Issue Station
Sign In
Tools hereToolbox

litellm - ✅(Solved) Fix [Bug]: Anthropic-to-OpenAI adapter does not round-trip thinking blocks in multi-turn conversations [1 pull requests, 1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
BerriAI/litellm#24985•Fetched 2026-04-08 02:35:14
View on GitHub
Comments
0
Participants
1
Timeline
9
Reactions
0
Author
yulangzyulangz
Participants
yulangzyulangz
Timeline (top)
labeled ×4subscribed ×2cross-referenced ×1referenced ×1

Fix Action

Fixed

PR fix notes

PR #25013: fix(anthropic): round-trip thinking blocks in multi-turn conversations

Description (problem / solution / changelog)

Summary

Fixes #24985. Two bugs in the Anthropic pass-through adapter cause thinking blocks to be lost or misformatted in multi-turn conversations.

Bug 1: Responses API path — thinking becomes output_text

translate_messages_to_responses_input() converts Anthropic thinking blocks into {"type": "output_text"} inside the assistant message. This means the downstream provider sees two output_text blocks and loses the thinking/response boundary entirely.

Fix: Convert thinking blocks to top-level {"type": "reasoning"} items with signature as id and thinking text as summary_text. This matches the Responses API spec where reasoning is a separate item, not embedded in the message.

Bug 2: Chat Completions path — missing reasoning_content

translate_anthropic_messages_to_chat_completion() sets thinking_blocks but not reasoning_content. Backends that rely on reasoning_content (e.g. Kimi) reject the second turn with: "thinking is enabled but reasoning_content is missing".

Fix: After setting thinking_blocks, also compute and set reasoning_content by joining all thinking block texts.

Changes

  • responses_adapters/transformation.py: thinking blocks → top-level reasoning items (not output_text)
  • adapters/transformation.py: also set reasoning_content when thinking_blocks are present
  • 5 new test cases covering both paths

Testing

# Responses API path
def test_thinking_becomes_reasoning_item()
def test_thinking_not_inside_assistant_message()
def test_empty_thinking_text_skipped()

# Chat Completions path
def test_reasoning_content_set_when_thinking_blocks_present()
def test_no_reasoning_content_without_thinking()

Changed files

  • litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py (modified, +10/-0)
  • litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py (modified, +16/-2)
  • tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py (modified, +40/-0)
  • tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_transformation.py (modified, +66/-0)

Code Example

{"type": "message", "role": "assistant", "content": [
  {"type": "output_text", "text": "my thinking..."},
  {"type": "output_text", "text": "my response"}
]}

---

{"type": "reasoning", "id": "", "summary": [{"type": "summary_text", "text": "my thinking..."}]},
{"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "my response"}]}

---

elif btype == "thinking":
    thinking_text = block.get("thinking", "")
    if thinking_text:
        input_items.append({
            "type": "reasoning",
            "id": block.get("signature") or "",
            "summary": [{"type": "summary_text", "text": thinking_text}],
        })

---

if len(thinking_blocks) > 0:
    assistant_message["thinking_blocks"] = thinking_blocks
    reasoning_text = "".join(
        block.get("thinking", "") for block in thinking_blocks
        if block.get("type") == "thinking"
    )
    if reasoning_text:
        assistant_message["reasoning_content"] = reasoning_text

---
RAW_BUFFERClick to expand / collapse

Check for existing issues

  • I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

When using LiteLLM to proxy Anthropic /v1/messages to OpenAI-compatible backends, thinking content from previous turns is not correctly converted back on subsequent requests. This affects both the Responses API path and the Chat Completions path.

Issue 1: Responses API path — thinking becomes output_text

In translate_messages_to_responses_input (responses_adapters/transformation.py ~line 168), Anthropic thinking blocks in assistant message history are converted to {"type": "output_text"} instead of a top-level {"type": "reasoning"} item.

What happens:

{"type": "message", "role": "assistant", "content": [
  {"type": "output_text", "text": "my thinking..."},
  {"type": "output_text", "text": "my response"}
]}

What should happen:

{"type": "reasoning", "id": "", "summary": [{"type": "summary_text", "text": "my thinking..."}]},
{"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "my response"}]}

Fix — replace asst_parts.append(output_text) with input_items.append(reasoning item):

elif btype == "thinking":
    thinking_text = block.get("thinking", "")
    if thinking_text:
        input_items.append({
            "type": "reasoning",
            "id": block.get("signature") or "",
            "summary": [{"type": "summary_text", "text": thinking_text}],
        })

Issue 2: Chat Completions path — missing reasoning_content

In translate_messages_to_chat_completion_messages (adapters/transformation.py ~line 658), assistant messages get thinking_blocks but not reasoning_content. Backends that use reasoning_content (e.g. Kimi) reject the second turn with: "thinking is enabled but reasoning_content is missing".

Fix — after setting thinking_blocks, also set reasoning_content:

if len(thinking_blocks) > 0:
    assistant_message["thinking_blocks"] = thinking_blocks
    reasoning_text = "".join(
        block.get("thinking", "") for block in thinking_blocks
        if block.get("type") == "thinking"
    )
    if reasoning_text:
        assistant_message["reasoning_content"] = reasoning_text

Environment

  • litellm 1.82.1 ~ 1.83.0, both affected
  • Tested with Claude Code → LiteLLM → qwen3.6-plus (Responses API) and kimi-k2.5 (Chat Completions)

Steps to Reproduce

as above

Relevant log output

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.82.1 ~ v1.83.0

Twitter / LinkedIn details

No response

extent analysis

TL;DR

To fix the issue with LiteLLM not correctly converting thinking content from previous turns, update the translate_messages_to_responses_input and translate_messages_to_chat_completion_messages functions to properly handle thinking blocks.

Guidance

  • In responses_adapters/transformation.py, replace asst_parts.append(output_text) with input_items.append(reasoning item) to correctly convert thinking blocks to reasoning items.
  • In adapters/transformation.py, after setting thinking_blocks, also set reasoning_content to include the thinking text.
  • Verify the fix by checking the output of the Responses API and Chat Completions API to ensure that thinking content is correctly converted.
  • Test the fix with different backends, such as Claude Code and Kimi, to ensure compatibility.

Example

elif btype == "thinking":
    thinking_text = block.get("thinking", "")
    if thinking_text:
        input_items.append({
            "type": "reasoning",
            "id": block.get("signature") or "",
            "summary": [{"type": "summary_text", "text": thinking_text}],
        })

Notes

The provided fixes assume that the issue is specific to the translate_messages_to_responses_input and translate_messages_to_chat_completion_messages functions. If the issue persists after applying these fixes, further investigation may be necessary.

Recommendation

Apply the provided workarounds to update the translate_messages_to_responses_input and translate_messages_to_chat_completion_messages functions, as they directly address the issue with handling thinking blocks.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#memory optimization#batch processing#GPU compatibility#latency issue

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

litellm - ✅(Solved) Fix [Bug]: Anthropic-to-OpenAI adapter does not round-trip thinking blocks in multi-turn conversations [1 pull requests, 1 participants]