Keep the First Token: content_block_start in Anthropic Streaming

What it does

When streaming Claude responses through LangChain, the content_block_start event was silently dropping the initial text chunk of the assistant message. langchain-anthropic 1.4.8 preserves the initial text that arrives in the first content_block_start callback, so the model no longer appears to jump in partway through its own sentence.

Why it matters

Streaming is what makes LLM apps feel alive. Users expect tokens to appear progressively, not in sudden bursts. If the opening words vanish, the output reads as truncated or incoherent right at the start, which is especially jarring for longer generations. Any app using stream_events, astream, or plain stream with langchain-anthropic benefits from this fix.

Example

$ Stream a Claude reply with astream_events and confirm the first word is present.

Full response: Quantum entanglement is a phenomenon where two particles become
interconnected so that the quantum state of one instantly influences the other,
regardless of the distance between them.

Common flags

ChatAnthropic.stream()

Synchronous streaming, affected by the fix

ChatAnthropic.astream()

Async streaming, affected by the fix

astream_events()

Low-level event stream; on_chat_model_stream now includes the full initial chunk

ChatAnthropic.invoke()

Non-streaming invocation, not affected