A bare /skill:<name> used to expand the skill block and then append "The user invoked this skill with no task text - ask what they want before executing any protocol inside it", so a prompt-only skill never ran its body. The floor is gone: expand_skill_submission builds the core session, expands through it, and returns the block unchanged, so the bare invocation runs as the prompt. Invocations with task text are unchanged. The four agent_engine skill-seam tests now assert the bare expansion is the block alone (no trailing user message), the with-args row keeps the args as the trailing message, the batched bare row admits on the expanded block, and a close-tagged body still expands.