@juicesharp/rpiv-pi

A skill-based development workflow for Pi Agent. 30 contract-carrying skills (discover → research → design → plan → implement → validate → code-review → commit, and more), the 15 named subagents they dispatch, and four built-in /wf workflows that chain th

Packages

Package details

extensionskill

Install @juicesharp/rpiv-pi from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:@juicesharp/rpiv-pi
Package
@juicesharp/rpiv-pi
Version
2.13.9
Published
Oct 17, 2026
Downloads
5,498/mo · 2,580/wk
Author
juicesharp
License
MIT
Types
extension, skill
Size
1.7 MB
Dependencies
1 dependency · 14 peers
Pi manifest JSON
{
  "image": "https://raw.githubusercontent.com/juicesharp/rpiv-mono/main/packages/rpiv-pi/docs/cover.png",
  "skills": [
    "./skills"
  ],
  "extensions": [
    "./extensions"
  ]
}

Security note

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

README

@juicesharp/rpiv-pi

npm version License: MIT

Turn "build this feature" into a run of named stages — discover → research → design → plan → implement → validate → code-review → commit — each one writing a reviewable Markdown artifact under .rpiv/artifacts/ that the next stage consumes. rpiv-pi adds 30 skills, 15 specialist subagents, and four ready-made /wf pipelines to Pi Agent, and runs every stage in a detached child session you watch from a lane dock below your editor. It is for developers who want structured, reviewable multi-stage work instead of one very long chat turn.

Install

pi install npm:@juicesharp/rpiv-pi

Restart your Pi session.

Quick start

rpiv-pi registers no tools of its own — it drives the ones its sibling extensions provide. Install them in one step, then restart Pi again:

/rpiv-setup

The dialog previews every change and applies nothing until you confirm.

You also need a model configured in Pi. If /login already works you are set; otherwise see docs/models-config.md.

Now run a pipeline:

/wf build "add a --json flag to the export command"

build is the default — run /wf "<task>" and it resolves here. It captures your brief, slices it, designs each slice in parallel, gates the plan, then implements and validates; see docs/workflows.md for the full stage list and the other pipelines. The run detaches immediately: a lane appears in the dock under your editor with live progress while you keep typing. Press ctrl+q — or ↓ on an empty prompt, or /lanes — to step in, replay the transcript, answer a parked question with ⏎, or stop the run with x.

Prefer one stage at a time? Every skill also stands alone:

/skill:research "how does session state survive /reload"
/skill:blueprint .rpiv/artifacts/research/<latest>.md
/skill:implement .rpiv/artifacts/plans/<latest>.md

What you get

  • Reviewable artifacts instead of one long turn — every skill declares a contract (produces / consumes) and writes timestamped Markdown into .rpiv/artifacts/<bucket>/, so you can read, edit, or reject a design before a line of code is written.
  • Runs that don't hold your session hostage — each /wf stage executes in its own detached child session with bounded parallel fan-out; your interactive session stays a launcher and an observer.
  • A dock that shows what is happening — one row per run with live progress, streaming thinking, per-unit fan-out sub-rows, token usage, and failure reasons. Step in for a faithful transcript replay with full tool rendering.
  • Questions that never get lost — when a detached stage needs input it parks the question on its lane with a badge instead of hijacking your prompt; ⏎ on the flagged lane arms it inline.
  • Four pipelines out of the box — build, vet, polish, and ship, so you never have to author a workflow to get value.
  • Cheap models for cheap stages — /rpiv-models sets model and reasoning-effort overrides per skill, per stage, per workflow, or per subagent, with a typo-catcher for keys that would otherwise silently never apply.
  • Your architecture docs, injected where they matter — touch a file and the nearest AGENTS.md, CLAUDE.md, or .rpiv/guidance/<sub>/architecture.md at each depth is sent to the model, once per session, hidden from your transcript.

Configuration

/rpiv-models writes ~/.config/rpiv-pi/models.json (or $XDG_CONFIG_HOME/rpiv-pi/… when that is set to an absolute path, or a ~ / ~/… path, which is tilde-expanded first), with mode 0600. The file is optional — missing or malformed JSON degrades to no overrides.

