AdversaryGate
Verify AI-generated code patches through executed test evidence, targeted mutation testing, and fail-closed validation.
AdversaryGate operates as a fail-closed verification gate for code patches written by autonomous AI coding agents, guarding pipelines against silent errors and false confidence. By running test suites across baseline and patched codebases, executing targeted mutation testing exclusively on modified lines, and verifying cryptographic diff and coverage evidence, it evaluates changes into MERGE, BLOCK, or INCONCLUSIVE states. Available as a CLI, GitHub Action, and MCP server, it ensures agentic workflows only merge changes backed by reproducible, measured execution evidence.
主な機能
014 GitHub stars
02Fail-closed decision engine providing MERGE, BLOCK, or INCONCLUSIVE verdicts with clear CI exit codes
03Line-targeted mutation testing focused strictly on code authored by the patch
04Cryptographic SHA-256 artifact verification for diffs and coverage reports
05Sandboxed execution boundaries using bubblewrap with network and process isolation
06Automated claim discovery identifying tests that exercised modified lines using coverage contexts
ユースケース
01Validating AI-generated pull requests and automated patches within CI/CD pipelines
02Empowering LLM agent workflows with a verification oracle via MCP integration
03Preventing test regressions and untested code merges in autonomous coding loops