What problem does it solve? Coding agents often declare work finished based only on their own statements, with no external evidence or recorded human decision. This Skill drives the Loopbench workflow so a task can only reach done when the accepted contract, external process evidence, and explicit human decisions are all current and recorded. ## Core Features & Use Cases - Contract-driven work loop: Drafts and revises task contracts, reconciles optional plan.toml files, and blocks implementation until the human accepts the contract revision. - Gated verification: Runs process gates via the loopbench CLI, diagnoses failures before retrying, and refuses to weaken gates or contracts to make failures disappear. - Human decision boundaries: Prepares evidence, comprehension checks, adversarial reviews, and completion reports before recording accept/reject decisions verbatim from the user. - Use Case: A researcher asks an agent to implement and validate a statistical model. The Skill isolates a worktree, drafts a contract with recovery gates, runs simulation checks, presents evidence at the human gate, and only reports completion after loopbench finish succeeds. ## Quick Start Ask the agent to work on a Loopbench task by mentioning the task ID, for example: continue task model-recovery and run its next verification gate.
