Post-patch-validation tests a security fix after it exists and before it is accepted or reported as fixed. The agent pins the vulnerable base and the patched input, scaffolds a plan, fills in checks, and a bundled Python runner executes them on both revisions in isolated Git worktrees.
The plan must hold at least one check of each kind: control, exploit, root-cause variant, behavior, regression, security and suite. The runner saves raw logs and writes a result and a report with findings and validation gaps, and passing checks still call for human review.
What You Get
- A pinned plan.json scaffolded from base and patched refs or a patch file
- A result.json with an assessment holding status, findings, gaps and a human review flag
- A report.md plus raw stdout and stderr for each check
- Archived scratch output and hashed helper files for review
- Findings named by check ID and assertion, ready to return to the patch author
Common Use Cases
- Validating an AI-generated security patch before human review
- Looking for missed variants when a fix covers only one exploit path
- Checking that a remediation commit keeps legitimate behavior byte-identical
- Producing a fresh validation run after the patch author revises
What Makes It Unique
Exploit and variant checks must print a PPV_REACHED marker on stdout before the assertion and run side-blind, so a broken harness or a check that can see which revision it runs on cannot fake a pass. Commands are argv arrays and the environment is fixed and stripped, with credential-like and code-changing variables refused.
Under The Hood
The agent reads SKILL.md and an evidence model reference, and runs one bundled Python script through uv with scaffold, validate-plan and run commands. The runner creates locked Git worktrees for the pinned revisions, executes the plan’s argv commands under a minimal fixed environment, and writes logs, hashes and reports to an output folder. It runs the repository’s own code and tests with the caller’s privileges and is not a sandbox, so untrusted check code needs a container. Submodules come from local Git objects and it makes no network calls of its own. A Claude workflow in the plugin can drive it.