The stated residual risk of keying on tools, made LOUD instead of silent.
If a future main-loop request ever arrives WITHOUT tool declarations it is
misrouted to the side channel, the script is never consumed, and the agent
loops on {} — a failure with no symptom except a test that mysteriously
asserts against an empty run. Every real run that reached the model at all
consumes at least one script turn, so "side-channel requests arrived and
script turns did not" cannot happen in the healthy case. Returns the warning
line, or undefined when there is nothing to say (including the run that
never reached the model at all — that has its own, visible, failure).
The stated residual risk of keying on
tools, made LOUD instead of silent.If a future main-loop request ever arrives WITHOUT tool declarations it is misrouted to the side channel, the script is never consumed, and the agent loops on
{}— a failure with no symptom except a test that mysteriously asserts against an empty run. Every real run that reached the model at all consumes at least one script turn, so "side-channel requests arrived and script turns did not" cannot happen in the healthy case. Returns the warning line, orundefinedwhen there is nothing to say (including the run that never reached the model at all — that has its own, visible, failure).