feat(agent): improve prompt cache hit rate

app/agent/__init__.py

@@ -129,9 +129,18 @@ class _SessionUsageSnapshot:
     last_output_tokens: int = 0
     last_total_tokens: int = 0
     last_context_usage_ratio: Optional[float] = None
+    last_cache_usage_available: bool = False
+    last_cache_read_input_tokens: int = 0
+    last_cache_write_input_tokens: int = 0
+    last_uncached_input_tokens: int = 0
+    last_cache_hit_ratio: Optional[float] = None
     total_input_tokens: int = 0
     total_output_tokens: int = 0
     total_tokens: int = 0
+    total_cache_read_input_tokens: int = 0
+    total_cache_write_input_tokens: int = 0
+    total_uncached_input_tokens: int = 0
+    cache_usage_available: bool = False
     model_call_count: int = 0
     last_updated_at: Optional[datetime] = None
 
@@ -144,9 +153,23 @@ class _SessionUsageSnapshot:
             "last_output_tokens": self.last_output_tokens,
             "last_total_tokens": self.last_total_tokens,
             "last_context_usage_ratio": self.last_context_usage_ratio,
+            "last_cache_usage_available": self.last_cache_usage_available,
+            "last_cache_read_input_tokens": self.last_cache_read_input_tokens,
+            "last_cache_write_input_tokens": self.last_cache_write_input_tokens,
+            "last_uncached_input_tokens": self.last_uncached_input_tokens,
+            "last_cache_hit_ratio": self.last_cache_hit_ratio,
             "total_input_tokens": self.total_input_tokens,
             "total_output_tokens": self.total_output_tokens,
             "total_tokens": self.total_tokens,
+            "total_cache_read_input_tokens": self.total_cache_read_input_tokens,
+            "total_cache_write_input_tokens": self.total_cache_write_input_tokens,
+            "total_uncached_input_tokens": self.total_uncached_input_tokens,
+            "cache_usage_available": self.cache_usage_available,
+            "total_cache_hit_ratio": (
+                self.total_cache_read_input_tokens / self.total_input_tokens
+                if self.cache_usage_available and self.total_input_tokens
+                else None
+            ),
             "model_call_count": self.model_call_count,
             "last_updated_at": self.last_updated_at.strftime("%Y-%m-%d %H:%M:%S")
             if self.last_updated_at
@@ -554,9 +577,33 @@ class MoviePilotAgent:
         self._session_usage.last_output_tokens = output_tokens
         self._session_usage.last_total_tokens = total_tokens
         self._session_usage.last_context_usage_ratio = usage.get("context_usage_ratio")
+        cache_usage_available = bool(usage.get("cache_usage_available"))
+        cache_read_input_tokens = self._coerce_int(
+            usage.get("cache_read_input_tokens")
+        ) or 0
+        cache_write_input_tokens = self._coerce_int(
+            usage.get("cache_write_input_tokens")
+        ) or 0
+        uncached_input_tokens = self._coerce_int(
+            usage.get("uncached_input_tokens")
+        )
+        if uncached_input_tokens is None:
+            uncached_input_tokens = max(
+                input_tokens - cache_read_input_tokens - cache_write_input_tokens,
+                0,
+            )
+        self._session_usage.last_cache_usage_available = cache_usage_available
+        self._session_usage.last_cache_read_input_tokens = cache_read_input_tokens
+        self._session_usage.last_cache_write_input_tokens = cache_write_input_tokens
+        self._session_usage.last_uncached_input_tokens = uncached_input_tokens
+        self._session_usage.last_cache_hit_ratio = usage.get("cache_hit_ratio")
         self._session_usage.total_input_tokens += input_tokens
         self._session_usage.total_output_tokens += output_tokens
         self._session_usage.total_tokens += total_tokens
+        self._session_usage.total_cache_read_input_tokens += cache_read_input_tokens
+        self._session_usage.total_cache_write_input_tokens += cache_write_input_tokens
+        self._session_usage.total_uncached_input_tokens += uncached_input_tokens
+        self._session_usage.cache_usage_available |= cache_usage_available
 
