refactor: 重构重试策略、条件函数与上下文注入逻辑

主要变更：
1. 替换旧retries参数为RetryPolicy配置
2. 重构条件函数，支持上下文参数与动态依赖判断
3. 更新上下文注入逻辑，支持软依赖与更清晰的注入描述
4. 新增sglang CLI命令与相关配置
5. 格式化代码统一列表与参数写法
6. 更新文档与测试用例适配新API

pyproject.toml

@@ -44,6 +44,7 @@ piptool     = "pyflowx.cli.piptool:main"
 pymake      = "pyflowx.cli.pymake:main"
 reseticon   = "pyflowx.cli.reseticoncache:main"
 scrcap      = "pyflowx.cli.screenshot:main"
+sglang      = "pyflowx.cli.llm.sglang:main"
 sshcopy     = "pyflowx.cli.sshcopyid:main"
 taskk       = "pyflowx.cli.taskkill:main"
 wch         = "pyflowx.cli.which:main"

src/pyflowx/__init__.py

@@ -4,9 +4,15 @@
 --------
 * :class:`TaskSpec` —— 不可变任务描述符（唯一需要配置的东西）。
 * :class:`Graph` —— 由一组 spec 构建的 DAG；负责校验、分层、可视化。
-* :func:`run` —— 以 ``sequential`` / ``thread`` / ``async`` 策略执行图。
+* :func:`run` ——以 ``sequential`` / ``thread`` / ``async`` / ``dependency``
+  策略执行图。
 * :class:`RunReport` —— 类型化、可查询的运行结果。
 * :class:`Context` —— 整体上下文注入的标注标记。
+* :class:`RetryPolicy` —— 重试策略（max_attempts/delay/backoff/jitter/retry_on）。
+* :class:`TaskHooks` —— 任务生命周期钩子（pre_run/post_run/on_failure）。
+* :class:`GraphDefaults` —— 图级默认值。
+* :func:`compose` —— 编程式组合多图。
+* :func:`task_template` —— 批量生成相似 TaskSpec 的工厂。
 * 状态后端：:class:`StateBackend`、:class:`MemoryBackend`、:class:`JSONBackend`。
 
 快速上手
@@ -18,7 +24,7 @@
 
     graph = px.Graph.from_specs([
         px.TaskSpec("extract", extract),
-        px.TaskSpec("double", double, ("extract",)),
+        px.TaskSpec("double", double, depends_on=("extract",)),
     ])
     report = px.run(graph, strategy="sequential")
     print(report["double"])  # [2, 4, 6]
@@ -29,23 +35,18 @@
     from pyflowx.conditions import IS_WINDOWS, BuiltinConditions
 
     graph = px.Graph.from_specs([
-        # 使用命令列表
         px.TaskSpec("list_files", cmd=["ls", "-la"]),
-        # 使用 shell 命令
         px.TaskSpec("check_git", cmd="git status"),
-        # 条件执行：仅在 Windows 上运行
         px.TaskSpec(
             "win_only",
             cmd=["dir"],
             conditions=(IS_WINDOWS,)
         ),
-        # 条件执行：仅在 git 已安装时运行
         px.TaskSpec(
             "git_check",
             cmd=["git", "--version"],
             conditions=(BuiltinConditions.HAS_INSTALLED("git"),)
         ),
-        # 命令不存在时自动跳过（而非失败）
         px.TaskSpec(
             "optional_build",
             cmd=["maturin", "build"],
@@ -58,6 +59,10 @@
 from __future__ import annotations
 
 from .conditions import (
+    IS_LINUX,
+    IS_MACOS,
+    IS_POSIX,
+    IS_WINDOWS,
     BuiltinConditions,
     Condition,
     Constants,
@@ -74,20 +79,33 @@ from .errors import (
     TaskTimeoutError,
 )
 from .executors import Strategy, run
-from .graph import Graph, GraphComposer
+from .graph import Graph, GraphComposer, GraphDefaults, compose
 from .report import RunReport
 from .runner import CliExitCode, CliRunner
 from .storage import JSONBackend, MemoryBackend, StateBackend
-from .task import TaskCmd, TaskEvent, TaskResult, TaskSpec, TaskStatus
+from .task import (
+    CacheKeyFn,
+    RetryPolicy,
+    TaskCmd,
+    TaskEvent,
+    TaskHooks,
+    TaskResult,
+    TaskSpec,
+    TaskStatus,
+    task_template,
+)
 
-__version__ = "0.2.6"
+__version__ = "0.3.0"
 
 __all__ = [
+    "IS_LINUX",
+    "IS_MACOS",
+    "IS_POSIX",
+    "IS_WINDOWS",
     "BuiltinConditions",
+    "CacheKeyFn",
     "CliExitCode",
-    # CLI 运行器
     "CliRunner",
-    # 条件判断
     "Condition",
     "Constants",
     "Context",
@@ -95,28 +113,28 @@ __all__ = [
     "DuplicateTaskError",
     "Graph",
     "GraphComposer",
+    "GraphDefaults",
     "InjectionError",
     "JSONBackend",
     "MemoryBackend",
     "MissingDependencyError",
-    # 错误
     "PyFlowXError",
+    "RetryPolicy",
     "RunReport",
-    # 状态后端
     "StateBackend",
     "StorageError",
     "Strategy",
     "TaskCmd",
     "TaskEvent",
     "TaskFailedError",
+    "TaskHooks",
     "TaskResult",
-    # 核心类型
     "TaskSpec",
     "TaskStatus",
     "TaskTimeoutError",
-    # 辅助（高级）
     "build_call_args",
+    "compose",
     "describe_injection",
-    # 执行
     "run",
+    "task_template",
 ]

