Root-cause analysis and fixes

Every failure gets a record

Healed, unhealed, or suppressed, every transaction produces a typed RcaRecord: a clean human-readable message, the root cause, the evidence collected, the actions attempted, and a suggested permanent fix. For unhealable failures, an RCA agent enriches the message — turning a raw TimeoutError stack into "the login button id changed from login-button to signin-btn."

Test-change context

When the test source is in a git repository, heal collects the failing line's last-modified date and recent history. That lets the RCA distinguish "test outdated" (the locator's line last changed 14 months ago; the app moved on) from "application changed" or "application defect" — a causality hint no single-run view can give.

Fix blast radius

A heal can become a permanent code change, but where the change lands decides how dangerous it is. The fix engine resolves the locator's origin through the RF parsing model and classifies the blast radius:

flowchart TD
    H[healed locator] --> O{origin?}
    O -->|literal at call site| LOCAL[local]
    O -->|variable / prefix-variable| VAR{used at<br/>N sites?}
    O -->|user-keyword argument| CALL[trace to call sites]
    VAR -->|1| LOCAL
    VAR -->|many| SHARED[shared]
    CALL -->|1| LOCAL
    CALL -->|many| SHARED
    LOCAL --> AUTO[eligible for in-place]
    SHARED --> REVIEW[patch / review only,<br/>never auto-applied]

A locator stored in a ${VARIABLE} used by 14 keywords is shared — heal will never auto-edit it in place, because that would change 14 call sites at once. Instead it produces a patch and lists every usage so you (or a coding agent) can decide. A literal at a single call site is local and safe to apply.

This is the difference between a trustworthy autofixer and one that breaks the other 13 tests. See Fix test files for the application tiers.