Prove-patch gives an independent agent the baseline commit, the final patch commit and the declared repair scope. It reviews the planned cases and the patch without the patcher’s conclusion, runs every case on clean baseline and patched trees with one harness, and keeps raw output as receipts.
It then runs the project’s relevant checks and a three-arm helper: the baseline must fail, the patch must pass, and a full revert must fail again. The result is a prove record with a manifest, listing findings, evidence gaps and unproven limits.
What You Get
- A prove.json record covering every planned case by unique check id
- Raw stdout, stderr, exit status, timing, command and harness source for each baseline and patched run
- A three-arm result with source receipts that pin the commits and show the reverted tree equals the baseline tree
- A list of findings, evidence gaps and specific unproven limits to return to the orchestrator
Common Use Cases
- Independently checking a patch written by another agent
- Confirming that a fix covers a bounded class of inputs and keeps valid behavior
- Deciding whether a failed attempt needs a new patch round
What Makes It Unique
The prover must be a different agent from the patcher, may not edit the patch, and refuses to validate an uncommitted tree. It needs at least two distinct failing representatives and a valid control for a bounded class, and an empty list of unproven limits leaves the result inconclusive.
Under The Hood
The agent reads a record contract and runs bundled Python scripts through uv: a three-arm helper that builds the pinned baseline, patched and reverted trees in temporary Git worktrees, a native proof assessor, and a record validator that writes manifests. It runs the project’s build and test commands and a retained harness, and writes receipts and records into the run directory. It does not edit the patch and makes no network calls of its own.