src/pyflowx/cli/envpy.py

@@ -112,9 +112,9 @@ def main() -> None:
     args = parser.parse_args()
 
     if args.command == "mirror":
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [px.TaskSpec("set_pip_mirror", fn=set_pip_mirror, args=(args.name,), kwargs={"token": args.token})]
+            px.TaskSpec("set_pip_mirror", fn=set_pip_mirror, args=(args.name,), kwargs={"token": args.token})
-        )
+        ])
     else:
         parser.print_help()
         return

src/pyflowx/cli/envqt.py

@@ -43,13 +43,13 @@ def main() -> None:
             px.TaskSpec(
                 "envqt_install",
                 cmd=["sudo", "apt", "install", "-y", *QT_LIBS],
-                conditions=(lambda: Constants.IS_LINUX,),
+                conditions=(lambda _: Constants.IS_LINUX,),
                 verbose=True,
             ),
             px.TaskSpec(
                 "envqt_fonts",
                 cmd=["sudo", "apt", "install", "-y", *CHINESE_FONTS],
-                conditions=(lambda: Constants.IS_LINUX,),
+                conditions=(lambda _: Constants.IS_LINUX,),
                 verbose=True,
             ),
         ],

src/pyflowx/cli/gittool.py

@@ -37,7 +37,7 @@ def init_sub_dirs() -> None:
                 px.TaskSpec(
                     "init",
                     cmd=["git", "init"],
-                    conditions=(not_has_git_repo,),
+                    conditions=(lambda _: not_has_git_repo(),),
                     cwd=subdir,
                 ),
                 px.TaskSpec("add", cmd=["git", "add", "."], depends_on=("init",)),
@@ -70,7 +70,7 @@ def main() -> None:
         graphs={
             # 添加并提交
             "a": px.Graph.from_specs([
-                px.TaskSpec("add", cmd=["git", "add", "."], conditions=(has_files,)),
+                px.TaskSpec("add", cmd=["git", "add", "."], conditions=(lambda _: has_files(),)),
                 px.TaskSpec("commit", cmd=["git", "commit", "-m", "chore: update"], depends_on=("add",)),
             ]),
             # 清理
@@ -80,10 +80,13 @@ def main() -> None:
             ]),
             # 初始化、添加并提交
             "i": px.Graph.from_specs([
-                px.TaskSpec("init", cmd=["git", "init"], conditions=(not_has_git_repo,)),
+                px.TaskSpec("init", cmd=["git", "init"], conditions=(lambda _: not_has_git_repo(),)),
-                px.TaskSpec("add", cmd=["git", "add", "."], depends_on=("init",), conditions=(has_files,)),
+                px.TaskSpec("add", cmd=["git", "add", "."], depends_on=("init",), conditions=(lambda _: has_files(),)),
                 px.TaskSpec(
-                    "commit", cmd=["git", "commit", "-m", "init commit"], depends_on=("add",), conditions=(has_files,)
+                    "commit",
+                    cmd=["git", "commit", "-m", "init commit"],
+                    depends_on=("add",),
+                    conditions=(lambda _: has_files(),),
                 ),
             ]),
             # 初始化子目录

src/pyflowx/cli/llm/sglang.py

@@ -0,0 +1,55 @@
+"""使用 SGLang 运行本地模型."""
+
+import argparse
+from pathlib import Path
+
+import pyflowx as px
+from pyflowx.conditions import BuiltinConditions
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Run a local model using SGLang.")
+    parser.add_argument("name", help="Model name.")
+    parser.add_argument("--dir", default=None, help="Directory of model.")
+    args = parser.parse_args()
+
+    if not args.name:
+        parser.error("name is required")
+
+    model_dir = Path(args.dir) if args.dir else Path.home() / ".models" / args.name.split("/")[-1]
+    if not model_dir.exists():
+        parser.error(f"Model directory {model_dir} does not exist.")
+
+    graph = px.Graph.from_specs([
+        px.TaskSpec(
+            name="download",
+            cmd=[
+                "uv",
+                "install",
+                "sglang[all]",
+            ],
+            conditions=(BuiltinConditions.NOT(BuiltinConditions.HAS_INSTALLED("sglang")),),
+            verbose=True,
+        ),
+        px.TaskSpec(
+            name="run",
+            cmd=[
+                "uvx",
+                "sglang",
+                "serve",
+                "--model-path",
+                str(model_dir),
+                "--host",
+                "0.0.0.0",
+                "--port",
+                "8000",
+                "--mem-fraction-static",
+                "0.88",
+                "--context-length",
+                "32768",
+            ],
+            verbose=True,
+        ),
+    ])
+
+    px.run(graph, verbose=True)

src/pyflowx/cli/piptool.py

@@ -21,12 +21,10 @@ PACKAGE_DIR = "packages"
 REQUIREMENTS_FILE = "requirements.txt"
 
 # 受保护的包名集合
-_PROTECTED_PACKAGES: frozenset[str] = frozenset(
+_PROTECTED_PACKAGES: frozenset[str] = frozenset({
-    {
+    "pyflowx",
-        "pyflowx",
+    "bitool",
-        "bitool",
+})
-    }
-)
 
