Claude Coordinator
A structured orchestration system for Claude Code that plans, delegates, reviews, and learns.
What is this?
Claude Coordinator is a set of Claude Code agent definitions that turn Claude into a structured project manager. Instead of having a single Claude session try to do everything, this system uses a team of specialized agents working under a pure-delegation control plane:
- Coordinator — The control plane. Plans work, maintains state, delegates everything, and writes context for the next session. Its only tool is
Agent. It does not read or write files directly.
- Briefer (Haiku) — Reads context files and returns structured briefings. Cheap, fast.
- Planner (Sonnet) — Produces task breakdowns with behavioral-test specifications.
- Worker (Sonnet) — Strict TDD implementation for
feature / bugfix tasks. Produces an auditable red → green → regression commit trail.
- Worker-Refactor (Sonnet) — Behavior-preserving refactors. Existing tests must pass before and after.
- Worker-Test (Sonnet) — Adds tests to existing untested code. Mutation-checks its own tests.
- Worker-Investigation (Sonnet) — Read-only research. Returns structured findings; writes no code.
- Reviewer (Opus) — Read-only code reviewer with severity ratings, plus an external GPT-5.4 review pass.
- UI / UX / System Testers — Validate the built product visually, experientially, and functionally.
- Intent-Validator (Opus) — Final quality gate. Compares what was built against the user's original intent. Runs foreground so it can ask the user questions.
- Learning-Extractor (Opus) — Analyzes task outputs, review findings, AND sub-agent JSONL transcripts to surface code learnings and process learnings (retries, dead ends, scope drift).
- Scribe (Haiku) — Writes all state files. Cheap, fast, precise.
The coordinator maintains state across sessions using two mechanisms: machine-readable files in .coord/ and human-readable files in docs/. This means a project can be picked up exactly where it left off, even after days away.
How it works
The coordinator operates as an explicit 10-phase state machine:
startup → intake → plan → delegate → integrate → review → test → promote-learnings → validate → close
| Phase | What happens |
|---|
| startup | Briefer reads .coord/ and docs/ to orient the session. Fresh sessions add .coord/ to .gitignore via scribe. |
| intake | Capture the user's request as a command-intent doc (verbatim words, interpreted intent, success criteria). User confirms before proceeding. |
| plan | Planner produces a task breakdown with behavioral test specs. User approves before delegation. |
| delegate | Coordinator routes each task to the right worker (worker, worker-refactor, worker-test, worker-investigation) and spawns them in isolated worktrees. No file overlap between concurrent workers. |
| integrate | Collect worker results. Validate output contracts, including audit-trail commit hashes for TDD task types. Reject and re-delegate anything missing the required evidence. |
| review | Reviewer (Opus + GPT-5.4) checks risky changes. Critical/high findings block progress. |
| test | UI tester, UX tester, and system tester validate the built product. UI/UX only run for user-facing changes. |
| promote-learnings | Learning-extractor analyzes task artifacts AND sub-agent transcripts to surface code + process learnings. Scribe records accepted candidates to .coord/learning-inbox.jsonl. |
| validate | Intent-validator (foreground) compares the work against the original intent doc. May ask the user clarifying questions. |
| close | Scribe updates the task ledger and writes a context packet for the next session. Coordinator summarizes for the user. |
The coordinator can revisit earlier phases when new information invalidates the current plan.
State Layers
Truth is maintained in three places, each with a different purpose:
1. .coord/ — Machine Operational State
Working memory. Ephemeral, structured, machine-readable.
.coord/
├── task-ledger.json # All tasks: pending, in-flight, blocked, done, failed
├── learning-inbox.jsonl # Candidate learnings (JSON lines) awaiting promotion
├── context-packet.md # Compressed session context for continuity
├── tasks/
│ └── TASK-XXX.json # Per-task artifacts: output, files changed, test results
├── reviews/
│ └── REVIEW-XXX.json # Per-review artifacts: findings, severity, recommendations
└── milestones/
└── M-XXX.json # Milestone summaries: scope, tasks, learnings promoted
2. GitHub Issues — Durable Public Tracker