     def get_session_status(self) -> dict[str, Any]:
         if not self._session_usage.model:
@@ -590,6 +637,17 @@ class MoviePilotAgent:
                 input_tokens=self._session_usage.total_input_tokens,
                 output_tokens=self._session_usage.total_output_tokens,
                 total_tokens=self._session_usage.total_tokens,
+                cache_read_input_tokens=self._session_usage.total_cache_read_input_tokens,
+                cache_write_input_tokens=self._session_usage.total_cache_write_input_tokens,
+                uncached_input_tokens=self._session_usage.total_uncached_input_tokens,
+                cache_hit_ratio=(
+                    self._session_usage.total_cache_read_input_tokens
+                    / self._session_usage.total_input_tokens
+                    if self._session_usage.cache_usage_available
+                    and self._session_usage.total_input_tokens
+                    else None
+                ),
+                cache_usage_available=self._session_usage.cache_usage_available,
                 model_call_count=self._session_usage.model_call_count,
                 success=success,
                 error=error,
@@ -819,7 +877,17 @@ class MoviePilotAgent:
         :param streaming: 是否启用流式输出
         """
         runtime_config = await self._resolve_llm_runtime_config()
-        return await LLMHelper.get_llm(streaming=streaming, **runtime_config)
+        return await LLMHelper.get_llm(
+            streaming=streaming,
+            prompt_cache_key=self._build_prompt_cache_key(),
+            **runtime_config,
+        )
+
+    def _build_prompt_cache_key(self) -> str:
+        """生成不暴露用户标识、且在同一会话内稳定的提示词缓存键。"""
+        cache_identity = f"{self.user_id or ''}\x00{self.session_id}"
+        digest = hashlib.sha256(cache_identity.encode("utf-8")).hexdigest()[:32]
+        return f"moviepilot-agent-{digest}"
 
     @classmethod
     def _has_image_input_content(cls, content: Any) -> bool:

app/agent/llm/helper.py

@@ -6,6 +6,7 @@ import json
 import time
 from functools import wraps
 from typing import TYPE_CHECKING, Any, List, Optional
+from urllib.parse import urlsplit
 
 from langchain_core.messages import AIMessage, AIMessageChunk
 
@@ -813,6 +814,91 @@ class LLMHelper:
             headers["User-Agent"] = normalized_user_agent
         return headers or None
 
