BugRep-AI

Bugs are expensive — not because they are hard to fix, but because the path from "something is broken" to "here is a proven repair" is slow, error-prone, and largely manual. A developer receives a bug report, reads it, tries to reproduce it, digs through the codebase, writes a fix, writes a test, runs the suite, and only then raises a PR. On a busy team that cycle takes hours or days. AI code assistants have started to shorten this loop, but they stop at the suggestion stage. They propose a change without reproducing the bug first, without locking regression tests before the fix is written, and without preventing the patch from being applied unless the evidence is solid. The result is AI-generated patches that look plausible but have never been proven correct by a real test runner. BugRep-AI is a full-pipeline automated repair agent built on IBM Bob. Target users are software developers and engineering teams who want to go from a bug report to a verified, human-approved patch with minimal manual effort. A developer pastes a bug description and points BugRep-AI at a codebase — a local folder, a GitHub URL, or a ZIP. BugRep-AI runs a strict two-stage AI pipeline: 1. Investigate — IBM Bob localizes the defect to a specific function and writes regression tests that fail right now (RED). BugRep-AI runs those tests with Jest or pytest to confirm the bug is reproduced before proceeding. 2. Repair — IBM Bob receives the confirmed failure evidence and the locked test suite, then proposes a corrected file. BugRep-AI runs the same locked tests against the candidate fix. Only if they all go GREEN does the run reach human approval. The developer sees the diff, test output, and a full run report. If they approve, the fix is applied. They can share the report to Jira, Slack, or Teams with one click. What makes this novel: evidence before action (bug must be reproduced by a real test runner before any fix is attempted), locked tests (SHA-256 hashed — cannot be changed to make a bad fix look green), exactly two Bob calls per run (no hidden retries), and an approval gate (source is never modified without explicit confirmation). The result is test-proven patches in minutes with a complete audit trail.