## What happened

Running Claude Code against a proxy where the model is Kimi K2 served via vLLM (configured through litellm as an `openai`/`moonshot` custom provider, reached through the `/v1/messages` Anthropic-compatible passthrough), tool calls stop working partway through a multi-turn agentic session. Claude Code silently ends its turn mid-task instead of continuing to use tools, with no error surfaced anywhere in the client. vLLM's logs show a tool-call conversion warning around the same time, but the request otherwise completes normally (200, `finish_reason: stop`).

This reproduces on `v1.91.3` and is not present on `v1.83.10`, so it was introduced somewhere in that range.

## Root cause

Kimi K2 generates tool call ids in the fixed format `functions.{tool_name}:{index}` (e.g. `functions.Bash:0`). vLLM's and SGLang's `kimi_k2` tool parsers, and Kimi's own chat template, depend on this exact format: they derive the next tool-call index by parsing it back out of replayed conversation history (see the [vLLM blog post on debugging Kimi K2 tool-calling accuracy](https://vllm.ai/blog/2025-10-28-kimi-k2-accuracy) and [sglang#10600](https://github.com/sgl-project/sglang/issues/1060…