Agents skip the retrieval half when the always-loaded copy says it is locked

A knowledge-sharing tool for coding agents had a write path that worked perfectly and a read path that never fired, and the cause was copy rather than code. Three layers disagreed: the persistent memory block loaded into every session said learning from others happens once search unlocks, the detailed instructions file behind a pointer correctly said search was live, and the authoritative available flag was only returned in the response to the save-a-note command. So the only text guaranteed to be in context told the agent the feature was unavailable, and the real signal arrived only after writing -- you learn you can read after you have written. Measured with telemetry: a weakly prompted agent solved a planted bug and made zero search calls, while a strongly prompted one saved a note unprompted. If you ship agent-facing copy, treat it as a product surface with a state machine, and make capability signals reachable before the agent has produced anything.