@juicesharp/rpiv-workflow

Chain the Pi skills you already have into multi-stage pipelines. rpiv-workflow adds a /wf command to Pi Agent that runs each stage in its own detached session, validates its output against a schema, routes between stages on predicates over the previous stage's data, and appends each step to a resumable JSONL trail. It ships zero workflows — write your own in .rpiv/workflows/config.ts, or install @juicesharp/rpiv-pi for five.

Install

pi install npm:@juicesharp/rpiv-workflow

Restart your Pi session.

Quick start

/wf runs skills you already have installed — this package ships none of its own. The fastest way to see a real run is the sibling that bundles both skills and workflows:

pi install npm:@juicesharp/rpiv-pi

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

The run starts detached and your prompt comes straight back. Every step appends to .rpiv/workflows/runs/<run-id>.jsonl, so you can walk away and resume later.

To write your own instead, put a workflow at .rpiv/workflows/config.ts naming skills already on your machine (check ~/.pi/agent/skills/):

import { defineWorkflow, acts, terminal } from "@juicesharp/rpiv-workflow";

export default defineWorkflow({
  name: "review-and-ship",
  start: "code-review",
  stages: { "code-review": acts(), commit: terminal() },
  edges: { "code-review": "commit", commit: "stop" },
});

Each stage key is the skill name; acts() runs a skill for its side effects, terminal() does the same for a stage that carries nothing forward. Schemas, artifact outcomes, and predicate routing are in the authoring reference.

Command	What it does
/wf	Preview every loaded workflow
/wf <name>	Preview one workflow's stage graph
/wf <name> "<input>"	Run it detached — your prompt returns immediately
/wf @<ref>	Resume a past run by run-id, --name alias, or .jsonl path

Add --name <slug> as the first or last token to give a run an alias you can resume by. With no workflows registered anywhere, /wf says so and stops.

What you get

• Chain skills you already have — the runner dispatches /skill:<name> through Pi's native loader, so anything in ~/.pi/agent/skills/, <cwd>/.pi/skills/, or a settings-declared skillPaths[] is fair game.
• Every run resumes from disk — each step appends to .rpiv/workflows/runs/<run-id>.jsonl, and /wf @<ref> re-enters at the pending step, replaying already-finished branches from the trail rather than re-running them.
• Stages run detached and in parallel — each stage, and each parallel branch, gets its own child session, up to the host's maxConcurrency.
• Bad stage output never reaches the next stage — outputSchema / inputSchema are Standard Schema v1 values (Zod, Valibot, ArkType, TypeBox), checked at both seams with bounded retries.
• Routing is data-driven and auditable — gate() on numeric fields, match() on enums, defineRoute() for arbitrary TypeScript, and every no-match writes a note on the audit row instead of falling through silently.
• Loops without hand-rolled state machines — repeat a stage across many items at once, or re-run it until a TypeScript or model check says it is done.
• Team-wide skill swaps in one line — skillAliases: { commit: "attributed-commit" } rewrites the skill across every loaded workflow, built-ins included.

Configuration

Workflows are TypeScript, loaded with jiti — no build step. Project config lives at <cwd>/.rpiv/workflows/config.ts, user config at ~/.config/rpiv-workflow/config.ts; project config wins where both define the same workflow name.

The stage knobs most people touch:

Key	What it does	Default
sessionPolicy	"fresh" starts a clean session; "continue" forks the predecessor's	"fresh"
onInvalid	What to do when outputSchema rejects: "retry" or "halt"	"retry"
maxRetries	Validation retries per stage, clamped to 1–3	1
validateTimeoutMs	Ceiling on one validate call, clamped to 1 s–30 min	300000
outcome	How the stage's artifacts are collected and parsed — required on produces()	none

Config and pack files are TypeScript modules whose top-level code is evaluated on load — the same trust boundary as an npm post-install script. Diff them before running /wf in a repository you do not control.

Reference

• Workflow basics — /wf invocation forms, --name and @resume, where every file resolves, layer merging, config vs pack files, skill aliases, run state, and the glossary.
• Authoring reference — the full DSL: stage factories, loops, verify / panel, script and prompt stages, edge targets, routing, outcomes with the collector and parser catalogs, multi-input stages, validators, and the load-time validation rules.
• Embedding — host ports, the entry-point map, registerBuiltIns / registerLifecycle, the programmatic runner and resume API, cancellation, triggers, and past-run inspection.

Requirements

Node ≥ 22. No native modules, no API key, and no model selection of its own — stages dispatch through your host's session.

Related

• @juicesharp/rpiv-pi — supplies the detached SDK executor, the terminal-keystroke abort, and five bundled workflows (ship, arch, vet, polish, build).

License

MIT — see LICENSE.