What problem does it solve? Queue items and PR prose assert things about the code that may be false, and downstream review never re-checks them because the diff and the item are wrong in the same direction. This skill catches false premises before work is built on them, preventing duplicate implementations, fixes in the wrong file, and unbacked behaviour claims in shipped documentation. ## Core Features & Use Cases - Pre-work premise verification: Extracts the load-bearing claims from a queue item and checks each against the actual code with file:line citations, returning PREMISES HOLD, PREMISE FALSE, or UNVERIFIABLE. - Pre-gate prose audit: A second entry point checks the run's own rulebook prose and PR description for behaviour claims nothing backs, returning UNMEASURED with exactly two exits: delete the sentence or point it at the test that proves it. - Structured verdict block: Emits a single fenced JSON verdict with blockers, evidence, and the head SHA of the tree examined, so the calling loop can validate the report mechanically. - Use Case: A queue item says "the schema does not reject an empty title". The skill reads the validator, finds it does reject empty titles, and reports PREMISE FALSE with the citation — the task becomes a caller fix instead of a duplicate validation rule. ## Quick Start Ask the agent to check the premises of this queue item against the code before starting the failing test.
