mend

Mend is a CI-oriented tool that repairs Playwright end-to-end tests broken by selector drift, such as a renamed id or class, a restructured DOM, or changed text. It uses an OpenAI tool-calling agent to investigate a DOM snapshot, propose a corrected selector, and verify the fix by re-running the test against a temporary copy of the spec file. The tool never accepts a fix without execution-based verification, and it enforces several invariants: the agent cannot weaken or skip assertions, the original spec file is never mutated, and the healing loop has a hard cap of 5 tool calls. Confidence in a fix is derived from observable signals like DOM match count and tool-call count, not from the model's self-reported certainty. High-confidence fixes are opened as GitHub pull requests, but a human always merges. Every attempt, including failures, is persisted to PostgreSQL with the full tool-call transcript for replay in a dashboard.