Check for existing issues

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

What happened?

When using LiteLLM with the Moonshot provider (moonshot/kimi-k2.6) for multi-turn tool calling, the Moonshot API returns a 400 Bad Request error because reasoning_content is missing from the assistant tool call message in subsequent turns.

This is the same issue as #21672, which was fixed by PR #23580 — but the fix only covers kimi-k2.5, kimi-k2-thinking, and kimi-k2-thinking-turbo. The newly released kimi-k2.6 model is not included in the model_prices_and_context_window.json file, so supports_reasoning returns False and fill_reasoning_content() is never called.

Kimi K2.6 has thinking/reasoning enabled by default. The Moonshot API requires that reasoning_content from the assistant response be preserved and sent back in the conversation history during multi-step tool calling. The official API documentation (https://platform.moonshot.ai/docs/api/chat) describes the thinking parameter with a keep field for K2.6 that controls whether reasoning_content from historical turns is preserved:

Controls whether reasoning_content from previous turns is preserved across a multi-turn conversation, i.e. whether to enable Preserved Thinking. Defaults to null, meaning historical thinking is NOT preserved.

• "all": Preserves reasoning_content from historical turns and provides it to the model as part of the context, enabling Preserved Thinking. When using this, keep the reasoning_content from every historical assistant message in messages as-is.

Steps to Reproduce

1. Configure moonshot/kimi-k2.6 in LiteLLM proxy config
2. Make a multi-turn tool calling request via /v1/messages (Anthropic format)
3. On the second turn (after tool_result), Moonshot API returns 400

Relevant log output

Error: MoonshotException - thinking is enabled but reasoning_content is missing in assistant tool call message at index 2
  (HTTPStatusError: Client error 400 Bad Request for url https://api.moonshot.cn/v1/chat/completions)

What part of LiteLLM is this about?

SDK (litellm Python package), Proxy (LiteLLM Proxy Server)

What LiteLLM version are you on?

v1.83.10

Suggested Fix

Add moonshot/kimi-k2.6 to model_prices_and_context_window.json with "supports_reasoning": true so that the existing fill_reasoning_content() logic in MoonshotChatConfig applies to this model as well.

Additionally, K2.6 introduces a new thinking.keep parameter (vs K2.5's simpler thinking.type) that should be mapped from Anthropic's cache_control / thinking params.

Workaround

Use a custom callback that intercepts outgoing HTTP requests and injects reasoning_content: " " into assistant tool-call messages that are missing it, as described in the comments of #21672.