What it does
Bug Reproduction Agent takes a bug context written in ordinary prose and turns it into evidence: a failing test, the minimum steps to trigger the failure, the file and line it most likely lives in, and a suggested fix. Evidence first, change second. It exists to end two specific loops. The bug that cannot be reproduced and bounces between support and engineering until it is quietly closed. And "works on my machine", which is not a disagreement about facts but the absence of a shared one. The failing test is the artefact that settles the argument, and its contract is worth reading before you buy. The emitted test must be red before the fix. When a concrete expected value can be derived from the report it is asserted directly; when it cannot, the test asserts a TODO placeholder that fails loudly until a human fills it in. It never goes green by accident — which is the failure mode that makes generated tests worthless. The confidence score is the other piece of honesty. The agent lowers its own confidence for environment-dependent bugs — anything hinging on a specific browser or a particular data state — and flags them explicitly. A reproduction tool that claims high confidence on a bug that only appears on one person's Safari with three-year-old local storage wastes more time than it saves. Each step is callable alone. `generate_failing_test` returns just the test. `repro_steps` returns the minimal sequence. `locate_affected_path` gives the file, line, triggering input and expected versus actual. `score_repro_confidence` answers "is this even reproducible" before anyone spends an afternoon on it. Everything in the plan is deterministic — the steps, the test, the affected path, the confidence, the risk signals. The language model only narrates, and the free-text report is sanitised and scanned for prompt injection before it reaches the model, which matters when the input is text a stranger wrote. No connectors and no credentials — the whole pipeline runs offline on the context you provide. Built for developers, QA and support engineers.
Example prompts
- Here is the bug report — write me a failing test for it
- What are the minimum steps to trigger this failure?
- Which file is this bug most likely in?
