The repo has no guidance files for coding agents: no AGENTS.md, CLAUDE.md or .github/copilot-instructions.md on main. More contributions are being written with LLM assistance, and each contributor's agent arrives knowing nothing about the project's rules.
Scope from the Zulip discussion:
AGENTS.md at the root with the project rules (testing philosophy: e2e first, unit only when e2e is too slow or flaky; keep changes focused; no drive-by refactors; no new dependencies without an issue), and CLAUDE.md as an @AGENTS.md import so Claude, Codex and Copilot all read the same file.- Commands (
npm run verify, mix test, mix format) stay in the README, humans read it and agents read the README too.
- Per-subdirectory files for rules scoped to
server/ and ui/, so the root file stays small.
- A code-review/LLM-use guidelines file: what review should catch, and keeping the docs free of hard-to-read generated noise.
- For the Elixir side, usage_rules as the existing precedent to lean on.
A first draft of the root files and a testing skill exists from the #1033 test-setup work and can seed the PR.
The repo has no guidance files for coding agents: no
AGENTS.md,CLAUDE.mdor.github/copilot-instructions.mdon main. More contributions are being written with LLM assistance, and each contributor's agent arrives knowing nothing about the project's rules.Scope from the Zulip discussion:
AGENTS.mdat the root with the project rules (testing philosophy: e2e first, unit only when e2e is too slow or flaky; keep changes focused; no drive-by refactors; no new dependencies without an issue), andCLAUDE.mdas an@AGENTS.mdimport so Claude, Codex and Copilot all read the same file.npm run verify,mix test,mix format) stay in the README, humans read it and agents read the README too.server/andui/, so the root file stays small.A first draft of the root files and a testing skill exists from the #1033 test-setup work and can seed the PR.