What problem does it solve? Running a multi-phase implementation plan in Claude Code normally requires an operator to open a session per chat, manually dispatch each batch, and track progress by hand. Autopilot removes that babysitting by chaining start-phase, execute-batch, and close-phase across one or more phases unattended, while keeping every run bounded, logged, and resumable. ## Core Features & Use Cases - Gated unattended execution: Requires explicit operator authorisation with a named scope and stop condition (max-phases, max-duration, or stop-at-end-of-EPIC-NN) before any subagent is dispatched, and writes a reviewable run plan that blocks dispatch if it cannot be created. - Checkpoint and resume: Records durable checkpoints after every atomic boundary, reconciles against board ground truth on resume, and distinguishes completed, halted, paused, and died-in-flight runs via named terminal boundaries. - Fail-safe guardrails: Halts on failed tests, branch drift, missing usage signals, or commit/status divergence, and requires written human acknowledgement before resuming a halted run. - Use Case: You have an epic with stories and paired testplans already planned. Authorise an autopilot run naming RUN-01 from the execution strategy sidecar, and it dispatches each chat as a fresh-context subagent, tiers models by story risk, captures usage and decisions, and stops at the declared condition. ## Quick Start Ask the AI to run the autopilot for EPIC-31 naming RUN-01 from the execution strategy with a stop condition of max-phases 2.
