@adddog/claude-plugin-pnpm

A Claude Code plugin that turns a single prompt into a coordinated multi-agent team — from requirements through implementation. Built for pnpm monorepos.

"as a team, build a cache layer for the API"

That one sentence triggers a full agentic pipeline: requirements gathering, codebase research, research-informed refinement, architecture planning, parallel implementation, code review, and verification — all coordinated by a lead agent with disk-backed artifacts at every step.

The Pipeline

Every team follows the same artifact chain. Each phase reads the previous phase’s output from disk — no context window pressure, full traceability.

prompt.md → clarify → explore → present → requirements.md
  → research(findings.md)
  → REFINE LOOP (semi-autonomous, grills with research findings)
  → plan(design.md + team-plan.md)
  → review → spawn → implement → verify

How It Works

Semi-Autonomous Refine

The refine phase is where research meets requirements. After the researcher explores the codebase, the designer:

• Self-resolves questions answerable by code exploration (no lead round-trip)
• Returns to lead only for human-judgment questions (scope, priority, preference)
• Tracks rounds, self-resolved vs human-resolved counts
• Max 10 rounds (configurable), every round writes to disk

Research found existing cache layer in @scope/utils covering 80% of requirements.
→ Designer self-resolves: "extend existing" instead of "build new"
→ Updates requirements.md inline
→ Asks human: "Should we add TTL support? Recommended: yes, existing pubsub supports it."

Agent Roster

Every agent has explicit tool scoping — researchers can’t edit code, reviewers can’t write source files.

Planning Phase

Execution Phase

Natural Language Triggers

No special syntax. The lead interprets intent:

Structured Coordination

Disk-Backed Artifacts

All state lives on disk in team-session/{team-name}/. No in-memory-only state. Agents are stateless — they read previous phases from disk, write their output, return. Context windows don’t limit history.

STATUS Protocol

Every agent ends with a structured status:

STATUS: CLEAN                              — done, no issues
STATUS: PARTIAL — completed 3/5 tasks      — progress, more to do
STATUS: ERRORS_REMAINING: 2 in @scope/api  — tried, issues remain
STATUS: BLOCKED — missing API schema       — can't proceed

Interrupt Protocol

Lead can interrupt long-running agents:

INTERRUPT: scope changed
Action: pause | abort | report_status

Agents complete their current atomic operation, write progress to disk, respond with status.

Phase Gating

Phases are sequential. Tasks within a phase run in parallel. blockedBy enforces ordering. QB reviews gate phase transitions.

Skills

Fork Mode (Cost Optimization)

For ~10x cost reduction on parallel agents:

export CLAUDE_CODE_FORK_SUBAGENT=1

Children inherit the lead’s context via prompt cache. First child pays full price, children 2-N pay ~10%.

"as a team (fork), implement the auth refactor"

Dependencies

This plugin integrates with four open-source tools. Use /third-party-manager to check versions, update, and verify.

npx claude-mem install
# Or: curl -fsSL https://install.cmem.ai/openclaw.sh | bash

pipx install cocoindex-code
ccc init && ccc index

Install

# Claude Code marketplace (recommended)
claude plugin add @adddog/claude-plugin-pnpm

# Or npm
npm install -g @adddog/claude-plugin-pnpm

Known Issues

{
  "CLAUDE_MEM_WORKER_PORT": "37701"
}