code-puppy 0.0.302__py3-none-any.whl → 0.0.323__py3-none-any.whl

code_puppy/chatgpt_codex_client.py

@@ -0,0 +1,283 @@
+"""HTTP client interceptor for ChatGPT Codex API.
+
+ChatGPTCodexAsyncClient: httpx client that injects required fields into
+request bodies for the ChatGPT Codex API and handles stream-to-non-stream
+conversion.
+
+The Codex API requires:
+- "store": false - Disables conversation storage
+- "stream": true - Streaming is mandatory
+
+Removes unsupported parameters:
+- "max_output_tokens" - Not supported by Codex API
+- "max_tokens" - Not supported by Codex API
+- "verbosity" - Not supported by Codex API
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from typing import Any
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+
+def _is_reasoning_model(model_name: str) -> bool:
+    """Check if a model supports reasoning parameters."""
+    reasoning_models = [
+        "gpt-5",  # All GPT-5 variants
+        "o1",  # o1 series
+        "o3",  # o3 series
+        "o4",  # o4 series
+    ]
+    model_lower = model_name.lower()
+    return any(model_lower.startswith(prefix) for prefix in reasoning_models)
+
+
+class ChatGPTCodexAsyncClient(httpx.AsyncClient):
+    """Async HTTP client that handles ChatGPT Codex API requirements.
+
+    This client:
+    1. Injects required fields (store=false, stream=true)
+    2. Strips unsupported parameters
+    3. Converts streaming responses to non-streaming format
+    """
+
+    async def send(
+        self, request: httpx.Request, *args: Any, **kwargs: Any
+    ) -> httpx.Response:
+        """Intercept requests and inject required Codex fields."""
+        force_stream_conversion = False
+
+        try:
+            # Only modify POST requests to the Codex API
+            if request.method == "POST":
+                body_bytes = self._extract_body_bytes(request)
+                if body_bytes:
+                    updated, force_stream_conversion = self._inject_codex_fields(
+                        body_bytes
+                    )
+                    if updated is not None:
+                        try:
+                            rebuilt = self.build_request(
+                                method=request.method,
+                                url=request.url,
+                                headers=request.headers,
+                                content=updated,
+                            )
+
+                            # Copy core internals so httpx uses the modified body/stream
+                            if hasattr(rebuilt, "_content"):
+                                setattr(request, "_content", rebuilt._content)
+                            if hasattr(rebuilt, "stream"):
+                                request.stream = rebuilt.stream
+                            if hasattr(rebuilt, "extensions"):
+                                request.extensions = rebuilt.extensions
+
+                            # Ensure Content-Length matches the new body
+                            request.headers["Content-Length"] = str(len(updated))
+
+                        except Exception:
+                            pass
+        except Exception:
+            pass
+
+        # Make the actual request
+        response = await super().send(request, *args, **kwargs)
+
+        # If we forced streaming, convert the SSE stream to a regular response
+        if force_stream_conversion and response.status_code == 200:
+            try:
+                response = await self._convert_stream_to_response(response)
+            except Exception as e:
+                logger.warning(f"Failed to convert stream response: {e}")
+
+        return response
+
+    @staticmethod
+    def _extract_body_bytes(request: httpx.Request) -> bytes | None:
+        """Extract the request body as bytes."""
+        try:
+            content = request.content
+            if content:
+                return content
+        except Exception:
+            pass
+
+        try:
+            content = getattr(request, "_content", None)
+            if content:
+                return content
+        except Exception:
+            pass
+
+        return None
+
+    @staticmethod
+    def _inject_codex_fields(body: bytes) -> tuple[bytes | None, bool]:
+        """Inject required Codex fields and remove unsupported ones.
+
+        Returns:
+            Tuple of (modified body bytes or None, whether stream was forced)
+        """
+        try:
+            data = json.loads(body.decode("utf-8"))
+        except Exception:
+            return None, False
+
+        if not isinstance(data, dict):
+            return None, False
+
+        modified = False
+        forced_stream = False
+
+        # CRITICAL: ChatGPT Codex backend requires store=false
+        if "store" not in data or data.get("store") is not False:
+            data["store"] = False
+            modified = True
+
+        # CRITICAL: ChatGPT Codex backend requires stream=true
+        # If stream is already true (e.g., pydantic-ai with event_stream_handler),
+        # don't force conversion - let streaming events flow through naturally
+        if data.get("stream") is not True:
+            data["stream"] = True
+            forced_stream = True  # Only convert if WE forced streaming
+            modified = True
+
+        # Add reasoning settings for reasoning models (gpt-5.2, o-series, etc.)
