Streaming Callbacks And Safety

Integrator wants token-level events from a complex chain but only uses .stream, which yields chain output chunks, not granular per-node events.

Use astream_events (v2) to receive granular events (on_chat_model_stream for tokens, on_tool_start/on_tool_end, on_chain_*) across the whole chain. Filter by event type/name; do not conflate the coarse .stream output with the event stream.

While streaming, the integrator tries to execute a tool call from the first chunk, but the tool_calls arguments arrive incrementally and are incomplete.

Accumulate AIMessageChunks (they add together) until the tool_calls / tool_call_chunks are fully assembled before executing. Do not act on partial tool-call arguments; wait for the complete, parsed tool_calls on the aggregated message.

Integrator subclasses BaseCallbackHandler and does blocking network I/O inside on_llm_new_token, stalling token streaming.

Keep callback handler methods fast and non-blocking; for async chains use the async handler methods. Offload heavy work (DB writes, network) outside the hot streaming path. A callback must not throw and abort the run unless that is the intent.