pi-subagents

Sub-agents extension for pi coding agent.

Delegates tasks to isolated pi child processes — each running with its own model, system prompt, and tool set. Sub-agents start with a fresh session by default; pass session: "fork" to branch from the current parent pi session.

Install

pi install npm:@johnnywu/pi-subagents

Quick start

1. Define an agent

Create a .pi/agents/code-reviewer.md file in your project:

---
name: code-reviewer
description: Reviews code changes for correctness and style
tools: read, grep, find, ls, bash
model: anthropic/claude-sonnet-4-6
thinking: low
---

You are a code reviewer. When given a diff or file list, read the relevant
files and provide a concise review covering:

- Logic errors and edge cases
- Style and consistency issues
- Performance concerns
- Test coverage gaps

2. Use it

The subagent tool is automatically registered. In a pi session:

Review the changes in src/auth.ts using the code-reviewer agent

Or instruct pi to delegate:

Run the code-reviewer agent on the last three commits

By default, sub-agents do not inherit parent conversation history. For tasks that need the current conversation, request a forked session:

subagent({ agent: "code-reviewer", task: "Review the approach we just discussed", session: "fork" })

session is optional and accepts:

Value	Behavior
none	Default. Start a new sub-agent session in the subagents session directory.
fork	Fork the current parent session at its active leaf and run the sub-agent with that branched session.

If fork is requested but unavailable, pi-subagents falls back to none and shows a warning in the tool details/rendering. Fallback happens when the parent session is not persisted, has no current leaf, the forked session file is not materialized, or the call uses a cwd different from the parent session cwd.

The tool prompt also guides the model to choose session: "fork" when a delegated task depends on the current conversation, prior discussion, or parent session history, and to keep session: "none" for self-contained tasks.

Available subagents in the prompt

The extension exposes discovered sub-agents to the model when the active tool set includes subagent:

• The subagent tool keeps a one-line prompt guideline: Available subagents: code-reviewer, refactor, test-writer
• At agent-start time, the system prompt gets an independent block:

Available subagents:
- code-reviewer
- refactor
- test-writer

The prompt file passed to the child process contains only the agent prompt plus skills. Runtime prompt assembly then depends on systemPrompt mode.

What goes into the sub-agent system prompt

Component	append	replace	replace-all
pi default system prompt	✅ kept	❌ replaced	❌ replaced
Project context files (AGENTS.md/CLAUDE.md, etc.)	✅ included	✅ included	❌ skipped
Agent body (.md file body)	✅ appended	✅ becomes the prompt	✅ becomes the prompt
Skills XML block	✅ appended	✅ appended	✅ appended
Available tools / Guidelines block	from default prompt	re-injected by before_agent_start hook	re-injected by before_agent_start hook
Available subagents block	injected at agent start	injected at agent start	injected at agent start

append keeps pi's full default prompt (with project context) and adds the agent body at the end. replace swaps out pi's default prompt for the agent body while keeping pi context files. replace-all is the fully isolated mode: it swaps out pi's default prompt and skips pi context files, then the runtime hook re-injects tool and guideline blocks to preserve tool visibility.

Breaking change: the old replace behavior is now replace-all. Existing agents that need to keep skipping AGENTS.md/CLAUDE.md should change systemPrompt: replace to systemPrompt: replace-all.

The Available subagents block is injected by the child process at agent-start time, after PI_SUBAGENT_ALLOWED and recursion depth filtering are applied.

Debug a sub-agent prompt

Set debug: true in an agent's frontmatter to export that sub-agent's effective runtime system prompt on each run:

---
name: scout
debug: true
---

The child process writes debug-system-prompt.md in the project cwd. The file contains the prompt visible during before_agent_start, including systemPrompt append/replace/replace-all behavior, tools/guidelines, skills, project context files in append and replace modes, and pi-subagents' runtime Available subagents block when applicable.

