Reasoning fixes

ccproxy/llms/formatters/anthropic_to_openai/responses.py

@@ -44,13 +44,10 @@ def convert__anthropic_message_to_openai_responses__response(
             text_parts.append(getattr(block, "text", ""))
         elif block_type == "thinking":
             thinking = getattr(block, "thinking", None) or ""
-            signature = getattr(block, "signature", None)
-            sig_attr = (
-                f' signature="{signature}"'
-                if isinstance(signature, str) and signature
-                else ""
-            )
-            text_parts.append(f"<thinking{sig_attr}>{thinking}</thinking>")
+            text_parts.append(f"<think>{thinking}</think>")
+        elif block_type == "redacted_thinking":
+            # Skip redacted thinking blocks
+            continue
         elif block_type == "tool_use":
             tool_contents.append(
                 {
@@ -113,14 +110,11 @@ def convert__anthropic_message_to_openai_chat__response(
                 parts.append(text)
         elif btype == "thinking":
             thinking = getattr(block, "thinking", None)
-            signature = getattr(block, "signature", None)
             if isinstance(thinking, str):
-                sig_attr = (
-                    f' signature="{signature}"'
-                    if isinstance(signature, str) and signature
-                    else ""
-                )
-                parts.append(f"<thinking{sig_attr}>{thinking}</thinking>")
+                parts.append(f"<think>{thinking}</think>")
+        # Skip redacted_thinking blocks
+        elif btype == "redacted_thinking":
+            continue
         elif btype == "tool_use":
             tool_calls.append(
                 build_openai_tool_call(

ccproxy/llms/formatters/anthropic_to_openai/streams.py

@@ -73,13 +73,11 @@ def _anthropic_delta_to_text(
     block_type = block_meta.get("type")
 
     if block_type == "thinking":
+        # Return just the thinking text - tags handled by block start/stop events
         thinking_text = delta.get("thinking")
-        if not isinstance(thinking_text, str) or not thinking_text:
-            return None
-        signature = block_meta.get("signature")
-        if isinstance(signature, str) and signature:
-            return f'<thinking signature="{signature}">{thinking_text}</thinking>'
-        return f"<thinking>{thinking_text}</thinking>"
+        if isinstance(thinking_text, str) and thinking_text:
+            return thinking_text
+        return None
 
     text_val = delta.get("text")
     if isinstance(text_val, str) and text_val:
@@ -1378,6 +1376,34 @@ async def generator() -> AsyncGenerator[
                 if not message_started:
                     continue
 
+                if event_type == "content_block_start":
+                    content_block = (
+                        event_payload.get("content_block", {})
+                        if isinstance(event_payload, dict)
+                        else {}
+                    )
+                    if (
+                        isinstance(content_block, dict)
+                        and content_block.get("type") == "thinking"
+                    ):
+                        # Emit opening <think> tag
+                        yield openai_models.ChatCompletionChunk(
+                            id="chatcmpl-stream",
+                            object="chat.completion.chunk",
+                            created=0,
+                            model=model_id,
+                            choices=[
+                                openai_models.StreamingChoice(
+                                    index=0,
+                                    delta=openai_models.DeltaMessage(
+                                        role="assistant", content="<think>"
+                                    ),
+                                    finish_reason=None,
+                                )
+                            ],
+                        )
+                    continue
+
                 if event_type == "content_block_delta":
                     block_index = int(event_payload.get("index", 0))
                     text_delta = _anthropic_delta_to_text(
@@ -1409,7 +1435,28 @@ async def generator() -> AsyncGenerator[
                     if not block_info:
                         continue
                     _, block_meta = block_info
-                    if block_meta.get("type") != "tool_use":
+                    block_type = block_meta.get("type")
+
+                    if block_type == "thinking":
+                        # Emit closing </think> tag
+                        yield openai_models.ChatCompletionChunk(
+                            id="chatcmpl-stream",
+                            object="chat.completion.chunk",
+                            created=0,
+                            model=model_id,
+                            choices=[
+                                openai_models.StreamingChoice(
+                                    index=0,
+                                    delta=openai_models.DeltaMessage(
+                                        role="assistant", content="</think>"
+                                    ),
+                                    finish_reason=None,
+                                )
+                            ],
+                        )
+                        continue
+
+                    if block_type != "tool_use":
                         continue
                     if block_index in emitted_tool_indices:
                         continue

ccproxy/llms/models/anthropic.py

@@ -219,7 +219,13 @@ class RedactedThinkingBlock(ContentBlockBase):
 
 
 RequestContentBlock = Annotated[
-    TextBlock | ImageBlock | ToolUseBlock | ToolResultBlock, Field(discriminator="type")
+    TextBlock
+    | ImageBlock
+    | ToolUseBlock
+    | ToolResultBlock
+    | ThinkingBlock
+    | RedactedThinkingBlock,
+    Field(discriminator="type"),
 ]
 
 ResponseContentBlock = Annotated[