# Aegis

A method pack that makes AI coding agents plan against a project's real baseline before editing, prove completion with fresh verification evidence, and skip ceremony on simple tasks, across Codex, Claude Code, and other skill-aware hosts.

Aegis is a method pack that aims to make AI coding agents behave like disciplined engineers: aligning with a project's real baseline of owners, contracts, and boundaries before touching code, backing a completion claim with fresh verification evidence, covered scope, and residual risk rather than an unverified claim of done, and tracking or removing retired fallbacks and old code paths instead of letting them accumulate silently. Trivial requests are meant to stay on a fast path, with more process appearing only when a task genuinely needs it.

The same method pack works across Codex, Claude Code, OpenCode, Kimi, and other skill-aware hosts. The project reports a frozen, held-out benchmark for one release in which the same client, prompts, projects, tool policy, and model setting were used in both arms and only the Aegis method pack differed, with contract pass rate rising from 61.67% to 93.33% and unsafe outcomes falling from 13.33% to 0% across 120 runs on 20 cases; the README describes this as bounded advisory evidence from that specific benchmark rather than a universal quality claim.

Website: https://github.com/GanyuanRan/Aegis