TabProof

TabProof helps web developers turn keyboard interaction failures into reproducible evidence and verify a correction. Give Meta Muse Spark a URL and a business task: it explores with keyboard presses and text entry, while TabProof records actions, screenshots and focus. The action boundary is enforced in code. When the agent reaches a configured dialog, TabProof extracts its executed route and replays that exact prefix in fresh browsers on both versions. A read-only checker measures dialog closure and focus restoration independently of the model’s completion claim. Route divergence, API errors and interrupted runs remain inconclusive. Developers can start/stop runs, inspect before/after evidence and download an offline report. In three new DispatchDesk pairs, all six real Meta sessions reached the target. First Escape failed in all three original sessions and passed in all three corrected sessions. Separate reproductions failed 9/9 before and passed 9/9 after. Original agents eventually recovered; that did not erase the first failure. DispatchDesk is a synthetic workflow with a seeded regression, not a claim of general accessibility conformance. The repository also contains attributed local reproductions of known React Aria and Radix keyboard issues. My original contributions include the restricted agent runtime, route capture/replay, independent checks, workbench and evidence exports. Development dates and third-party attribution are retained in Git and PROVENANCE.md.