+    @staticmethod
+    def _matches_endpoint_host(base_url: str | None, expected_host: str) -> bool:
+        """严格匹配官方 API 主机，避免向兼容端点发送供应商专属参数。"""
+        try:
+            return (urlsplit(str(base_url or "")).hostname or "").lower() == expected_host
+        except ValueError:
+            return False
+
+    @classmethod
+    def _build_openai_prompt_cache_options(
+            cls,
+            *,
+            provider: str,
+            base_url: str | None,
+            use_responses_api: bool | None,
+            prompt_cache_key: str | None,
+            default_headers: dict[str, str] | None,
+            model_kwargs: dict[str, Any],
+    ) -> tuple[dict[str, str] | None, dict[str, Any]]:
+        """为 OpenAI 与 xAI 官方端点构造稳定提示词缓存路由参数。"""
+        cache_key = str(prompt_cache_key or "").strip()
+        headers = dict(default_headers or {})
+        kwargs = dict(model_kwargs)
+        provider_name = str(provider or "").strip().lower()
+        if not cache_key:
+            return headers or None, kwargs
+
+        is_openai = provider_name in {"chatgpt", "openai"} and cls._matches_endpoint_host(
+            base_url,
+            "api.openai.com",
+        )
+        is_xai = provider_name == "xai" and cls._matches_endpoint_host(
+            base_url,
+            "api.x.ai",
+        )
+        if not is_openai and not is_xai:
+            return headers or None, kwargs
+
+        if is_xai and use_responses_api is not True:
+            headers["x-grok-conv-id"] = cache_key
+            return headers, kwargs
+
+        extra_body = dict(kwargs.get("extra_body") or {})
+        extra_body["prompt_cache_key"] = cache_key
+        kwargs["extra_body"] = extra_body
+        return headers or None, kwargs
+
+    @staticmethod
+    def _with_prompt_cache_control(
+            model_cls: type,
+            cache_control: dict[str, str],
+    ) -> type:
+        """创建在最终模型绑定阶段保留缓存控制参数的适配类。"""
+
+        class PromptCachingModel(model_cls):
+            """在 LangChain 工具绑定后仍保留提示词缓存参数的模型适配器。"""
+
+            def bind(self, **kwargs: Any) -> Any:
+                """绑定调用参数，并补入当前 Provider 的默认缓存控制。"""
+                kwargs.setdefault("cache_control", dict(cache_control))
+                return super().bind(**kwargs)
+
+        PromptCachingModel.__name__ = f"PromptCaching{model_cls.__name__}"
+        return PromptCachingModel
+
+    @classmethod
+    def _use_anthropic_prompt_cache(
+            cls,
+            *,
+            provider: str,
+            runtime: dict[str, Any],
+            prompt_cache_key: str | None,
+    ) -> bool:
+        """判断当前运行时是否为可安全启用缓存的 Anthropic 官方端点。"""
+        return (
+            bool(str(prompt_cache_key or "").strip())
+            and str(provider or "").strip().lower() == "anthropic"
+            and str(runtime.get("runtime") or "").strip().lower()
+            == "anthropic_compatible"
+            and cls._matches_endpoint_host(
+                runtime.get("base_url"),
+                "api.anthropic.com",
+            )
+        )
+
     @classmethod
     def _should_use_openai_responses_api(
             cls,
@@ -983,6 +1069,7 @@ class LLMHelper:
             use_proxy: bool | None = None,
             api_protocol: str | None = None,
             web_search_mode: str | None = None,
+            prompt_cache_key: str | None = None,
     ):
         """
         获取LLM实例
@@ -1006,6 +1093,7 @@ class LLMHelper:
         :param web_search_mode: 联网搜索模式
             （local/builtin/auto/disabled）。未显式传入时使用配置项
             ``LLM_WEB_SEARCH_MODE``。
+        :param prompt_cache_key: 同一 Agent 会话内稳定且脱敏的提示词缓存路由键。
         :return: LLM实例
         """
         provider_name = str(provider if provider is not None else settings.LLM_PROVIDER).lower()
@@ -1093,6 +1181,14 @@ class LLMHelper:
             runtime=runtime,
             api_protocol=effective_api_protocol,
         )
+        default_headers, openai_model_kwargs = cls._build_openai_prompt_cache_options(
+            provider=provider_name,
+            base_url=runtime.get("base_url"),
+            use_responses_api=use_responses_api,
+            prompt_cache_key=prompt_cache_key,
+            default_headers=default_headers,
+            model_kwargs=thinking_kwargs,
+        )
         llm_proxy = _resolve_llm_proxy(use_proxy)
 
         if runtime["runtime"] == "google":
@@ -1146,6 +1242,16 @@ class LLMHelper:
 
             from app.agent.llm.provider import LLMProviderManager
 
+            bedrock_model_cls = ChatBedrockConverse
+            if (
+                    str(prompt_cache_key or "").strip()
+                    and runtime.get("supports_prompt_cache")
+            ):
+                bedrock_model_cls = cls._with_prompt_cache_control(
+                    ChatBedrockConverse,
+                    {"type": "default"},
+                )
+
             aws_region = runtime.get("aws_region") or "us-east-1"
             aws_auth = runtime.get("aws_auth") or {}
             # Bearer 认证需要跳过 SigV4 签名并注入 Authorization 头，SigV4 认证
@@ -1158,7 +1264,7 @@ class LLMHelper:
                 use_proxy=use_proxy,
                 read_timeout=settings.LLM_TOOL_TIMEOUT,
             )
-            model = ChatBedrockConverse(
+            model = bedrock_model_cls(
                 model_id=model_name,
                 client=bedrock_client,
                 temperature=temperature_value,
@@ -1167,7 +1273,18 @@ class LLMHelper:
         elif runtime["runtime"] in {"anthropic_compatible", "copilot_anthropic"}:
             from langchain_anthropic import ChatAnthropic
 
-            model = ChatAnthropic(
+            anthropic_model_cls = ChatAnthropic
+            if cls._use_anthropic_prompt_cache(
+                    provider=provider_name,
+                    runtime=runtime,
+                    prompt_cache_key=prompt_cache_key,
+            ):
+                anthropic_model_cls = cls._with_prompt_cache_control(
+                    ChatAnthropic,
+                    {"type": "ephemeral"},
+                )
+
+            model = anthropic_model_cls(
                 model=model_name,
                 api_key=runtime["api_key"],
                 base_url=runtime["base_url"],
@@ -1208,7 +1325,7 @@ class LLMHelper:
                 default_headers=default_headers,
                 use_responses_api=use_responses_api,
                 output_version=("responses/v1" if use_responses_api else None),
-                **thinking_kwargs,
+                **openai_model_kwargs,
             )
 
