@jerryan/pi-subagent-lite

A minimal pi extension that delegates tasks to isolated subagent processes.

What makes this different?

Most subagent extensions ship with heavy abstractions: agent definition files, configurable models, working-directory overrides, and a kitchen sink of rarely-used parameters. This one doesn't.

• Zero setup: Install via pi and use it in the next session. No agent directories to manage, no agent definitions to write.
• Minimal interface: Only task and optional skills. We removed model, cwd, agent, and other parameters that add more confusion than value.
• No agent definitions: Unlike almost every other subagent tool, we don't use ~/.pi/agent/agents/*.md or any custom agent discovery. If you need specialization, reuse your existing pi skills via the skills parameter.
• One focused system prompt: Every subagent gets the same lean, task-oriented prompt designed for delegation and clear reporting.
• Transparent long-task handling: Tasks longer than 4000 chars are automatically spilled to a temp file so they never hit CLI length limits.

Features

• Isolated context: Each subagent runs in a separate pi process
• Live progress: See turn-by-turn updates as the subagent works
• Optional skills: Preload capabilities via --skill flags
• Auto-spill: Long tasks (>4000 chars) are automatically written to a temp file to avoid CLI limits
• Clean result rendering: Final output is clearly marked with a ✓ --- Result --- separator
• No recursive nesting: When running inside a subagent process, the tool automatically unregisters itself so subagents cannot spawn further subagents

Installation

pi install npm:@jerryan/pi-subagent-lite

The extension will be available the next time you start a pi session.

To try it without installing permanently:

pi -e npm:@jerryan/pi-subagent-lite

For local development, run inside the repo:

pi -e .

Usage

Once installed, the subagent tool is available:

Run a subagent to find all test files in the project

With skills:

Run a subagent with skills ["code-review"] to review src/auth.ts

You can also invoke multiple subagents in parallel by making separate tool calls in the same turn.

Tool Parameters

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | task | string | Yes | The task to delegate to the subagent | | skills | string[] | No | Optional skill paths or names to load via --skill |

License

MIT © jerryan