 
 # ============================================================================
@@ -161,37 +159,33 @@ def main() -> None:
     if args.command == "i":
         graph = px.Graph.from_specs([px.TaskSpec("pip_install", cmd=["pip", "install", *args.packages], verbose=True)])
     elif args.command == "u":
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [px.TaskSpec("pip_uninstall", fn=pip_uninstall, args=(args.packages,), verbose=True)]
+            px.TaskSpec("pip_uninstall", fn=pip_uninstall, args=(args.packages,), verbose=True)
-        )
+        ])
     elif args.command == "r":
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [
+            px.TaskSpec(
-                px.TaskSpec(
+                "pip_reinstall",
-                    "pip_reinstall",
+                fn=pip_reinstall,
-                    fn=pip_reinstall,
+                args=(args.packages,),
-                    args=(args.packages,),
+                kwargs={"offline": args.offline},
-                    kwargs={"offline": args.offline},
+                verbose=True,
-                    verbose=True,
+            )
-                )
+        ])
-            ]
-        )
     elif args.command == "d":
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [
+            px.TaskSpec(
-                px.TaskSpec(
+                "pip_download",
-                    "pip_download",
+                fn=pip_download,
-                    fn=pip_download,
+                args=(args.packages,),
-                    args=(args.packages,),
+                kwargs={"offline": args.offline},
-                    kwargs={"offline": args.offline},
+                verbose=True,
-                    verbose=True,
+            )
-                )
+        ])
-            ]
-        )
     elif args.command == "up":
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [px.TaskSpec("pip_upgrade", cmd=["python", "-m", "pip", "install", "--upgrade", "pip"], verbose=True)]
+            px.TaskSpec("pip_upgrade", cmd=["python", "-m", "pip", "install", "--upgrade", "pip"], verbose=True)
-        )
+        ])
     elif args.command == "f":
         graph = px.Graph.from_specs([px.TaskSpec("pip_freeze", fn=pip_freeze, verbose=True)])
     else:

src/pyflowx/conditions.py

@@ -1,7 +1,12 @@
 """条件判断模块.
 
-提供平台条件、应用安装条件等预定义条件判断函数,
+所有条件均为 ``Callable[[Context], bool]``，接收依赖上下文映射（可能为空）。
-用于 TaskSpec 的条件执行功能.
+这使得条件可基于上游任务的运行时返回值做决策，实现动态分支。
+
+内置条件分两类：
+1. *静态条件* —— 不依赖上下文（平台/环境变量/安装检查），通过 ``_static``
+   包装忽略传入的 context，便于作为模块级常量使用。
+2. *上下文条件* —— 基于上游结果判断，如 :meth:`BuiltinConditions.DEP_EQUALS`。
 """
 
 from __future__ import annotations
@@ -11,10 +16,11 @@ import shutil
 import subprocess
 import sys
 from pathlib import Path
-from typing import Callable
+from typing import Any, Callable
 
-# 条件判断函数类型
+from .task import Condition, Context
-Condition = Callable[[], bool]
+
+__all__ = ["BuiltinConditions", "Condition", "Constants"]
 
 
 class Constants:
@@ -26,65 +32,56 @@ class Constants:
     IS_POSIX: bool = sys.platform != "win32"
 
 
+def _static(predicate: Callable[[], bool], name: str) -> Condition:
+    """将无参谓词包装为忽略上下文的 :class:`Condition`。"""
+
+    def _cond(_ctx: Context) -> bool:
+        return predicate()
+
+    _cond.__name__ = name
+    return _cond
+
+
+# ---------------------------------------------------------------------- #
+# 模块级静态条件常量
+# ---------------------------------------------------------------------- #
+IS_WINDOWS: Condition = _static(lambda: Constants.IS_WINDOWS, "IS_WINDOWS")
+IS_LINUX: Condition = _static(lambda: Constants.IS_LINUX, "IS_LINUX")
+IS_MACOS: Condition = _static(lambda: Constants.IS_MACOS, "IS_MACOS")
+IS_POSIX: Condition = _static(lambda: Constants.IS_POSIX, "IS_POSIX")
+
+
 class BuiltinConditions:
-    """内置条件判断函数集合."""
+    """内置条件判断函数集合.
 
+    静态条件工厂返回忽略上下文的 :class:`Condition`；上下文条件工厂返回
+    会读取依赖结果的 :class:`Condition`。
+    """
+
+    # ------------------------------------------------------------------ #
+    # 静态条件
+    # ------------------------------------------------------------------ #
     @staticmethod
-    def PYTHON_VERSION(major: int, minor: int | None = None) -> bool:
+    def PYTHON_VERSION(major: int, minor: int | None = None) -> Condition:
-        """检查 Python 版本是否匹配.
+        """检查 Python 版本是否匹配."""
-
-        Parameters
-        ----------
-        major : int
-            主版本号.
-        minor : int | None
-            次版本号, 若为 None 则仅检查主版本.
-
-        Returns
-        -------
-        bool
-            版本是否匹配.
-        """
         if minor is None:
-            return sys.version_info.major == major
+            return _static(lambda: sys.version_info.major == major, f"PYTHON_VERSION({major})")
-        return sys.version_info.major == major and sys.version_info.minor == minor
+        return _static(
+            lambda: sys.version_info.major == major and sys.version_info.minor == minor,
+            f"PYTHON_VERSION({major},{minor})",
+        )
 
