maestro v0.1.0

Initial release — Claude conducts Codex.

Claude Maestro is an orchestration harness where Claude acts as the conductor (planning, judgment, verification) and Codex as the performer (implementation labor). One invocation runs the full loop: analyze → dispatch → observe → verify → rework/escalate.

Install

curl -fsSL https://raw.githubusercontent.com/animepics/maestro/main/install.sh | sh

Requires the Codex CLI with codex app-server running, and Node (≥23.6 guaranteed; 22.18+ typically works) or Bun. Then in Claude Code: /maestro "your task".

Highlights

• Acceptance-criteria-first dispatch — every Codex prompt embeds testable criteria (Goal/Do/Don't/Expected/Test shape, guide + real examples included)
• Parallel execution with real isolation — per-unit git worktree + maestro/<slug> branch, hard cap 4, deterministic merge order
• Evidence-based verification — the session's final answer is treated as a claim; git diff vs a recorded baseline + build/tests are the evidence
• Rework loop — concrete-defect rework messages to the same session (≤3 rounds), then escalation with per-criterion status
• Crash-safe — .maestro/state.json + resume/reattach; stale app-server systemError detection with troubleshooting
• Self-contained — transport CLI (WebSocket JSON-RPC to codex app-server) vendored in-repo; MIT licensed
• AGENTS.md — the contract documented from the Codex side too

All flows verified against real Codex sessions: single-unit, parallel-isolation, and rework/escalation smoke tests. See CHANGELOG.md.