loomwright

Standard output format for all agents. Use to ensure consistent, structured communication with clear sections.

Agent Teams patterns for parallel coordination using Claude Code native teams feature. Use when coordinating multiple agents on cross-layer changes, competing hypotheses, or research tasks.

Background dispatch patterns, non-blocking polling, parallelism decisions, and git worktree lifecycle. Use when running parallel workers in Supervisor workflows. Part 2 — the Supervisor Phase 4 FINALIZE protocol (pre-merge safety gate, sequential merge, worktree cleanup, commit/push/PR creation, PR-base self-verify), the verbatim Subagent Spawn Contracts, and the worktree-lifecycle phase sequence, moved from agents/supervisor.md.

Protocol authority for `/automate` — the generic automation engine that converts ANY source (a prompt via /product-owner, a requirements folder, or a backlog-doc) into a FULL Queue with a per-run processing cap inside ONE markdown run file (`.supervisor/automate/<run_id>.md` — the contract, dashboard, and resume state), then drives each Queue item through the per-item loop (`/autonomous --single-iteration` → owned inline `/review-pr --until-mergeable` → trusted-merge-or-park → pull main → check off + append `## Progress`). Smart resume = glob `*.md` for not-done + reconcile-vs-ground-truth. Use when implementing or invoking `/automate`.

Outer-loop protocol for `/autonomous` — v14 continuous (multi-iteration default with stacked branches), single-iteration opt-in, EVALUATE PR-base verification + Signal-1 stacked rubric gate + no-rubric gate, --notify gate webhooks via send-webhook.sh, CI / non-TTY fail-closed protection, and the AUTONOMOUS_RUN summary format. Use when implementing or invoking the `/autonomous` command.

Beads CLI commands for task lifecycle management. Use when creating tasks, updating status, or syncing team visibility.

Read-on-demand context enrichment ladder — orientation memos → owned repo-map → nothing (plain grep/read). Advisory only — a missing/broken/stale store or map silently falls back down the ladder; NEVER blocks a run and NEVER changes a decision. Deliberately NOT preloaded into any agent frontmatter; agents read it at context-setup / analysis points.

Multi-mind brainstorming framework with 5 expert lenses, structured debate, and idea scoring. Use when generating options, exploring solutions, or evaluating strategic directions.

CI/CD patterns for GitHub Actions. Covers pipeline stages, caching, deployment strategies, secrets management, and quality gates. Use when setting up or reviewing CI/CD workflows.

Validate CLAUDE.md freshness and pattern accuracy. Use when establishing project context to ensure documented patterns are current.

Standard procedure for agents to establish project context. Use when starting a new task to load CLAUDE.md, Beads state, and git history.

Summarize agent outputs to minimize context usage. Use after each subagent call to compress results.

On-demand external library documentation lookup with strict token budgets. Use when CLAUDE.md knowledge is insufficient for specific library features or when you need detailed API documentation.

Template for documenting domain expertise in CLAUDE.md. Use when setting up domain context for a project.

Error handling patterns across the stack. Covers custom error hierarchies, NestJS exception filters, React error boundaries, retry with backoff, and graceful degradation. Use when implementing or reviewing error handling.

Enforce design system components, accessibility (WCAG 2.1 AA), responsive design, and component reusability. Use when reviewing frontend/UI code in React, Vue, Angular, or Svelte projects.

When to use Anthropic's Memory Tool vs the plugin's file-based agent memory; both are persistent stores — choice depends on runtime (direct API vs Claude Code plugin) and access model

Monitoring and observability patterns. Covers structured logging, distributed tracing with OpenTelemetry, metrics collection, and alerting thresholds. Use when implementing or reviewing observability.

Prioritize features into MVP, Phase 2, and Nice-to-have. Use when scoping work or making trade-off decisions.

Identify and propose new patterns for CLAUDE.md updates. Use when discovering reusable code patterns that appear 3+ times, or when reviewing code for pattern consistency.

Write, run, and debug Playwright E2E tests using CLI. Covers test authoring, execution, debugging, reports, and CI integration.

Inline, READ-ONLY post-hoc analysis protocol for `/pr-postmortem <pr-url>` — gathers an existing PR's review/churn signals via pr-postmortem-gather.sh, categorizes each review round into one of 6 root-cause classes, attributes it to a flow stage, prints a human-readable root-cause report, and appends one fail-safe POSTMORTEM_RESULT trend line to .supervisor/postmortem/results.jsonl. Use when implementing or invoking the `/pr-postmortem` command.