     @staticmethod
-    def PYTHON_VERSION_AT_LEAST(major: int, minor: int = 0) -> bool:
+    def PYTHON_VERSION_AT_LEAST(major: int, minor: int = 0) -> Condition:
-        """检查 Python 版本是否 >= 指定版本.
+        """检查 Python 版本是否 >= 指定版本."""
-
+        return _static(lambda: sys.version_info >= (major, minor), f"PYTHON_VERSION_AT_LEAST({major},{minor})")
-        Parameters
-        ----------
-        major : int
-            主版本号.
-        minor : int
-            次版本号.
-
-        Returns
-        -------
-        bool
-            当前版本是否 >= 指定版本.
-        """
-        return sys.version_info >= (major, minor)
 
     @staticmethod
     def IS_RUNNING(app_name: str) -> Condition:
-        """检查指定应用是否正在运行.
+        """检查指定应用是否正在运行."""
-
-        Parameters
-        ----------
-        app_name : str
-            应用名称 (如 "explorer", "chrome", "python").
-
-        Returns
-        -------
-        Condition
-            条件判断函数.
-        """
 
         def _check() -> bool:
             if Constants.IS_WINDOWS:
-                # Windows: 使用 tasklist 命令
                 result = subprocess.run(
                     ["tasklist", "/nh", "/fi", f"imagename eq {app_name}"],
                     capture_output=True,
@@ -93,148 +90,119 @@ class BuiltinConditions:
                 )
                 return app_name.lower() in result.stdout.lower()
             else:
-                # Linux/macOS: 使用 pgrep 命令
+                result = subprocess.run(["pgrep", "-x", app_name], capture_output=True, check=False)
-                result = subprocess.run(
-                    ["pgrep", "-x", app_name],
-                    capture_output=True,
-                    check=False,
-                )
                 return result.returncode == 0
 
-        _check.__name__ = f"IS_RUNNING({app_name!r})"
+        return _static(_check, f"IS_RUNNING({app_name!r})")
-        return _check
 
     @staticmethod
     def HAS_INSTALLED(app_name: str) -> Condition:
-        """检查指定应用是否已安装.
+        """检查指定应用是否已安装."""
-
+        return _static(lambda: shutil.which(app_name) is not None, f"HAS_INSTALLED({app_name!r})")
-        Parameters
-        ----------
-        app_name : str
-            应用名称 (如 "git", "python", "pytest").
-
-        Returns
-        -------
-        Condition
-            条件判断函数.
-        """
-
-        def _check() -> bool:
-            return shutil.which(app_name) is not None
-
-        _check.__name__ = f"HAS_INSTALLED({app_name!r})"
-        return _check
 
     @staticmethod
-    def DIR_EXISTS(dir: Path) -> Condition:
+    def DIR_EXISTS(path: Path) -> Condition:
         """路径是否存在."""
-        return dir.exists
+        return _static(path.exists, f"DIR_EXISTS({path!r})")
 
     @staticmethod
     def ENV_VAR_EXISTS(var_name: str) -> Condition:
-        """检查环境变量是否存在.
+        """检查环境变量是否存在."""
-
+        return _static(lambda: var_name in os.environ, f"ENV_VAR_EXISTS({var_name!r})")
-        Parameters
-        ----------
-        var_name : str
-            环境变量名.
-
-        Returns
-        -------
-        Condition
-            条件判断函数.
-        """
-
-        def _check() -> bool:
-            return var_name in os.environ
-
-        _check.__name__ = f"ENV_VAR_EXISTS({var_name!r})"
-        return _check
 
     @staticmethod
     def ENV_VAR_EQUALS(var_name: str, value: str) -> Condition:
-        """检查环境变量是否等于指定值.
+        """检查环境变量是否等于指定值."""
+        return _static(
+            lambda: os.environ.get(var_name) == value,
+            f"ENV_VAR_EQUALS({var_name!r},{value!r})",
+        )
 
-        Parameters
+    # ------------------------------------------------------------------ #
-        ----------
+    # 上下文条件：基于上游依赖结果
-        var_name : str
+    # ------------------------------------------------------------------ #
-            环境变量名.
+    @staticmethod
-        value : str
+    def DEP_EQUALS(dep_name: str, value: Any) -> Condition:
-            期望的值.
+        """上游任务 ``dep_name`` 的返回值等于 ``value`` 时为真。
 
-        Returns
+        若依赖未在上下文中（被跳过或未执行），返回 ``False``。
-        -------
-        Condition
-            条件判断函数.
         """
 
-        def _check() -> bool:
+        def _cond(ctx: Context) -> bool:
-            return os.environ.get(var_name) == value
+            return dep_name in ctx and ctx[dep_name] == value
 
-        _check.__name__ = f"ENV_VAR_EQUALS({var_name!r}, {value!r})"
+        _cond.__name__ = f"DEP_EQUALS({dep_name!r},{value!r})"
-        return _check
+        return _cond
 
     @staticmethod
-    def NOT(condition: Condition) -> Condition:
+    def DEP_MATCHES(dep_name: str, predicate: Callable[[Any], bool]) -> Condition:
-        """对条件取反.
+        """上游任务 ``dep_name`` 的返回值满足 ``predicate`` 时为真。
 
-        Parameters
+        依赖不存在时返回 ``False``。
-        ----------
-        condition : Condition
-            原始条件.
-
-        Returns
-        -------
-        Condition
-            取反后的条件.
         """
 