         # 优先使用 provider / models.dev 目录中的上下文上限，减少用户手填成本。

app/agent/llm/provider.py

@@ -1665,6 +1665,20 @@ class LLMProviderManager(metaclass=Singleton):
             await self.get_models_dev_data(use_proxy=use_proxy)
         ).get(models_dev_provider_id, {}) or {}
 
+    @staticmethod
+    def _models_dev_model_candidates(
+            provider_id: str,
+            model_id: str,
+    ) -> tuple[str, ...]:
+        """生成模型目录查询候选，兼容 Provider 添加的透明模型前缀。"""
+        candidates = [model_id]
+        if model_id.startswith("models/"):
+            candidates.append(model_id.removeprefix("models/"))
+        if provider_id == "amazon-bedrock" and "." in model_id:
+            # Cross-region Inference Profile 会增加 us./eu./global. 等前缀。
+            candidates.append(model_id.split(".", 1)[1])
+        return tuple(dict.fromkeys(candidates))
+
     async def _models_dev_model(
             self,
             provider_id: str,
@@ -1684,15 +1698,32 @@ class LLMProviderManager(metaclass=Singleton):
         if not isinstance(models, dict):
             return None
 
-        candidates = [model_id]
-        if model_id.startswith("models/"):
-            candidates.append(model_id.removeprefix("models/"))
-
-        for candidate in candidates:
+        for candidate in self._models_dev_model_candidates(provider_id, model_id):
             if candidate in models:
                 return models[candidate]
         return None
 
+    @staticmethod
+    def _metadata_supports_prompt_cache(metadata: Any) -> bool:
+        """从统一模型元数据中判断是否声明了提示词缓存能力。"""
+        if not isinstance(metadata, dict):
+            return False
+
+        explicit_capability = metadata.get("prompt_cache")
+        if isinstance(explicit_capability, bool):
+            return explicit_capability
+
+        capabilities = metadata.get("capabilities")
+        if isinstance(capabilities, dict):
+            explicit_capability = capabilities.get("prompt_cache")
+            if isinstance(explicit_capability, bool):
+                return explicit_capability
+
+        cost = metadata.get("cost")
+        return isinstance(cost, dict) and any(
+            key in cost for key in ("cache_read", "cache_write")
+        )
+
     def _cached_models_dev_model(
             self,
             provider_id: str,
@@ -1719,11 +1750,7 @@ class LLMProviderManager(metaclass=Singleton):
         if not isinstance(models, dict):
             return None
 
-        candidates = [model_id]
-        if model_id.startswith("models/"):
-            candidates.append(model_id.removeprefix("models/"))
-
-        for candidate in candidates:
+        for candidate in self._models_dev_model_candidates(provider_id, model_id):
             if candidate in models:
                 return models[candidate]
         return None
@@ -3112,6 +3139,9 @@ class LLMProviderManager(metaclass=Singleton):
             "model_id": model,
             "model_record": model_record,
             "model_metadata": model_metadata,
+            "supports_prompt_cache": self._metadata_supports_prompt_cache(
+                model_metadata
+            ),
             "default_headers": None,
             "use_responses_api": None,
             "auth_mode": "api_key",

app/agent/middleware/activity_log.py

@@ -3,7 +3,7 @@
 
 按日期存储在 CONFIG_PATH/agent/activity/YYYY-MM-DD.md 中，
 每次 Agent 执行完毕后自动调用 LLM 对本轮对话生成简洁的活动摘要，
-并在每次 Agent 启动时注入轻量索引，完整日志由工具按需查询。
+系统提示词只注入稳定的检索规则，完整日志由工具按需查询。
 """
 
 import asyncio
@@ -459,12 +459,8 @@ async def _summarize_with_llm(conversation_text: str) -> Optional[str]:
 
