Back to Claude Scientific Skills

pi-subagents Package

skills/pi-agent/references/pi-subagents.md

2.48.07.3 KB
Original Source

pi-subagents Package

Source: https://pi.dev/packages/pi-subagents

Extension for delegating tasks to focused child agents with sequential chains, parallel execution, dynamic fanout, worktree isolation, and acceptance gates.

bash
pi install npm:pi-subagents

Built-in Agents

AgentPurpose
scoutFast local codebase recon: files, entry points, data flow, risks
researcherWeb/docs research with sources and a concise brief
plannerConcrete implementation plan from existing context
workerImplementation: file editing and validation
reviewerCode review and small fixes against task/plan
context-builderSetup pass gathering code context and handoff material
oracleSecond opinion challenging assumptions; no edits
delegateLightweight general delegate close to parent behavior

Packaged planner, worker, oracle default to context: "fork" (branch from parent session state); others default to fresh.

Commands

bash
/run <agent> [task]                  # single agent; --bg detached, --fork branch session
/run reviewer[model=anthropic/claude-sonnet-4] summarize this code
/chain scout "scan the codebase" -> planner "create an implementation plan"
/parallel scanner "find security issues" -> reviewer "check code style"
/run-chain <chainName> -- <task>     # saved workflow
/subagents-doctor                    # setup diagnostics

Per-step config uses [key=value,...] on the agent name: output=file.md, outputMode=file-only|inline, reads=a.md+b.md, model=..., skills=a+b, thinking=high, progress.

Natural language also works: "Use reviewer to review this diff", "Run parallel reviewers: one for correctness, one for tests".

Packaged prompt shortcuts: /parallel-review, /review-loop, /parallel-research, /parallel-context-build, /parallel-handoff-plan, /gather-context-and-clarify, /parallel-cleanup (add autofix to apply synthesized fixes).

Programmatic API (subagent tool)

javascript
{ agent: "worker", task: "refactor auth" }
{ tasks: [{ agent: "scout", task: "audit frontend" }, { agent: "reviewer", task: "audit backend" }] }   // parallel; count: N duplicates a task
{ chain: [{ agent: "scout", task: "Gather context" }, { agent: "planner" }, { agent: "worker" }, { agent: "reviewer" }] }
{ chain: [...], timeoutMs: 30000 }
{ agent: "worker", task: "...", maxRuntimeMs: 600000 }
{ action: "list" | "get" | "create" | "update" | "delete" | "status" | "interrupt" | "resume" | "doctor" }
{ action: "resume", id: "<run-id>", message: "follow-up" }

Key parameters: output (file or false), outputMode (inline/file-only), skill (string/array/false), model, concurrency (default 4), worktree, context (fresh/fork), chainDir, clarify (default true for chains), async, cwd, maxOutput (default 200KB/5000 lines), share (Gist upload, off by default), acceptance.

Dynamic fanout: a chain step with expand: { from: { output: "name", path: "/items" }, item: "target", maxItems: N } plus parallel: { agent, task: "Review {target.path}" } and collect: { as: "reviews" } fans out over a prior step's structured output (as + outputSchema).

Agent Definition Files

Markdown with YAML frontmatter. Precedence: project .pi/agents/**/*.md > user ~/.pi/agent/agents/**/*.md > builtin.

yaml
---
name: scout
description: Fast codebase recon
model: claude-haiku-4-5
fallbackModels: openai/gpt-5-mini, anthropic/claude-sonnet-4
thinking: high
tools: read, grep, find, ls, bash      # allowlist; mcp: prefix for direct MCP tools
extensions: mcp:chrome-devtools        # omitted = all; empty = none
skills: safe-bash
systemPromptMode: replace              # or append
inheritProjectContext: false
inheritSkills: false
defaultContext: fork                   # or fresh
output: context.md
defaultReads: context.md
defaultProgress: true
completionGuard: false                 # false for non-implementation validators
interactive: true
maxSubagentDepth: 1
maxExecutionTimeMs: 600000
maxTokens: 50000
---
Your system prompt goes here.

Subagents only receive direct MCP tools when listed in tools: frontmatter (requires pi-mcp-adapter); a global directTools: true is insufficient.

Chain Files

Reusable workflows: project .pi/chains/**/*.chain.md|.chain.json > user ~/.pi/agent/chains/. Markdown chains use ## agent headings with per-step keys (phase, label, as, output, outputMode, reads, model, skills, progress, outputSchema) and a task body. Template variables: {task}, {previous}, {chain_dir}, {outputs.name}.

Configuration

Builtin agent overrides in ~/.pi/agent/settings.json or .pi/settings.json:

json
{ "subagents": { "agentOverrides": { "reviewer": { "model": "anthropic/claude-sonnet-4", "thinking": "high" } }, "disableBuiltins": false } }

Override fields: model, fallbackModels, thinking, systemPromptMode, inheritProjectContext, inheritSkills, defaultContext, disabled, skills, tools, systemPrompt.

Extension config at ~/.pi/agent/extensions/subagent/config.json: asyncByDefault, forceTopLevelAsync, parallel: { maxTasks, concurrency }, defaultSessionDir, maxSubagentDepth, intercomBridge: { mode: "always"|"fork-only"|"off", instructionFile }, worktreeSetupHook (+ worktreeSetupHookTimeoutMs; hook gets stdin JSON with repo/worktree paths and must print { "syntheticPaths": [...] }).

Nesting depth defaults to 2 levels; tighten/relax via PI_SUBAGENT_MAX_DEPTH env var, config maxSubagentDepth, or per-agent frontmatter (per-agent can only tighten). Children never get the subagent tool unless their resolved tools explicitly include it.

Worktree Isolation

worktree: true on parallel tasks or chain steps runs each agent in an isolated git worktree. Requires a git repo with a clean working tree; node_modules/ is symlinked in; task-level cwd overrides must match the shared cwd.

Acceptance Gates

Attach explicit contracts to any run/step:

javascript
{ agent: "worker", task: "Implement the fix", acceptance: {
  criteria: ["Patch the bug without widening scope"],
  evidence: ["changed-files", "tests-added", "commands-run", "residual-risks", "no-staged-files"],
  verify: [{ id: "focused", command: "npm test", timeoutMs: 120000 }],
  maxFinalizationTurns: 3
} }

Provenance levels reported: attested, checked, verified, reviewed, rejected.

Async, Clarify, Observability

--bg / async: true detaches runs; status files under <tmpdir>/pi-subagents-<scope>/async-subagent-runs/<id>/ (status.json, events.jsonl, logs). Chains open a clarify TUI by default to preview/edit steps (e edit, m model, t thinking, s skills, b background, Enter run, Esc cancel). Debug artifacts land in {sessionDir}/subagent-artifacts/ with input/output/jsonl/meta per run; chain artifacts in <tmpdir>/pi-subagents-<scope>/chain-runs/{runId}/. Directories older than 24h are cleaned on startup. Events: subagent:async-started, subagent:async-complete, subagent:control-intercom, subagent:result-intercom.

Optional companion pi install npm:pi-intercom lets children call contact_supervisor (reasons: need_decision, progress_update) and groups completion delivery back to the parent.

Recommended implementation pattern: clarify → planner → worker → fresh reviewers → worker.