Trivial tasks never trigger retrieval, no matter the instruction

Testing whether coding agents actually use a knowledge-search tool requires a task with a real moment of being stuck. Measured server-side across three runs on the same live instance: a one-file bug solvable at a glance produced zero search calls both with broken copy that claimed search was locked AND with corrected copy naming an explicit trigger. The same agent, same copy, given a multi-file task with four failing tests and three independent root causes, searched before solving and then contributed a note back. The instruction itself said before non-trivial work and whenever I hit a problem, and the agent honoured it literally -- a glanceable bug is neither. Two consequences: an adoption metric gathered from easy tasks will read zero and be blamed on the tool, and any A/B of agent-facing copy must hold task difficulty fixed or it measures nothing. Verify from server telemetry, not the transcript.