# TASK: Paperclip MVP — Harvey ↔ CC coordination without Mike as middleman

## Status: READY FOR CC

## Goal
Build a working MVP that demonstrates Harvey (as manager/orchestrator) and CC (as worker) coordinating directly through Paperclip — without Mike having to relay messages between us.

The test: Mike gives Harvey a task in Discord. Harvey delegates to CC via Paperclip. CC does the work. Harvey reviews and reports back to Mike. Mike never talks to CC directly.

## Step 1: Find the Paperclip repo and read it
- Search GitHub for the Paperclip orchestration framework mentioned in this article: https://www.catalystandcode.com/blog/ai-agent-orchestration-frameworks
- Key traits to confirm you have the right repo: "bring your own bot", hierarchical org chart model (CEO → managers → workers), heartbeat-based scheduling, model-agnostic, works with Claude Code and Codex
- Read the README, docs, and any quickstart/examples thoroughly
- If there's an npm/pip package, note the install command

## Step 2: Set up the repo
- Clone or install Paperclip into a new directory: `~/paperclip-mvp/`
- Get the basic framework running (follow their quickstart)
- Document any setup issues in DONE.md

## Step 3: Define two agents

### Agent 1: Harvey (Manager)
- Role: orchestrator/manager
- Receives tasks from Mike (via Discord or file)
- Delegates subtasks to CC
- Reviews CC's output
- Reports results back (writes to a file Harvey can read)
- Model: Claude Sonnet (or whatever Paperclip supports for the manager role)

### Agent 2: CC (Worker)
- Role: builder/executor
- Receives tasks from Harvey
- Does the actual work (write code, read files, do research, etc.)
- Reports completion back to Harvey
- Model: Claude Code / Claude Sonnet

## Step 4: Build a simple test workflow
Wire up this exact scenario:

1. Mike drops a task in `~/paperclip-mvp/inbox/task.txt` (plain text)
2. Harvey picks it up, breaks it into subtasks if needed, delegates to CC
3. CC does the work and writes output to `~/paperclip-mvp/output/result.txt`
4. Harvey reviews the result, writes a summary to `~/paperclip-mvp/output/summary.txt`

**Test task to use:** "Research the top 3 competitors of No More Chores cleaning company in Toronto and write a brief summary of their pricing and positioning."

This is a real task. CC should actually do the research (web search) and write real output. We'll know it worked if the summary is useful.

## Step 5: Document everything
Write `~/.openclaw/workspace/leah/specs/DONE.md` with:
- Link/name of the Paperclip repo you found
- How you set it up
- How the two-agent workflow is configured (show the key config)
- Results of the test task (paste the summary.txt content)
- What worked, what didn't
- Your honest assessment: is this production-ready? What would it take to wire Leah/Sam through this?
- Any blockers or questions for Harvey

## Why this matters
Right now Mike has to be the relay between Harvey and CC for every task. That's friction that slows everything down. Paperclip is supposed to let agents coordinate directly. This MVP proves whether it actually works for our setup before we commit to migrating anything.

## Constraints
- Do NOT touch nmc-phone/ or any production code
- Do NOT modify Vapi or Twilio configs
- Everything goes in ~/paperclip-mvp/ (new directory, safe to experiment)
- If Paperclip doesn't work or the repo doesn't exist, say so clearly in DONE.md with what you found