The chain opens by recording the complete engineering context: the system prompt the agent operates under, its tool manifest, the active policy pack, the pinned model snapshot, and default sampling parameters. Everything downstream is interpretable — and replayable — only because this is on the record first.