 
 ACTIVITY_LOG_SYSTEM_PROMPT = """<activity_log>
-<activity_log_index>
-{activity_log_index}
-</activity_log_index>
-
 <activity_log_guidelines>
-    The index only shows recent dates and entry counts, not full log contents.
+    Activity log contents and indexes are not included in the default context.
     Use `query_activity_log` only when the user references previous work, asks to continue a prior task, or recent activity is clearly relevant.
     Activity logs are read-only and retained for {retention_days} days; use MEMORY.md for durable preferences.
 </activity_log_guidelines>
@@ -473,10 +469,10 @@ ACTIVITY_LOG_SYSTEM_PROMPT = """<activity_log>
 
 
 class ActivityLogMiddleware(AgentMiddleware[ActivityLogState, ContextT, ResponseT]):  # noqa
-    """自动记录 Agent 活动日志并注入轻量索引的中间件。
+    """自动记录 Agent 活动日志并注入稳定检索规则的中间件。
 
     - abefore_agent: 加载近几天的活动日志索引
-    - awrap_model_call: 将活动日志索引和检索规则注入系统提示词
+    - awrap_model_call: 将固定的活动日志检索规则注入系统提示词
     - aafter_agent: 从本次对话中提取摘要并追加到当日日志文件
 
     参数：
@@ -516,31 +512,9 @@ class ActivityLogMiddleware(AgentMiddleware[ActivityLogState, ContextT, Response
         """获取指定日期的日志文件路径。"""
         return AsyncPath(self.activity_dir) / f"{date_str}.md"
 
-    def _format_activity_log(self, contents: dict[str, str]) -> str:
-        """格式化活动日志索引用于系统提示词注入。"""
-        if not contents:
-            return ACTIVITY_LOG_SYSTEM_PROMPT.format(
-                activity_log_index="(近期暂无活动日志索引。需要历史上下文时可调用 query_activity_log。)",
-                retention_days=self.retention_days,
-            )
-
-        # 按日期排序（最近的在前）
-        sorted_dates = sorted(contents.keys(), reverse=True)
-        sections = []
-        for date_str in sorted_dates:
-            content = contents[date_str].strip()
-            if content:
-                sections.append(f"### {date_str}\n{content}")
-
-        if not sections:
-            return ACTIVITY_LOG_SYSTEM_PROMPT.format(
-                activity_log_index="(近期暂无活动日志索引。需要历史上下文时可调用 query_activity_log。)",
-                retention_days=self.retention_days,
-            )
-
-        log_body = "\n".join(sections)
+    def _format_activity_log(self, _contents: dict[str, str]) -> str:
+        """生成不受活动日志内容变化影响的系统提示词。"""
         return ACTIVITY_LOG_SYSTEM_PROMPT.format(
-            activity_log_index=log_body,
             retention_days=self.retention_days,
         )
 

app/agent/middleware/usage.py

@@ -51,6 +51,18 @@ class UsageMiddleware(AgentMiddleware):
 
         return None
 
+    @classmethod
+    def _first_int(
+        cls,
+        candidates: tuple[tuple[Any, tuple[str, ...]], ...],
+    ) -> int | None:
+        """按优先级返回首个可用的 usage 整数值。"""
+        for container, keys in candidates:
+            value = cls._lookup_int(container, *keys)
+            if value is not None:
+                return value
+        return None
+
     @classmethod
     def _extract_model_name(cls, model: Any) -> str | None:
         return (
@@ -82,6 +94,131 @@ class UsageMiddleware(AgentMiddleware):
             or {}
         )
 
