dsh-reprolab

Fixing a bug is easy; proving the fix is where most teams lose the thread. A nonzero exit code is not RED. A test that never gets discovered, a syntax error, a timeout—none of these constitute a genuine failure. Without a frozen environment and pinned inputs, what you call "fix evidence" is just a story nobody can replay. ReproLab turns that narrative into a deterministic execution contract.

The core is a nine-tool pipeline: freeze the baseline, spawn isolated control and candidate worktrees, lock down the test patch and command hash, then walk three gates in order—RED (prove the test genuinely fails), GREEN (prove the identical frozen command passes after the fix), and REGRESSION (compare the candidate failure set against the recorded baseline). Every step emits structured JSON with input hashes, a state machine, and legal nextActions. Reports are regenerated from a single shared model with a recomputable SHA-256, so any party can verify authenticity without trusting prose.

It fits teams that treat "fixed" as an audit requirement: open-source maintainers who need machine-verifiable evidence before merging, CI pipelines that demand gate signals free of human interpretation, and deep Harness users who want ReproLab to close the loop with session-level discipline plugins like dsh-doublecheck—writing the test in one layer, proving the fix in the other.