--- name: bug-hunt description: Find confirmed bugs in someone's recent research code and make it faster with identical outputs, then report through the outreach rules. Use for a lane that works on other people's code (friends, collaborators, a lab). --- The goal is something the author would want to fix or merge today: a confirmed bug with a repro and a patch, or a speedup that gives the same answers. Not a code review, and not a list of style notes. 1. **Target.** Only recent, active code: `rh bughunt target-check LANE OWNER/REPO` must PASS (a commit or release within the lane's limit, not archived, not marked unmaintained). A paper alone does not make code a target. Someone outside the project must use it too: outside issue, PR or fork accounts, stars, or downloads (`--package pypi:NAME`), or it fails as `TARGET_UNUSED`. Hunt their latest commit. 2. **Bugs.** Each bug gets a minimal repro that fails on their latest commit, a patch that makes it pass, and the effect on a result they published or would publish. Write the claims as `claims.json` (file, function, root cause, observed, expected) and run `rh bughunt claims-check LANE claims.json`: a claim a referee refuted before is refused unless the source changed or it says how it differs. Record what the round covered with `rh bughunt coverage LANE record FILE`; one pass finds about half, so a target is not done until a gap-fill round finds nothing new. The referee gets the standard prompt from `rh bughunt referee-prompt LANE claims.json WORKDIR`, which mixes in planted false claims, and writes `verdict.json` (CONFIRMED, CONFIRMED-MINOR, KNOWN, INTENDED, REFUTED, BLOCKED). Run `rh bughunt verdicts LANE WORKDIR/verdict.json`: malformed output is run again, never repaired, and a referee that confirms a planted claim is rejected whole. The claims file names the hunter (`"hunter": {"model", "family"}`, or `--hunter-family`), and a referee from the same maker's family is rejected whole too (`REFEREE_SAME_FAMILY`). Report only what it confirmed. A BLOCKED claim is never a bug in an email; its one missing fact can be the question the email opens with. 3. **Speedups.** Before starting, read their open pull requests and branches (the target check lists any about speed) and, if someone there is already optimizing, ask them first. Profile on a representative input (theirs where public) and name the hot spot with numbers. Write the change as a patch against their latest commit, describe it in `SPEEDUP.json`, and run `rh bughunt speedup-verify LANE DIR` on the compute host. It checks equal outputs (byte-identical or within a stated tolerance) and a real margin, and records the benchmark. 4. **Report.** One email per person per day, batched with `rh outreach batch`: draft with `--kind code-bug` or `--kind code-speedup`, `--target`, and `--recheck DIR` or `--speedup DIR`; lint; read; approve; send; then `sent`. Lint enforces the target check, the verified speedup, their own speed work, the contact's address and scrutiny, and the lane's required questions. The rules and their codes are in docs/outreach.md, "Bug-hunt lane". A finding lint refuses is fixed or dropped, never argued around.