LLM 0.32 records the whole agent conversation

LLM 0.32 can stream reasoning, text, tool calls and results as distinct Python events. In the CLI, reasoning is sent to standard error so it does not contaminate piped output.

The release also changes how conversations are stored. Repeatedly sending a full history used to duplicate it in the logs. A Git-inspired, content-addressable message store keeps shared messages once while allowing readable conversations to be reconstructed.

Explicit message-list inputs and tool chains that can pause for approval and resume from saved history complete the shift. An agent run has intermediate actions and interruptions worth preserving, even when the final output is a short answer.