From 56faf4fcf1cd37b8144ef1bd908a1403cb465fd3 Mon Sep 17 00:00:00 2001
From: h88782481 <54714341+h88782481@users.noreply.github.com>
Date: Sun, 15 Mar 2026 16:26:07 +0800
Subject: [PATCH] =?UTF-8?q?=E4=BC=98=E5=8C=96=E7=BC=93=E5=AD=98=E5=91=BD?=
 =?UTF-8?q?=E4=B8=AD=E9=97=AE=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 adapters/cc_anthropic_adapter.py |  8 ++++++
 adapters/responses_cc_adapter.py | 44 ++++++++++++++++++++++++--------
 routes/chat.py                   |  3 ++-
 3 files changed, 44 insertions(+), 11 deletions(-)

diff --git a/adapters/cc_anthropic_adapter.py b/adapters/cc_anthropic_adapter.py
index 512b3de..7848d05 100644
--- a/adapters/cc_anthropic_adapter.py
+++ b/adapters/cc_anthropic_adapter.py
@@ -261,6 +261,12 @@ def _convert_request_message(message: Any) -> tuple[JsonDict | None, str | None]
     anthropic_role = 'assistant' if role == 'assistant' else 'user'
     anthropic_content = _convert_content(message)
 
+    if role == 'assistant' and message.get('reasoning_content'):
+        thinking_block = {'type': 'thinking', 'thinking': message['reasoning_content']}
+        blocks = _to_blocks(anthropic_content)
+        blocks.insert(0, thinking_block)
+        anthropic_content = blocks
+
     if role == 'assistant' and 'tool_calls' in message:
         anthropic_content = _append_tool_use_blocks(anthropic_content, message.get('tool_calls', []))
 
@@ -463,6 +469,8 @@ def _convert_content_part(part: Any) -> JsonDict | None:
         return {'type': 'text', 'text': part.get('text', '')}
     if part_type == 'image_url':
         return _convert_image(part)
+    if part_type == 'image':
+        return part
     if part_type in ('tool_use', 'tool_result'):
         return part
     return None
diff --git a/adapters/responses_cc_adapter.py b/adapters/responses_cc_adapter.py
index 0f35c52..e6c864a 100644
--- a/adapters/responses_cc_adapter.py
+++ b/adapters/responses_cc_adapter.py
@@ -703,7 +703,11 @@ def _append_responses_input_item(
     instructions: list[str],
     input_items: list[JsonDict],
 ) -> None:
-    """将单条 Chat Completions 消息追加为 Responses `input` 项。"""
+    """将单条 Chat Completions 消息追加为 Responses `input` 项。
+
+    尽量使用 EasyInputMessage 格式（{role, content}）以减少 token 开销，
+    提高上游 prompt caching 的前缀匹配命中率。
+    """
     if not isinstance(message, dict):
         return
 
@@ -724,21 +728,26 @@ def _append_responses_input_item(
         })
         return
 
-    item: JsonDict = {
-        'type': 'message',
-        'role': role or 'user',
-        'content': _content_to_responses_parts(content, role),
-    }
-    input_items.append(item)
+    text = _content_to_text(content)
+    has_tool_calls = bool(message.get('tool_calls'))
 
-    if role == 'assistant':
+    if role == 'assistant' and has_tool_calls:
+        if text:
+            input_items.append({
+                'type': 'message',
+                'role': 'assistant',
+                'content': [{'type': 'output_text', 'text': text}],
+            })
         for tool_call in message.get('tool_calls') or []:
             input_items.append(_build_responses_function_call_item(tool_call))
+    else:
+        input_items.append({'role': role or 'user', 'content': text or ''})
 
 
 def _convert_input_items(items: list[Any], messages: list[JsonDict]) -> None:
     """将 Responses `input` 数组重建为 Chat Completions `messages` 列表。"""
     index = 0
+    pending_reasoning: str | None = None
     while index < len(items):
         item = items[index]
 
@@ -754,20 +763,35 @@ def _convert_input_items(items: list[Any], messages: list[JsonDict]) -> None:
         item_type = item.get('type', '')
         role = item.get('role', '')
 
+        if item_type == 'reasoning':
+            pending_reasoning = _extract_reasoning_text(item)
+            index += 1
+            continue
+
         if role and not item_type:
-            messages.append({
+            msg: JsonDict = {
                 'role': role,
                 'content': _normalize_simple_content(item.get('content', '')),
-            })
+            }
+            if role == 'assistant' and pending_reasoning:
+                msg['reasoning_content'] = pending_reasoning
+                pending_reasoning = None
+            messages.append(msg)
             index += 1
             continue
 
         if item_type == 'message':
             consumed = _append_message_item(items, start=index, messages=messages)
+            if item.get('role') == 'assistant' and pending_reasoning and messages:
+                messages[-1]['reasoning_content'] = pending_reasoning
+                pending_reasoning = None
             index += consumed
             continue
 
         if item_type == 'function_call':
+            if pending_reasoning and messages and messages[-1].get('role') == 'assistant':
+                messages[-1]['reasoning_content'] = pending_reasoning
+                pending_reasoning = None
             _append_function_call_item(item, messages)
             index += 1
             continue
diff --git a/routes/chat.py b/routes/chat.py
index 1ca6f81..be4f775 100644
--- a/routes/chat.py
+++ b/routes/chat.py
@@ -127,7 +127,8 @@ def chat_completions():
     log_route_context('聊天补全', ctx, extra=f'消息数={message_count}')
     _log_messages(payload)
 
-    payload['messages'] = thinking_cache.inject(payload.get('messages', []))
+    if ctx.backend != 'responses':
+        payload['messages'] = thinking_cache.inject(payload.get('messages', []))
 
     if ctx.backend == 'openai':
         return _handle_openai_backend(ctx, payload, turn)