Skip to main content
Debug with AI hands a failing test run or test case to your AI coding assistant with the scope already filled in. The assistant reads the failure from TestDino, proposes a fix in your repository, and confirms whether the fix held once a new test run lands.

Prerequisites

No MCP setup is needed beforehand. The panel connects the TestDino remote MCP server for the editor you pick, and you sign in to TestDino with OAuth (open authorization). An assistant with TestDino connected reads the test run or test case the prompt names. To set up the server by hand, see Remote MCP Server.

Where the button appears

The button stays hidden while a test run is still executing, and on a test run where nothing failed. A test run reporting passed still shows the button when it carries flaky test cases.

Hand off a failure

1

Open the failing test run or test case

Click Debug with AI. A panel opens with 3 steps: Pick your AI IDE, Connect TestDino MCP, Send the prompt.
2

Pick your AI IDE

Choose Cursor, Claude Code, VS Code Copilot, Codex app, or Other. The panel remembers the choice in this browser, so the next hand-off opens on the step you left.
3

Connect TestDino MCP

The panel shows that editor’s setup: Add to Cursor or Add to VS Code Copilot installs the server in 1 click, Claude Code and the Codex app get a terminal command to run, and Other gets your MCP config to paste. Your editor signs you in to TestDino, so there is no token to paste. In Claude Code, run /mcp to sign in. For the Codex app, the setup includes codex mcp login testdino, which signs you in. Tick I’ve connected it in Cursor when it is done and the step collapses.
4

Send the prompt

The prompt arrives filled in and stays editable. Click the send button, which names your editor (Send to Cursor, Send to Claude Code), or Copy prompt to paste into any other assistant. The instruction names the project, the test run, and the test case, so you do not describe what you are looking at.
5

Review what it proposes

The assistant reads the evidence, then proposes a change. It asks before editing a file or running a command.
6

Verify the fix

Push the change and let CI upload a new test run. Ask the assistant to verify the fix, and it compares the new test run against the one you started from.
The panel’s 3 steps: Debug with AI panel with Cursor picked, showing the Add to Cursor button, the editable prompt, and the Send to Cursor and Copy prompt buttons
The send button appears only when the prompt is short enough for your editor’s link. A prompt over that limit, and the Other choice, show no send button: use Copy prompt, with a hint such as Too long for a Cursor link. Copy it instead. After a send that opens nothing in Claude Code, the panel says Nothing opened? Send one prompt in Claude Code first so it can open links, or copy the prompt.

What the assistant reads

The assistant reads current data from TestDino each time, so a hand-off you copied earlier is never out of date. It pulls 3 things before proposing anything:
  • Whether the failure repeats across retry attempts, or passed on one of them
  • The last test run the test case passed, and the first one it failed
  • The trace, screenshots, and the expected, actual, and diff images on a visual failure
Knowing the last test run a test case passed turns “why is this broken” into “what changed between these 2 test runs”, which is a smaller question to answer. Every hand-off asks the assistant to show its proposed fix as a diff and not to commit or push it, so you review the change before it lands.

Fix a test case or check a flake

The single-test-case hand-off checks the retry evidence before proposing anything. A test case that fails identically on every attempt is not treated as flaky, and a test case that passed on retry does not get its assertion rewritten. Verdicts describe behavior, not cause. They report which fixes the evidence rules out rather than naming a root cause:

Triage a whole test run

The test run hand-off groups failures by shared cause and puts the largest group first, because one fix there often closes many test cases at once. On a large test run the assistant tells you when it is working from a sample rather than every affected test case.

Verify the fix held

This is the step that closes the loop. After a new test run lands, ask the assistant to verify the fix. It compares that test run against the one you started from and reports one of: A test case that passes only on a retry never reports fixed. That is the most common way a fix looks successful without being one.
Fix suggestions are proposals to review, not changes to apply unread. Approving one action does not approve the next, and an assistant asks again before it commits.

Debugging Basics

Evidence types and where to find them

MCP Tools Reference

Parameters for every debugging tool

Remote MCP Server

Connect the TestDino MCP server by hand

Flaky Test Detection

How TestDino identifies flaky test cases
Last modified on October 13, 2026