The priority order is explicit: human readability comes first; coding-agent traceability is the minimum gate. A human should understand code in one pass, and an agent must be able to trace a feature from CLI flag to executed branch, tensor/record, metric, and test without reconstructing hidden control flow.
- Over-complex or hard-to-follow code is a bug, not a style issue.
- Reduce complexity and line count — prefer deleting code over adding it.
- No over-encapsulation: a helper needs 3+ real call sites AND nontrivial logic; never wrap trivial code or add classes/files for one call site.
- Do NOT remove features, performance knobs, or observability in the name of simplicity — knobs default ON stay ON.
- A "bug" that cannot trigger under the shipped recipes is not worth fixing.
Details: .claude/skills/simplicity-first (invoke before any code change).
Concise "why" only, 2-4 lines, written for an external reader: no job ids, commit hashes, single-run metrics, or internal cluster paths; keep upstream issue/PR links.
- Reviews report findings only; fixes ship as one minimal diff per issue after approval.
- Verify with
python -m compileall -q openrlhf examples testsandpython -m pytest -q; shell scripts withbash -n.