GanyuanRan/Aegis

Aegis – Method‑Pack for Trustworthy AI Coding Agents

What it is – Aegis is a method pack (a collection of prompts, configuration files, and helper scripts) that can be added to a variety of AI‑code‑generation hosts such as Codex, Claude Code, DeepSeek Harness, GitHub Copilot, etc. It does not ship a new model or a daemon; instead it adds a disciplined workflow layer that makes the host behave more like a careful software engineer.

Why it matters – AI coding agents are powerful but can produce unsafe or incomplete changes that need a human to review. Aegis tries to reduce that “babysitting” by:

• Planning against the real project baseline before any edit is made.
• Requiring fresh verification evidence (tests, contract checks) before declaring a task “done”.
• Keeping simple requests on a fast path while automatically adding ceremony only for risky work.
• Tracking retired code paths to avoid hidden technical debt.

Key claims (backed by the repo’s own benchmark)

Metric	Without Aegis	With Aegis
Contract‑pass rate	61.7 %	93.3 %
Unsafe outcomes	13.3 %	0 %
These numbers come from a frozen A/B benchmark (120 runs, 20 cases) using a gpt‑5.6‑sol / xhigh setting. The README stresses that the evidence is advisory and not a universal guarantee.

How you use it

1. Install – Give your AI coding agent a single prompt that tells it to read the repo, detect the host you are using, and run the host‑specific install command (e.g., dsh plugin --profile <profile> add "git+https://github.com/GanyuanRan/Aegis.git").
2. Verify – Run the supplied scripts/aegis-doctor.py script; a successful JSON output contains "ok": true, "workspaceSupport": "available", and "configStatus": "configured".
3. Activate – The pack defaults to automatic activation and TDD off. You can switch to explicit activation or auto‑TDD with the same doctor script.
4. Interact – Once installed, you talk to the agent as usual, but you can add Aegis‑specific triggers:
   • Aegis goal: … – defines scope, success evidence, and boundaries.
   • Grill me … – starts a single‑question decision interview.
   • TDD Route: strict / test‑first – forces a test‑first workflow.
   • aegis:update – updates the method pack via the host‑aware path.

Supported hosts – The pack ships installation guides for many popular code‑generation tools. Some hosts have only a guide (no fresh smoke test yet), while others (Codex, OpenCode) already have benchmark evidence.

Development & testing – The repo includes a suite of end‑to‑end tests (tests/e2e/*.sh) that verify installation, workflow quality, and host‑specific compliance.

Community – Discussions, issue tracking, and a DEV.to article are linked from the README. The project is MIT‑licensed and builds on the earlier Superpowers framework.

Bottom line – If you already use an AI coding assistant and want a plug‑in‑style way to make it behave more like a disciplined engineer—adding planning, verification, and risk‑aware routing—Aegis provides a ready‑made, host‑agnostic method pack to do that.