task-orchestrator

name	task-orchestrator
description	Pick a task from docs/backlog, implement it using dev-skill, and verify with a QA subagent. Use when you want to autonomously complete a backlog task end-to-end for team michele-luca's Exercise One.

task-orchestrator

Autonomous task-completion agent for team michele-luca's budgeting app. Reads a pending task from docs/backlog/, implements it via dev-skill, then spawns a QA subagent with qa-skill to verify the result.

When to use this skill

• User says "pick a task and implement it"
• User wants end-to-end automation: read → code → verify
• User wants to delegate the full dev cycle to Claude

How it works

1. Read task: Scan teams/michele-luca/excercise_one/docs/backlog/ for pending tasks. Pick one (or the user specifies which). Task files should be markdown with frontmatter containing at least status: pending or similar.

2. Implement: Use the dev-skill skill to write code, start the dev server, drive the app with Playwright, and screenshot the result. Follow the domain model and architecture in CLAUDE.md.

3. QA handoff: Spawn an Agent with subagent_type qa-agent (or similar) that invokes the qa-skill skill. The QA agent runs all gates (lint, typecheck, build), reviews the diff against the checklist, and reports pass/fail.

4. Update task: If QA passes, mark the task completed (e.g., update frontmatter to status: completed and add a completion timestamp). If QA fails, report the issue and optionally iterate.

Prerequisites

• Tasks must exist in docs/backlog/ as markdown files with frontmatter (status, title, etc.).
• dev-skill and qa-skill must be available (they are).
• Node 24, npm 11, and deps installed (npm install from teams/michele-luca/excercise_one/).

Usage

# From Claude Code CLI or chat:
/task-orchestrator
# or with a specific task file:
/task-orchestrator task-001-add-transaction-form.md

The skill is designed to be invoked by Claude Code's Skill tool — it does not run as a standalone script.

Task file format (expected)

---
id: task-001
title: Add transaction input form
status: pending
priority: high
created: 2026-05-26
---

# Task: Add transaction input form

Allow users to input a new transaction (amount, category, date). Store in
localStorage via the storage layer. Display in a transaction list below the form.

## Acceptance criteria

- Form has fields: amount (cents), category, date
- Submit button adds transaction to storage
- Transaction list re-renders showing the new entry
- Type-safe: Transaction type from src/types/

The orchestrator reads the task description, implements it, and updates the frontmatter status field when done.

QA subagent

The orchestrator spawns an Agent with a prompt like:

Run QA on the current diff for task-001. Use the qa-skill to run lint, typecheck, and build. Review the diff against the money-correctness checklist (integer cents, no floats, storage layer abstraction). Report pass/fail.

The subagent has access to qa-skill and reports back. If it fails, the orchestrator decides whether to fix and retry or escalate to the user.

Gotchas

• Tasks must be well-scoped. A vague task ("improve the UI") will produce vague code. The skill works best with concrete acceptance criteria.
• No commit by default. The orchestrator implements and verifies but does NOT commit unless explicitly told to. You can extend it to auto-commit green QA runs if desired.
• Dev server lifecycle. The dev-skill starts the server in the background (npm run dev &). The orchestrator should ensure it's running before implementing and optionally kill it after QA (or leave it running for the next task).
• Task status tracking. This skill assumes tasks have a status field in frontmatter. If your tasks use a different tracking system (e.g., GitHub issues, Linear), adapt the read/update logic.

Extension ideas

• Auto-commit + push when QA passes
• Slack/webhook notification on completion
• Multi-task mode: process all pending tasks in priority order
• Fallback to user on QA failure after N retries