What problem does it solve? Code review based on reading diffs often misses real defects because claims of correctness go untested. This Skill independently verifies a supplied change to the CG-agent-harness by actually running tests and hostile probes, producing evidence-backed PASS/FAIL verdicts instead of storytelling. ## Core Features & Use Cases - Adversarial verification: Attempts to break a supplied change with at least one negative probe (hostile argv, missing confirm, non-loopback Host, CSRF absent, jail escape) before any PASS verdict. - Surface-mapped evidence: Maps change classes (HTTP guards, shim whitelist, write-refused paths, clone sandbox, config gates, live serve, Chrome browser, desktop packaging) to the specific locking tests that prove or break them. - Structured verdict reporting: Emits per-check command/output/result blocks plus exactly one overall verdict of PASS, FAIL, or PARTIAL, with PARTIAL reserved for checks that could not run. - Use Case: After a teammate patches the harness CSRF guard, invoke this Skill to run tests/auth_guards.rs, probe a request without a CSRF token, and report a verdict with verbatim command output. ## Quick Start Ask the agent to adversarially verify the supplied CG-agent-harness change set, listing the files touched and the original task, and require executed command output for every claimed check.
