@protokol11/pi-deep-research

Multi-agent web research swarm for pi: planner, researchers, skeptic, merger, gate — one cited report. Bundles the deep-research skill and the agent-directory extension.

Packages

Package details

extension

Install @protokol11/pi-deep-research from npm and Pi will load the resources declared by the package manifest.

npm repo home report

$ pi install npm:@protokol11/pi-deep-research

Package: @protokol11/pi-deep-research
Version: 0.1.0
Published: Oct 29, 2026
Downloads: 246/mo · 246/wk
Author: protokol11
License: MIT
Types: extension
Size: 101.4 KB
Dependencies: 0 dependencies · 2 peers

Pi manifest JSON

{
  "extensions": [
    "./extensions/agent-directory/index.ts"
  ]
}

Security note

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

README

@protokol11/pi-deep-research

Multi-agent web research swarm for pi: a planner fans out researchers over a question, a skeptic attacks every finding against its source, and a merger returns one cited report. This package bundles the runtime wiring: the deep-research skill plus the agent-directory extension that resolves the swarm's agents from declarative markdown.

The skill itself is runtime-agnostic and lives in its own repository: protokol11/deep-research. This package is the pi install path; the sources of truth live in the agent-governance monorepo.

Why

A single agent doing research grades its own work: it judges findings with the same context that produced them. The swarm fixes this with three controls — no agent scores its own coverage, every finding lands on disk before context compacts, and a skeptic pass writes a ternary verdict (confirmed / needs-revisit / rejected) per claim. Rejected claims never reach the report.

Quick start

npm install @protokol11/pi-deep-research

Or add the package to packages in your pi settings. See pi.dev for the extension API.

Then, in pi:

/skill:deep-research

Or say: "run deep research on X". The skill asks for five policy answers (question, output path, model split, researcher count, compaction directives), then runs phases P0–P5. The final report.md carries a source: line under every claim.

What is inside

Path	What it is
`extensions/deep-research/SKILL.md`	The skill procedure (P0–P5 phases, gates)
`extensions/deep-research/REFERENCE.md`	State contract: file schemas, brief templates, gate commands
`extensions/agent-directory/`	pi extension: typed catalog of subagents loaded from `.pi/agents/.md` and `.agents/agents/.md`; provides the `agent(name, task)` tool the swarm spawns through
`extensions/custom-compaction.reference.ts`	Reference adaptation of pi's `session_before_compact` for compaction directives — reference material, not loaded

Prerequisites

pi with subagent and web tools.
Model ids for researchers and skeptic/merger (defaults assume deepseek/deepseek-flash:low and kimi-coding/kimi-for-coding:low; resolve yours with subagent({action:"models"}) before a run).

Tests

npm test

Runs the agent-directory loader and spawn tests (node --experimental-strip-types).

Contributing

File issues in the agent-governance monorepo. Skill changes belong in protokol11/deep-research first.

License

MIT — see LICENSE.