@themuuln/pi-research
v0.1.0
Published
Deep research extensions for pi coding agents: /research (deep multi-agent dives), /research-general (any topic), /dig (fast adaptive research). Works in both pi-coding-agent and OMP.
Readme
@themuuln/pi-research
Deep research extensions for pi-coding-agent (and OMP, the Oh My Pi fork). Three commands that turn a prompt into an evidence-backed, publication-quality research run:
| Command | What it does |
|---|---|
| /research <query> | Deep multi-agent research with planning, parallel workers, quality checkpoints, synthesis and audit |
| /research-general <query> | Same orchestrator for ANY topic — markets, health, geopolitics, consumer decisions, science, news. Source tiers, freshness and bias checks built in |
| /dig <query> | Fast adaptive research — 3-4 calls (~30s) for simple facts, auto-escalates to deeper rounds or subagents when evidence is thin |
Install
pi install @themuuln/pi-researchOMP users can install the same package; it works in both runtimes.
Usage
/research compare Bun and Node.js runtime tradeoffs for CLI tools
/research-general is a carnivore diet backed by current evidence?
/dig mongolian payment sms automation
/dig --deep carbon capture tax credits
/dig --lang=mn cэрэг мэдээHow it works
/research and /research-general
The command injects an orchestrator prompt that:
- Boots a run directory (default
~/.omp/research/when~/.ompexists, otherwise~/.pi/research/; override withPI_RESEARCH_HOME) - Writes a plan: complexity class, orthogonal sub-questions, source-quality bar
- Parallel mode — when a
tasktool is available, spawns one subagent per sub-question in a single batch, each writing notes tonotes/*.mdInline mode — without atasktool, does the searching itself and writes the same notes sequentially (same workflow, no subagents) - Synthesizes claims + sources, then calls the
research_checkpoint/research_general_checkpointtool — a code-enforced quality gate (min sources, avg confidence, unresolved contradictions) - Iterates with targeted gap-fillers up to 3 rounds, then critic pass,
synthesis.md, andmeta.json
/dig
- Triage → Phase 1 (fast search, 3-4 calls)
- Calls
dig_checkpoint— the tool returns DONE / ITERATE / ESCALATE - DONE: synthesize. ITERATE: one more inline round. ESCALATE: spawns 1-2 read-only subagents (or, inline mode, does 3-6 more targeted searches)
- Answers in the query's language, with evidence, caveats and exact sources
Configuration
| Env var | Default | Purpose |
|---|---|---|
| PI_RESEARCH_HOME | ~/.omp/research (if ~/.omp exists) else ~/.pi/research | Where run directories are created |
Requirements
- Node.js >= 20
- A pi-coding-agent-compatible runtime (
pior OMP) with web access tools configured (search, page fetch). Research quality depends on the tools the runtime exposes — the orchestrator uses whatever is available.
License
MIT
