[Hye-Dev]

Enforces upstream review on changes to files that could modify upstream policy.

Additional note on AI policy alignment (changes not drafted, requesting feedback):

To be context-efficient with AI tooling, the durable rules of CONTRIBUTING.md were loaded into as-needed skills and the always needed AGENTS.md rules were placed in skill descriptions (which are always in context), with further information embedded to guarantee alignment when loaded.

I would recommend keeping this structure. Keeping CONTRIBUTING.md targeted for humans and the skills for AGENTS guidance maintains better alignment and prevents bloating of the context window. Many of the additional points not yet covered (e.g. test rules) can also be embedded into skills to save context.

Folder-specific rules (like .github/) I think should be moved to directory-local AGENTS.md files.

TLDR: Some things still certainly belong in AGENTS.md, but things that should offer a signal for further context or are path-specific could save context by being placed in other structures and achieve the same effect.

[stgraber]

@Hye-Dev So I'm not sure that I really buy the context saving here.

AGENTS.md is the file that every tool is basically guaranteed to read from the start and include in their context. Skills are lazy loaded based on their description, so they work well to give instructions on the "how" of implementing what's required in AGENTS.md but shouldn't themselves be the policy as they may be loaded too late or not at all.

So I'd say all AI/LLM rules should be in AGENTS.md, then the .claude/skills and .agents/skills stuff can be extended with skills that make AI agents save time for specific common tasks.

I agree the folder specific AGENTS.md make sense and could save context but we're not at a point where it's worth it as it's currently a single instruction for a single directory. Once we get to having an actual list of directories with multiple instructions, it will naturally make sense to move that over to a per-directory AGENTS.md.

One real bit of duplication is between the guidelines in the agent commit-message skill and the human facing equivalent in CONTRIBUTING.md. The duplication there means that content will easily get out of sync as I for one will never think to go update the guidelines.md file deep inside the agent stuff. We should make CONTRIBUTING.md authoritative for this stuff.