Tool call arguments may contain invalid JSON #17750

Summary

When using SiliconFlow's Qwen3-8B model, tool call arguments may contain invalid JSON due to the </tool_call> XML tag being included in the function.arguments field. This affects both streaming and non-streaming modes because neither validates JSON before emitting tool-call events.

Environment

• OpenCode Version: v1.2.20+
• Provider: SiliconFlow (China)
• Model: Qwen/Qwen3-8B
• File: packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts

Problem Description

Root Cause

Qwen3-8B internally uses <tool_call>...</tool_call> XML format to represent tool calls. When SiliconFlow's API converts this to OpenAI-compatible format, the </tool_call> tag may be included in function.arguments.

Code Analysis

Non-streaming mode (doGenerate) - Line 245-256:

if (choice.message.tool_calls != null) {
  for (const toolCall of choice.message.tool_calls) {
    content.push({
      type: "tool-call",
      toolCallId: toolCall.id ?? generateId(),
      toolName: toolCall.function.name,
      input: toolCall.function.arguments!,  // No JSON validation!
      // ...
    })
  }
}

Streaming mode (flush) - Line 640-653:

flush(controller) {
  // ...
  for (const toolCall of toolCalls.filter((toolCall) => !toolCall.hasFinished)) {
    controller.enqueue({
      type: "tool-call",
      toolCallId: toolCall.id ?? generateId(),
      toolName: toolCall.function.name,
      input: toolCall.function.arguments,  // No JSON validation!
    })
  }
}

Note: In streaming mode, there is JSON validation during chunk processing (line 603-606), but unfinished tool calls in flush bypass this check.

Error Message

Invalid input for tool question: JSON parsing failed: Text: {"questions": [...], "multiple": true}}
</tool_call>.
Error message: JSON Parse error: Expected ']'

Steps to Reproduce

1. Configure OpenCode to use SiliconFlow with Qwen3-8B
2. Send a message that triggers a tool call (e.g., ambiguous request that needs clarification)
3. Observe that the tool call arguments contain </tool_call> tag

Minimal reproduction with curl:

curl -s -N https://api.siliconflow.cn/v1/chat/completions \
  -H "Authorization: Bearer $SILICONFLOW_CN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-8B",
    "messages": [{"role": "user", "content": "帮我写一个Python爬虫"}],
    "tools": [{"type": "function", "function": {"name": "question", "description": "Ask user questions", "parameters": {"type": "object", "properties": {"questions": {"type": "array", "items": {"type": "object", "properties": {"question": {"type": "string"}, "options": {"type": "array", "items": {"type": "string"}}, "multiple": {"type": "boolean"}}}}}, "required": ["questions"]}}}],
    "tool_choice": "auto",
    "stream": true
  }'

Expected Behavior

Tool call arguments should always be valid JSON in both streaming and non-streaming modes.

Suggested Fix

Option 1: Add JSON validation in both modes

For non-streaming (doGenerate):

if (choice.message.tool_calls != null) {
  for (const toolCall of choice.message.tool_calls) {
    let args = toolCall.function.arguments!
    // Strip XML artifacts from providers like SiliconFlow
    args = args.replace(/<\/?tool_call>/g, '').trim()

    if (!isParsableJson(args)) {
      console.warn(`Invalid tool call arguments: ${args}`)
      continue
    }

    content.push({
      type: "tool-call",
      toolCallId: toolCall.id ?? generateId(),
      toolName: toolCall.function.name,
      input: args,
      // ...
    })
  }
}

For streaming (flush):

flush(controller) {
  // ...
  for (const toolCall of toolCalls.filter((toolCall) => !toolCall.hasFinished)) {
    let args = toolCall.function.arguments
    // Strip XML artifacts from providers like SiliconFlow
    args = args.replace(/<\/?tool_call>/g, '').trim()

    if (!isParsableJson(args)) {
      console.warn(`Invalid tool call arguments: ${args}`)
      continue
    }

    controller.enqueue({
      type: "tool-call",
      toolCallId: toolCall.id ?? generateId(),
      toolName: toolCall.function.name,
      input: args,
    })
  }
}

Option 2: Strip XML tags during chunk accumulation (streaming only)

if (toolCallDelta.function?.arguments != null) {
  let args = toolCallDelta.function?.arguments ?? ""
  // Strip XML tags that some providers include
  args = args.replace(/<\/?tool_call>/g, '')
  toolCall.function!.arguments += args
}

Note: Option 2 only fixes streaming mode. Non-streaming mode still needs separate handling.

Related Code

• packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts:245-256 - doGenerate tool_calls handling (non-streaming)
• packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts:640-653 - flush function (streaming)
• packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts:591-592 - arguments accumulation (streaming)
• packages/opencode/src/session/llm.ts:178-197 - experimental_repairToolCall (current workaround redirects to invalid tool)