fix: retry without tools when OpenAI-compatible server returns empty response

[thanhnnict]

Problem

Some OpenAI-compatible servers (NIM, vLLM, TGI) return content: null with an empty tool_calls: [] array when tools are provided in the request but the model decides not to call any tool. This results in blank responses rendered in the Continue GUI.

Solution

Detect empty response (content: null + tool_calls: []) and retry the same request without the tools parameter. The model then responds with text content normally.

Behavior Matrix

Model Response	tools in request	Retry?	Result
content: "Hello", tool_calls: []	Yes	❌ No	Text shown normally
content: null, tool_calls: [{...}]	Yes	❌ No	Tool call executed
content: null, tool_calls: []	Yes	✅ Yes	Retry → text shown
content: null, tool_calls: []	No	❌ No	Empty (no tools to remove)

Affected Models

• nvidia/nemotron-ultra-253b (NIM) — frequently
• deepseek-ai/DeepSeek-V4-Flash (vLLM) — occasionally
• Any OpenAI-compatible server with tool support

Trade-offs

• Latency: Adds one extra API call when empty response is detected (~2-5s)
• Correctness: Model genuinely wants to say nothing → forced to respond (acceptable for coding assistant)
• Scope: Only affects chatCompletionNonStream (non-streaming path)

Related

• Agent Mode: Streamed tool_calls from vLLM ignored - no action performed #5508

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[thanhnnict]

I have read the CLA Document and I hereby sign the CLA

[thanhnnict]

Hi @continuedev/maintainers,

The failing checks (jetbrains-tests, CLI PR Checks / test macos-latest) appear unrelated to this change — this PR only modifies packages/openai-adapters/src/apis/OpenAI.ts to handle empty responses from NIM/vLLM servers.

Could you confirm if these are pre-existing flaky tests or re-run the CI? Thanks!

[thanhnnict]

Hi @sestinj - Friendly ping on this PR. This fixes a real issue we hit in production with NIM (Nemotron Ultra 253B) and vLLM (DeepSeek V4 Flash) serving ~300 developers. The empty response causes blank GUI output which is confusing for users. The fix is minimal (single file, defensive retry) and references the existing issue #5508. Happy to add unit tests or address any feedback if needed. Thanks for your time.