- Start with structured repository truth: the nearest manifest, Bazel owner targets, dependency metadata, and and rust-analyzer/LSP definition, implementation, and reference queries when available.
- Use targeted text search only after identifying the owning package or symbol. Do not begin with broad filesystem scans, tree dumps, or indiscriminate documentation reading.
- After identifying the owner, read its
README.md, the relevant linked design documents, and the nearest nestedAGENTS.md.
- Manifests, Bazel targets, lints, tests, and generators are authoritative. Do not duplicate their dependency, formatting, generated-code, or command rules here.
- When an architectural lint or repository check rejects a design, follow it to the owning contract. Do not bypass it with a broad allow or suppression.
- Any necessary exception must be narrow and justified beside the code.
- Change generated output through its owning source or generator, never by hand.
- Prefer names, types, structure, and tests over narration. Keep comments for non-obvious constraints or workarounds, and do not add TODOs unless requested.
- Iterate with the narrowest verification exposed by the owning package, then use the repository gates exposed by
BUILD.bazelbefore handoff.
- When practical, run the exact affected test, then the nearest owner's fast gate while iterating. Use aggregate slow only for final acceptance; run an exact slow test only while actively debugging that boundary.
- Root
./bazel test //:testis the complete non-privileged final gate. Run it once, not per slice or review. - Builders report actual evidence. Reviewers reuse it unless it is missing, stale, contradictory, or needed to reproduce a finding; orchestrators request missing owning evidence rather than duplicating broad runs.
- Final slow evidence is reusable until a repair invalidates it. At renewed final acceptance, rerun affected slow owners, including every affected browser boundary for shared changes, and reuse unaffected passing evidence.
- Privileged perf/network operations and advisory seed sweeps remain outside ordinary test routing.