Drift Detector

Find UI inconsistency across a frontend codebase, propose a consolidation plan, and never lose an archetype on the way to consistency.

On Trigger

Start the first progress update with:

Using drift-detector ...

Design

Two layers that stay in their lanes:

1. Deterministic scanners (scripts) — emit facts as JSON. No opinions, no prose, no LLM in the loop. Reproducible on the same SHA.
2. LLM interpretation — reads the JSON + source, clusters findings into design families (buttons, cards, dropdowns, spacing scale, color scale), proposes a canonical form per family, ranks families by impact × frequency.

The agent never guesses what is duplicated. Scanners surface candidate duplicates; the agent names the family and picks the canonical form.