### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

Server-side MCP tool auto-execution (`require_approval: "never"`) works when the underlying model routes through a cloud provider, but **silently does not execute** when the underlying model routes through `ollama_chat/`. The model's tool call is returned raw to the client with `finish_reason: "tool_calls"` instead of being executed by the proxy's auto-exec loop.

This makes local Ollama models unusable as the base for proxy-managed agents (`litellm_agent/` models with MCP tools attached): chat UIs like Open WebUI receive an unexpected tool call for tools they never sent and render an empty assistant message.

### A/B evidence (same request, only the base model differs)

Agent defined via `litellm_agent/<base>` with an MCP server attached (`require_approval: "never"`), prompt requiring a web search:

| Base model route | Result | |---|---| | `anthropic/claude-haiku-4-5` | tools executed server-side, `finish_reason: "stop"`, final grounded answer ✅ | | `ollama_chat/gemma4:12b` (`api_base` = Ollama) | `finish_reason: "tool_calls"`, raw `tool_call…