+        model = data.get("model", "")
+        if "reasoning" not in data and _is_reasoning_model(model):
+            data["reasoning"] = {
+                "effort": "medium",
+                "summary": "auto",
+            }
+            modified = True
+
+        # Remove unsupported parameters
+        # Note: verbosity should be under "text" object, not top-level
+        unsupported_params = ["max_output_tokens", "max_tokens", "verbosity"]
+        for param in unsupported_params:
+            if param in data:
+                del data[param]
+                modified = True
+
+        if not modified:
+            return None, False
+
+        return json.dumps(data).encode("utf-8"), forced_stream
+
+    async def _convert_stream_to_response(
+        self, response: httpx.Response
+    ) -> httpx.Response:
+        """Convert an SSE streaming response to a complete response.
+
+        Consumes the SSE stream and reconstructs the final response object.
+        """
+        logger.debug("Converting SSE stream to non-streaming response")
+        final_response_data = None
+        collected_text = []
+        collected_tool_calls = []
+
+        # Read the entire stream
+        async for line in response.aiter_lines():
+            if not line or not line.startswith("data:"):
+                continue
+
+            data_str = line[5:].strip()  # Remove "data:" prefix
+            if data_str == "[DONE]":
+                break
+
+            try:
+                event = json.loads(data_str)
+                event_type = event.get("type", "")
+
+                if event_type == "response.output_text.delta":
+                    # Collect text deltas
+                    delta = event.get("delta", "")
+                    if delta:
+                        collected_text.append(delta)
+
+                elif event_type == "response.completed":
+                    # This contains the final response object
+                    final_response_data = event.get("response", {})
+
+                elif event_type == "response.function_call_arguments.done":
+                    # Collect tool calls
+                    tool_call = {
+                        "name": event.get("name", ""),
+                        "arguments": event.get("arguments", ""),
+                        "call_id": event.get("call_id", ""),
+                    }
+                    collected_tool_calls.append(tool_call)
+
+            except json.JSONDecodeError:
+                continue
+
+        logger.debug(
+            f"Collected {len(collected_text)} text chunks, {len(collected_tool_calls)} tool calls"
+        )
+        if final_response_data:
+            logger.debug(
+                f"Got final response data with keys: {list(final_response_data.keys())}"
+            )
+
+        # Build the final response body
+        if final_response_data:
+            response_body = final_response_data
+        else:
+            # Fallback: construct a minimal response from collected data
+            response_body = {
+                "id": "reconstructed",
+                "object": "response",
+                "output": [],
+            }
+
+            if collected_text:
+                response_body["output"].append(
+                    {
+                        "type": "message",
+                        "role": "assistant",
+                        "content": [
+                            {"type": "output_text", "text": "".join(collected_text)}
+                        ],
+                    }
+                )
+
+            for tool_call in collected_tool_calls:
+                response_body["output"].append(
+                    {
+                        "type": "function_call",
+                        "name": tool_call["name"],
+                        "arguments": tool_call["arguments"],
+                        "call_id": tool_call["call_id"],
+                    }
+                )
+
+        # Create a new response with the complete body
+        body_bytes = json.dumps(response_body).encode("utf-8")
+        logger.debug(f"Reconstructed response body: {len(body_bytes)} bytes")
+
+        new_response = httpx.Response(
+            status_code=response.status_code,
+            headers=response.headers,
+            content=body_bytes,
+            request=response.request,
+        )
+        return new_response
+
+
+def create_codex_async_client(
+    headers: dict[str, str] | None = None,
+    verify: str | bool = True,
+    **kwargs: Any,
+) -> ChatGPTCodexAsyncClient:
+    """Create a ChatGPT Codex async client with proper configuration."""
+    return ChatGPTCodexAsyncClient(
+        headers=headers,
+        verify=verify,
+        timeout=httpx.Timeout(300.0, connect=30.0),
+        **kwargs,
+    )