Repository and Subsystem Review

Read only the sections matching the diff. Load the corresponding implementation skill for implementation contracts; this reference adds reviewer-specific checks.

Apply checks to changed code and affected integrations; tie findings to contracts, not heuristics. Do not expand scope or run implementation workflows.

Repository Map

Generated files identify themselves. Review the human-authored source and confirm required checked-in counterparts are present without running generators.

Lint Rules and Assists

Load lint-rule-development. Prioritize these reviewer checks:

• A rule matching a known global must reject locally shadowed bindings.
• Cover every relevant AST union member and framework-specific value shape.
• Verify suppression and option defaults preserve existing behavior.
• FixKind::Safe must not change semantics in any reachable case.
• Applying an action must produce code the rule no longer reports.
• New rules belong in nursery; established rules remain in their declared group.
• An option needs an actual conflicting preference or semantic mode, not speculative flexibility.
• A group/path mismatch can make a test silently exercise no diagnostics.

Parser

Load parser-development.

• A presence test must not consume input before returning Absent.
• Recovery must produce a BOGUS_* node allowed by the grammar at that position.
• Recovery sets should stop at the relevant separator or enclosing boundary without swallowing following valid syntax.
• Parser behavior changes need both successful and malformed fixtures when recovery is involved.

Formatter

Load formatter-development.

• Dropped comments or tokens and idempotency failures are high severity.
• Formatting composition belongs in a type implementing Format; a cluster of free functions passing &mut Formatter has no clear owner for layout invariants.
• A multi-way layout should be selected once and represented explicitly rather than recomputed at write sites.
• fmt_fields should destructure every generated field explicitly. .. can hide a newly added token, and a bound field that is never written drops source content.
• A Prettier snapshot disappearing proves agreement only for that input. Behavior changes still need a focused internal spec.

Diagnostics

Load diagnostics-development.

• The message identifies the problem rather than restating a construct or rule name.
• Advice explains why the condition matters and gives a concrete next action when no code action exists.
• The highlighted range points to the smallest useful source span.
• Category and link registration should come from the supported generator or scaffolding workflow, not manual edits to generated registries.

Types and Module Graph

Load type-inference.

• Unknown means the system cannot prove the type; treating it as a negative answer creates false positives.
• Start from the narrowest query that can answer the caller instead of inferring an entire module by default.
• A Salsa query must track every input that can change its result.
• Cross-module data must remain reference-based so edits cannot leave copied data stale.
• Trace changed resolution paths from caller result through each query, reference, import, fallback, and owning module.
• Require semantic tests for the type shapes actually traversed. Require query-event tests only for claims about dependency or inference scope.

Code Generation and Registration

• Verify artifacts, not claims that a command was run.
• A grammar change requires matching generated syntax and factory changes.
• A formatter source change requires formatter generation for that language.
• A lint-rule change requires generated rule registration and configuration.
• Do not report bindings or other full analyzer outputs that AGENTS.md explicitly leaves to CI Autofix.
• Search parallel registration, serialization, migration, preset, and documentation sites for new enum variants, options, and manifest fields.

AGENTS.md and testing-codegen define required artifacts and CI exceptions. Flag missing or stale outputs, not merely an unchanged generated diff.

Tests and Snapshots

Load testing-codegen.

• Read snapshot diffs as behavior, not generated noise.
• Verify a regression fixture reaches the changed branch and fails without the fix.
• Confirm rule fixture paths match the declared group.
• A manually deleted snapshot is suspicious; the supported pruning workflow determines whether it is orphaned.
• Name the exact untested scenario and defect it would catch instead of requesting generic additional coverage.