Optional toolJudges the calls the GMI runs itself before any call of their model step runs: each call alone as a
TOOL_CALL_REQUEST chunk holding it, then the step's calls that pass alone together, as the stream's chunk for the
step holds them, each time with the sources the turn retrieved; a blocking evaluation refuses the calls it judged.
Unset when no output guardrail sets config.evaluateToolCalls.
What AgentOS hands a turn, a continuation or a resumed turn from its output guardrails.