Rule-File Code Review Gates

PROJECT CONTEXT:
Reviewing agent output by eye does not scale, and asking another LLM to review is probabilistic on top of probabilistic. Guardrail research shows the deterministic layer: versioned rule files (AST checks, patterns, diff analysis) gating agent-generated changes.

GOAL:
Build a rule-file gate: versioned deterministic rules evaluated against agent plans and diffs, wired into the loop and the merge path.

REQUIREMENTS:
- Rules as versioned files: pattern rules, AST-level checks, and diff-shape rules
- Gate points: post-plan (before execution) and post-diff (before completion/merge)
- Deterministic verdicts with rule IDs, locations, and fix guidance
- Severity levels: block, warn, and info - each handled distinctly
- Rule authoring workflow: incidents become rules with tests

CONSTRAINTS:
- Rules are deterministic - no LLM in the verdict path
- Every blocking rule needs a documented rationale and at least one test case

DEFINITION OF DONE:
- A diff violating a blocking rule cannot complete/merge until fixed
- Verdicts cite rule ID, file:line, and remediation guidance
- The plan gate catches policy violations before execution wastes work
- The last incident produced a new rule with a test proving it fires

COMMON FAILURES TO AVOID:
- Prose guidelines nobody enforces mechanically
- Regex-only rules missing structural violations an AST check catches
- Verdicts without locations or guidance, teaching the model nothing
- Rules accumulating without tests, rotting into false-positive noise