maestro-plan

Multi-agent codebase exploration automates this by running parallel discovery agents across architecture and integration points, then synthesizing findings into a verified execution plan with task definitions and convergence criteria.

A CSV-based source-of-truth workflow uses a tasks.csv file to orchestrate multi-agent explorations, enforce append-only discovery boards, manage concurrency controls, and produce deterministic plan.json outputs for phase-based software planning.

You generate task definitions by running a two-phase pipeline where parallel exploration agents append discoveries to a board, then a sequential planning agent consolidates findings into TASK-*.json files with grep-verifiable convergence criteria.

Yes, multi-agent orchestration supports concurrency controls through spawn_agents_on_csv, allowing you to scale parallel exploration agents and manage gap-mode handling during the codebase discovery phase.

Plan validation enforces grep-verifiable convergence criteria, read_first requirements, and a revision loop to ensure the sequential planner produces a high-quality, actionable execution plan from parallel agent discoveries.

Automated phase-based planning uses a two-wave pipeline where Wave 1 runs parallel exploration agents for discovery and Wave 2 runs a sequential planner to emit plan.json and task files, ensuring verified plan generation without manual synthesis.