Auto Orchestrate

name: auto-orchestrate description: Goal-based development pipeline. Classifies complexity, plans, implements, runs an automated code review (GATE 1), then waits for human approval (GATE 2). Works with any project — uses Claude Code's built-in subagent for review; no external tooling required. trigger: "auto-orchestrate" auto-orchestrate $ARGUMENTS

Orchestrate a development goal end-to-end: classify → plan → implement → gate1 → gate2.

Step 1 — Classify Before anything else, classify the goal.

Level Signal Pipeline LOW Single file, clear scope, no design decisions implement → gate1 → done MEDIUM 2–5 files, minor design, no architecture change plan → implement → gate1 → gate2 → done HIGH New system, architecture change, cross-cutting concern spec → plan → implement → gate1 → gate2 → done State the classification out loud, then proceed.

Step 2 — Spec (HIGH only) Write a short design doc (one page max):

Goal and scope Approach chosen and why Key trade-offs Stop and wait for explicit user confirmation before continuing to Plan.

Step 3 — Plan (MEDIUM / HIGH only) Write a task list using TodoWrite:

Files to create or change Tests to add or update Migration or config changes Keep it concrete. No tasks without a file target.

Step 4 — Implement Work through the plan. Mark each todo complete as you finish it.

Do not move to GATE 1 until all todos are done.

Step 5 — GATE 1 (automated code review) Spawn a code-reviewer subagent:

"Review these changed files for bugs, logic errors, and security issues: . Convergence target: 0 CRITICAL, 0 MAJOR findings."

Fix every CRITICAL and MAJOR finding. Repeat until the reviewer reports 0/0.

Never self-approve. The subagent must confirm convergence.

LOW tasks end here — report done after GATE 1 passes.

Step 6 — GATE 2 (human approval) — MEDIUM / HIGH only Present a summary:

Ready for approval What changed: <1–2 sentences> Files: GATE 1: 0 CRITICAL, 0 MAJOR ✓

Approve to proceed?

Wait for an explicit "yes" / "approved" / "go" before closing.

Rules Never skip GATE 1 for any complexity level. Never skip GATE 2 for MEDIUM or HIGH. Never self-approve at GATE 1 — subagent only. Fix all GATE 1 findings before presenting GATE 2. If GATE 2 is denied, address the feedback and restart from GATE 1.