What problem does it solve? Delivery claims like "it works" or "tests pass" are not proof. This Skill makes an AI builder self-certify its work: it authors or discovers a verification plan, exercises the real product on the correct surface, captures concrete evidence (screenshots, logs, recordings), and publishes an immutable acceptance round that a reviewer can accept or reject per check. ## Core Features & Use Cases - Plan authoring or discovery: Write your own checks when no verify plan exists, or consume a frozen plan via a LobeHub operation id; rounds without an operation are recorded as standalone acceptances. - Surface-matched verification: Pick the cheapest proving surface per outcome — CLI output for backend logic, agent-browser for web and Electron, osascript/screencapture for native macOS, and AXe/simctl for iOS Simulator. - Evidence enforcement: Checks declaring requiredEvidence cannot pass on text alone; missing artifacts hold the delivery at uncertain, and programmatic gates (tests, lint, type-check) are rejected as acceptance checks. - Immutable multi-round history: Repairs publish as new rounds with supersedes lineage, preserving provenance of reused evidence. - Use Case: After fixing a bug where a web app's dashboard fails to load new cards, use this Skill to drive the running app with agent-browser, capture a screenshot and network trace proving the cards render, and publish an acceptance URL the reviewer can inspect. ## Quick Start Verify the task I just completed by exercising the changed behavior on the appropriate surface, capturing evidence for each check, and publishing an acceptance round with the lh CLI.