+        input_token_details = None
+        if usage_metadata:
+            getter = getattr(usage_metadata, "get", None)
+            input_token_details = (
+                getter("input_token_details")
+                if callable(getter)
+                else getattr(usage_metadata, "input_token_details", None)
+            )
+
+        cache_read_tokens = cls._first_int(
+            (
+                (
+                    input_token_details,
+                    (
+                        "cache_read",
+                        "cached_tokens",
+                        "cache_read_input_tokens",
+                        "cacheReadInputTokens",
+                    ),
+                ),
+                (
+                    token_usage,
+                    (
+                        "prompt_cache_hit_tokens",
+                        "cache_read_input_tokens",
+                        "cacheReadInputTokens",
+                    ),
+                ),
+                (
+                    response_metadata,
+                    (
+                        "prompt_cache_hit_tokens",
+                        "cache_read_input_tokens",
+                        "cacheReadInputTokens",
+                        "cached_tokens",
+                    ),
+                ),
+            )
+        )
+        if cache_read_tokens is None:
+            cache_read_tokens = cls._first_int(
+                (
+                    (
+                        token_usage.get("prompt_tokens_details", {}),
+                        ("cached_tokens", "cache_read"),
+                    ),
+                    (
+                        token_usage.get("input_tokens_details", {}),
+                        ("cached_tokens", "cache_read"),
+                    ),
+                )
+            )
+
+        cache_write_tokens = cls._first_int(
+            (
+                (
+                    input_token_details,
+                    (
+                        "cache_creation",
+                        "cache_write",
+                        "cache_write_tokens",
+                        "cache_write_input_tokens",
+                        "cacheWriteInputTokens",
+                    ),
+                ),
+                (
+                    token_usage,
+                    (
+                        "cache_creation_input_tokens",
+                        "cache_write_tokens",
+                        "cache_write_input_tokens",
+                        "cacheWriteInputTokens",
+                    ),
+                ),
+                (
+                    response_metadata,
+                    (
+                        "cache_creation_input_tokens",
+                        "cache_write_tokens",
+                        "cache_write_input_tokens",
+                        "cacheWriteInputTokens",
+                    ),
+                ),
+            )
+        )
+        if cache_write_tokens is None:
+            cache_write_tokens = cls._first_int(
+                (
+                    (
+                        token_usage.get("prompt_tokens_details", {}),
+                        ("cache_write_tokens", "cache_creation"),
+                    ),
+                    (
+                        token_usage.get("input_tokens_details", {}),
+                        ("cache_write_tokens", "cache_creation"),
+                    ),
+                )
+            )
+        cache_write_ttl_tokens = sum(
+            cls._lookup_int(
+                input_token_details,
+                ttl_key,
+            )
+            or 0
+            for ttl_key in (
+                "ephemeral_5m_input_tokens",
+                "ephemeral_1h_input_tokens",
+            )
+        )
+        if cache_write_ttl_tokens:
+            cache_write_tokens = cache_write_ttl_tokens
+
+        cache_miss_tokens = cls._first_int(
+            (
+                (
+                    token_usage,
+                    ("prompt_cache_miss_tokens", "cache_miss_input_tokens"),
+                ),
+                (
+                    response_metadata,
+                    ("prompt_cache_miss_tokens", "cache_miss_input_tokens"),
+                ),
+            )
+        )
+
         if input_tokens is None:
             input_tokens = cls._lookup_int(
                 token_usage,
@@ -94,6 +231,27 @@ class UsageMiddleware(AgentMiddleware):
                 "prompt_token_count",
                 "input_tokens",
             )
+        if input_tokens is None:
+            bedrock_input_tokens = cls._lookup_int(token_usage, "inputTokens")
+            if bedrock_input_tokens is not None:
+                input_tokens = (
+                    bedrock_input_tokens
+                    + (cache_read_tokens or 0)
+                    + (cache_write_tokens or 0)
+                )
+        if input_tokens is None and any(
+            value is not None
+            for value in (
+                cache_read_tokens,
+                cache_write_tokens,
+                cache_miss_tokens,
+            )
+        ):
+            input_tokens = (
+                (cache_read_tokens or 0)
+                + (cache_write_tokens or 0)
+                + (cache_miss_tokens or 0)
+            )
 
         if output_tokens is None:
             output_tokens = cls._lookup_int(
@@ -113,8 +271,24 @@ class UsageMiddleware(AgentMiddleware):
         if total_tokens is None:
             total_tokens = cls._lookup_int(response_metadata, "total_token_count")
 
