coordinator

Coordinator

Launch a structured orchestration session using a 10-phase state machine. The coordinator delegates ALL work to specialized subagents — it never reads or writes files directly.

Usage

/coordinator [task description]

If invoked with no arguments, the coordinator enters the intake phase and asks what you want to work on.

Agent Team

Agent	Model	Tools	Purpose
briefer	Haiku	Read, Glob, Grep	Reads context files, returns structured briefings
planner	Sonnet	Read, Glob, Grep, Agent	Produces task breakdowns with behavioral test specs
worker	Sonnet	Full toolset	Strict TDD implementation (feature/bugfix). Red → green → regression commits.
worker-refactor	Sonnet	Full toolset	Behavior-preserving refactors. Existing tests must pass before and after.
worker-test	Sonnet	Full toolset	Adds tests to existing untested code. Mutation-checks its own tests.
worker-investigation	Sonnet	Read, Bash, Glob, Grep	Read-only research. Returns structured findings.
reviewer	Opus	Read, Bash, Glob, Grep	Code review with severity ratings.
ui-tester	Sonnet	Read, Bash, Glob, Grep	Visual quality inspector via browser automation.
ux-tester	Opus	Read, Bash, Glob, Grep	Usability evaluator via browser automation.
system-tester	Sonnet	Read, Bash, Glob, Grep	Full test suites, regression coverage, integration.
intent-validator	Opus	Read, Glob, Grep	Validates work matches user's original intent. Foreground.
learning-extractor	Opus	Read, Glob, Grep, Bash	Surfaces learnings from task artifacts and transcripts.
scribe	Haiku	Read, Write	Writes all state files.
issue-implementer	Opus	Agent	Autonomous groomed-backlog → tested-PR loop. Claims status:ready issues, implements each via TDD workers in dedicated worktrees, opens a PR, then loops to the next ready issue. Started via --agent issue-implementer; runs unattended after invocation. See docs/issue-implementer-runbook.md for preconditions, label setup, stop conditions, and first-run guidance.
issue-groomer	Opus	Agent	Autonomous backlog groomer. Claims open status-less issues, researches each deeply (codebase + product docs + external APIs), fills the target repo's issue template exhaustively with behavior contracts, assumptions, and alternatives, then applies status:ready (clear implementation path) or status:blocked (exhaustively groomed, genuine product/path decision needed). Loops; in WATCH mode re-checks every poll_interval for new issues. Started via --agent issue-groomer; runs unattended after invocation. See docs/issue-groomer-runbook.md for preconditions, label setup, stop conditions, and first-run guidance.
researcher	Sonnet	Read, Bash, Glob, Grep, WebFetch, WebSearch, Write	Deep external research on libraries, frameworks, SDKs, APIs, CLI tools, or cloud services. Reads official docs, source code, examples, runtime probes. Returns structured findings — capability map, mental model, failure modes, expectation gaps. No code commits.
distiller	Opus	Read, Glob, Grep, Write	Compresses raw evidence (research notes, command/error logs, debug transcripts, code samples) into citable knowledge artifacts: gotchas, patterns, anti-patterns, playbooks, before-you-build guides, decision records. Every claim must cite evidence. Returns a distillation manifest.
knowledge-packager	Sonnet	Read, Glob, Grep, Write	Assembles distilled knowledge into a navigable, self-contained artifact (skill pack, runbook, onboarding guide, internal wiki). Writes a cross-linked directory with index, quickstart, lessons, labs, recipes, checklists, troubleshooting, and agent instructions. Returns a navigability-verified manifest.
evidence-auditor	Opus	Read, Bash, Glob, Grep	Read-only auditor of claim-vs-evidence integrity. Verifies documented claims have supporting evidence; evaluates named quality gates (research-quality, distillation-quality, skill-pack-quality). Returns per-gate verdicts and unverified-claim findings with severity ratings. No file modifications.

State Machine

startup → intake → plan → delegate → integrate → review → test → promote-learnings → validate → close

Hard Gates (require user approval)

• intake → plan: User must confirm the intent document.
• plan → delegate: User must approve the plan before work begins.

Worker Selection

Task type	Agent	TDD enforced?
feature	worker	Yes — red/green/regression commits
bugfix	worker	Yes — red commit reproduces bug
refactor	worker-refactor	No — existing tests must pass before & after
test	worker-test	No — tests must be meaningful (mutation-checked)
investigation	worker-investigation	N/A (no code written)

Task Contract

Every delegated task includes:

{
  "title": "Clear task name",
  "type": "feature | bugfix | refactor | test | investigation",
  "scope": "Precise description",
  "allowed_files": ["list of files/directories"],
  "forbidden_files": ["files NOT to touch"],
  "dependencies": ["task IDs"],
  "behavioral_tests": ["When [condition], then [result]"],
  "regression_test_requirements": "What regression tests must exist"
}

State Layers

.coord/ — Machine State (ephemeral, gitignored)

File	Purpose
task-ledger.json	All tasks and statuses
learning-inbox.jsonl	Candidate learnings
context-packet.md	Session continuity
tasks/TASK-XXX.json	Per-task artifacts
reviews/REVIEW-XXX.json	Review artifacts

docs/ — Durable Human-Readable Memory

File	Purpose
docs/context/command-intent.md	Captured user intent
docs/context/repo-practices.md	Conventions and patterns
docs/context/known-issues.md	Known problems
docs/plans/active-plan.md	Current plan

Instructions

When this skill is invoked, load and follow the full coordinator agent definition at agents/coordinator.md in this skill's directory. The agent definitions for all subagents are in agents/. Output schemas are in schemas/.

References

Reference	Purpose
agents/coordinator.md	Full coordinator system prompt
agents/	All agent definitions
schemas/	JSON Schema contracts for agent outputs
coordinator-settings.json	Permission deny list
templates/.worktreeinclude	Files to copy into worktrees