Setting What it does Default
/rpiv-models Pick model + reasoning effort for the global default, a subagent, a skill, a workflow stage, or a preset stage no overrides
RPIV_LANES_HOTKEY Rebind the lane-browser hotkey to any Pi key id, or set off to register none ctrl+q
RPIV_BASH_TIMEOUT_MS Per-command bash watchdog inside detached child sessions, clamped to 5 s–30 min 180000 (3 min)

Reference

  • docs/skills.md — all 30 skills: what each consumes, what it writes, and which ones the model may pick on its own.
  • docs/workflows.md — the four /wf pipelines stage by stage, plus hand-driven recipes for when you don't want a whole pipeline.
  • docs/lanes.md — the dock, the lane browser's full key map, parked questions, and the lane environment variables.
  • docs/models-config.md — giving Pi a provider, then the complete models.json schema, cascade, and worked examples.
  • docs/agents.md — the 15 bundled subagents and how the on-disk sync decides what to overwrite.
  • docs/architecture.md — what gets registered, guidance and git-context injection, sibling coupling, and uninstall.

Requirements

Requirement Why
Pi Agent 0.80.x or 1.x The host. Install it globally so pi is on your PATH. Workflow child sessions are tested against Pi 0.80.6, 1.0.4, and 1.1.0
A model provider Pi's own /login or ~/.pi/agent/models.json — see docs/models-config.md
@tintinweb/pi-subagents Supplies the Agent tool. Without it the parallel-analysis skills cannot dispatch. Installed by /rpiv-setup
@juicesharp/rpiv-workflow Supplies /wf. Without it the four built-in workflows do not register. Installed by /rpiv-setup
git (optional) Branch, commit, and user context. If it fails, injection is skipped and nothing else changes
Node.js 22+ Runtime for Pi and rpiv-pi's bundled scripts

You need no build step and no native dependency.

Troubleshooting

Symptom Cause Fix
Session-start banner: N sibling extension(s) missing Sibling plugins not installed Run /rpiv-setup, then restart Pi
Session-start banner: bundled agents need attention A bundled agent changed upstream and your on-disk copy was edited Run /rpiv-update-agents — it overwrites rpiv-managed files, and never touches agents you added yourself
/rpiv-setup or /rpiv-models says it requires interactive mode Running headless Install manually with pi install npm:<pkg>, or hand-edit models.json
/rpiv-setup fails on one package Network or registry issue Retry that package with pi install npm:<pkg>, then re-run /rpiv-setup
/lanes reports No in-flight runs (and ctrl+q does nothing) Nothing is running Expected — the browser only opens over live lanes
ctrl+q does nothing RPIV_LANES_HOTKEY is set to off, empty, or an unrecognized key id Unset it, or use /lanes
A stage stalls, then reports a per-command timeout A bash command wedged past the watchdog Raise RPIV_BASH_TIMEOUT_MS (max 1800000) before starting Pi
/rpiv-models reports No models available Pi has no provider configured Run /login, or add a provider to ~/.pi/agent/models.json

Related

/rpiv-setup installs the ones marked auto; the rest are standalone pi install npm:@juicesharp/rpiv-<name>.

Package Role Auto
@juicesharp/rpiv-workflow /wf runner — chains skills into typed multi-stage pipelines ✓
@juicesharp/rpiv-args $1 / $ARGUMENTS placeholders in skill bodies ✓
@juicesharp/rpiv-ask-user-question Structured questionnaires back to you ✓
@juicesharp/rpiv-todo Live task overlay surviving /reload ✓
@juicesharp/rpiv-advisor Escalate to a stronger reviewer model ✓
@juicesharp/rpiv-web-tools Web search + fetch with pluggable providers ✓
@juicesharp/rpiv-i18n Localization SDK for the rpiv TUI strings ✓
@juicesharp/rpiv-warp Warp Blocked badge on parked lane questions —

License

MIT