fix(anthropic/adapter): emit thinking block for reasoning_content-only streaming chunks

[Anai-Guo]

Summary

Fixes #29518. On the /v1/messages Anthropic-shape streaming path, chain-of-thought from OpenAI-compatible reasoning backends (vLLM/SGLang reasoning parsers — DeepSeek-R1, Qwen3-reasoning, gpt-oss, etc.) is silently dropped: the visible answer is preserved but no thinking content block / thinking_delta reaches the client.

Root cause

The streaming content-block classifier _translate_streaming_openai_chunk_to_anthropic_content_block decides the content_block_start type. It only recognizes thinking_blocks. These backends populate Message.reasoning_content and explicitly set thinking_blocks=None, so for a reasoning-only chunk the classifier falls through to the default text block.

Meanwhile the delta translator _translate_streaming_openai_chunk_to_anthropic already has a reasoning_content fallback and correctly returns a thinking_delta. The net effect in streaming_iterator.py: a content_block_start of type text is opened (block type never changes, so no new block is started), and the thinking_delta events are emitted inside that text block. Anthropic streaming clients (Claude Code, SDK .stream()) ignore thinking_delta outside a thinking block, so the reasoning is discarded.

The non-stream translator (_translate_openai_content_to_anthropic) and the streaming delta translator already fall back to reasoning_content; only the content-block classifier was missing the same fallback.

Fix

Add the matching reasoning_content fallback to the classifier so it opens a thinking block when thinking_blocks is absent. With the classifier and the delta translator now agreeing, the iterator emits a proper thinking content_block_start followed by thinking_delta events.

Test

Adds test_translate_streaming_openai_chunk_to_anthropic_reasoning_content_only_content_block, which feeds a reasoning_content-only streaming chunk (thinking_blocks=None) and asserts the classifier returns a thinking block. The existing thinking-blocks tests still cover the Claude-backed path.

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing Anai-Guo:fix/anthropic-adapter-reasoning-content-thinking-block (fe12408) with main (5be0797)}

[codecov]

Codecov Report

❌ Patch coverage is 80.00000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...perimental_pass_through/adapters/transformation.py	80.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fixes a silent drop of chain-of-thought reasoning on the /v1/messages Anthropic-shaped streaming path for OpenAI-compatible backends (vLLM/SGLang) that populate reasoning_content instead of thinking_blocks. The content-block classifier was missing the same reasoning_content fallback that the delta translator already had, causing thinking_delta events to be emitted inside a text block and ignored by Anthropic streaming clients.

• Adds an elif branch to _translate_streaming_openai_chunk_to_anthropic_content_block that opens a thinking block when thinking_blocks is absent/empty but reasoning_content is present, making the classifier consistent with the delta translator.
• Adds a unit test that feeds a reasoning_content-only chunk with thinking_blocks=None and asserts the classifier returns a thinking block type.

Confidence Score: 4/5

The change is narrow and well-targeted, touching only the content-block classifier's missing reasoning_content fallback and adding a companion test.

The fix correctly aligns the classifier with the delta translator for the described use case. The only minor inconsistency is the truthiness check vs. is not None guard — if a backend emits an empty reasoning_content string on the first chunk, the classifier still opens a text block while later delta events would carry thinking content, reproducing the original mismatch in that edge case.

transformation.py — specifically the truthiness check on reasoning_content at the new elif branch.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py	Adds a reasoning_content fallback in the content-block classifier so it opens a thinking block instead of the default text block when vLLM/SGLang-style backends supply reasoning_content with thinking_blocks=None.
tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py	Adds one targeted unit test exercising the new classifier branch with reasoning_content set and thinking_blocks=None; uses mocks only, no real network calls.

_{Reviews (1): Last reviewed commit: "fix(anthropic/adapter): open thinking bl..." | Re-trigger Greptile}

[greptile-apps]

The truthiness check for reasoning_content (getattr(..., None)) will evaluate to False for an empty string, so a chunk where reasoning_content="" and thinking_blocks=None still falls through to opening a text block. The delta translator at line 1506 uses is not None and accumulates "", ultimately also returning text_delta for empty content — so there's no mismatch in that specific case. However, if a provider sends an empty reasoning_content on the first chunk (to signal block start) and non-empty content on subsequent chunks, the classifier would incorrectly open a text block while the delta translator would emit thinking_delta. Mirroring the delta translator's is not None guard here makes the two codepaths consistent.

Suggested change

	elif getattr(choice.delta, "reasoning_content", None):
	return "thinking", ChatCompletionThinkingBlock(
	type="thinking", thinking="", signature=""
	)
	elif getattr(choice.delta, "reasoning_content", None) is not None:
	return "thinking", ChatCompletionThinkingBlock(
	type="thinking", thinking="", signature=""
	)

[Sameerlite]

@Anai-Guo Can you please get this to 5/5?

[Anai-Guo]

@Sameerlite pushed a fix. The failing check was my own new test — root cause: Delta.__init__ does del self.thinking_blocks when it's unset, so hasattr(delta, "thinking_blocks") is False for reasoning-only chunks and my nested branch was unreachable in exactly the vLLM/SGLang case. Moved it to a sibling elif (mirroring the delta translator's existing reasoning_content handling), so the content block is now classified as thinking. The other red check (test_token_counter::test_tokenizers) is an unrelated HF tokenizer-download flake.

🤖 Generated with Claude Code

[mateo-berri]

This has been merged via internal copy. Thank you for the contribution!