feat: strip markdown code blocks from JSON responses

[rothnic]

Problem

AI SDK generateObject fails with "Invalid JSON response" when using kimi models through 9router.

Root Cause Investigation

Tested: Direct API calls to kimi with response_format: {type: "json_object"}

Findings:

• kimi models natively support JSON mode via Moonshot API
• However, kimi wraps JSON responses in markdown code blocks (```json...```)
• Example response: ```json\n{"key": "value"}\n```
• AI SDK expects raw JSON, not markdown-wrapped JSON
• This causes AI_APICallError: Invalid JSON response in AI SDK 5.x

Why kimi outputs markdown:

• Despite explicit system prompts saying "Respond ONLY with raw JSON"
• kimi still wraps in markdown, possibly training data influence
• This happens in BOTH json_object and json_schema modes

Related Work

This PR builds on previous response_format support:

• feat: Add OpenAI API response_format support for structured JSON output #285 - feat: Add OpenAI API response_format support for structured JSON output
• Fix: SSE data: [DONE] sentinel for non-streaming requests (complements PR #285) #286 - Fix SSE streaming for response_format compatibility

Those PRs added the foundation for translating response_format to Claude-compatible system prompts. This PR adds the final layer: stripping markdown that some models (kimi, etc.) add despite explicit instructions.

Solution

Strip markdown code block markers from JSON responses only when response_format is set.

Implementation Details

Conditional Application:

// Only process if response_format is specified
if (body?.response_format) {
  const isStructuredOutput = 
    responseFormat.type === "json_schema" || 
    responseFormat.type === "json_object";
  
  if (isStructuredOutput) {
    // Strip markdown markers
    text = text.replace(/^\s*\`\`\`\s*json\s*\n?/i, "")
               .replace(/\n?\s*\`\`\`\s*$/i, "");
  }
}

Safety Measures

1. Mode-gated: Only applies when response_format.type is json_schema or json_object
2. Regex boundaries: Only matches at start/end of content with word boundaries
3. Case-insensitive: Handles json, JSON, Json variations
4. Optional whitespace: Accounts for formatting variations
5. Preserves inline: Does NOT strip backticks from inline code or explanations

What This Protects

✅ Strips: Code block wrappers in structured output mode

{"key": "value"}  // Result after stripping

❌ Preserves: Inline backticks in normal chat

• "Use the config.json file" stays intact
• "The auth header should contain..." stays intact

Changes

Files Modified

1. open-sse/translator/request/openai-to-claude.js

   • Translates OpenAI response_format to Claude-compatible system prompts
   • Explicitly instructs: "Return ONLY raw JSON without markdown"

2. open-sse/translator/response/claude-to-openai.js

   • Strips markdown from streaming text_delta chunks
   • Applied only when response_format present in request

3. open-sse/handlers/chatCore/nonStreamingHandler.js

   • Strips markdown from non-streaming Claude format responses
   • Applied only to final assembled content

4. open-sse/utils/stream.js & jsonExtractor.js (new)

   • Centralized utility for consistent markdown extraction
   • Validates JSON structure before/after stripping

Testing

Verified Scenarios

• ✅ response_format: {type: "json_object"} → markdown stripped
• ✅ response_format: {type: "json_schema", json_schema: {...}} → markdown stripped
• ✅ Normal chat (no response_format) → backticks preserved
• ✅ Inline code in explanations → backticks preserved
• ✅ Multiple models: kimi/kimi-k2.5-thinking, gh/gpt-4o

Test Command

curl -X POST http://localhost:20128/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d "{\"model\": \"kimi/kimi-k2.5-thinking\", \"messages\": [{\"role\": \"user\", \"content\": \"Return JSON: {\\\"test\\\": \\\"value\\\"}\"}], \"response_format\": {\"type\": \"json_object\"}}"

Before fix: Returns \``json\n{"test": "value"}\n```**After fix:** Returns{"test": "value"}` ✓

Global Application

These changes are conditionally applied based on request context:

• Only active when client explicitly requests structured JSON output
• Safe for all models (idempotent operation)
• No performance impact on normal chat requests

Related

• feat: Add OpenAI API response_format support for structured JSON output #285 - feat: Add OpenAI API response_format support for structured JSON output
• Fix: SSE data: [DONE] sentinel for non-streaming requests (complements PR #285) #286 - Fix SSE streaming for response_format compatibility
• Fixes compatibility with AI SDK 5.x generateObject
• Enables structured output with kimi, minimax, and other Claude-compatible providers