What problem does it solve? Autonomous coding agents often claim success without proof, ship stubs, or settle for partial implementations. This Skill builds a Kanban board where a separate judge process re-executes the tests itself and rejects any work lacking real evidence, so tasks only reach done when they genuinely pass. ## Core Features & Use Cases - Strict judge process: A separate judge.py re-runs pytest in the card's workspace, applies seven hard gates (tests passed, no stubs/TODO/eval, spec compliance, edge coverage, reproducibility), and defaults to REJECT with a fixed verdict block. - File-based board with browser UI: board_state.json as the single source of truth, a stdlib http.server API, and a browser UI with columns, an inline code editor, live test console, and frozen read-only test files. - Free-provider resilience: Tests make zero model calls; the loop survives rate limits via raised agent.api_max_retries, a supervisor that sleeps out recorded cooldowns, and a block-attempt counter orthogonal to throttling. - Use Case: A user on rate-limited free models wants a self-correcting coding loop driven from a browser, where cards move backlog to done only after the judge independently verifies passing tests. ## Quick Start Ask the agent to scaffold a coding Kanban board with a strict pytest judge and browser UI, then verify it end-to-end with the included verification script.
