修复缓存问题

adapters/responses_cc_adapter.py

@@ -654,6 +654,10 @@ class ResponsesToCCStreamConverter:
             'completion_tokens': self._usage.get('output_tokens', 0),
             'total_tokens': self._usage.get('total_tokens', 0),
         }
+        if isinstance(self._usage.get('input_tokens_details'), dict):
+            chunk['usage']['prompt_tokens_details'] = dict(self._usage['input_tokens_details'])
+        if isinstance(self._usage.get('output_tokens_details'), dict):
+            chunk['usage']['completion_tokens_details'] = dict(self._usage['output_tokens_details'])
         return [chunk]
 
     def _make_chunk(self, delta: JsonDict, finish_reason: str | None = None) -> JsonDict:
@@ -678,20 +682,44 @@ def _copy_request_options(payload: JsonDict, result: JsonDict) -> None:
     """将 Responses 请求中的通用选项复制到 CC 请求体。"""
     if 'tools' in payload:
         result['tools'] = _convert_tools(payload['tools'])
-    for key in ('temperature', 'top_p'):
+    for key in (
+        'temperature',
+        'top_p',
+        'tool_choice',
+        'parallel_tool_calls',
+        'truncation',
+        'store',
+        'metadata',
+        'conversation',
+        'previous_response_id',
+        'prompt_cache_key',
+        'service_tier',
+        'user',
+    ):
         if key in payload:
             result[key] = payload[key]
     if 'max_output_tokens' in payload:
         result['max_tokens'] = payload['max_output_tokens']
-    if 'tool_choice' in payload:
-        result['tool_choice'] = payload['tool_choice']
 
 
 def _copy_responses_request_options(payload: JsonDict, result: JsonDict) -> None:
     """将聊天补全请求中的通用选项复制到原生 Responses 请求体。"""
     if 'tools' in payload:
         result['tools'] = _convert_cc_tools_to_responses(payload['tools'])
-    for key in ('temperature', 'top_p', 'tool_choice'):
+    for key in (
+        'temperature',
+        'top_p',
+        'tool_choice',
+        'parallel_tool_calls',
+        'truncation',
+        'store',
+        'metadata',
+        'conversation',
+        'previous_response_id',
+        'prompt_cache_key',
+        'service_tier',
+        'user',
+    ):
         if key in payload:
             result[key] = payload[key]
     if 'max_tokens' in payload:
@@ -914,11 +942,18 @@ def _make_function_call_output_item(tool_call: JsonDict) -> JsonDict:
 
 def _build_responses_usage(usage: JsonDict) -> JsonDict:
     """将 Chat Completions 的 usage 字段映射为 Responses usage 结构。"""
-    return {
+    result = {
         'input_tokens': usage.get('prompt_tokens', 0),
         'output_tokens': usage.get('completion_tokens', 0),
         'total_tokens': usage.get('total_tokens', 0),
     }
+    prompt_details = usage.get('prompt_tokens_details')
+    if isinstance(prompt_details, dict):
+        result['input_tokens_details'] = dict(prompt_details)
+    completion_details = usage.get('completion_tokens_details')
+    if isinstance(completion_details, dict):
+        result['output_tokens_details'] = dict(completion_details)
+    return result
 
 
 def _collect_cc_parts_from_responses_output(output_items: Any) -> tuple[str, str, list[JsonDict]]:

routes/chat.py

@@ -42,6 +42,7 @@ from routes.common import (
     build_responses_target,
     build_route_context,
     chat_error_chunk,
+    ensure_responses_cache_control,
     inject_instructions_anthropic,
     inject_instructions_cc,
     inject_instructions_responses,
@@ -311,6 +312,7 @@ def _handle_responses_backend(ctx: RouteContext, payload: dict[str, Any], turn:
     responses_payload = cc_to_responses_request(payload)
     responses_payload['model'] = ctx.upstream_model
     responses_payload = inject_instructions_responses(responses_payload, ctx.custom_instructions, ctx.instructions_position)
+    responses_payload = ensure_responses_cache_control(responses_payload)
     _dbg(
         '已转换为 Responses 请求：字段=' + str(list(responses_payload.keys()))
         + f' 输入项数={len(responses_payload.get("input", []))}'

routes/common.py

@@ -195,6 +195,23 @@ def inject_instructions_responses(payload: dict[str, Any], instructions: str, po
     return payload
 
 
+def ensure_responses_cache_control(payload: dict[str, Any]) -> dict[str, Any]:
+    """为 Responses 请求补齐自动 prompt caching 开关。
+
+    一些支持 `/v1/responses` 的上游会参考顶层 `cache_control` 来自动放置缓存断点。
+    Cursor 侧通常不会主动携带这个字段，因此这里在缺失时补一个保守的默认值，
+    同时允许调用方通过 body_modifications 或显式字段自行覆盖/关闭。
+    """
+    if not isinstance(payload, dict):
+        return payload
+    cache_control = payload.get('cache_control')
+    if isinstance(cache_control, dict) and cache_control.get('type'):
+        return payload
+    payload['cache_control'] = {'type': 'ephemeral'}
+    logger.info('已为 Responses 请求自动启用 cache_control=ephemeral')
+    return payload
+
+
 def inject_instructions_anthropic(payload: dict[str, Any], instructions: str, position: str = 'prepend') -> dict[str, Any]:
     """向 Anthropic Messages 请求注入自定义指令（写入 system 字段）。
 

routes/responses.py

@@ -27,6 +27,7 @@ from routes.common import (
     build_openai_target,
     build_responses_target,
     build_route_context,
+    ensure_responses_cache_control,
     inject_instructions_anthropic,
     inject_instructions_cc,
     inject_instructions_responses,
@@ -247,6 +248,7 @@ def _handle_responses_backend(ctx: RouteContext, payload: dict[str, Any], turn:
     payload = dict(payload)
     payload['model'] = ctx.upstream_model
     payload = inject_instructions_responses(payload, ctx.custom_instructions, ctx.instructions_position)
+    payload = ensure_responses_cache_control(payload)
     url, headers = build_responses_target(ctx)
     payload = apply_body_modifications(payload, ctx.body_modifications)
     headers = apply_header_modifications(headers, ctx.header_modifications)
@@ -629,4 +631,22 @@ def _finalize_responses_response(
     attach_client_response(turn, response_data)
     finalize_turn(turn, usage=response_data.get('usage'))
 
+    output_items = response_data.get('output', [])
+    if isinstance(output_items, list):
+        for item in output_items:
+            if not isinstance(item, dict) or item.get('type') != 'reasoning':
+                continue
+            summary = item.get('summary', [])
+            if not isinstance(summary, list):
+                continue
+            reasoning_text = ''.join(
+                part.get('text', '')
+                for part in summary
+                if isinstance(part, dict) and part.get('type') == 'summary_text'
+            )
+            if reasoning_text:
+                cc_messages = responses_to_cc(request.get_json(silent=True, force=True) or {}).get('messages', [])
+                thinking_cache.store_from_response(cc_messages, reasoning_text)
+                break
+
     return jsonify(response_data)