What problem does it solve? AI agents often stop working because they are tired or looping, not because the task is actually done. This Skill makes "done" verifiable by requiring completion gates written before work starts, where each gate is only counted as passed when a command was executed, exited with code zero, and produced the expected output. ## Core Features & Use Cases - Executable Completion Gates: Write a GATES.md ledger before work begins, with each gate declaring a CHECK command, an EXPECT output, and an EVIDENCE field recording exit code, shell, and output hash. - Four Explicit Abort Reasons: Declares budget exhaustion, impasse (repeated tool calls with no progress), unrecoverable error, and human interruption as the only non-acceptance exits, each requiring a written report. - Anti-Tautology Rules: Prevents gates that cannot fail (e.g., echo ok), requiring negative controls, direct measurement, and command-by-command review of inherited ledgers as untrusted input. - Use Case: Before starting a multi-hour refactoring task, define gates like "pytest -k reserva passes" and "ruff check passes", then only declare completion when each command has been re-executed and its output verified. ## Quick Start Ask the agent to write completion gates with CHECK and EXPECT commands before starting the task, and to only mark the task done after re-running each gate command and recording the evidence.