Protocol authority for Supervisor Phase 1.5 PRE-FLIGHT SYNC — CLEAR/OVERLAP/SUPERSEDED remote-state reconciliation of the requested work against recent origin/$BASE_BRANCH commits and open PRs, before Phase 2 PLAN spawns anything. Read on demand at phase entry, deliberately not preloaded.

Understand business problems before defining solutions. Use at start of feature work or when requirements are unclear.

Quality gates, gap analysis tiers, and enforcement checklist for QA Executor. Contains Phase 4.5 (4-tier gap analysis), Phase 4.6 (dry-run), Phase 4.7 (13 quality gates with PASS CRITERIA), and quality checklist.

Session management for large-app QA testing. Covers plan.json schema, coverage.json schema, scope clustering, and multi-session workflow patterns.

Risk-based QA strategy framework, structured output formats, debate loop protocol, and discovery engine specification. Foundation skill for QA Strategist and QA Executor agents.

Test Pattern Library, assertion rules, and generation rules for QA Executor. Contains all test generation patterns including signal-driven security, UI interaction, and API test patterns.

Pre-task and post-task quality gates extracted from AGENT_GUIDELINES.md. Use when starting implementation, during development, or before completing code review.

Shared loop contract for the standalone PR review-and-heal workflow (`/review-pr <pr-url>` + `loomwright:review-pr-runner`). Single source of truth for the bounded review→fix→re-review loop, PR-URL→branch resolution, the REVIEW_HEAL_RESULT block, and the pinned canonical names consumed by the dispatcher script, the runner agent, and the autonomous EVALUATE step. Use when implementing or invoking standalone PR review-and-heal.

Supervisor Phase 4.5 protocol authority. Part 1 — advisory-only machinery (pre-review enrichments, System Twin conformance/benchmark/ground-truth, contract-builder WRITE path, delta line, hard-signal dual emission; never changes heal_decision or blocks the PR). Part 2 — the full Phase 4.5 SELF_HEAL loop protocol (on-entry actions, base-mismatch cleanup, bounded review-and-fix loop, rubric grading, red-team lens, completion-tail procedure), Read on demand at Phase 4.5 entry, deliberately not preloaded.

Module protocol authority for the /setup umbrella command — the check/report/offer/apply/verify contract, the module registry, the jq deep-merge rules for ~/.claude/settings.json (backup-first, abort-on-parse-failure, idempotent), the .gitignore managed-block rules for the memory module, the statusline module's user-scope write class, the ui module's local-serve write class, the seeded-vs-learned provenance contract for the rules module, and the OTLP smoke-test recipe. Use when running /setup or modifying any setup module.

State file schema, .supervisor/ directory setup, checkpoint and resume protocols. Use when managing Supervisor session state across phases.

Protocol authority for Supervisor Phase 0 INIT configuration resolution — flag/defaults table, cost-profile resolution, base-branch handling, environment detection. Read on demand at phase entry, deliberately not preloaded.

Pre-flight checklist, Supervisor-Ready Brief format, jobs folder convention, and failure prevention. Use before launching autonomous workflows or when diagnosing Supervisor failures.

Opt-in GitHub Issues telemetry for loomwright. Enables longitudinal analysis of agent quality by emitting structured, scored, redacted issues from `SUPERVISOR_RESULT`, `CODE_REVIEW_RESULT`, and `QA_RESULT` blocks. Use when a project wants to feed run outcomes back into the maintainer repo (or its own analytics repo) without standing up a backend.

Unit testing patterns with Jest and Vitest. Covers AAA pattern, mocking, coverage thresholds, fixtures, and snapshot testing. Use when writing or reviewing tests.

Write clear user stories with testable acceptance criteria. Use for requirements definition and feature scoping.

Protocol authority for `/verify <ticket>` and the QA Executor's `--verify <run_dir>` mode — AC extraction, AC → Playwright spec derivation, the four observation-derived verdicts (PASS / FAIL / BLOCKED / NOT_VERIFIABLE), the V7 mutation carve-out, evidence-per-AC, and budget. Read on demand at mode entry, deliberately not preloaded.

Patterns for managing autonomous workflows including context, checkpoints, parallel execution, and permissions. Use when orchestrating multi-stage agent workflows.