Open a URL in a real browser window owned by this server. The window is parked off-screen unless hidden=false. Returns a stage_id. Does not choose what to test.

Close a stage and its browser context.

Move a stage window from off-screen to the visible desktop and focus it.

Report URL, viewport, hidden flag, last capture time, capture staleness, and runtime error counts for a stage.

Execute the given click/fill/press/hover/select/scroll actions in order. No extra steps, no retries, no healing. Returns per-action ok/error/timing.

Take a PNG of the current stage and return it as an image block plus metadata. Reports when it was taken; it does not judge the frame.

Wait until a selector is visible or a text appears, up to timeout_ms. There is no implicit 'ready' heuristic.

Run rule-based measurements on the live page: horizontal overflow, broken images, missing alt, small tap targets, clipped text, plus collected console/page/network errors. Every rule returns its definition and threshold.

Open a URL, execute the supplied steps exactly as written, capture the requested frames, and store a run under qa/runs/<run_id>. Steps are never invented, repaired, or retried.

Compare a named capture from a run with the promoted baseline for that scenario. Returns changed pixel counts and a bounding box, and writes a diff image. Does not decide if the change is acceptable.

Persist findings supplied by the caller. Each finding needs a claim and at least one evidence reference. This stores; it does not grade, dedupe, or resolve.

Open a visible window with a drawing overlay so the user can mark parts of the app and attach notes. Returns immediately; poll annotations_list until status is ready.

List annotation sessions and their status. Marks and notes are returned verbatim.

Return an annotation bundle with each mark's element, rect, note, and the cropped images of the marked regions.

Mark an annotation as handled and record how it was addressed.