What is Pattern Learning?
Last updated
Pattern learning in code review is the ability of a review system to identify recurring code patterns, anti-patterns, and codebase-specific conventions from past reviews, and apply them in later reviews, so that fewer findings flag what the team does on purpose.
Why does pattern learning matter for engineering teams?
Every codebase has conventions that aren't written down: 'we always use guard clauses,' 'this service never makes external calls directly,' 'error handling goes through the middleware.' Generic review tools flag these as issues because they don't know the codebase's conventions. Pattern learning turns the codebase's own review history into context for later reviews.
How does Argus handle pattern learning?
Argus stores findings as review memories, and patterns separately. Patterns come from high-scoring findings (score 80 or higher on Deep Review, 90 or higher single-pass), up to 3 reusable patterns an LLM extracts per review, up to 3 conventions read from each diff's added lines, and patterns people teach with @argus-eye remember (the handle is your GitHub App's slug; argus-eye is the default). They reach later reviews through the memory briefing and through citations on matching findings, such as 'Matches a prior fix in PR #N.' When a new convention contradicts a stored one, Argus posts a PR comment and a write-access user picks which one stands. Dismissals become suppression signals. Nothing promotes a pattern to a rule; rules are written by the team in the dashboard.