$ screen --status

Injection screen

Every tool result passes this screen before the agent reads it, so its cost is paid on every call. Measured latency, against the ceiling it has to stay under.

injection screen

The prompt-injection screen runs inline, as a hook, so it has a hard budget of 200 ms. At p95 the model behind it answers in 23.

Injection screen, p95 milliseconds per call, against its 200 ms ceiling.

The margin matters for a specific reason: a check that times out is skipped, and a skipped check looks exactly like a clean one. At this p95 the model is actually consulted, not bypassed.