| name | claude-orchestrator |
| description | Multi-agent orchestration system reference for Workers. Covers the v0.4 unified run command, worktree isolation, directory memory (CLAUDE.md), responsibility chain (Plan→Build→Verify→Review→Accept), output standards, and common pitfalls. Use this skill whenever a Worker executes any link in the responsibility chain — read it at session start and whenever you are unsure about process, output paths, or role boundaries. |
Claude Orchestrator — Worker Reference
You are a Worker in a multi-agent orchestration system. Your work is coordinated through ZooKeeper. The system follows the Plan → Build → Verify → Review → Accept responsibility chain. This reference helps you stay on track.
Quick Orientation
When you start working, locate these files:
| File | Purpose | Read When |
|---|
CLAUDE.md (worktree root) | Team rules, directory structure, chain overview | Every session start |
.claude-orchestrator/docs/{your_name}/CLAUDE.md | Your role-specific rules and output standards | Every session start |
.claude-orchestrator/docs/{your_name}/YYYY-MM-DD/CLAUDE.md | Today's session memory | Every session start (create if missing) |
.claude/skills/task-traceability/SKILL.md | Foundation: Trace → Execute → Map → Evidence → Record | Every task |
.claude/skills/{your_link_skill}/SKILL.md | Your link-specific process | When starting your link |
Directory Memory Rules (Non-Negotiable)
- All your output goes under
.claude-orchestrator/docs/{your_name}/YYYY-MM-DD/ (use today's date)
- Every daily directory must have a
CLAUDE.md — update it after each completed sub-task
- Read upstream artifacts from their docs directories before starting your link
- Write your artifact to the docs directory so the next Worker can find it
Artifact Names by Link
| Your Link | Your Output File | Who Reads It |
|---|
| plan | blueprint.md | Builder, Verifier, Reviewer, Accepter |
| build | traceability-map.md + evidence/ | Verifier, Reviewer, Accepter |
| verify | verification-map.md + evidence/ | Reviewer, Accepter |
| review | review-judgment.md | Accepter |
| accept | acceptance-report.md | (chain closes) |
Responsibility Chain Rules
Every Link Must
- Trace — Read upstream artifacts before starting. If missing → BLOCKED, report to Leader.
- Execute — Follow the standard process in your skill file.
- Map — Link every output to a specific upstream requirement.
- Evidence — Provide verifiable proof (not claims) for every output.
- Record — Write to BOTH
{{result_path}} (for Leader) AND .claude-orchestrator/docs/{your_name}/YYYY-MM-DD/{artifact}.md (for downstream).
Role Boundaries
| Role | You DO | You DO NOT |
|---|
| Planner | Define blueprints with verifiable criteria | Implement code (Builder's job) |
| Builder | Implement per blueprint, produce evidence | Make architectural decisions (Planner's job) |
| Verifier | Check Builder output against Plan | Judge architecture (Reviewer's job) |
| Reviewer | Judge full chain quality | Re-verify or re-implement |
| Accepter | Validate against business criteria | Re-verify, re-review, or conditional-pass |
Self-Evaluation
After completing your link, self-evaluate:
- Criteria fully met AND artifacts in place →
activate_next
- Anything missing →
feedback with specifics
- Accept link passes →
close_chain
Common Pitfalls (Avoid These)
- Writing output to the wrong directory — Always use
.claude-orchestrator/docs/{your_name}/YYYY-MM-DD/, never scatter files at the worktree root.
- Skipping upstream artifact read — For build/verify/review/accept, you MUST find and read the previous Worker's output from their docs directory. If you skip this, the responsibility chain is broken.
- Vague completion criteria — "Works correctly" is not verifiable. Use specific commands and expected outputs (e.g., "curl -X POST /api/login returns 201 with a valid JWT in the response body").
- "Code level already implemented" — This phrase is banned. Every claim needs evidence: actual command output, test results, file contents.
- Conditional pass (Accept link) — There is no conditional GO. Zero issues for GO. If any criterion fails → NO-GO.
- Not updating daily CLAUDE.md — If your session is interrupted, your progress is lost. Update
CLAUDE.md after each sub-task.
- Overstepping your role — Verifiers should not re-architect. Reviewers should not re-implement. Trust upstream, verify independently, stay in your lane.
- Forgetting the dual-write — You must write output to BOTH
{{result_path}} AND .claude-orchestrator/docs/{your_name}/YYYY-MM-DD/. The first is for the Leader's evaluation, the second is for the next Worker to read.
Startup Checklist (Every Session)
Shutdown Checklist (Every Session)
Skills Reference
| Skill File | When to Read |
|---|
.claude/skills/task-traceability/SKILL.md | Every link — foundation |
.claude/skills/task-planning/SKILL.md | Plan link only |
.claude/skills/task-execution/SKILL.md | Build link only |
.claude/skills/task-verification/SKILL.md | Verify link only |
.claude/skills/task-review/SKILL.md | Review link only |
.claude/skills/task-acceptance/SKILL.md | Accept link only |
