pi-subagent-factory

Auto-scan any codebase and generate a tailored pi subagent team (architect / implementer / reviewer / tester + stack specialists) with ready-to-use subagent_spawn args. 自动扫描项目技术栈，自动创建专属 subagent 团队，一键并行派活。

Packages

Package details

extensionskill

Install pi-subagent-factory from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:pi-subagent-factory
Package
pi-subagent-factory
Version
0.1.0
Published
Oct 21, 2026
Downloads
338/mo · 17/wk
Author
flying_dragon
License
MIT
Types
extension, skill
Size
36.5 KB
Dependencies
0 dependencies · 4 peers
Pi manifest JSON
{
  "skills": [
    "./skills"
  ],
  "extensions": [
    "./extensions/subagent-factory.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-subagent-factory

Scan any codebase → auto-generate a tailored subagent team for pi, with ready-to-use subagent_spawn args. 全自动扫描项目技术栈，自动创建并出具合适的 subagent 团队。

TL;DR — one tool call scans up to 6,000 files (languages, frameworks, databases, testing, build tooling, monorepo, Docker, CI), then deterministically generates 4–10 specialized subagents with per-agent tool allowlists and thinking levels. The scan is a rules engine: 0 model tokens. Smoke-tested end to end on a React + Prisma + Vitest project: 8 agents generated in under 5 seconds.

pi package npm License: MIT

Install

pi install npm:pi-subagent-factory

Try once without installing:

pi -e npm:pi-subagent-factory

How it works

  1. Scan — walks the repo (skips node_modules, .git, build output; caps at 6,000 entries), reads package.json / pyproject.toml / go.mod / Cargo.toml / pom.xml, and produces a ProjectProfile: language mix, frameworks, databases, test runners, package manager, test/build commands, monorepo / Docker / CI flags, entry points, key directories, LOC.
  2. Generate — a deterministic roster builder turns the profile into a team. Every agent carries spawnArgs that map 1:1 to @narumitw/pi-subagents subagent_spawn parameters (task / tools / thinkingLevel / timeout).
  3. Dispatch — the model spawns agents in parallel or sequence (subagent_wait, subagent_send, subagent_inspect, subagent_cancel manage the lifecycle). The roster persists at .pi/team/subagents.json.

The base team (every project)

Agent thinking Tools Mission
architect high read, grep, find, ls, codemode structure analysis, implementation plan, acceptance criteria — read-only
implementer medium read, bash, edit, write, grep, find, ls executes the plan, behavior-preserving, step summaries
reviewer high read, grep, find, bash reviews diffs: behavior drift, edge cases, over-design, security debt
tester medium bash, read, edit, write, find, grep runs the project's detected test command, guards regression

Stack specialists (added only when detected)

Detection Added agent
React / Vue / Svelte / Angular / Next.js frontend — components, state, render performance
api/ server/ services/ dirs, Express / NestJS / Fastify / FastAPI / Spring… api — routes, auth, error handling, contracts
Prisma / Drizzle / TypeORM / SQLAlchemy / pg / Redis… data — schema, migrations (reversible), query slimming
Dockerfile / .github/workflows devops — build pipeline, deploy scripts
docs/ dir or >200 files docs — README/architecture/change log sync (中文)
pnpm-workspace / turbo / nx / lerna monorepo-coordinator — dependency direction, build order

Each task template embeds the project's measured context — stack, key directories, entry points, test command, build command, scale — so spawned agents start warm, not cold.

Usage

Model side (recommended):

扫描这个项目，生成 subagent 团队，把 architect 和 tester 派出去

The model calls the subagent_factory tool:

action behavior
scan scan only, return the ProjectProfile
generate scan + generate team, write .pi/team/subagents.json + .pi/team/SUBAGENTS.md
load read the previously generated team

Command side (no model tokens at all):

/team          # scan + generate (same as /team gen)
/team scan     # show the tech-stack profile only
/team show     # check whether a team already exists

How it compares

pi-subagent-factory ciscoittech/claude-agent-framework keysersoose/claude-agent-builder
Runtime pi (any model) Claude Code Claude Code
Scan deterministic rules engine, 0 tokens LLM prompt-driven interview + LLM research
Output subagent_spawn-ready args (JSON) .claude/agents/*.md + registry + validator full agent/skill/hook/command file set
Speed seconds minutes (model turn) minutes (conversation)
Determinism same repo → same team model-dependent model-dependent

The trade-off is deliberate: no LLM in the loop means the roster is reproducible and free, and every generated task is plain text you can edit before dispatch.

FAQ

Does it call an LLM during scan/generate? No. Scanning and roster generation are a pure rules engine — the only model calls happen later, when spawned subagents do their jobs.

Which agents get write access? Only implementer, tester, frontend, api, data, devops, and docs. architect, reviewer, and monorepo-coordinator are read-only by construction (their tool allowlists exclude edit/write).

Does it work without @narumitw/pi-subagents? The spawnArgs are plain JSON. Without pi-subagents you can still read the roster and run agents through any mechanism that accepts task text + tool lists.

Where is the generated team stored? <project>/.pi/team/subagents.json (structured) and .pi/team/SUBAGENTS.md (human-readable table + per-agent dispatch JSON). Re-running generate overwrites them.

What languages/frameworks does detection cover? TypeScript/JavaScript (React, Next.js, Vue, Svelte, Angular, Express, NestJS, Fastify, Koa, Hono), Python (FastAPI, Django, Flask, SQLAlchemy), Go (Gin), Rust (Axum, Actix), Java (Spring, Maven), plus Vite/Webpack/esbuild/Rollup/Turborepo, Vitest/Jest/Mocha/Playwright/Cypress/pytest, and monorepo/Docker/CI flags.

Related

License

MIT