Engi Autopilot

Engi Autopilot is an explicit task-custody overlay for long-running EngiFoundry engineering work.

It keeps the authorized task goal and acceptance outcome fixed while allowing the Agent to revise peripheral execution means, explore materially different continuation paths, and proceed across many task packages without low-intensity user supervision. It is intended for work measured in hours rather than a short interactive session.

Registered invocation: $engi-autopilot.

Why Engi Autopilot

Long tasks often stop for reasons that do not make the goal unreachable: an Executor behaves unexpectedly, an original file boundary proves incomplete, a preferred validation environment is unavailable, a real-time interval is inconveniently long, or an external signal cannot be produced on development hardware.

Engi Autopilot changes the default response to those conditions:

• the task goal is the sole contract invariant;
• continuous execution is a first-class strategy;
• plans, PAKs, Jobs, file lists, role assignments, tools, and validation methods remain revisable means;
• the Agent must actively generate different non-destructive paths instead of waiting or repeating an unchanged failure;
• material contract exceptions are recorded independently and disclosed at completion.

The overlay does not create another EngiFoundry execution mode or a lighter Package contract. It works alongside the applicable Router, Orch, Exec, Review, Verify, and Deliver contracts.

Restricted Activation

The overlay has exactly two activation paths:

1. Explicitly invoke $engi-autopilot.
2. Use unequivocal absolute custody language whose direct meaning is that execution must not pause before completion, for example: “除非确凿证明不可达，否则持续执行” or “执行所有任务包，若无重大阻碍绝对不要中断”.

Ordinary requests such as “continue”, “retry”, “finish it”, or “make progress” do not activate the Skill. Task size, urgency, elapsed time, and previous failures do not activate it either.

The second path requires implicit Skill discovery to remain enabled, but it is still a conscious user instruction rather than broad automatic routing.

Custody Behavior

While active, the Agent:

• turns the complete goal into durable Phases, PAKs, and Jobs at the scale the work genuinely needs;
• executes eligible work in dependency order and proceeds directly to the next unit;
• keeps enough durable state to survive context compaction and recoverable process or tool failure;
• treats progress reports as non-blocking updates rather than approval requests;
• takes over unreliable delegated execution when that preserves progress;
• expands insufficient contracts and adds causally necessary prerequisite, diagnostic, integration, repair, or verification work;
• removes reversible environmental obstacles and changes validation methods when useful;
• accelerates time- or scale-dependent evidence when the accelerated condition exercises the relevant behavior;
• substitutes or simulates unavailable external inputs at a controllable boundary;
• records material departures and reports persistent effects and unverified boundaries at closure.

Custody does not end because the task has taken hours, contains dozens of packages, exhausted its initial plan, encountered repeated failures, or lacks a preferred dependency. It ends when the goal is accepted, the user stops or replaces it, all plausible non-destructive paths are conclusively unavailable, or a host boundary has no compliant substitute.

Example Request

$engi-autopilot

Build the requested system. At minimum it must provide:
1. ...
2. ...
3. ...

Plan as many task packages as the complete goal requires and execute them in
dependency order. Unless a major evidenced blocker makes the goal unreachable,
do not interrupt execution before every requirement is accepted.

The bundled operational examples are optional references. They are not loaded into the default Skill prompt and do not turn particular tools, commands, devices, or timing values into universal rules.

Installation

Node.js 18 or newer is required. Choose either npm or GitHub installation.

npm

Install the published package outside an engi-autopilot source checkout:

npx --yes engi-autopilot install

GitHub

Install directly from the public repository:

npm exec --yes --package=github:caoyuan-fire/engi-autopilot -- engi-autopilot install

From a cloned source checkout:

npm run skill:install -- --host codex

Target hosts

Without target options, the installer presents detected hosts for selection. Use explicit targets in non-interactive environments:

npx --yes engi-autopilot install --host codex
npx --yes engi-autopilot install --host codex,claude
npx --yes engi-autopilot install --host antigravity
npx --yes engi-autopilot install --all
npx --yes engi-autopilot install --target /custom/skills

Default locations:

Codex       ~/.codex/skills/engi-autopilot
Claude      ~/.claude/skills/engi-autopilot
Kimi Code   ~/.kimi/skills/engi-autopilot
Antigravity ~/.gemini/config/skills/engi-autopilot

Manage an installation

npx --yes engi-autopilot check --host codex
npx --yes engi-autopilot update --host codex
npx --yes engi-autopilot uninstall --host codex

Add --dry-run to preview changes. Replacing or removing an unmanaged installation requires --force.

Project Structure

bin/engi-autopilot.mjs               Multi-host npm installer
skills/engi-autopilot/
├── SKILL.md                         Custody overlay contract
├── agents/openai.yaml               Skill interface and invocation policy
└── references/
    ├── continuity-patterns.md       Optional non-normative analogies
    └── exception-records.md         Append-only exception record contract
tests/test-installer.mjs             Installer test suite
package.json

Development

Run the focused checks and inspect the publishable package:

npm test
python3 /path/to/skill-creator/scripts/quick_validate.py skills/engi-autopilot
npm pack --dry-run

Status

Engi Autopilot is an early 0.1.0 release. Its custody, divergence, exception-recording, and progressive-disclosure contracts are available; multi-hour behavior still depends on the continuation and process-lifetime capabilities of the active Agent host.

License

This project is licensed under the Apache License, Version 2.0. See LICENSE for details.