LLM 0.32 Adds Pause-and-Resume to Agent Chains

LLM 0.32 lets tools pause mid-chain and wait for a human to approve before continuing, per Simon Willison's August 4 release notes. He's calling it the most significant version since the project launched, and Willison doesn't reach for that word lightly. For anyone wiring LLM into a script that touches real systems, that's the difference between a demo and something safe to leave unattended.

A tool can raise llm.PauseChain to stop a chain, and it resumes later from the pending call instead of replaying steps that already ran. Every call now carries a tool_call_id for tracking.

Prompts and responses are now typed Part objects, covering text, reasoning, tool calls, tool results, and attachments, with messages= replacing prompt= and system= arguments. Provider-hosted tools invoke straight from the command line, like -T WebSearch or -T 'CodeInterpreter(memory_limit="4g")', and llm tools -m MODEL lists what a given model supports.

The SQLite log behind every session now stores messages once, content-addressed by hash, instead of duplicating JSON on each append, per the release notes. A new message_tree view renders whole threads as indented outlines.

The quieter change is reasoning output moving to stderr, suppressible with -R. That's the fix Willison undersells. A reasoning model can now sit inside a chain of piped shell commands without dumping its thinking into stdout. That's what makes 0.32 worth adopting, not the approval prompts.