feat: refine job handling by filtering active jobs and updating date context in prompts

app/agent/__init__.py

@@ -22,7 +22,11 @@ from app.agent.callback import StreamingHandler
 from app.agent.llm import LLMHelper
 from app.agent.memory import memory_manager
 from app.agent.middleware.activity_log import ActivityLogMiddleware
-from app.agent.middleware.jobs import JobsMiddleware
+from app.agent.middleware.jobs import (
+    JobsMiddleware,
+    filter_active_jobs,
+    load_jobs_metadata,
+)
 from app.agent.middleware.memory import MemoryMiddleware
 from app.agent.middleware.patch_tool_calls import PatchToolCallsMiddleware
 from app.agent.middleware.runtime_config import RuntimeConfigMiddleware
@@ -160,6 +164,9 @@ class ReplyMode(str, Enum):
     CAPTURE_ONLY = "capture_only"
 
 
+HEARTBEAT_SESSION_PREFIX = "__agent_heartbeat_"
+
+
 class MoviePilotAgent:
     """
     MoviePilot AI智能体（基于 LangChain v1 + LangGraph）
@@ -288,6 +295,16 @@ class MoviePilotAgent:
         """
         return self.reply_mode == ReplyMode.DISPATCH
 
+    @property
+    def is_heartbeat_session(self) -> bool:
+        """
+        是否为后台心跳会话。
+
+        心跳场景只负责检查并执行待处理 job，不需要携带近期活动日志，
+        否则会让这类高频后台调用持续带入无关动态上下文，影响缓存命中率。
+        """
+        return self.session_id.startswith(HEARTBEAT_SESSION_PREFIX)
+
     def _should_stream(self) -> bool:
         """
         判断是否应启用流式输出：
@@ -430,10 +447,6 @@ class MoviePilotAgent:
                 RuntimeConfigMiddleware(),
                 # 记忆管理
                 MemoryMiddleware(memory_dir=str(agent_runtime_manager.memory_dir)),
-                # 活动日志
-                ActivityLogMiddleware(
-                    activity_dir=str(agent_runtime_manager.activity_dir),
-                ),
                 # 上下文压缩
                 SummarizationMiddleware(
                     model=non_streaming_model, trigger=("fraction", 0.85)
@@ -444,6 +457,14 @@ class MoviePilotAgent:
                 UsageMiddleware(on_usage=self._record_usage),
             ]
 
+            if not self.is_heartbeat_session:
+                middlewares.insert(
+                    4,
+                    ActivityLogMiddleware(
+                        activity_dir=str(agent_runtime_manager.activity_dir),
+                    ),
+                )
+
             # 工具选择
             if max_tools > 0:
                 middlewares.append(
@@ -1085,8 +1106,17 @@ class AgentManager:
         由定时调度器周期性调用，每次使用独立的会话避免上下文干扰。
         """
         try:
+            active_jobs = filter_active_jobs(
+                await load_jobs_metadata([str(agent_runtime_manager.jobs_dir)])
+            )
+            # 先在本地判断是否存在活跃任务。没有任务时直接短路，避免一次完整
+            # 的后台 Agent/LLM 空调用。
+            if not active_jobs:
+                logger.info("智能体心跳唤醒：没有活跃任务，跳过模型调用")
+                return
+
             # 每次使用唯一的 session_id，避免共享上下文
-            session_id = f"__agent_heartbeat_{uuid.uuid4().hex[:12]}__"
+            session_id = f"{HEARTBEAT_SESSION_PREFIX}{uuid.uuid4().hex[:12]}__"
             user_id = SYSTEM_INTERNAL_USER_ID
 
             logger.info("智能体心跳唤醒：开始检查待处理任务...")

app/agent/middleware/jobs.py

@@ -21,6 +21,7 @@ from app.log import logger
 
 # JOB.md 文件最大限制为 1MB
 MAX_JOB_FILE_SIZE = 1 * 1024 * 1024
+ACTIVE_JOB_STATUSES = ("pending", "in_progress")
 
 
 class JobMetadata(TypedDict):
@@ -143,6 +144,9 @@ async def _alist_jobs(source_path: AsyncPath) -> list[JobMetadata]:
     if not job_dirs:
         return []
 
+    # 显式按目录名排序，避免文件系统返回顺序不稳定时破坏提示词缓存命中。
+    job_dirs.sort(key=lambda p: p.name.casefold())
+
     # 解析 JOB.md
     for job_path in job_dirs:
         job_md_path = job_path / "JOB.md"
@@ -161,6 +165,31 @@ async def _alist_jobs(source_path: AsyncPath) -> list[JobMetadata]:
     return jobs
 
