feat: add response_format json_object support

CONTRIBUTORS.md

@@ -14,3 +14,4 @@ Thank you to all the contributors who have helped improve this project!
 - [@PAzter1101](https://github.com/PAzter1101) — Docker containerization with CI/CD (#55)
 - [@Ry-DS](https://github.com/Ry-DS) — Images in tool results support for Anthropic MCP servers (#57)
 - [@saaj](https://github.com/saaj) — Regional endpoint fix for eu-central-1 and other non-us-east-1 regions (#58)
+- [@lukasBlckIT](https://github.com/lukasBlckIT) — response_format json_object support (#123)

kiro/converters_openai.py

@@ -296,6 +296,33 @@ def convert_openai_tools_to_unified(tools: Optional[List[Tool]]) -> Optional[Lis
 # Main Entry Point
 # ==================================================================================================
 
+def _wants_json_response(request_data: ChatCompletionRequest) -> bool:
+    """
+    Check if the request has response_format with type 'json_object'.
+
+    Since ChatCompletionRequest uses extra='allow', response_format
+    is accepted but not explicitly modeled. We check it dynamically.
+    """
+    rf = getattr(request_data, "response_format", None)
+    if rf is None:
+        return False
+    if isinstance(rf, dict):
+        return rf.get("type") == "json_object"
+    # Pydantic model or other object
+    return getattr(rf, "type", None) == "json_object"
+
+
+JSON_SYSTEM_PROMPT_ADDITION = (
+    "\n\n---\n"
+    "# Response Format: JSON\n\n"
+    "The caller has requested JSON output (response_format: json_object).\n"
+    "You MUST respond with raw, valid JSON only.\n"
+    "Do NOT wrap the response in markdown code blocks (```json ... ``` or ``` ... ```).\n"
+    "Do NOT add any text before or after the JSON.\n"
+    "The response must be parseable by JSON.parse() directly."
+)
+
+
 def build_kiro_payload(
     request_data: ChatCompletionRequest,
     conversation_id: str,
@@ -324,6 +351,11 @@ def build_kiro_payload(
     # Convert tools to unified format
     unified_tools = convert_openai_tools_to_unified(request_data.tools)
 
+    # If response_format is json_object, inject JSON instruction into system prompt
+    if _wants_json_response(request_data):
+        system_prompt = (system_prompt + JSON_SYSTEM_PROMPT_ADDITION) if system_prompt else JSON_SYSTEM_PROMPT_ADDITION.strip()
+        logger.info("response_format=json_object detected, injecting JSON system prompt")
+
     # Get model ID for Kiro API (normalizes + resolves hidden models)
     # Pass-through principle: we normalize and send to Kiro, Kiro decides if valid
     model_id = get_model_id_for_kiro(request_data.model, HIDDEN_MODELS)

kiro/json_response_format.py

@@ -0,0 +1,39 @@
+# -*- coding: utf-8 -*-
+"""
+Utilities for handling response_format: json_object.
+
+Strips markdown code block wrappers from responses when the caller
+requested raw JSON output via response_format.
+"""
+
+import re
+
+# Matches ```json\n...\n``` or ```\n...\n``` wrapping the entire response
+_MD_JSON_BLOCK_RE = re.compile(
+    r'^\s*```(?:json)?\s*\n(.*?)\n\s*```\s*$',
+    re.DOTALL
+)
+
+
+def strip_markdown_json_wrapper(content: str) -> str:
+    """
+    Strip markdown code block wrapper from JSON content.
+
+    If the content is wrapped in ```json ... ``` or ``` ... ```,
+    extract the inner content. Only strips if the entire response
+    is a single code block.
+
+    Args:
+        content: Response content that may be markdown-wrapped
+
+    Returns:
+        Unwrapped content, or original content if no wrapper found
+    """
+    if not content:
+        return content
+
+    m = _MD_JSON_BLOCK_RE.match(content)
+    if m:
+        return m.group(1).strip()
+
+    return content

kiro/routes_openai.py

@@ -46,8 +46,9 @@
 from kiro.auth import KiroAuthManager, AuthType
 from kiro.cache import ModelInfoCache
 from kiro.model_resolver import ModelResolver
-from kiro.converters_openai import build_kiro_payload
+from kiro.converters_openai import build_kiro_payload, _wants_json_response
 from kiro.streaming_openai import stream_kiro_to_openai, collect_stream_response, stream_with_first_token_retry
+from kiro.json_response_format import strip_markdown_json_wrapper
 from kiro.http_client import KiroHttpClient
 from kiro.utils import generate_conversation_id
 
@@ -393,6 +394,17 @@ async def stream_wrapper():
 
             await http_client.close()
 
+            # Strip markdown JSON wrapper if response_format is json_object
+            if _wants_json_response(request_data):
+                try:
+                    content = openai_response["choices"][0]["message"]["content"]
+                    stripped = strip_markdown_json_wrapper(content)
+                    if stripped != content:
+                        openai_response["choices"][0]["message"]["content"] = stripped
+                        logger.info("Stripped markdown JSON wrapper from non-streaming response")
+                except (KeyError, IndexError):
+                    pass
+
             # Log access log for non-streaming success
             logger.info(f"HTTP 200 - POST /v1/chat/completions (non-streaming) - completed")