autopilot

Autonomous iterative loop for the GitHub Copilot CLI — a standalone TUI that re-feeds the same prompt to copilot -p ... until the agent declares completion or a hard cap fires.

What it is

autopilot is a standalone TUI app that drives the GitHub Copilot CLI through autonomous iteration loops. You arm it with a prompt — or pick one of three baked SDLC modes (--self-improve, --grow-project, --prompt) — and the driver re-spawns copilot -p ... every iteration until either the agent emits a configurable completion phrase, you call autopilot run --stop <runId>, or --max is hit.

Out-of-band control is first-class: --pause, --resume, --stop, and --status can all be invoked from a second terminal, write to the run’s state.json under a CAS-protected lockfile, and never kill an in-flight copilot -p subprocess mid-iteration. Plain mode is zero-dep — only Node ≥ 20 is required to drive a loop, tail events, list runs, or replay history. The Ink-rendered watch UI for autopilot watch and autopilot run adds Ink + React + Yoga + Commander; install it via cd packages/tui && npm install when you want the live TTY view.

!!! tip “Bare autopilot runs --self-improve --fresh” Running the binary with no arguments is the same as autopilot run --self-improve --fresh. It starts a fresh self-improvement loop on the current repo against the baked prompt — no flags to remember on the happy path.

Why a loop?

autopilot drives a coding agent through three levels of decomposition, then loops:

  1. Level 1 — find what to do. Scan the repo (red CI, stale PRs, open issues, SDLC hardening rotation) and pick one concrete work item. If nothing is queued, ideate a new feature and file it as a GitHub issue first, then pick it up.
  2. Level 2 — split into stages. Plan the SDLC stages this work item needs — generated per work item, not a fixed pipeline. Typical shape: orient → critique → baseline → implement → test → commit → push, but a doc fix may skip several and a feature may add more (acceptance, demo, close).
  3. Level 3 — split into tasks. Break each stage into the smallest executable steps the agent can deliver in one turn.

Orchestrate the tasks until every stage is delivered, then loop back to Level 1 for the next work item. Each iteration is a brand-new copilot -p subprocess, so context never poisons the next pass. The driver owns the spawn / capture / decide-to-stop trichotomy plus out-of-band control, structured event logs (events.jsonl), CAS-protected run state, completion / abort / stagnation triggers, and an adaptive iteration budget for --self-improve.

Repository

This site is open source. Improve this page.