 
+def filter_active_jobs(jobs_metadata: list[JobMetadata]) -> list[JobMetadata]:
+    """筛选需要参与心跳检查的活跃任务。
+
+    这里严格以任务状态为准，只保留 `pending` / `in_progress`。
+    `recurring` 任务执行完成后按约定应回写为 `pending`，因此无需再额外放宽
+    到 `completed`，避免已结束任务被重复注入后台心跳。
+    """
+    return [
+        job for job in jobs_metadata if job.get("status") in ACTIVE_JOB_STATUSES
+    ]
+
+
+async def load_jobs_metadata(source_paths: list[str]) -> list[JobMetadata]:
+    """按顺序加载多个 jobs 目录下的任务元数据。"""
+    all_jobs: list[JobMetadata] = []
+    for source_path_str in source_paths:
+        source_path = AsyncPath(source_path_str)
+        if not await source_path.exists():
+            await source_path.mkdir(parents=True, exist_ok=True)
+            continue
+        source_jobs = await _alist_jobs(source_path)
+        all_jobs.extend(source_jobs)
+    return all_jobs
+
+
 JOBS_SYSTEM_PROMPT = """
 <jobs_system>
 You have a **scheduled jobs** system that allows you to track and execute long-running or recurring tasks.
@@ -289,13 +318,8 @@ class JobsMiddleware(AgentMiddleware[JobsState, ContextT, ResponseT]):  # noqa
         """将任务文档注入模型请求的系统消息中。"""
         jobs_metadata = request.state.get("jobs_metadata", [])  # noqa
 
-        # 过滤：只展示活跃任务（pending / in_progress / recurring）
-        active_jobs = [
-            j
-            for j in jobs_metadata
-            if j["status"] in ("pending", "in_progress")
-            or (j["schedule"] == "recurring" and j["status"] not in ("cancelled",))
-        ]
+        # 仅注入真正活跃的任务，避免把已完成任务继续塞进心跳上下文。
+        active_jobs = filter_active_jobs(jobs_metadata)
 
         jobs_list = self._format_jobs_list(active_jobs)
         jobs_location = self.sources[0] if self.sources else ""
@@ -322,18 +346,9 @@ class JobsMiddleware(AgentMiddleware[JobsState, ContextT, ResponseT]):  # noqa
         if "jobs_metadata" in state:
             return None
 
-        all_jobs: list[JobMetadata] = []
-
-        # 遍历源加载任务
-        for source_path_str in self.sources:
-            source_path = AsyncPath(source_path_str)
-            if not await source_path.exists():
-                await source_path.mkdir(parents=True, exist_ok=True)
-                continue
-            source_jobs = await _alist_jobs(source_path)
-            all_jobs.extend(source_jobs)
-
-        return JobsStateUpdate(jobs_metadata=all_jobs)
+        return JobsStateUpdate(
+            jobs_metadata=await load_jobs_metadata(self.sources)
+        )
 
     async def awrap_model_call(
         self,
@@ -347,4 +362,10 @@ class JobsMiddleware(AgentMiddleware[JobsState, ContextT, ResponseT]):  # noqa
         return await handler(modified_request)
 
 
-__all__ = ["JobMetadata", "JobsMiddleware"]
+__all__ = [
+    "ACTIVE_JOB_STATUSES",
+    "JobMetadata",
+    "JobsMiddleware",
+    "filter_active_jobs",
+    "load_jobs_metadata",
+]

app/agent/middleware/skills.py

@@ -227,6 +227,9 @@ async def _alist_skills(source_path: AsyncPath) -> list[SkillMetadata]:
     if not skill_dirs:
         return []
 
+    # 显式按目录名排序，避免文件系统返回顺序不稳定时破坏提示词缓存命中。
+    skill_dirs.sort(key=lambda p: p.name.casefold())
+
     # 解析已下载的 SKILL.md
     for skill_path in skill_dirs:
         skill_md_path = skill_path / "SKILL.md"

app/agent/prompt/__init__.py

@@ -286,8 +286,10 @@ class PromptManager:
                 f"{settings.DB_POSTGRESQL_TARGET}/{settings.DB_POSTGRESQL_DATABASE})"
             )
 
+        # 保留日期用于提供“今天是哪天”的稳定上下文，但不再注入秒级时间，
+        # 避免每次请求都生成不同的 system prompt，影响 provider 侧 cache 命中率。
         info_lines = [
-            f"- 当前时间: {strftime('%Y-%m-%d %H:%M:%S')}",
+            f"- 当前日期: {strftime('%Y-%m-%d')}",
             f"- 运行环境: {SystemUtils.platform} {'docker' if SystemUtils.is_docker() else ''}",
             f"- 主机名: {hostname}",
             f"- IP地址: {ip_address}",
@@ -426,11 +428,11 @@ class PromptManager:
             return text
 
         context = cls._normalize_template_context(template_context)
-        missing_fields = sorted(field for field in required_fields if field not in context)
+        missing_fields = sorted(f for f in required_fields if f not in context)
         if missing_fields:
             raise PromptConfigError(
                 f"系统任务定义 `{task_type}` 的 `{field_name}` 缺少变量: "
-                + ", ".join(f"`{field}`" for field in missing_fields)
+                + ", ".join(f"`{f}`" for f in missing_fields)
             )
 
         # 这里统一做字符串替换，让 YAML 成为后台任务文案的唯一行为来源。