-        def _check() -> bool:
+        def _cond(ctx: Context) -> bool:
-            return not condition()
+            if dep_name not in ctx:
+                return False
+            try:
+                return predicate(ctx[dep_name])
+            except Exception:
+                return False
 
-        _check.__name__ = f"NOT({getattr(condition, '__name__', repr(condition))})"
+        _cond.__name__ = f"DEP_MATCHES({dep_name!r},{getattr(predicate, '__name__', 'pred')})"
-        return _check
+        return _cond
+
+    @staticmethod
+    def DEP_PRESENT(dep_name: str) -> Condition:
+        """上游任务 ``dep_name`` 存在于上下文（即已成功执行）时为真。"""
+
+        def _cond(ctx: Context) -> bool:
+            return dep_name in ctx and ctx[dep_name] is not None
+
+        _cond.__name__ = f"DEP_PRESENT({dep_name!r})"
+        return _cond
+
+    @staticmethod
+    def DEP_TRUTHY(dep_name: str) -> Condition:
+        """上游任务 ``dep_name`` 的返回值为真值时为真。"""
+
+        def _cond(ctx: Context) -> bool:
+            return bool(ctx.get(dep_name))
+
+        _cond.__name__ = f"DEP_TRUTHY({dep_name!r})"
+        return _cond
+
+    # ------------------------------------------------------------------ #
+    # 逻辑组合
+    # ------------------------------------------------------------------ #
+    @staticmethod
+    def NOT(condition: Condition) -> Condition:
+        """对条件取反."""
+
+        def _cond(ctx: Context) -> bool:
+            return not condition(ctx)
+
+        _cond.__name__ = f"NOT({getattr(condition, '__name__', repr(condition))})"
+        return _cond
 
     @staticmethod
     def AND(*conditions: Condition) -> Condition:
-        """多个条件的逻辑与.
+        """多个条件的逻辑与."""
 
-        Parameters
+        def _cond(ctx: Context) -> bool:
-        ----------
+            return all(c(ctx) for c in conditions)
-        *conditions : Condition
-            条件列表.
-
-        Returns
-        -------
-        Condition
-            组合条件.
-        """
-
-        def _check() -> bool:
-            return all(c() for c in conditions)
 
         names = [getattr(c, "__name__", repr(c)) for c in conditions]
-        _check.__name__ = f"AND({', '.join(names)})"
+        _cond.__name__ = f"AND({', '.join(names)})"
-        return _check
+        return _cond
 
     @staticmethod
     def OR(*conditions: Condition) -> Condition:
-        """多个条件的逻辑或.
+        """多个条件的逻辑或."""
 
-        Parameters
+        def _cond(ctx: Context) -> bool:
-        ----------
+            return any(c(ctx) for c in conditions)
-        *conditions : Condition
-            条件列表.
-
-        Returns
-        -------
-        Condition
-            组合条件.
-        """
-
-        def _check() -> bool:
-            return any(c() for c in conditions)
 
         names = [getattr(c, "__name__", repr(c)) for c in conditions]
-        _check.__name__ = f"OR({', '.join(names)})"
+        _cond.__name__ = f"OR({', '.join(names)})"
-        return _check
+        return _cond

src/pyflowx/context.py

@@ -1,18 +1,16 @@
 """上下文注入：把上游结果转换为函数参数。
 
 本机制让用户可以编写普通函数，其参数名*就是*依赖声明，从而消除其他
-DAG 库中泛滥的样板包装器（如 ``def wrapper(): return fn(workflow.get_task_result('x'))``）。
+DAG 库中泛滥的样板包装器。
 
 注入规则（按顺序求值）
 ----------------------
-1. **标注为** :class:`Context` 的参数接收完整结果映射。适用于需要遍历
+1. **标注为** :class:`Context` 的参数接收完整结果映射（含硬依赖与软依赖）。
-   所有输入的任务。
+2. **名称匹配某个依赖**（硬或软）的参数接收该依赖的结果。
-2. **名称匹配某个依赖**的参数接收该依赖的结果。
 3. ``**kwargs`` 参数以 dict 形式接收*所有*依赖结果。
 4. ``TaskSpec.args`` / ``TaskSpec.kwargs`` 为*非依赖*参数提供静态值。
 
-若某参数无法解析且无默认值，则抛出 :class:`~pyflowx.errors.InjectionError`，
+若某参数无法解析且无默认值，则抛出 :class:`~pyflowx.errors.InjectionError`。
-并附带精确错误信息。
 """
 
 from __future__ import annotations
@@ -27,21 +25,11 @@ __all__ = ["Context", "_is_context_annotation", "build_call_args", "describe_inj
 
 
 def _is_context_annotation(annotation: Any) -> bool:
-    """判断参数标注是否为（或指向）``Context``。
+    """判断参数标注是否为（或指向）``Context``。"""
-
-    处理三种形式：
-    * ``Context`` 别名对象本身；
-    * ``__name__``/``_name`` 为 ``Context`` 或 ``Mapping`` 的 typing 别名；
-    * *字符串*标注（``from __future__ import annotations`` 会在运行时
-      把所有标注变为字符串），如 ``"Context"`` 或 ``"px.Context"``。
-    """
     if annotation is Context:
         return True
-    # `from __future__ import annotations` 产生的字符串标注。
     if isinstance(annotation, str):
-        # 匹配 "Context"、"px.Context"、"pyflowx.Context" 等。
         return annotation == "Context" or annotation.endswith(".Context")
-    # 按限定名匹配，支持 ``from pyflowx import Context`` 再导出。
     name = getattr(annotation, "__name__", None) or getattr(annotation, "_name", None)
     return name in ("Context", "Mapping")
 
