Skip to main content

Heal bindings when the interface changes

Healing is model-free relocalization: when no candidate of a binding resolves, every element now on the page is scored against the recorded fingerprint, and a clear winner, above the threshold and clear of the runner-up, is proposed. A run that only passed because a binding was healed is healed, never passed, and a repair reaches the store only through a diff a person reads.

Inline, in a Playwright project​

YAM_MODE=heal npx playwright test

On a LocatorError the fixture relocalizes, retries the step once against the proposal, annotates the test healed, and stages the repair in .yam/heal-proposals.jsonl. The store is untouched.

As a step, in a flow project​

yam run --host playwright || yam heal --run "$(cat .yam/last-run)"
yam heal --from-bind-failures            # from a Playwright project's .yam/
yam heal --run <id> --apply              # write the accepted repairs

yam heal reads the failures of a run, relocalizes each, verifies the proposal by resolving it live, and writes a diff and a report. --apply writes the repairs to the store; without it nothing changes, which is the right default for CI. --no-model is the default in every published number; a model may be consulted for re-grounding when a gateway is configured, and then the repair says so in its provenance.

Reading the numbers​

The healing claim is published with its method in reports/eval-healing.md: over twenty deliberate interface changes to the sample application, relocalization alone recovers 92.3 percent of the bindings that degraded, with zero repairs onto the wrong element. A repair counts only when the proposed element carries the same ground-truth key as the one recorded, and that key is hidden from the healer. Regenerate it with pnpm eval:healing.

What healing does not do​

It never invents an element. If nothing scores above heal.relocalizeThreshold (default 0.72) with a margin of heal.margin (default 0.10) over the next candidate, the failure stands and the report says why. A control that was removed stays a failure, which is what a test is for.