fix(anthropic): correct content block type detection when content and reasoning_content share a chunk

[Yan233th]

Problem

When an OpenAI streaming chunk carries both delta.content and reasoning_content (or thinking_blocks), _translate_streaming_openai_chunk_to_anthropic_content_block() returns "text" instead of "thinking". This is because the elif content > 0 check preceded both reasoning-related checks in the elif chain, so mixed chunks always hit the text branch first.

As a result, thinking_delta events are emitted into a text content block (SSE block type mismatch), causing thinking content to leak into the user-visible text output that Claude Code processes.

Root cause

In litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py, method _translate_streaming_openai_chunk_to_anthropic_content_block, the elif chain ordered content before both reasoning branches:

elif choice.delta.content is not None and len(choice.delta.content) > 0:
    return "text", TextBlock(type="text", text="")       # caught mixed chunks

# ... never reached when content was also present:
elif isinstance(choice, StreamingChoices) and hasattr(choice.delta, "thinking_blocks"):
    ...
    return "thinking", ...

elif isinstance(choice, StreamingChoices) and getattr(choice.delta, "reasoning_content", None):
    return "thinking", ...

A duplicate reasoning_content check at the very bottom of the chain was also unreachable for the same reason.

Fix

Reordered the elif chain to check thinking/reasoning branches before text:

tool_use → thinking_blocks → reasoning_content → content → text (fallback)

This ensures:

• Chunks with thinking_blocks (the richest signal) open a thinking block containing the actual thinking text and signature
• Chunks with only reasoning_content (e.g. vLLM/SGLang reasoning backends) open a thinking placeholder block
• Text-only chunks fall through to "text" as before

The now-dead duplicate reasoning_content check was also removed.

Related

• fix(anthropic-adapter): re-emit first delta on streaming content-block transitions #30024 — fixed first delta being dropped during streaming content-block transitions (same streaming path, different root cause)
• fix: prevent streaming thinking blocks split into multiple content_bl… #30191 — addresses eager text block creation in streaming_iterator.py (complementary fix)

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Integration test via AnthropicStreamWrapper with 3 streaming chunks (mixed content+reasoning, pure text, finish_reason):

Before fix:

Block types: ['text']

No thinking block was created — thinking_delta leaked into a text content block.

After fix:

Block types: ['text', 'thinking', 'text']

Correct block separation: thinking in its own thinking block (index 1), text in its own text block (index 2).

Type

🐛 Bug Fix

Changes

• litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py: reorder elif chain in _translate_streaming_openai_chunk_to_anthropic_content_block — priority order is now thinking_blocks → reasoning_content → content; remove now-dead duplicate reasoning_content check
• tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py: add integration test test_text_not_dropped_when_reasoning_content_shares_chunk; add direct unit test test_content_block_type_for_mixed_reasoning_and_content; add thinking_deltas assertion

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR fixes Anthropic pass-through streaming block detection for mixed reasoning and text chunks. The main changes are:

• Reasoning and thinking signals are checked before text content when opening streaming content blocks.
• Mixed content plus reasoning_content chunks now open a thinking block instead of a text block.
• Tests cover the direct block classifier and the wrapper-level streaming output.

Confidence Score: 5/5

The changes are narrowly scoped to Anthropic pass-through streaming block classification and include focused test coverage for the mixed reasoning/text scenario.

The implementation directly addresses the ordering issue described in the streaming content block classifier, and the accompanying tests cover both the direct classifier behavior and wrapper-level streaming output.

T-Rex Logs

What T-Rex did

• Ran the mixed streaming reasoning repro test and confirmed both runs exited with code 0.
• Verified the after-state shows block_types: text, thinking, text with thinking_delta and text_delta emitted in separate blocks.
• Ran the classifier-priority repro test and observed the before-state shows mixed content in a single text block, while the after-state separates content+reasoning_content with thinking and content-only as text.
• Observed that content+thinking_blocks containing both thinking and a signature raised a ValueError in the after-state.