Agent configuration

Agents are Markdown files with YAML frontmatter.

Field	Required	Default	Description
name	yes	—	Unique agent identifier
description	no	—	Human-readable summary
tools	no	none	Comma-separated tool whitelist (read, write, bash, grep, etc.)
model	no	parent's model	Provider/model-id (anthropic/claude-sonnet-4-6)
thinking	no	off	Reasoning level: off, minimal, low, medium, high, xhigh
systemPrompt	no	append	How the body is applied: append (append to pi default system prompt and project context), replace (replace pi default prompt while keeping project context), or replace-all (replace pi default prompt and skip project context)
skills	no	none	Comma-separated skill names or simple wildcard patterns (*, obsidian-*) to load (resolved from project .agents/skills/, .pi/skills/, global ~/.pi/agent/skills/, or npm packages)
allowedAgents	no	all	Comma-separated list of sub-agents this agent may spawn
maxDepth	no	10	Maximum recursion depth (0 = no sub-agents, 1 = one level, etc.)
debug	no	false	When true, export the effective runtime system prompt to debug-system-prompt.md

The Markdown body after the frontmatter is the agent's system prompt.

skills entries match the skill frontmatter name. Use skills: * to load all available skills, or prefix-style patterns such as skills: obsidian-* to load matching skills. Only * is supported as a wildcard; glob features like ? or {a,b} are not supported.

Example with all fields

---
name: orchestrator
description: High-level planner that delegates to specialists
tools: subagent, read, grep, find
model: anthropic/claude-sonnet-4-6
thinking: high
systemPrompt: append
skills: tdd, obsidian-*
allowedAgents: code-reviewer, refactor, test-writer
maxDepth: 2
debug: false
---

You are an orchestrator. Break complex tasks into sub-tasks and delegate
them to specialist agents. Combine their results and report a summary.

Agent discovery

Agents are discovered from two locations (project overrides global):

Scope	Path
Global	~/.pi/agent/agents/*.md
Project	.pi/agents/*.md

Only .md files are scanned. Files are parsed at extension load time; parse errors produce warnings but don't block other agents.

Recursion control

Sub-agents can spawn their own sub-agents (if the subagent tool is in their whitelist). Two mechanisms prevent unbounded recursion:

maxDepth — Hard limit counting from the originating agent. maxDepth: 0 means the agent cannot spawn sub-agents. maxDepth: 1 allows one level, etc. Defaults to 10 when the agent has subagent in tools.

allowedAgents — Whitelist enforced by the parent before spawning. A child process never sees agent names outside its parent's whitelist.

The available-subagents prompt entries respect the same filtering: parent sessions use the currently visible agents, and child sessions only list agents allowed by their parent.

These are passed via environment variables (PI_SUBAGENT_DEPTH, PI_SUBAGENT_MAX_DEPTH, PI_SUBAGENT_ALLOWED). Child processes also receive PI_SUBAGENT_NAME, PI_SUBAGENT_SYSTEM_PROMPT_MODE, and PI_SUBAGENT_SESSION so runtime hooks can distinguish prompt mode and effective session mode.

Session storage

Sub-agent sessions are saved as .jsonl files for post-hoc debugging:

~/.pi/agent/sessions/--{safe-project-path}--/subagents/
  ├── 2026-05-31T03-47-16-502Z_019e7c24-a395-707a-a262-ec5b1664ffa7.jsonl
  └── ...

Each file contains one JSON object per line — session headers, messages, tool calls, and usage data. Parent pi sessions live in the same project directory (no subagents/ subdirectory). When session: "fork" succeeds, the branched child session is still written under subagents/ and points back to the parent session in its header.

Development

# Install dependencies
bun install

# Run tests
bun test

# Type-check
bun run typecheck

# Format
bun run format

# Release (requires GH_TOKEN and NPM_TOKEN)
bun run release

This project uses semantic-release with conventional commits.

License

MIT