@@ -52,39 +40,22 @@ def build_call_args(
 ) -> tuple[tuple[Any, ...], dict[str, Any]]:
     """解析用于调用 ``spec.fn`` 的 ``(args, kwargs)``。
 
-    参数
+    ``context`` 必须已包含所有硬依赖与软依赖的结果（软依赖被跳过时由
-    ----
+    执行器填入 :attr:`TaskSpec.defaults` 中的默认值）。
-    spec:
-        任务 spec，提供 ``fn``、``depends_on``、``args``、``kwargs``。
-    context:
-        依赖名 -> 结果值的映射。仅保证本任务自身的 ``depends_on`` 条目
-        存在；其他任务的结果被排除，以保持注入的确定性。
-
-    返回
-    ----
-    (args, kwargs)
-        可直接展开为 ``spec.fn(*args, **kwargs)``。
-
-    抛出
-    ----
-    InjectionError
-        若必需参数无法满足，或静态 ``kwargs`` 与注入依赖名冲突。
     """
-    # 使用 effective_fn 而不是 fn，以支持 cmd 参数
     fn = spec.effective_fn
     sig = inspect.signature(fn)
     params = sig.parameters
 
-    # 检测特殊参数类型。
     var_keyword = next(
         (p for p in params.values() if p.kind == inspect.Parameter.VAR_KEYWORD),
         None,
     )
 
-    # 与本任务相关的上下文子集。
+    # 本任务相关的上下文子集：硬依赖 + 软依赖。
-    dep_context: dict[str, Any] = {name: context[name] for name in spec.depends_on if name in context}
+    all_deps = set(spec.depends_on) | set(spec.soft_depends_on)
+    dep_context: dict[str, Any] = {name: context[name] for name in all_deps if name in context}
 
-    # 检测静态 kwargs 与依赖名的冲突。
     collisions = set(spec.kwargs) & set(dep_context)
     if collisions:
         raise InjectionError(
@@ -96,8 +67,6 @@ def build_call_args(
     injected_kwargs: dict[str, Any] = {}
     leftover_dep_results: dict[str, Any] = dict(dep_context)
 
-    # 被 spec.args 消费的位置参数。记录哪些参数名已被位置填充，
-    # 以便在基于名称的注入（依赖 / Context / 静态 kwargs）时跳过。
     positional_params: list[str] = []
     positional_kinds = (
         inspect.Parameter.POSITIONAL_ONLY,
@@ -106,33 +75,25 @@ def build_call_args(
     for pname, param in params.items():
         if param.kind in positional_kinds:
             positional_params.append(pname)
-    # 前 len(spec.args) 个位置参数由 spec.args 填充。
     args_filled: set[str] = set(positional_params[: len(spec.args)])
 
     for pname, param in params.items():
-        # 跳过已被位置 spec.args 填充的参数。
         if pname in args_filled:
             continue
 
-        # 规则 1：标注为 Context -> 完整映射。
         if _is_context_annotation(param.annotation):
             injected_kwargs[pname] = dep_context
             continue
 
-        # 规则 2：名称匹配某个依赖。
         if pname in dep_context:
             injected_kwargs[pname] = dep_context[pname]
             leftover_dep_results.pop(pname, None)
             continue
 
-        # 规则 3：在循环后通过 **kwargs 处理。
-
-        # 规则 4：静态 kwargs 填充其余参数。
         if pname in spec.kwargs:
             injected_kwargs[pname] = spec.kwargs[pname]
             continue
 
-        # 该参数无来源：必须有默认值，否则报错。
         if param.default is inspect.Parameter.empty and param.kind not in (
             inspect.Parameter.VAR_POSITIONAL,
             inspect.Parameter.VAR_KEYWORD,
@@ -142,9 +103,7 @@ def build_call_args(
                 f"parameter {pname!r} has no dependency, static value, or default.",
             )
 
-    # 规则 3：**kwargs 吞掉剩余依赖结果。
     if var_keyword is not None and leftover_dep_results:
-        # 先合并静态 kwargs，再合并依赖结果（冲突已在上方拒绝）。
         merged = dict(spec.kwargs)
         merged.update(injected_kwargs)
         merged.update(leftover_dep_results)
@@ -154,14 +113,9 @@ def build_call_args(
 
 
 def describe_injection(spec: TaskSpec[Any]) -> str:
-    """生成任务参数注入方式的人类可读描述。
+    """生成任务参数注入方式的人类可读描述。供 ``dry_run`` 使用。"""
-
-    供 ``dry_run`` 使用，在不执行的情况下展示执行计划。
-    """
-    # 使用 effective_fn 而不是 fn，以支持 cmd 参数
     fn = spec.effective_fn
     sig = inspect.signature(fn)
-    # 确定哪些位置参数由 spec.args 填充。
     positional_params = [
         p
         for p, param in sig.parameters.items()
@@ -172,6 +126,7 @@ def describe_injection(spec: TaskSpec[Any]) -> str:
         )
     ]
     args_filled = set(positional_params[: len(spec.args)])
+    all_deps = set(spec.depends_on) | set(spec.soft_depends_on)
     parts = []
     for pname, param in sig.parameters.items():
         if pname in args_filled:
@@ -179,8 +134,9 @@ def describe_injection(spec: TaskSpec[Any]) -> str:
             parts.append(f"{pname}={spec.args[idx]!r}")
         elif _is_context_annotation(param.annotation):
             parts.append(f"{pname}=<Context>")
-        elif pname in spec.depends_on:
+        elif pname in all_deps:
-            parts.append(f"{pname}=<result:{pname}>")
+            tag = "soft" if pname in spec.soft_depends_on else "dep"
+            parts.append(f"{pname}=<{tag}:{pname}>")
         elif pname in spec.kwargs:
             parts.append(f"{pname}={spec.kwargs[pname]!r}")
         elif param.default is not inspect.Parameter.empty:

src/pyflowx/examples/etl_pipeline.py

@@ -55,7 +55,9 @@ def main() -> None:
             depends_on=("extract_customers", "extract_orders"),
             tags=("transform",),
         ),
-        px.TaskSpec("load", load, depends_on=("transform",), retries=1, tags=("load",)),
+        px.TaskSpec(
+            "load", load, depends_on=("transform",), retry=px.RetryPolicy(max_attempts=1, delay=1.0), tags=("load",)
+        ),
     ])
 
     print("=== Execution plan ===")

src/pyflowx/executors.py

@@ -1,15 +1,26 @@
 """执行器与公共 :func:`run` 入口。
 
-三种执行策略共享一个逐层驱动器：
+四种执行策略：
 
 * ``sequential`` —— 确定性、一次一个任务。最适合调试。
 * ``thread``     —— 通过线程池实现层内并发。最适合 I/O 密集型同步任务。
 * ``async``      —— 通过 ``asyncio.gather`` 实现层内并发。同步任务被
                     卸载到线程池；异步任务运行在事件循环上。最适合
                     I/O 密集型异步任务。
+* ``dependency`` —— 依赖驱动调度：任务在其所有硬依赖完成后立即启动，
+                    无需等待同层其他任务。最大化并行度。
 
-三者都遵循 ``retries``、``timeout``、上下文注入、状态后端（续跑），
+所有策略共享统一异步内核，支持：
-并向观察者发出 :class:`~pyflowx.task.TaskEvent`。
+* :class:`RetryPolicy`（max_attempts/delay/backoff/jitter/retry_on）
+* 软依赖注入与默认值
+* :class:`TaskHooks`（pre_run/post_run/on_failure）
+* 按任务策略覆盖
+* 优先级排序（同层内）
+* 并发限制（concurrency_key + concurrency_limits）
+* ``continue_on_error``
+* ``cache_key`` 存储键
+* 条件判断（上下文感知）
+* 状态后端（续跑）
 """
 
 from __future__ import annotations
@@ -18,6 +29,7 @@ import asyncio
 import concurrent.futures
 import inspect
 import logging
+import threading
 from datetime import datetime
 from typing import Any, Awaitable, Callable, Literal, Mapping, cast
 
@@ -26,24 +38,24 @@ from .errors import TaskFailedError, TaskTimeoutError
 from .graph import Graph
 from .report import RunReport
 from .storage import StateBackend, resolve_backend
-from .task import TaskEvent, TaskResult, TaskSpec, TaskStatus
+from .task import TaskEvent, TaskHooks, TaskResult, TaskSpec, TaskStatus
 
 logger = logging.getLogger("pyflowx")
 
 # 观察者回调类型。
 EventCallback = Callable[[TaskEvent], None]
-Strategy = Literal["sequential", "thread", "async"]
+Strategy = Literal["sequential", "thread", "async", "dependency"]
 
 
+# ---------------------------------------------------------------------- #
+# 辅助
+# ---------------------------------------------------------------------- #
 def _is_async_fn(spec: TaskSpec[Any]) -> bool:
     """判断 ``spec.effective_fn`` 是否为协程函数。"""
     return inspect.iscoroutinefunction(spec.effective_fn)
 
 
-def _emit(
+def _emit(on_event: EventCallback | None, result: TaskResult[Any]) -> None:
-    on_event: EventCallback | None,
-    result: TaskResult[Any],
-) -> None:
     """若注册了回调则触发一个观察者事件。"""
     if on_event is None:
         return
@@ -59,71 +71,60 @@ def _emit(
     )
 
 
-def _log_retry(spec: TaskSpec[Any], attempts: int, max_attempts: int, exc: BaseException) -> None:
+def _log_retry(spec: TaskSpec[Any], attempt: int, max_attempts: int, exc: BaseException) -> None:
-    """记录重试日志（sync 与 async 共享，便于测试覆盖）。"""
+    """记录重试日志。"""
     logger.warning(
         "task %r failed (attempt %d/%d): %r; retrying",
         spec.name,
-        attempts,
+        attempt,
         max_attempts,
         exc,
     )
 
