vfpf-orchestrate

name	vfpf-orchestrate
version	0.1.0
description	Portable orchestration skill for pi. Use when Claude and Codex should work as peer agents with bounded scopes, stable task cards, and evidence-backed handoffs. Use it to drive the orchestrator -> builder -> reviewer loop and keep all delegated work tied to plans/<feature>/ artifacts.
allowed-tools	["Read","Write","Edit"]

vfpf-orchestrate

Goal

Coordinate multi-agent work without losing scope, evidence, or ownership.

Rules

• Every delegated unit is a task card, not a conversational suggestion.
• Define objective, risk class, approval owner, scope, allowed paths, allowed tools, acceptance checks, and escalation conditions before work starts.
• Pair each active builder task with one explicit reviewer.
• Keep Claude and Codex on the same artifact model.
• Use plans/<feature>/tasks/<task-id>.md for task cards and handoffs.
• Mark Effect-touching tasks explicitly and require effect-best-practices plus plans/<feature>/effect-review-rubric.md for those loops.
• Do not merge plan notes with execution evidence.
• For regulated healthcare work, default to explicit approval and auditability instead of silent autonomy.
• Use plans/<feature>/sprint-contract.md to lock the current red-green loop.
• Use plans/<feature>/review-report.md and plans/<feature>/findings-absorption.md to track blocked review cycles.
• Assemble plans/<feature>/external-review-bundle.md before any human or external review boundary.

Required artifacts

• plans/<feature>/plan.md
• plans/<feature>/tasks/<task-id>.md
• plans/<feature>/sprint-contract.md
• plans/<feature>/review-report.md
• plans/<feature>/findings-absorption.md
• plans/<feature>/decisions.md
• plans/<feature>/verification.md

Exit criteria

• Every active task has one owner.
• Every handoff points to concrete artifacts.
• Escalations are explicit instead of implicit.