coordinator-claude

coordinator-claude

A plugin system that turns Claude Code into a structured engineering team — you're PM, Claude's EM.

• 8 plugins, 24 agents, 37 skills — a coherent orchestration stack built on every Claude Code extension primitive (hooks, subagents, skills, commands, Agent Teams, MCP)
• Agent Teams for research and planning — tiered pipelines (Haiku scouts → Sonnet specialists → Opus synthesizer) that run autonomously; staff sessions where persona-based engineers debate and converge on plans without intervention. Pipeline design is research-backed — derived from published guidance (OpenAI, Perplexity, Google, Anthropic, Stanford STORM) and validated through controlled experiments
• Prospective handoff artifacts — structured baton-passing before compaction fires, not retrospective summarization after. Research shows this beats automatic summarization for chained agent work
• Inverted capability delegation — the coordinator sees ~8 thin tools; domain agents access 40+ via proxy. The orchestrator is intentionally less capable than its delegates, saving ~40K tokens for judgment instead of tool schemas
• Sequential persona-based review — domain expert first, all fixes applied, then generalist reviews a clean artifact. Research supports both the persona mechanism and multi-agent review gains
• 6-layer project knowledge — structure, architecture, activity, temporal, intent, state — none bulk-loaded, all maintained by an 11-phase doc pipeline that fights staleness automatically

Quick Start

git clone https://github.com/oduffy-delphi/coordinator-claude.git
cd coordinator-claude
bash setup/install.sh

Restart Claude Code, then run /session-start. See docs/getting-started.md for the full walkthrough.

How It Works

You sit in the captain's chair:

Role	Who	Responsibility
PM (Captain)	You	Vision, priorities, judgment calls, design approval
EM (First Officer)	Claude (main session)	Orchestration, delegation, spec compliance, pipeline flow
Executor	Claude (subagent, Sonnet)	Faithful implementation of well-specified work
Reviewers	Claude (subagents, Opus)	Code quality, UX, domain expertise — each with a rich behavioral profile

A typical feature flows through:

Brainstorm  →  Plan  →  Execute  →  Review  →  Ship
   (skill)    (skill     (executor    (domain       (finish
               or staff   agents)     → generalist   branch)
               session)               sequential)

See docs/architecture.md for the full model — agent roles, pipeline checkpoints, review routing, and the design philosophy behind it.

Agent Teams for research and planning — not just "build together"

Claude Code's Agent Teams (still experimental) enables multiple Claude sessions that communicate via messaging and coordinate via shared task lists. Most early adopters use it for collaborative coding. This system uses it differently: for structured research and multi-perspective planning. Four research pipelines (internet, repository, structured, NotebookLM media) follow a tiered pattern — Haiku scouts gather sources cheaply, Sonnet specialists analyze and cross-pollinate findings via messaging, an Opus sweep agent checks coverage adversarially and fills gaps. Internet research (Pipeline A v2.2) adds iterative deepening — a second, smaller team is dispatched to fill significant gaps when warranted. Repository research (Pipeline B) supports --deeper (dependency-weighted repomap) and --deepest (architecture atlas generation) modes. Staff sessions use the same infrastructure for planning: persona-based debaters form independent positions, challenge each other, and a synthesizer cross-references into a consensus plan. The coordinator scopes the work, spawns the team, and is freed — the team runs autonomously.

Prospective handoff artifacts — structured baton-passing that beats compaction