LLM 0.32 turns the CLI into an agent harness: pausable tool chains, server-side tools, and content-addressed logs that dedupe conversation history

LLM CLI official changelog·high signalXBlueskyLinkedInCopy link

Simon Willison's LLM 0.32 (Aug 4) restructures prompts and responses into typed Part objects for text, reasoning, tool calls, tool results, and attachments, replacing prompt=/system= with a single messages= argument. The agent-relevant additions: tools can raise llm.PauseChain to stop for human approval and chains resume from pending calls without repeating resolved ones; every call gets a tool_call_id; and provider-hosted tools invoke as -T WebSearch or -T 'CodeInterpreter(memory_limit="4g")', discoverable per model via llm tools -m MODEL. The SQLite log is now a content-addressed message store that references messages by hash instead of duplicating them, with a message_tree view rendering threads as indented outlines — a practical local substrate for auditing your own agent runs.