# Self-healing tests: what works, what is marketing

> A test that quietly rewrites itself to pass is not healing, it is lying. What a repair should do, what it should never do, and the question to ask a vendor.

Published 2026-10-08, about FlowQA.

A self-healing test is a good idea with a dangerous default.
What works: when a test fails on a page that changed, a tool proposes one small change to the test, replays the changed test on a fresh session, and waits for a person to approve it.
What is marketing: a test that quietly picks a new element and goes green, so a real regression and a renamed button look the same.

## The failure self-healing is meant to fix

Most failures after a deliberate change are not bugs.
A label was renamed, a button moved into a menu, a step was added to a flow.
The product is fine; the test is out of date.
Fixing each one by hand is the maintenance cost that kills suites, and it is the cost self-healing promises to remove.

## Where the quiet version goes wrong

Say a test clicks "Place order".
After a change the button is gone and a "Continue" button sits where it was.
A quiet healer finds the nearest match, clicks "Continue", and the test passes.
Was the order placed?
Nobody knows, and the green says yes.

The same mechanism hides real regressions.
If a required field disappears and the healer skips the step that filled it, the test passes on a page that no longer collects the field.
A suite that can heal its way past a bug is a suite you cannot trust, and a suite you cannot trust is not worth running.

## What a repair should do

**One small change.**
A new target for one step, a new value, a corrected assertion, or a new path to the same outcome.
Not a rewrite.

**Replayed before proposed.**
The repaired test replays on a fresh session, more than once, and only a repair that is green each time is shown.
A repair that works once is a coincidence.

**Shown with its reason.**
The step before and after, and why the model chose the change.
"The button labelled Place order is now labelled Continue inside the same form" is a reason a person can check at a glance.

**Approved by a person.**
Nothing changes until someone says yes.
Rejection leaves the test as it was.
The approval is the moment a human confirms the product changed on purpose.

**Declined when no repair fits.**
If the outcome the test asserts is simply gone, the honest answer is that the test cannot be repaired, and that is a finding about the product, not the test.

## A worked example

*A composed example: the team, the timings and the counts show the shape of the work and are not measurements from one customer.*

A team renames "Save" to "Save changes" across their settings pages.
Fourteen tests fail that night.
Their tool proposes fourteen repairs, each the same shape: the click target changes from the old label to the new one, replayed green twice on fresh sessions.
A reviewer opens the list, sees the same reason fourteen times, and approves them together, because one check covers all fourteen.

A week later one test fails again on the same page.
This time the tool declines: the assertion "the saved banner appears" cannot be satisfied because no banner appears at all.
That one is a bug, filed with the screenshot, and it would have been invisible under a quiet healer.

## The question to ask a vendor

"Show me a repair being refused."
If every failure can be healed, the tool is not testing anything.
