# Fix Tests with AI

> When a test fails, use Fix with AI to get a proposed patch you can review and apply—plus a fix-suggestion lifecycle that classifies whether the failure is a bad test or a product bug.

Source: https://totalshiftleft.ai/help-center/product-documentation/fix-tests-with-ai

## Overview

**Fix with AI** repairs a saved test on demand. It analyzes the test (using its last failed run when available), then proposes a patch with a **classification**, a **confidence** score, and a plain-language **explanation** shown as a per-field before/after diff. Nothing is written until you apply it. When the failure looks like a real API defect, the platform deliberately proposes *no* patch so the failure stays visible. Explore the capability at [AI test fix](/features/ai-test-generation/ai-test-fix).

## Before you begin

- The test must be **saved** first—**Fix with AI** only appears for existing tests. On a new test you'll see *"Save the test before using Fix with AI."*
- Fix with AI is available to all users for saved tests; it's gated only by the `AI_TEST_FIX_ENABLED` flag (on by default) and a valid license.
- Nav path: open a test in the test editor (from the endpoint tests table, click **Edit** on a row). The **Fix with AI** control sits at the bottom-left of the editor's action bar.

## Step 1 — Invoke Fix with AI

1. In the test editor, click the purple **Fix with AI** button (sparkle icon). While it works the label reads **Fixing…**.
2. To repair only part of the test, click the chevron next to the button and choose a target: **Fix everything**, **Fix data**, **Fix body**, **Fix parameters**, **Fix assertions**, or **Fix authentication**. The main button runs **Fix everything**.
3. A modal titled **Fix with AI** opens showing *"Analyzing the test and proposing a fix…"* while the suggestion is generated.

## Step 2 — Review the proposed patch

When the analysis returns, the modal shows a meta row and the diff:

| Element | What it tells you |
|---------|-------------------|
| **Classification chip** | **Test defect**, **Data issue**, **Possible product bug**, **Environment issue**, or **Flaky**. |
| **Confidence: N%** | How confident the model is in the proposed fix. |
| **Used last failed run** / **Spec-based (no run)** | Whether a prior failing run was used as evidence, or the fix is spec-only. |
| **Explanation** | A sentence describing what changed and why. |
| **Diff** | Per field: the `path`, the old value (red **−**) and new value (green **+**), with an optional reason line. |

If the classification is **Possible product bug** (or otherwise "do not fix"), the modal shows a warning instead of a diff: *"This looks like a real API defect, not a test problem. The test was left unchanged so the failure stays visible. Investigate the API rather than editing the test."* No Apply action is offered in that state.

## Step 3 — Apply or discard

At the footer of the modal:

- **Apply to form** — loads the patched values into the editor so you can review, then save normally with **Save Test**.
- **Apply & Save** — applies the patch and saves the test in one step.
- **Discard** (or **Close** when there's no patch) — dismisses the suggestion without changing anything.

Identity fields (test ID, endpoint, endpoint ID, project ID, feature ID, API type) are never overwritten by a patch.

## Step 4 — Handle a surfaced pending suggestion

When you open a test that already has a pending automated suggestion, the same **Fix with AI** diff modal appears with its classification, confidence, and diff. Applying it is resolved server-side (authoritative apply that also marks the suggestion **applied**), after which the editor closes so the list refreshes.

If the test was edited after the suggestion was created, the stored patch is **stale**: applying it returns HTTP **409** with *"This test was modified after the suggestion was created. Re-run \"Fix with AI\" to get a fresh suggestion."* (code `SUGGESTION_STALE`), and the suggestion is marked **superseded**. Re-run **Fix with AI** to get a current one. Applying or dismissing a suggestion that is already resolved also returns 409 (*"Suggestion is already applied/dismissed"*).

## Step 5 — Suggestion lifecycle and metrics

Fix suggestions are tracked so a team can triage them across a project:

- **List** — pending suggestions per project (or per test); a badge marks tests that have one.
- **Apply** — writes the patched test back and is audit-logged (recorded as an `ai_fix_suggestion_apply` update).
- **Dismiss** — closes a suggestion with an optional reason (up to 500 characters); captured as a learning signal.
- **Metrics** — totals by status and classification, **acceptance rate** (applied ÷ reviewed), and how many suspected **product bugs** were surfaced.

Applying and dismissing suggestions are learning signals that feed [Continuous learning insights](/help-center/product-documentation/continuous-learning-insights).

## Troubleshooting

- **"Save the test before using Fix with AI."** — the test isn't saved yet; save it, then retry.
- **409 "This test was modified after the suggestion was created…" (SUGGESTION_STALE)** — the patch is stale; re-run **Fix with AI**.
- **Warning instead of a diff** — the failure was classified as a product bug or environment issue and no safe patch was produced. Route product bugs to engineering rather than editing the test.
- **Error toast (bottom-right)** — the repair request failed; dismiss it and try again, or check AI configuration in [AI Settings](/help-center/product-documentation/ai-settings).

## Best practices

- Read the classification first: a *Possible product bug* means the test may be correct and the API is wrong—investigate the API.
- Prefer applying high-confidence patches; review low-confidence ones closely before saving.

## Related articles

- [Generate endpoint test cases](/help-center/product-documentation/ai-generated-tests)
- [Continuous learning insights](/help-center/product-documentation/continuous-learning-insights)
- [Test case](/help-center/product-documentation/test-case)
- [Test run](/help-center/product-documentation/test-run)