henyo-pi-subagent

Subagent extension for Pi — delegate tasks to specialized agents with isolated context windows, configured for a dual Qwen3.6 model setup (35B-A3B + 27B) served via llama.cpp router.

Models

| Agent | Model | Context | Role | |---|---|---|---| | scout | qwen3.6-27b | 58k | Fast codebase exploration, compressed findings | | planner | qwen3.6-35b-a3b | 200k | Implementation planning with deep reasoning | | worker | qwen3.6-27b | 58k | Deterministic code implementation | | reviewer | qwen3.6-35b-a3b | 200k | Insightful code review and quality analysis | | researcher | qwen3.6-35b-a3b | 200k | Web research, fact-finding, and information synthesis |

Installation

Via npm (recommended)

pi install npm:henyo-pi-subagent

Via git

pi install git:github.com/<user>/henyo-pi-subagent@v1

Local development

pi install /workspace/henyo-pi-subagent

Try without installing

pi -e /workspace/henyo-pi-subagent

Agents and prompts are bundled with the extension — no manual file copying needed. User-level agents in ~/.pi/agent/agents/ override bundled agents with the same name.

Usage

Via prompt templates (recommended)

# Full implement chain: scout → planner → worker
pi -p "/template implement 'Add a new API endpoint for user profiles'"

# Scout and plan only: scout → planner
pi -p "/template scout-and-plan 'Explore the auth module'"

# Implement with review: worker → reviewer → worker
pi -p "/template implement-and-review 'Refactor the auth module'"

Direct subagent tool calls

The main agent can invoke subagents directly:

{ "agent": "scout", "task": "Find all authentication-related code" }

{
  "chain": [
    { "agent": "scout", "task": "Explore the codebase for: <query>" },
    { "agent": "planner", "task": "Create a plan for <query> using: {previous}" },
    { "agent": "worker", "task": "Implement the plan: {previous}" }
  ]
}

How It Works

Each subagent runs as a separate pi subprocess with its own context window. This prevents the main agent's context from bloating with task-specific details.

Prompt composition (applied in order):

1. Qwen3.6 preamble — injected by model-system-prompt.ts extension for all qwen3* models
2. Pi default system prompt — standard tool-use instructions
3. Agent role prompt — from the agent's .md file (bundled or user override)

Model swap: With MODELS_MAX=1, the llama.cpp router swaps models on demand (~5-15s). Chains are the recommended pattern over parallel execution.

Customization

Override an agent

Create ~/.pi/agent/agents/scout.md with your own frontmatter + body. It will take precedence over the bundled version.

Add new agents

Create .md files in ~/.pi/agent/agents/ with frontmatter:

---
name: my-agent
description: What this agent does
model: qwen3.6-27b
tools: bash,read,edit,write
---

Your role-specific system prompt here.

File Structure

henyo-pi-subagent/
├── index.ts              # Extension entry point
├── agents.ts             # Agent discovery (package + user + project)
├── package.json          # Pi package manifest
├── agents/               # Bundled agent definitions
│   ├── scout.md
│   ├── planner.md
│   ├── worker.md
│   ├── reviewer.md
│   └── researcher.md
├── prompts/              # Prompt templates
│   ├── implement.md
│   ├── scout-and-plan.md
│   └── implement-and-review.md
└── README.md