The Problem
Anthropic's Claude Code reads CLAUDE.md. OpenAI's Codex reads AGENTS.md. Both files describe the same thing — "how this repo wants to be edited" — but each vendor evolved its own filename and conventions independently.
If your team uses both tools (and many do, even unintentionally), you have a fork problem. Either the two files drift, or one tool gets a worse experience because somebody forgot to update its file, or you copy-paste between them and one quietly goes stale.
You also have a CI problem. Both vendors teach you to write the conventions file. Neither teaches you the gates that catch the failure modes their own tools are most likely to introduce: hallucinated dependencies, mock-only tests that pretend coverage, fabricated APIs, deprecated patterns the model still loves.
The Core Insight
The conventions belong in one file. The two readers (or three, or five) consume that file via a small adapter.
The same pattern applies to CI: the gates that matter for AI-augmented code are mostly tool-agnostic, but they need to be wired in once and run on every PR — especially the AI-authored ones, which are the majority for fast-moving teams in 2026.