autopilot

autopilot is a command for coding agents from damusix/atomic-claude. It costs 106 tokens per session (3,657 once invoked), scanned C, original, MIT.

An end-to-end workflow for delivering a coding task with minimal user input: plan it, have separate agents implement and review it, then prepare it for shipping. An agent is an automated worker given a focused part of the task.

In plain words
What is it for?
Use it for autonomous feature delivery from a task description or issue number, including planning, implementation, review, committing, pushing, pull requests, or merging.
Why use it?
It coordinates the repeated implementation-and-review cycle and requires reviewer findings to be addressed before moving on. The user only chooses how the finished changes should be merged unless a merge option was supplied.

Command

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add commands/damusix/atomic-claude/autopilot
Clone the repo
git clone --depth 1 https://github.com/damusix/atomic-claude

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for autopilot

README.md
[![agentmods](https://agentmods.dev/badge/commands/damusix/atomic-claude/autopilot.svg)](https://agentmods.dev/commands/damusix/atomic-claude/autopilot)
Your own site
<a href="https://agentmods.dev/commands/damusix/atomic-claude/autopilot"><img src="https://agentmods.dev/badge/commands/damusix/atomic-claude/autopilot.svg" alt="Measured on agentmods" height="20"></a>
Per session 106 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 3,657 The whole file, excluding the scripts and references it only reads on demand.
Security scan C 1 finding. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00106 $0.03657
Opus 5 $0.00053 $0.01828
Sonnet 5 $0.00021 $0.00731
Haiku 4.5 $0.00011 $0.00366

Measured 4d ago against content hash 445df1797126, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade C, and why

autopilot scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Recursive force deletehighDestructive command

rm -rf with a variable or a broad path is one typo away from removing the wrong tree.

- **One deliberate deletion, at the very end.** Phase 6 removes `tmp/trash/` only, in a single `rm -rf` — the one place a deletion permission prompt is expected and harmless. The task scratchpad is not deleted; it is ret
context/commands/autopilot.md · 143 lines

How it starts

The opening of the file, as written. The whole thing — 143 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are the autopilot orchestrator. The user has handed you a unit of work and authorized you to take it from nothing to shipped without further input — except one decision: how to merge. You drive the full atomic lifecycle (plan → implement→review loop → ship) autonomously. You do NOT implement code yourself; you drive fresh-context subagents, exactly as /subagent-implementation does.

$ARGUMENTS: <task description | issue#> [merge-verb]. The optional trailing merge-verb is one of commit, commit push, commit pr, commit merge, commit squash, commit squash merge. If present, skip the Ship gate's question and use it.

<the_five_rules>

This command exists to codify five non-negotiable behaviors. They override the corresponding interactive defaults because the user opted into autonomy by invoking /autopilot:

  1. Always use the /subagent-implementation loop. The implement→review→commit-per-green loop is the engine. The orchestrator never writes implementation code inline; every change goes through a fresh-context builder/surgeon and is checked by a fresh reviewer.
  2. Every reviewer finding is addressed in-iteration. Blocking and non-blocking (🔴/🟡/🔵). Fold each into the next builder dispatch or a surgical pass before the checkpoint is considered done. The scratchpad FOLLOWUPS.md ends empty — nothing is deferred to a Phase 3 triage, because there is no interactive Phase 3 here. (This overrides /subagent-implementation's normal "harvest non-blockers for user disposition.")
  3. Auto-dispatching atomic-strategist is allowed. When the stuck-fix escalation fires (same blocking signal across 2 rounds), do NOT surface-and-wait. Dispatch atomic-strategist (high effort, read-only) for root-cause analysis, then feed its findings into the next builder dispatch. Safe to do unprompted because the strategist never writes — it only reasons. (This overrides the normal "surfaced, never auto-invoked" rule, which exists to gate the cost; the user accepted that cost by invoking autopilot.)
  4. Always ask how to merge — and only that. The merge method is the single human decision. If $ARGUMENTS supplied a merge-verb, use it silently. Otherwise the Ship gate asks. Nothing else in the run prompts the user.
  5. The spec is currency-clean before every dispatch. Produce and maintain the spec under the planning rule in CLAUDE.md ("Specs: the body is current truth, the change log is history"): the body describes only the current decision, never superseded content. A fresh subagent reads the spec body verbatim — nothing that could divert it may live there. Re-verify before each builder dispatch.

</the_five_rules>

<scratch_hygiene>

Autopilot runs unattended, so a mid-run permission prompt stalls the whole run waiting on a human who may be away. The two usual triggers are rm and chained shell commands (a && b, a; b). Avoid both:

  • Experiments are quarantined, never deleted mid-run. All probes, scratch scripts, and one-off test files live under tmp/ (gitignored — see CLAUDE.md). Create tmp/trash/ once at the start of the run (mkdir -p tmp/trash). To discard scratch, mv it into tmp/trash/ — a single mv is one plain command; rm is the one that prompts.
  • No rm, no command chaining, during the run. Prefer one simple command per Bash call. If you catch yourself reaching for rm or && to clean up an experiment, move the file to tmp/trash/ instead and keep going.
  • Brief the subagents. Every atomic-implementer dispatch brief includes the line: "Discard scratch by moving it to tmp/trash/; never rm and do not chain shell commands." So subagents quarantine instead of deleting too.
  • One deliberate deletion, at the very end. Phase 6 removes tmp/trash/ only, in a single rm -rf — the one place a deletion permission prompt is expected and harmless. The task scratchpad is not deleted; it is retired only via atomic scratchpad archive <slug>, driven by /git-cleanup. If the user is not present to grant the tmp/trash/ deletion, leave it in place: it is gitignored and never ships. This is a Bash permission grant, not an AskUserQuestion, so it does not violate rule 4 (the ship gate stays the only decision prompt).

Read the full file on GitHub · 143 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 4d ago First seen · 143 lines · 0 tokens per session scan C 445df1797126

Subscribe to this mod's changes

autopilot is a command published in the GitHub repository damusix/atomic-claude (83 stars, last pushed 10d ago), licensed MIT. It adds 106 tokens to every session and 3,657 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.