Anthropic → OpenAI conversion drops reasoning_content, breaks multi-turn with reasoning models

[Biogod2020]

Description

When LiteLLM converts Anthropic /v1/messages assistant responses (with thinking blocks) to OpenAI Chat Completions format, the thinking blocks are stored in a custom thinking_blocks field. The standard reasoning_content field is not set.

This causes multi-turn requests to reasoning models to fail with:

The `reasoning_content` in the thinking mode must be passed back to the API.

Root cause

In litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py:

if len(thinking_blocks) > 0:
    assistant_message["thinking_blocks"] = thinking_blocks

The upstream API (DeepSeek, OpenAI o-series) expects reasoning_content at the top level of the assistant message dict in multi-turn conversations.

Fix

if len(thinking_blocks) > 0:
    assistant_message["thinking_blocks"] = thinking_blocks
    first_thinking = thinking_blocks[0]
    assistant_message["reasoning_content"] = first_thinking.get("thinking", "")

Steps to reproduce

1. LiteLLM proxy with Anthropic /v1/messages + LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES=true
2. Model: deepseek/* pointing to a reasoning endpoint via api_base
3. Client sends a multi-turn conversation
4. First turn succeeds, second turn fails

Environment

• LiteLLM: main (commit 867470f)
• Model: DeepSeek reasoning model via OpenAI-compatible proxy

[Biogod2020]

PR: #27947 (closed — CI requires targeting litellm_oss_branch instead of main for fork contributions, but that branch does not exist in the upstream repo yet. If you create it, let me know and I will retarget.)

Patch on branch fix/anthropic-reasoning-content at https://github.com/Biogod2020/litellm

[Biogod2020]

Related: #24985 — same root cause reported 6 weeks ago with 0 responses.

[anisen943]

Opened #28258 with a clean, isolated fix targeting shin_agent_oss_staging_05_19_2026:

• Concatenates all thinking-type blocks into reasoning_content (addresses the single-block truncation Greptile flagged on fix(anthropic): add reasoning_content when converting thinking blocks to OpenAI format #27947)
• Skips opaque redacted_thinking blocks
• Ships 2 regression tests; full transformation.py suite (70 tests) passes
• Independently reproduced via the Claude Agent SDK path with Moonshot Kimi K2.5, where it cut a representative multi-turn from 47.7 s / 4 round-trips down to 13.6 s / 1 round-trip

[tomaskir]

Still reproduces on v1.87.1, and it's not just DeepSeek's own endpoint - it also breaks via openrouter/deepseek/*, where OpenRouter forwards to a DeepSeek host (SiliconFlow, in my case) that enforces the same rule:

OpenrouterException - Provider returned error … "code":20015,
"message":"The `reasoning_content` in the thinking mode must be passed back to the API.",
"provider_name":"SiliconFlow"

So this isn't a deepseek.com quirk - any DeepSeek-family host that enforces the passback rule hits it, through any provider route. The root cause you identified is correct: the adapter sets thinking_blocks but never reasoning_content, so the replayed assistant turn carries no reasoning_content. (Confirmed v1.87.1's adapter still only sets thinking_blocks.)

Three notes toward a fix:

A minimal fix exists but isn't merged. #27947 (open, targets main) implements this, but its diff has picked up a lot of unrelated files; the minimal version - just the transformation.py change - is in #28258 (closed). A clean, minimal PR against main is probably the shortest path.

It needs to be model-aware. Per #26395, DeepSeek-R1 has the opposite requirement - reasoning_content must not be passed back - so setting it unconditionally would break R1 multi-turn. Guarding the back-fill to models that require passback avoids that, and is likely the missing piece.

Set it from all thinking blocks, not just [0] - otherwise the chain is truncated when there are multiple blocks:

if len(thinking_blocks) > 0:
    assistant_message["thinking_blocks"] = thinking_blocks
    if <model requires reasoning passback>:   # guard out R1
        reasoning = "".join(
            tb.get("thinking", "")
            for tb in thinking_blocks
            if isinstance(tb, dict) and tb.get("type") == "thinking"
        )
        if reasoning:
            assistant_message["reasoning_content"] = reasoning

I'm running this (guarded to deepseek) as a local patch and it clears the 400 for both deepseek/ and openrouter/deepseek/* multi-turn.