Skip to content

The agent engineering platform.

fix(anthropic): preserve empty thinking field in signature_delta streaming (#38765)

Yuchen Fan313223347f84
diff --git a/libs/partners/anthropic/langchain_anthropic/chat_models.py b/libs/partners/anthropic/langchain_anthropic/chat_models.py
index ff43703b77..370171472e 100644
--- a/libs/partners/anthropic/langchain_anthropic/chat_models.py
+++ b/libs/partners/anthropic/langchain_anthropic/chat_models.py
@@ -627,14 +627,14 @@ def _format_messages(
if text.strip():
content.append(_format_text_block(block))
elif block["type"] == "thinking":
- content.append(
- {
- k: v
- for k, v in block.items()
- if k
- in ("type", "thinking", "cache_control", "signature")
- },
- )
+ formatted_thinking = {
+ k: v
+ for k, v in block.items()
+ if k in ("type", "thinking", "cache_control", "signature")
+ }
+ if "signature" in formatted_thinking:
+ formatted_thinking.setdefault("thinking", "")
+ content.append(formatted_thinking)
elif block["type"] == "redacted_thinking":
content.append(
{
@@ -1801,6 +1801,8 @@ class ChatAnthropic(BaseChatModel):
content_block = event.delta.model_dump()
content_block["index"] = event.index
content_block["type"] = "thinking"
+ if event.delta.type == "signature_delta":
+ content_block.setdefault("thinking", "")
message_chunk = AIMessageChunk(content=[content_block])
# Tool input JSON (streaming tool arguments)
diff --git a/libs/partners/anthropic/tests/unit_tests/test_chat_models.py b/libs/partners/anthropic/tests/unit_tests/test_chat_models.py
index 019784c250..2ee0b58ff2 100644
--- a/libs/partners/anthropic/tests/unit_tests/test_chat_models.py
+++ b/libs/partners/anthropic/tests/unit_tests/test_chat_models.py
@@ -2444,6 +2444,161 @@ def test_empty_thinking_content_block_start_emits_no_chunk() -> None:
assert block_start_event is not None
+def test_signature_delta_preserves_empty_thinking_field() -> None:
+ """A `signature_delta` chunk must include `thinking` even when empty.
+
+ When an empty adaptive-thinking block streams only a signature_delta
+ (no thinking_delta), the aggregated block must still carry the required
+ `thinking` key so the message round-trips to the API without a 400 error.
+ """
+ from anthropic.types import (
+ RawContentBlockDeltaEvent,
+ RawContentBlockStartEvent,
+ SignatureDelta,
+ ThinkingBlock,
+ )
+
+ llm = ChatAnthropic(model=MODEL_NAME) # type: ignore[call-arg]
+
+ # Step 1: process the empty content_block_start (no chunk emitted).
+ chunk, block_start_event = llm._make_message_chunk_from_anthropic_event(
+ RawContentBlockStartEvent(
+ content_block=ThinkingBlock(thinking="", signature="", type="thinking"),
+ index=0,
+ type="content_block_start",
+ ),
+ stream_usage=True,
+ coerce_content_to_string=False,
+ block_start_event=None,
+ )
+ assert chunk is None
+ assert block_start_event is not None
+
+ # Step 2: process the signature_delta with the tracked start event.
+ chunk, _ = llm._make_message_chunk_from_anthropic_event(
+ RawContentBlockDeltaEvent(
+ delta=SignatureDelta(signature="sig_abc", type="signature_delta"),
+ index=0,
+ type="content_block_delta",
+ ),
+ stream_usage=True,
+ coerce_content_to_string=False,
+ block_start_event=block_start_event,
+ )
+ assert chunk is not None
+ assert isinstance(chunk.content, list)
+ assert len(chunk.content) == 1
+ block = chunk.content[0]
+ assert isinstance(block, dict)
+ assert block.get("type") == "thinking"
+ assert block.get("thinking") == ""
+ assert block.get("signature") == "sig_abc"
