What problem does it solve? Executing a large multi-phase engineering plan by hand means repeatedly spawning agents, monitoring their progress, resolving merge conflicts, and escalating failures — a process that quickly exhausts an interactive chat's context window and demands constant human babysitting. ## Core Features & Use Cases - Autonomous phase execution: Spawns detached headless runners (Codex or Claude) per phase in parallel lanes across separate git worktrees, capped at 3 concurrent, each claiming its own task and reporting via status.json. - Two-tier orchestration: A thin interactive supervise session delegates heavy scheduling, merging, and escalation to self-recycling sub-orchestrators that measure their own token occupancy and hand off before hitting context limits. - Merge and escalation pipeline: Serializes merges into per-repository integration branches, resolves conflicts via dedicated merge-runners, and escalates stuck work through a ladder (retry → stronger model → human-in-the-loop pause). - Use Case: After publishing a multi-phase plan with the multiphase-plan skill, run the loop to have every phase implemented, verified, merged, and reviewed — ending with one draft PR per repository ready for human review. ## Quick Start Ask the assistant to run the loop on the published plan for this project and execute all phases autonomously.