_{Ran code and verified through T-Rex}

Comments Outside Diff (1)

1. General comment

   content+thinking_blocks with thinking and signature raises instead of returning a thinking content block

   • Bug
     • The requested contract case constructs a streaming delta with non-empty content plus a thinking_blocks entry containing both thinking text and signature. On head commit 204f6a5, _translate_streaming_openai_chunk_to_anthropic_content_block raises ValueError: Both thinkingandsignature in a single streaming chunk isn't supported. The validation objective expected this mixed thinking_blocks+content case to return block_type: thinking and a content block preserving the actual thinking/signature fields.
   • Cause
     • In the changed thinking_blocks branch of the classifier, the helper prioritizes thinking_blocks before content, but then explicitly rejects a single thinking block containing both thinking and signature before returning the Anthropic thinking block.
   • Fix
     • If this contract is intended, adjust the thinking_blocks classifier path to return a thinking content block that preserves both fields for the block-start classification case, or update the contract/tests to avoid constructing a single thinking block with both fields when that shape is unsupported.

   _{Ran code and verified through T-Rex}

_{Reviews (8): Last reviewed commit: "refactor(anthropic): order thinking_bloc..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Yan233th]

@greptileai

[Yan233th]

@greptileai

[Yan233th]

@greptileai

[Yan233th]

@Sameerlite this is ready for review, would appreciate a look. Thanks!

[Yan233th]

Greptile Summary

This PR fixes Anthropic pass-through streaming block detection for mixed reasoning and text chunks. The main changes are:

• Reasoning and thinking signals are checked before text content when opening streaming content blocks.
• Mixed content plus reasoning_content chunks now open a thinking block instead of a text block.
• Tests cover the direct block classifier and the wrapper-level streaming output.

Confidence Score: 5/5

The changes are narrowly scoped to Anthropic pass-through streaming block classification and include focused test coverage for the mixed reasoning/text scenario.

The implementation directly addresses the ordering issue described in the streaming content block classifier, and the accompanying tests cover both the direct classifier behavior and wrapper-level streaming output.

T-Rex Logs

What T-Rex did

• Ran the mixed streaming reasoning repro test and confirmed both runs exited with code 0.
• Verified the after-state shows block_types: text, thinking, text with thinking_delta and text_delta emitted in separate blocks.
• Ran the classifier-priority repro test and observed the before-state shows mixed content in a single text block, while the after-state separates content+reasoning_content with thinking and content-only as text.
• Observed that content+thinking_blocks containing both thinking and a signature raised a ValueError in the after-state.

Ran code and verified through T-Rex

Comments Outside Diff (1)

1. General comment
   content+thinking_blocks with thinking and signature raises instead of returning a thinking content block

   • Bug

     • The requested contract case constructs a streaming delta with non-empty content plus a thinking_blocks entry containing both thinking text and signature. On head commit 204f6a5, _translate_streaming_openai_chunk_to_anthropic_content_block raises ValueError: Both thinkingandsignature in a single streaming chunk isn't supported. The validation objective expected this mixed thinking_blocks+content case to return block_type: thinking and a content block preserving the actual thinking/signature fields.

   • Cause

     • In the changed thinking_blocks branch of the classifier, the helper prioritizes thinking_blocks before content, but then explicitly rejects a single thinking block containing both thinking and signature before returning the Anthropic thinking block.

   • Fix

     • If this contract is intended, adjust the thinking_blocks classifier path to return a thinking content block that preserves both fields for the block-start classification case, or update the contract/tests to avoid constructing a single thinking block with both fields when that shape is unsupported.

   Ran code and verified through T-Rex

Reviews (8): Last reviewed commit: "refactor(anthropic): order thinking_bloc..." | Re-trigger Greptile

This ValueError is pre-existing upstream logic — Anthropic SSE doesn't put thinking and signature in the same chunk. It was previously masked by the content bug; now that the ordering is fixed, it surfaces correctly. Not a regression introduced by this PR.

[Yan233th]

@Sameerlite Hey bro