Claude Plan Execute runs a fixed lifecycle for every task: Plan agent writes a structured plan; a Review loop checks it (APPROVED / NEEDS_REVISION) and iterates a revise agent up to the configured round cap; only an approved plan reaches the Implement agent, which works in committed chunks with a resumable progress checklist. Role specialization is prompt-swapping, not separate binaries. A dual-backend driver lets the same code run either the claude -p print backend or a real interactive Claude TUI inside tmux; the latter keeps usage on the Claude subscription pool rather than the metered API. The project dogfoods itself: its own roadmap is the tasks.yaml it executes.
Claude Plan Execute: Autonomous Task Orchestrator
2 driver backends
3 agent roles (plan, review, implement)
01What it is
A plan→review→implement orchestrator that drives the Claude Code CLI through a multi-phase, multi-agent workflow over a declarative task slate, with quality gates, branch-per-task, cross-task memory, and a subscription-pool-preserving tmux backend.
02Engineering
03 Architecture
Roadmap (dogfooded)
tasks.yaml
Plan
Plan agentstructured plan
Review loop
APPROVED / NEEDS_REVISIONrevise agentround cap
Implement
Implement agentcommitted chunksresumable checklist
Driver backend
claude -ptmux Claude TUIsubscription pool
04 Highlights
- Every task runs a fixed lifecycle: Plan writes a structured plan, a Review loop iterates a revise agent until approval, then Implement builds it
- Only an APPROVED plan reaches the Implement agent, which works in committed chunks with a resumable progress checklist
- Role specialization is prompt-swapping, not separate binaries: the same code becomes Plan, Review, or Implement
- A dual-backend driver runs either the claude -p print backend or an interactive Claude TUI in tmux, keeping usage on the subscription pool
05Stack
PythonClaude CodetmuxYAMLpytestFastAPI
06Status
active