pi-rnd#
Scientific-method orchestration extension for PI Coding Agent.
Ports the R&D framework methodology (pre-registration, information-barrier verification, multi-judge consensus, KISS/FP discipline, reality auditing) onto the PI runtime. Ships 11 slash commands, 14 subagent definitions, 35 methodology skills, 6 gates, and 4 lifecycle hooks.
The /rnd-start orchestrator runs on an Effect.ts substrate (src/effect/): the full pipeline is one Effect program, with a typed error channel, Schema-validated data shapes, a composable gate algebra, and subagent spawns that cancel cleanly via fiber interruption.
Requirements#
- Node ≥ 20
@earendil-works/pi-coding-agent(official PI runtime) installed and working
Subagent functionality is built into pi-rnd directly via PI's public SDK (createAgentSession) — no external subagent extension is required.
Develop#
pi -e ./src/index.ts
PI loads .ts directly via jiti — no build needed for development. Then in the pi session:
/rnd-doctor
/rnd-start <task description>
Install (built)#
npm install
npm run build
pi install ./
Subagent files in agents/ are NOT auto-discovered from npm packages. Copy them into .pi/agents/ (project-local, primary) or ~/.pi/agent/agents/ (global) before running /rnd-start:
mkdir -p .pi/agents
cp agents/*.md .pi/agents/
Commands#
| Command | Description |
|---|---|
/rnd-start <task> |
Full plan → discover → design → build → audit → verify → iterate → cleanup → polish → integrate pipeline |
/rnd-doctor |
Health and discovery diagnostics |
/rnd-status |
Show the current state of the R&D pipeline for the active project |
/rnd-history |
List past R&D pipeline sessions for this project with SHIP/NO-SHIP verdicts |
/rnd-scan |
Scan the project for facts (package manager, env, CI, external URLs) and write project-facts.md |
/rnd-brainstorm <idea> |
Brainstorm a vague idea into a focused plan; writes brainstorm.md |
/rnd-resume |
Identify the most recent incomplete R&D session and the wave to resume at |
/rnd-audit [--scope=src/] |
Audit the project tree for issues; writes a timestamped report to $RND_DIR/audits/ |
/rnd-review [--against=<ref>] |
Run a pre-PR code review via git diff against a base ref; writes a timestamped review report |
/rnd-debug <bug> [--apply] |
Debug a reported bug — spawns rnd-debugger, writes a diagnosis report, halts before modifying code |
/rnd-roadmap [--list|--next|--check Mn] |
List, navigate, and mark milestones in roadmap.md |
CLI flags#
7 registered flags control pipeline behavior. See docs/UX-DEGRADATIONS.md for the full list (--rnd-iteration-cap, --rnd-skip-failing, --rnd-reality-audit, and others).
What /rnd-doctor checks#
~/.pi/agent/(or$PI_CODING_AGENT_DIR) is writable$RND_DIRis writable- Native subagent module responds to
subagents:rpc:pingwithin 2 s - Each of the 11 subagent
.mdfiles is discoverable - Each of the 35 skill directories is discoverable
- Gate registry has registered seed gates (
gates: N) - A no-op
general-purposesubagent round-trips within 30 s
Documentation#
docs/PI-API.md— verified reference for PI's extension API surface.docs/PORTING.md— Claude Code → PI translation reference (frontmatter, vocabulary, install locations, hook model).docs/UX-DEGRADATIONS.md— canonical CLI flags that replace Claude Code's multi-option prompts (--tier,--amend,--design,--iteration-cap, etc.).docs/UNPORTABLE.md— inventory of Claude Code features with no PI surface (multi-option UI, output styles, lifecycle hooks).
License#
MIT.