turn-conduct

What problem does it solve? AI coding agents often end turns prematurely: declaring intent instead of doing work, summarizing someone else's step, stopping on exploration, or making claims about the repository state without running a verifying command. This Skill defines the four lawful ways a turn may end and the guard-enforced checks that keep every turn honest. ## Core Features & Use Cases - Turn exit discipline: A turn ends only with a question the rules cannot answer, a guard refusal, a filled context window without compaction, or handed-in work with the next task already begun. - Statement verification: Every claim about the repository state in a reply must be backed by a command run in the same turn and its output. - Guard integration: Maps each rule article to concrete hook scripts (turn-exit guard, waiting guard, claim guard, window-fill guard) that enforce behavior mechanically rather than relying on memory. - Use Case: While an epic runs, the agent finishes a task, opens a PR, and immediately starts the next task instead of ending the turn with a status summary — because the waiting guard refuses a turn that only reports on someone else's step. ## Quick Start Load the turn-conduct rule before ending any work turn so the reply follows the four lawful turn exits and every claim about the tree is backed by a same-turn command.