three-panel review. Not for reviewing others' code or spec/planning.

Use when reviewing a diff that changes control flow, a state machine, auth/permission checks, async or concurrent code, or input boundaries, and you need exhaustive enumeration of unhandled paths. A method (mechanical path-walk), not an attitude. Do not use for writing code or for style/quality opinions.

Use in non-deterministic testing — Mode-2 real-model or Mode-3 real-platform — when exact-string asserts would flake. Supplies the determinism split (mock = assert exact, real-model = assert a predicate over k runs), pass@k vs pass^k reliability metrics, the grader taxonomy (code / model / human), and baseline-SHA regression framing. Do not use for deterministic mock-LLM scenarios (assert exactly) or for scenario generation (your eager QA scenario contract owns that).

Use when reviewing a diff that touches LLM prompt construction, tool/function-calling, RAG/retrieval, rendering of model output, AI credentials, or any path where model or tool output flows onward — the LLM-gateway attack class. Supplies the sink patterns and severity/FP rules for that lens. Do not use for generic web vulns the base review already covers, or for writing code.

Use when a sub-issue adds or changes an MCP tool/server surface — the agent-centric tool-design rules plus the Zod-strict validation and tool-annotation checklist. Always defer current @modelcontextprotocol/sdk signatures to context7 (don't trust memory). Do not use for non-MCP code, or for merely consuming an existing MCP tool.

Use when a diff adds or changes a data/state migration — reviewing ({{CONTRACT_SEAT_NAME}}) or authoring ({{BUILDER_SEAT_NAME}}) one. Classify the surface first — the {{STACK_DESCRIPTION}} scan tells you which of state/session migrations, relational-DB DDL, or application-level data backfills actually exist here; apply only the iron rules for the surface(s) present. Do not use for ordinary feature diffs with no migration.

Use when the diff plus the modules it touches exceed what you can load in one read — your context-budgeting aid as the architecture reviewer. Plain `npx repomix` CLI that complements this seat's code-graph tool (code-graph tool = targeted caller-impact; repomix = whole-repo packing + token budgeting); `--token-count-tree` to size the surface, `--include` to scope to the diff's blast radius, then carry what you actually packed into an honest File manifest. Do not use for writing code, packing third-party libraries, or as a separate finding lens.

Use when {{MENTOR_SEAT_NAME}} runs a retrospective and wants a QUANTITATIVE read on the reviewer bench — per-reviewer addressed-rate, a separate refuted-with-named-evidence count, and signal-to-noise — computed over the parent + sub-issue verdict comments and the Lead's review_metrics_pr{N} records. Pairs with skill-harvest. Read-only measurement; never changes a reviewer seat; suppresses drift flags on tiny samples.

Use when {{LEAD_SEAT_NAME}} finds a summoned reviewer dead or churned at the synthesis barrier (a model/availability/runtime error, or the reconciler flags it), OR a subtask PR was already merged before review — the degrade-vs-collapse-vs-retroactive recovery rules. Not for a healthy panel.

Use when you are about to fix a bug, a failing test, or unexpected behavior in the target repo. Forces a deterministic repro and ranked falsifiable hypotheses BEFORE any code change. Use this on {{BUILDER_SEAT_NAME}} and {{QA_SEAT_NAME}} debugging work. Do not use for green-field feature authoring with no observed defect.

Use when a sub-issue is a refactor/restructure/extract/rename — a tool-agnostic safety discipline: map the blast radius with grep, gate on test coverage (NONE blocks an aggressive refactor — add characterization tests first), change in small steps, verify after each ({{TYPECHECK_CMD}} + {{TEST_CMD}} + {{BUILD_VERIFY_CMDS}}), never commit a broken tree. Do not use for new features or any behavior change.

Use when composing a durable lesson to save: the what-to-save bar (symptom/root-cause/fix-loc/guard) + confidence/freshness/contradiction rules. The recall-before/save-after loop is eager in the constitution MEMORY section, not here.

Use when {{MENTOR_SEAT_NAME}} runs a retrospective on already-executed tasks ({{LEAD_SEAT_NAME}} delegates it a finished parent at delivery, or a "harvest from done issues" window). Four outputs: PROPOSE new skills, evaluate + PROPOSE improvements to skills the run actually used, write durable LESSONS to the memory backend ({{MEMORY_BACKEND_URL_API}}), and hand the human a review packet. Supplies the look-back read-path (incl. loaded-skill detection), the Googleable/specific/effort filter, content-aware create-safe dedup, the rejection ledger, a secret screen, the efficacy-evaluation method, and the lesson-write + handoff rules. Skills are proposed, then applied only once a human member authorizes on the harvest issue (attach + overwrite are the gated live actions, §9); lessons are the always-allowed direct write. Not for per-task memory saves (that is self-improve) and not for execution work.

The measurement telemetry SCHEMAS the Lead records, and their recording cadence — validation_metrics + the durable validation_ledger, review_metrics_pr{N} + cost_pr{N}, and merge_gate_pr{N}. Use when a seat needs the exact field shape to WRITE a measurement record, or when a retrospective needs to know WHERE each stream lives and WHO wrote it. All three streams GATE NOTHING at record time (no threshold, no memory_save on write). INTERPRETATION — the per-seat reviewer scorecard, drift flags, and the validation-phase sunset aggregates — lives in review-quality-metrics, which reads these streams; this skill holds only the schemas + the recording cadence and does not duplicate that method.

Use when {{LEAD_SEAT_NAME}} must (1) check a backlog doc is execution-ready and (2) decompose it into turnkey, risk-routed sub-issues for {{BUILDER_SEAT_NAME}}. Supplies the Definition-of-Ready gate, the tracer-bullet vertical-slice method, and the writing-plans bar (exact files + change-per-file + AC-ids + verify command, zero placeholders) that makes a focused Builder safe (it removes judgment from the slice and keeps each slice reviewable — independent of the Builder's model strength). Do not use for product discovery, tech design, or implementation — those happen upstream in native Claude Code.