Sub-agents for pi. Schema-first, minimal token overhead.
Spawn specialized agents with isolated sessions, custom tools, and per-type models. Three tools, no bloated descriptions, no prompt snippets, no usage guidelines. Names like Agent, run_in_background, and worktree_path are the schema.
Foreground and background agents, custom agent types, per-model concurrency, steering, cost tracking, and a live status widget.
pi install npm:pi-subagents-lite
pi install -l npm:pi-subagents-lite # project-local
pi -e npm:pi-subagents-lite # try without installingThe LLM calls Agent like any other tool. Foreground agents return inline with stats. Background agents acknowledge immediately and auto-deliver on completion.
◈ Agents
├─ ⠙ Agent Write model precedence unit tests 6🛠︎ ·3⟳ ·↑6.8k↓1.3k 6%·12s
│ │ tail -f /tmp/pi-agent-outputs/bb3382a9-1f7e-474.log
│ └ The file already exists but is ~175 lines. The user wants a …
└─ ⠙ Explore Explore codebase architecture 13🛠︎ ·4⟳ ·↑16.1k↓2.9k 15%·12s
└ ## Architecture Summary: pi-subagents-lite
Manage everything from /agents: running agents (view, steer, stop), manual spawn without an LLM round-trip, model settings, concurrency, widget layout.
Agent— spawn a sub-agent.promptis required;agentselects a type,run_in_backgroundfor fire-and-forget,worktree_pathto run in a git worktree.model,thinking,max_turns, andmax_tokensare injected from config and frontmatter, never passed by the LLM.StopAgent— stop a running agent by ID. IDs come from the spawn result, the stop error, or/agents.AgentStatus— list all agents with type, short ID, and status.
Drop a .md file into .pi/agents/ (project), .agents/agents/ (shared), or ~/.pi/agent/agents/ (global). Frontmatter configures the agent, the body is its system prompt. The name auto-populates the agent parameter's enum, no registration needed. Built-ins general-purpose and Explore are always available; on name clash, project > shared > user > built-in.
---
name: security-review
description: Review code for security issues
tools: [read, bash, grep]
extensions: false
skills: false
model: zai/glm-5.2
thinking: high
max_turns: 80
---
You are a security review specialist. Analyze code for vulnerabilities,
focusing on injection flaws, auth bypasses, and insecure defaults.A minimal agent with just name and description gets everything, same as general-purpose. Set restrictions only when you want them.
| Field | Type | Default | Description |
|---|---|---|---|
name |
string | filename | Agent type name. Must be unique. |
display_name |
string | name |
Label in the UI. |
description |
string | "" |
One-sentence description. |
tools |
true | string[] | false |
true |
Tool whitelist. Mutually exclusive with exclude_tools. |
exclude_tools |
string[] |
none | Tool blacklist. Mutually exclusive with tools. |
extensions |
true | string[] | false |
true |
Which extensions load (hooks and commands). Does not control tool visibility. |
exclude_extensions |
string[] |
none | Extension blacklist. |
skills |
true | string[] | false |
true |
Skill whitelist (metadata-only in system prompt). |
preload_skills |
string[] | false |
false |
Dump full SKILL.md content into the system prompt. Expensive. |
model |
string | inherit parent | "provider/model-id". See Model Resolution. |
thinking |
string | inherit parent | off, minimal, low, medium, high, xhigh. |
max_turns |
number | unlimited | Soft turn limit, then grace turns before hard abort. |
max_tokens |
number | unlimited | Max output tokens per LLM response. |
hidden |
boolean | false |
Hide from the enum. Still callable by name. |
Tool and extension lists accept built-in names (read, bash, edit, write, grep), extension tool names (web_search), and ext/* globs (tavily/*). exclude_tools: [tavily/*] hides the tools but the extension still loads; use exclude_extensions: [tavily] to prevent loading.
loadSkillsImplicitly and loadExtensionsImplicitly (config, default ON) decide what an agent gets when frontmatter omits skills or extensions. Turn them OFF to default new agents to nothing and opt in explicitly.
Precedence, highest first:
- Session per-type override (
/agents> Model settings) - Session global default
- Config per-type override (
~/.pi/agent/subagents-lite.json) - Config global default
- Agent frontmatter
model - Parent model
The LLM never passes model. Set it once in config or frontmatter and forget.
systemPromptMode (default replace):
replace: minimal generic prompt plus the agent's instructions. Lowest cost, most isolated.inherit: parent's system prompt plus the agent's instructions.custom:~/.pi/agent/subagents-lite-prompt.mdplus the agent's instructions.
When includeContextFiles is true (default), AGENTS.md files load as shared context before agent instructions, which improves KV cache prefix hits.
~/.pi/agent/subagents-lite.json, managed via /agents or edited directly. Per-type model overrides are dynamic keys alongside the special fields.
{
"agent": {
"default": "zai/glm-5.2",
"forceBackground": true,
"graceTurns": 6,
"showCost": true,
"showTools": false,
"showTurns": true,
"showInput": true,
"showOutput": true,
"showContext": true,
"showTime": true,
"widgetMaxLines": 12,
"widgetMaxLinesCompact": 6,
"widgetDescLengthFull": 50,
"widgetCompact": true,
"widgetShortcut": false,
"systemPromptMode": "inherit",
"includeContextFiles": true,
"loadSkillsImplicitly": false,
"loadExtensionsImplicitly": false,
"disableDefaultAgents": false,
"Explore": "xiaomi/mimo-v2.5",
"builder": "xiaomi/mimo-v2-pro",
"architecture-reviewer": "zai/glm-5.2",
"planner": "zai/glm-5.2"
},
"concurrency": {
"default": 4,
"providers": {
"llamacpp": 1,
"ai.lan": 2
},
"models": {}
}
}Widget, stats visibility, and spawn defaults are all under /agents > Settings.
Output logs land in /tmp/pi-agent-outputs/<agentId>.log, append-only and tail -f friendly. Logs and completed results survive on disk even if a session reload (/reload, extension reload) kills running agents.
- Node.js >= 18
- pi >= 0.82.0
MIT