🤝 Top Collaborators
🤓 Latest Submissions
Reprise
Problem: About 17% of bug reports can't be reproduced by the developer who picks them up (Rahman et al., 2022), and finding that out costs hours. Real bugs then still need a root cause, a fix, and proof that the fix works and broke nothing else. All of this is done by hand, and for desktop and mobile apps it is repeated on each platform. Flaky bugs are the worst case: one passing run proves nothing. Solution: Reprise is a web app that connects a GitHub repository to IBM Bob and to the developer's own machine. The developer signs in with GitHub, sees the repository's open bug reports, and starts the Reprise Runner (a single Node.js file) on their local clone, paired with a 6-digit code. Then: 1. Acknowledge a report. IBM Bob structures it, checks it against earlier reports for duplicates, reads the code and writes a reproduction test. 2. The runner runs that test 10 to 20 times with the project's own test command and returns a verdict: reproduced, reproduced sometimes (with a measured failure rate), duplicate, or one precise question for the reporter. 3. Propose a fix. Bob identifies where the bug is and why, and proposes a change, shown as a normal diff with a pull request draft. 4. Test the fix locally. The runner applies the diff in a separate git worktree, reruns the reproduction test, and runs the full test suite with and without the fix. The result is Fix verified, Fix incomplete or Regression detected. 5. Records are published to a public dashboard. Target users: open-source maintainers and app teams who triage GitHub issues, especially for apps on Windows, macOS, Linux, Android and iOS, where whether a bug reproduces depends on the platform. What makes it creative and unique: most AI coding tools stop at a plausible patch. Reprise treats the AI's answer as a hypothesis and lets tests decide. A bug counts as reproduced only when a test fails for the reported reason. A fix counts only when that test passes repeatedly and nothing else in the suite breaks. The work is split on purpose: Bob reasons online with read-only access to a snapshot of the code, and all code runs on the developer's machine in throwaway worktrees, so their working copy is never modified. Efficiency: every Bob call has cost and turn caps. Each answer is checked against a schema, with one repair attempt. Secrets are removed before anything is sent to Bob. The whole system runs on free tiers.
27 Sep 2026
