proof-agent

Verify AI-generated code changes by applying adversarial separation between a worker and an independent verifier. The verifier performs static analysis on the git diff to check security, correctness, and factual claims, producing a PASS, FAIL, or PARTIAL verdict with concrete file and line evidence.

Adversarial verification is a static code analysis method that prevents a worker agent from validating its own output. It spawns an independent verifier to check the git diff for security and correctness issues, ensuring code changes are backed by evidence rather than self-approval.

Review a git diff for secrets and sensitive edits by running static analysis without executing code or tests. The verifier automatically triggers when sensitive patterns like auth, permissions, Dockerfiles, or .env files appear, requiring specific file and line evidence for its verdicts.

Yes, this static analysis tool applies to pull-request and agent workflows within GitHub Actions. It enforces adversarial verification of git diffs to block unsafe AI changes, automatically triggering when enough files change or sensitive patterns appear.

No, the verifier does not require test execution to check pull request correctness. It relies entirely on static analysis of the provided git diff, reviewing security, correctness, and factual claims without executing any code or sharing the worker's self-assessment.

Proof Agent blocks unsafe AI changes by enforcing an independent verifier to review git diffs. It checks security, correctness, and factual claims using concrete evidence, requiring exactly one verdict heading to ensure changes are safe for production.