What problem does it solve? Running a Terminal-Bench task repeatedly until it passes often devolves into unbounded retries, silent stalls, and untracked code changes. This Skill imposes a bounded, auditable loop: each iteration runs a smoke test against an isolated Paperclip App worktree, diagnoses the exact stop point, and requires board confirmation before any product fix is implemented. ## Core Features & Use Cases - Bounded iteration loop: Runs a single Terminal-Bench task through Paperclip with an explicit iteration budget, wall-clock caps, and defined stop rules (pass, board rejection, budget exhaustion, or named blocker). - Structured issue topology: Organizes the loop as a tree of Paperclip issues — a top-level loop parent, iteration children, diagnosis documents, and board confirmation interactions — wired with blockedByIssueIds dependencies. - Worktree isolation: Pins the benchmark command to an isolated Paperclip App worktree via PAPERCLIPAI_CMD and preserves the Harbor dispatch runner config so every iteration tests the same workspace where fixes land. - Use Case: An engineer is asked to "drive Terminal-Bench fix-git until it passes." The Skill opens a loop issue, runs a smoke, diagnoses that the benchmark issue was never assigned a heartbeat-enabled agent, proposes a product rule, obtains board approval, applies the fix in the same worktree, and reruns until the smoke passes. ## Quick Start Ask the agent to run the Terminal-Bench task fix-git through Paperclip in a bounded loop with an iteration budget of three, diagnosing each stop and requesting board confirmation before any product fix.
