fix: sanitize malformed tool call arguments to prevent vLLM 400 errors

[thanhnnict]

Problem

When LLMs generate malformed JSON in tool_calls.function.arguments (e.g., missing commas, truncated due to max_tokens), Continue stores this in conversation history and sends it back to the server on subsequent turns. vLLM validates JSON in tool_calls arguments during request preprocessing and rejects malformed input with a 400 error, breaking the conversation.

Solution

Validate JSON before sending tool_calls arguments back to server. If invalid, wrap in a valid JSON envelope ({"_raw": original}) to prevent server rejection while preserving the original content.

Behavior

Input arguments	Valid JSON?	Output
{"path": "/tmp/x"}	✅	Pass-through unchanged
{"path": "/tmp/x" "content": "hi"}	❌	{"_raw": "..."}
"" (empty)	—	"{}" (fallback)
null / undefined	—	"{}" (fallback)

Root Causes of Malformed JSON

1. max_tokens truncation — cuts off mid-JSON
2. Reasoning token leak — thinking tokens leak into arguments field
3. Model hallucination — LLM occasionally produces invalid syntax
4. Encoding issues — special characters not properly escaped

Trade-offs

• Performance: JSON.parse() on every tool_call argument — negligible cost (microseconds)
• Semantics: Wrapped {"_raw": "..."} may confuse model, but strictly better than 400 error
• No data loss: Original malformed string preserved in _raw field

Related

• [Bug]: vllm-project/vllm#43995

[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, core-checks) appear unrelated to this change — this PR only modifies core/llm/openaiTypeConverters.ts to sanitize malformed tool call arguments.

The core-checks failure is in llm/llm-pre-fetch.vitest.ts which tests Anthropic/Gemini/OpenAI auth flows — completely unrelated to the tool argument sanitization fix here.

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

[thanhnnict]

Hi @sestinj - Friendly ping on this PR. We have been hitting vLLM 400 errors in production when models produce truncated JSON in tool_calls arguments (especially under max_tokens pressure with Nemotron Ultra 253B). The sanitization is defensive - validates JSON before sending it back, zero impact on valid responses. Related upstream vLLM issue: vllm-project/vllm#43995 - Happy to add unit tests or adjust the wrapping strategy. Let me know if this aligns with the project direction. Thanks.