Why AST evidence belongs in AI-assisted code review

Language models are useful at explaining code, but explanation is not the same as evidence. An AST-based check can establish facts about syntax and selected patterns before a model writes review prose.

Separate collection from interpretation

The static-analysis step should produce structured findings with file locations, rule identifiers, and severity. The model can then summarize those findings, suggest a fix, or ask for more context without inventing the initial observation.

Bound the workflow

A review agent needs limits on tool iterations, repository scope, output size, and provider retries. Those limits make failure visible and prevent one malformed event from becoming an unbounded background job.

Evidence is not certainty

AST checks do not prove runtime correctness, and a model summary can still be wrong. The honest contract is narrower: collect repeatable signals first, preserve their provenance, and present uncertainty where the tools cannot decide.