+
+
+def test_thinking_delta_preserves_thinking_field() -> None:
+ """A `thinking_delta` with text is unaffected by the empty-thinking fix.
+
+ The trailing empty string from a signature_delta must not clobber
+ thinking text already accumulated from earlier thinking_delta chunks.
+ """
+ from anthropic.types import (
+ RawContentBlockDeltaEvent,
+ RawContentBlockStartEvent,
+ RawContentBlockStopEvent,
+ RawMessageStartEvent,
+ SignatureDelta,
+ ThinkingBlock,
+ ThinkingDelta,
+ )
+
+ msg = Message(
+ id="msg_thinking",
+ content=[],
+ model=MODEL_NAME,
+ role="assistant",
+ stop_reason=None,
+ stop_sequence=None,
+ usage=Usage(input_tokens=10, output_tokens=0),
+ type="message",
+ )
+ events: list = [
+ RawMessageStartEvent(message=msg, type="message_start"),
+ RawContentBlockStartEvent(
+ content_block=ThinkingBlock(thinking="", signature="", type="thinking"),
+ index=0,
+ type="content_block_start",
+ ),
+ RawContentBlockDeltaEvent(
+ delta=ThinkingDelta(thinking="Let me ", type="thinking_delta"),
+ index=0,
+ type="content_block_delta",
+ ),
+ RawContentBlockDeltaEvent(
+ delta=ThinkingDelta(thinking="think.", type="thinking_delta"),
+ index=0,
+ type="content_block_delta",
+ ),
+ RawContentBlockDeltaEvent(
+ delta=SignatureDelta(signature="sig_xyz", type="signature_delta"),
+ index=0,
+ type="content_block_delta",
+ ),
+ RawContentBlockStopEvent(index=0, type="content_block_stop"),
+ ]
+
+ llm = ChatAnthropic(model=MODEL_NAME) # type: ignore[call-arg]
+ aggregate = _aggregate_anthropic_events(llm, events, coerce_content_to_string=False)
+ assert aggregate is not None
+ assert isinstance(aggregate.content, list)
+ thinking_blocks = [
+ b
+ for b in aggregate.content
+ if isinstance(b, dict) and b.get("type") == "thinking"
+ ]
+ assert len(thinking_blocks) == 1
+ assert thinking_blocks[0]["thinking"] == "Let me think."
+ assert thinking_blocks[0]["signature"] == "sig_xyz"
+
+
+def test_format_messages_backfills_missing_thinking_field() -> None:
+ """Replaying a thinking block missing `thinking` must not 400.
+
+ Thinking blocks persisted before the streaming fix can carry a signature
+ but no `thinking` key. `_format_messages` must restore the canonical empty
+ string so the block round-trips to the API instead of failing with
+ ``400 thinking.thinking: Field required``.
+ """
+ llm = ChatAnthropic(model=MODEL_NAME) # type: ignore[call-arg]
+
+ # A corrupted block as produced by the pre-fix streaming path: signature
+ # present, required `thinking` key absent.
+ corrupted = AIMessage(
+ content=[{"type": "thinking", "signature": "sig_abc", "index": 0}]
+ )
+ payload = llm._get_request_payload(
+ [HumanMessage("hi"), corrupted, HumanMessage("continue")]
+ )
+ block = payload["messages"][1]["content"][0]
+ assert block["type"] == "thinking"
+ assert block["thinking"] == ""
+ assert block["signature"] == "sig_abc"
+
+
+def test_format_messages_preserves_nonempty_thinking_field() -> None:
+ """The backfill must not clobber real thinking text on replay."""
+ llm = ChatAnthropic(model=MODEL_NAME) # type: ignore[call-arg]
+
+ msg = AIMessage(
+ content=[
+ {"type": "thinking", "thinking": "Let me think.", "signature": "sig_xyz"}
+ ]
+ )
+ payload = llm._get_request_payload([HumanMessage("hi"), msg])
+ block = payload["messages"][1]["content"][0]
+ assert block["thinking"] == "Let me think."
+ assert block["signature"] == "sig_xyz"
+
+
def test_strict_tool_use() -> None:
model = ChatAnthropic(
model=MODEL_NAME, # type: ignore[call-arg]