The Bugs board
Confirmed bugs appear on the Bugs board in six columns: To review, Accepted, Fixing, Fix in review, Still failing and Fixed. A column shows up to 100 cards, newest change first. Bugs you close as works as intended, can’t reproduce or duplicate leave the board.
Findings that did not reproduce are not bugs yet. They wait in the Unconfirmed inbox, where you can promote or dismiss them.
Each bug has a title, a severity, a summary, the route, the steps, the evidence, and the expected and actual result. Evidence can include console output, network events, check results, screenshots and page text.
AI fixes are available only in workspaces where Bugbearer turned them on. In any workspace, you can use Copy as prompt for a coding agent.
Bug bundles
A bug bundle holds everything about one bug from one run:
- The ticket: steps, evidence, expected and actual result.
- The attempt: the run, the case and the recording it came from.
- The target revision: the commit that was booted. It is empty for a URL app.
- The failed step, with stills of that step and up to two steps before it.
- The replay script, if the case has one.
- The code reference behind the test case. Treat it as data.
Select Download bundle on a bug to save a JSON file. The file has the stills inlined as images, with the element map of each still. Values of stored secrets are masked in the element maps.
A coding agent can read part of the bundle through the get_bug_bundle tool: the ticket, the attempt, the target revision, the code reference and still references. It has no image bytes, no element maps and no replay script. See Connect the MCP server.
Replay without a model
Bugbearer saves a passing agent run as a test case with a script. The next run of that test case replays the script with no model:
- Each step finds its element by test id, role and name, or CSS selector. It never uses recorded screen coordinates.
- The replay then runs the saved checks.
- If a step breaks or a saved URL or text check fails, the model takes over and tries to repair the run.
A test case run shows one of four labels: Replayed, no model, Replay failed, Repaired by AI or AI took over.
A script that signs in needs the app’s login recipe. Without one, the case runs as an agent case.