fix(vllm): add tool support detection so agent mode sends tools parameter

[rodboev]

Fixes #5508

Summary

Agent mode with the vLLM provider never executes tools: the model replies with plain text or raw tool-call syntax leaks into chat. The request sent to vLLM never includes the tools parameter, so the model has nothing to call, regardless of how well the server supports OpenAI-compatible tool calling.

Root cause

PROVIDER_TOOL_SUPPORT in core/llm/toolSupport.ts has no vllm entry. modelSupportsNativeTools() (core/llm/toolSupport.ts:513) returns false for any provider absent from the map, so gui/src/redux/thunks/streamNormalInput.ts:116 sets useNativeTools = false, routes the request through the system-message code-block framework, and never populates completionOptions.tools. The downstream streaming path (fromChatCompletionChunk in core/llm/openaiTypeConverters.ts, addToolCallDeltaToState in gui/src/util/toolCallState.ts) already reassembles streamed tool_calls deltas correctly; it is simply never reached.

Changes

• core/llm/toolSupport.ts: add a vllm entry to PROVIDER_TOOL_SUPPORT, delegating to the existing ollama model heuristic (the same delegation pattern lmstudio uses), since vLLM commonly serves the same model families (Qwen3, Llama 3.x, DeepSeek, Hermes)
• core/llm/toolSupport.test.ts: new describe("vllm") block covering Qwen3, Llama 3.1, DeepSeek (true) and a non-tool model (false)

What this doesn't change

• The SSE chunk reassembly path: fromChatCompletionChunk and addToolCallDeltaToState are untouched
• Vllm.ts: it inherits the OpenAI adapter path, which correctly sends tools once populated
• Other providers' PROVIDER_TOOL_SUPPORT entries
• The capabilities: [tools] config override remains the escape hatch for unusual vLLM deployments

Checklist

• I've read the contributing guide
• The relevant docs, if any, have been updated or created
• The relevant tests, if any, have been updated or created

Screen recording or screenshot

N/A — no UI change.

Tests

• cd core && npm test -- llm/toolSupport.test.ts — 78/78 passing (73 existing + 5 new). Covers: vLLM provider tool-support detection for Qwen3 (the model cited in the issue), Llama 3.1, and DeepSeek model names, plus the negative case for llama2.

---

Summary by cubic

Fixes agent mode with vllm so tool calls work. We now detect tool support and send the tools parameter, enabling proper tool execution.

• Bug Fixes
  • Added vllm to PROVIDER_TOOL_SUPPORT, reusing the ollama heuristic so compatible models (Qwen3, Llama 3.1, DeepSeek) include tools in requests.
  • Added tests for Qwen3, Llama 3.1, DeepSeek, and a negative case for Llama 2.

^{Written for commit 7ad76ae. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 2 files

_{Re-trigger cubic}