+        has_cache_usage = any(
+            value is not None
+            for value in (
+                cache_read_tokens,
+                cache_write_tokens,
+                cache_miss_tokens,
+            )
+        )
         has_usage = any(
-            value is not None for value in (input_tokens, output_tokens, total_tokens)
+            value is not None
+            for value in (
+                input_tokens,
+                output_tokens,
+                total_tokens,
+                cache_read_tokens,
+                cache_write_tokens,
+                cache_miss_tokens,
+            )
         )
         resolved_input = input_tokens or 0
         resolved_output = output_tokens or 0
@@ -123,12 +297,32 @@ class UsageMiddleware(AgentMiddleware):
             if total_tokens is not None
             else resolved_input + resolved_output
         )
+        resolved_cache_read = cache_read_tokens or 0
+        resolved_cache_write = cache_write_tokens or 0
+        uncached_input_tokens = (
+            cache_miss_tokens
+            if cache_miss_tokens is not None
+            else max(
+                resolved_input - resolved_cache_read - resolved_cache_write,
+                0,
+            )
+        )
+        cache_hit_ratio = (
+            resolved_cache_read / resolved_input
+            if has_cache_usage and resolved_input
+            else None
+        )
 
         return {
             "has_usage": has_usage,
+            "cache_usage_available": has_cache_usage,
             "input_tokens": resolved_input,
             "output_tokens": resolved_output,
             "total_tokens": resolved_total,
+            "cache_read_input_tokens": resolved_cache_read,
+            "cache_write_input_tokens": resolved_cache_write,
+            "uncached_input_tokens": uncached_input_tokens,
+            "cache_hit_ratio": cache_hit_ratio,
         }
 
     async def awrap_model_call(
@@ -157,9 +351,14 @@ class UsageMiddleware(AgentMiddleware):
                 if ai_message
                 else {
                     "has_usage": False,
+                    "cache_usage_available": False,
                     "input_tokens": 0,
                     "output_tokens": 0,
                     "total_tokens": 0,
+                    "cache_read_input_tokens": 0,
+                    "cache_write_input_tokens": 0,
+                    "uncached_input_tokens": 0,
+                    "cache_hit_ratio": None,
                 }
             )
             context_window_tokens = self._extract_context_window_tokens(request.model)

app/chain/message.py

@@ -1349,6 +1349,33 @@ class MessageChain(ChainBase):
             f"排队消息数: {status.get('pending_messages', 0)}",
             f"最后更新: {status.get('last_updated_at') or '暂无'}",
         ]
+        if status.get("cache_usage_available"):
+            last_cache_ratio = status.get("last_cache_hit_ratio")
+            total_cache_ratio = status.get("total_cache_hit_ratio")
+            lines.insert(
+                6,
+                "最近一次缓存: "
+                f"命中 {cls._format_token_count(status.get('last_cache_read_input_tokens'))} / "
+                f"写入 {cls._format_token_count(status.get('last_cache_write_input_tokens'))} / "
+                f"未命中 {cls._format_token_count(status.get('last_uncached_input_tokens'))}"
+                + (
+                    f" ({last_cache_ratio * 100:.2f}%)"
+                    if last_cache_ratio is not None
+                    else ""
+                ),
+            )
+            lines.insert(
+                8,
+                "当前会话累计缓存: "
+                f"命中 {cls._format_token_count(status.get('total_cache_read_input_tokens'))} / "
+                f"写入 {cls._format_token_count(status.get('total_cache_write_input_tokens'))} / "
+                f"未命中 {cls._format_token_count(status.get('total_uncached_input_tokens'))}"
+                + (
+                    f" ({total_cache_ratio * 100:.2f}%)"
+                    if total_cache_ratio is not None
+                    else ""
+                ),
+            )
         return "\n".join(lines)
 
     def remote_session_status(

app/schemas/event.py

@@ -119,6 +119,11 @@ class AgentTokensUsageEventData(BaseEventData):
     input_tokens: int = Field(default=0, description="输入 tokens")
     output_tokens: int = Field(default=0, description="输出 tokens")
     total_tokens: int = Field(default=0, description="总 tokens")
+    cache_read_input_tokens: int = Field(default=0, description="从提示词缓存读取的输入 tokens")
+    cache_write_input_tokens: int = Field(default=0, description="写入提示词缓存的输入 tokens")
+    uncached_input_tokens: int = Field(default=0, description="未命中缓存的输入 tokens")
+    cache_hit_ratio: Optional[float] = Field(default=None, description="提示词缓存命中率")
+    cache_usage_available: bool = Field(default=False, description="供应商是否返回缓存用量明细")
     model_call_count: int = Field(default=0, description="模型调用次数")
     success: bool = Field(default=False, description="Agent 执行是否成功")
     error: Optional[str] = Field(default=None, description="失败原因")