simonw/llm 0.32a3

on GitHub

Driven by the needs of Datasette Agent's human-in-the-loop ask_user() feature, made the following improvements to how tool calls work:

• Tool implementations can declare a parameter named llm_tool_call in order to be passed the llm.ToolCall object for the current invocation. This allows them to access the current llm_tool_call.tool_call_id. See Accessing the tool call from inside a tool. #1480
• Every tool call is now guaranteed a unique tool_call_id - providers that do not supply one get a synthesized tc_-prefixed ULID. #1481
• Tools can raise a llm.PauseChain exception to cleanly pause the tool chain, useful for things like waiting for human approval. The exception propagates to the caller with .tool_call and .tool_results (completed sibling results) attached, and no model call is made with a placeholder result. See Pausing a chain from inside a tool. #1482
• Failure semantics for concurrent tool execution: async sibling tool calls always run to completion before a pause or hook exception propagates. #1482
• Chains can now resume from a messages= history ending in unresolved tool calls: the calls are executed through the normal before_call/after_call machinery before the first model call, skipping any that already have results. The execute_tool_calls() method also accepts a new optional tool_calls_list= argument for executing an explicit list of ToolCall objects in place of the calls requested by the response. See Resuming a chain with pending tool calls. #1482
• Fixed a bug where the async tool executor silently dropped calls to tools not present in tools= - these now return Error: tool "..." does not exist results, matching the sync executor. #1483