Reproduce-bug takes one finding and a local baseline checkout and tries to make the bug happen before anything is edited. For a runtime bug it builds the project and runs a proof of concept against the built artifact. For a compile failure it records the target compiler diagnostic from the failed invocation.
The agent saves raw stdout, stderr, exit status and timing, writes a reproduce record with the outcome (reproduced, not reproduced or blocked), updates the finding status and writes a manifest of the whole step directory.
What You Get
- A reproduce.json with baseline commit, command argument arrays, artifact path, observed result, outcome and agent identity
- An unchanged PoC saved under reproduce/poc for later stages
- Raw receipts for the build and each run
- A manifest and a record check before handoff
Common Use Cases
- Confirming that a reported runtime crash or path-escape happens on the current baseline
- Capturing the exact compiler diagnostic for a reported build failure
- Reporting a blocker when setup cannot exercise the claim
What Makes It Unique
A unit test that imitates the bug does not count. A runtime PoC must record that it loaded the built artifact and print PATCH_REACHED right before the assertion, and a nonzero exit alone is not accepted as proof. A compile case needs the target diagnostic.
Under The Hood
The agent reads SKILL.md and a record contract, reads project guidance files from Git objects, runs the project’s own build command and the PoC it writes, and saves outputs into a run directory. It uses a bundled Python helper through uv to write the manifest and check records. It writes files only inside the run directory, and its commands are those of the project under test.