 
-def _finalize_failure(
+def _run_hooks(hooks: TaskHooks, fn_name: str, *args: Any) -> None:
-    result: TaskResult[Any],
+    """安全调用钩子（异常仅记录，不影响任务状态）。"""
-    layer_idx: int | None,
+    hook: Callable[..., None] | None = getattr(hooks, fn_name, None)
-    on_event: EventCallback | None = None,
+    if hook is None:
-) -> None:
+        return
-    """标记任务为 FAILED 并抛出 TaskFailedError。"""
+    try:
-    result.status = TaskStatus.FAILED
+        hook(*args)
-    result.finished_at = datetime.now()
+    except Exception as exc:
-    _emit(on_event, result)
+        logger.warning("hook %s raised: %r", fn_name, exc)
-    raise TaskFailedError(
-        task=result.spec.name,
-        cause=result.error if result.error is not None else RuntimeError("unknown"),
-        attempts=result.attempts,
-        layer=layer_idx,
-    )
 
 
 def _check_upstream_skipped(
     spec: TaskSpec[Any],
     report: RunReport | None,
 ) -> tuple[bool, str | None]:
-    """检查上游任务是否被 SKIPPED。
+    """检查硬依赖上游任务是否被 SKIPPED 或 FAILED。
 
-    Returns
+    软依赖不影响本检查——软依赖被跳过时注入默认值。
-    -------
-    tuple[bool, str | None]
-        (是否应该跳过, 跳过原因)
     """
     if report is None:
         return False, None
 
-    # 若任务允许上游跳过，则不检查上游状态
     if spec.allow_upstream_skip:
         return False, None
 
     for dep in spec.depends_on:
-        if dep in report.results and report.results[dep].status == TaskStatus.SKIPPED:
+        if dep not in report.results:
-            return True, f"上游任务 '{dep}' 被跳过"
+            continue
+        dep_status = report.results[dep].status
+        if dep_status in (TaskStatus.SKIPPED, TaskStatus.FAILED):
+            return True, f"上游任务 '{dep}' 状态为 {dep_status.value}"
     return False, None
 
 
-def _evaluate_skip_reason(spec: TaskSpec[Any]) -> str | None:
+def _evaluate_conditions(spec: TaskSpec[Any], context: Mapping[str, Any]) -> str | None:
-    """单次求值所有条件与 skip_if_missing，返回跳过原因或 None。
+    """求值所有条件，返回跳过原因或 ``None``。
 
-    与旧实现不同：条件只求值一次。`should_execute()` 内部会调用所有条件，
+    条件接收上下文映射（硬依赖 + 软依赖结果）。
-    若再分支调用 `_is_cmd_available` 之外的逻辑会二次求值（如
-    ``IS_RUNNING`` 会 spawn 两次 subprocess）。此处显式逐个求值并记录结果，
-    失败原因直接来自求值过程，无需二次调用。
     """
-    # 1. 逐个求值条件，记录失败项。
     failed_conditions: list[str] = []
     for condition in spec.conditions:
         try:
-            ok = condition()
+            ok = condition(context)
         except Exception:
             ok = False
             name = getattr(condition, "__name__", None) or "匿名条件(执行错误)"
@@ -135,7 +136,6 @@ def _evaluate_skip_reason(spec: TaskSpec[Any]) -> str | None:
     if failed_conditions:
         return f"条件不满足: {', '.join(failed_conditions)}"
 
-    # 2. skip_if_missing 检查（仅对 list[str] 命令有效）。
     if spec.skip_if_missing and not spec._is_cmd_available():
         cmd_name = spec.cmd[0] if isinstance(spec.cmd, list) and spec.cmd else "unknown"
         return f"命令不存在: {cmd_name}"
@@ -148,10 +148,7 @@ def _make_skipped_result(
     reason: str,
     on_event: EventCallback | None,
 ) -> TaskResult[Any]:
-    """构造 SKIPPED 的 TaskResult 并发出事件、打印日志。
+    """构造 SKIPPED 的 TaskResult。"""
-
-    sync 与 async 执行路径共用，消除重复的 result 构造/emit/print 代码。
-    """
     result: TaskResult[Any] = TaskResult(
         spec=spec,
         status=TaskStatus.SKIPPED,
@@ -165,31 +162,118 @@ def _make_skipped_result(
     return result
 
 
+def _build_context(
+    spec: TaskSpec[Any],
+    global_context: Mapping[str, Any],
+    report: RunReport | None = None,  # noqa: ARG001
+) -> dict[str, Any]:
+    """构建本任务的上下文：硬依赖 + 软依赖（含默认值回退）。
+
+    硬依赖：若上游 SKIPPED/FAILED 则不注入（本任务通常也会被跳过）。
+    软依赖：上游成功则注入其值；否则注入 ``spec.defaults`` 中的默认值（或 ``None``）。
+    """
+    ctx: dict[str, Any] = {}
+
+    for dep in spec.depends_on:
+        if dep in global_context:
+            ctx[dep] = global_context[dep]
+
+    for dep in spec.soft_depends_on:
+        if dep in global_context:
+            ctx[dep] = global_context[dep]
+        elif dep in spec.defaults:
+            ctx[dep] = spec.defaults[dep]
+        else:
+            ctx[dep] = None
+
+    return ctx
+
+
+def _apply_cached(
+    name: str,
+    spec: TaskSpec[Any],
+    context: dict[str, Any],
+    report: RunReport,
+    backend: StateBackend,
+    on_event: EventCallback | None,
+) -> bool:
+    """若 ``name`` 命中缓存，写入 context/report 并返回 True。"""
+    storage_key = spec.storage_key(context)
+    if not backend.has(storage_key):
+        return False
+    cached = backend.get(storage_key)
+    context[name] = cached
+    result = TaskResult(spec=spec, status=TaskStatus.SKIPPED, value=cached, reason="缓存命中")
+    report.results[name] = result
+    _emit(on_event, result)
+    logger.info("task %r skipped (cached)", name)
+    return True
+
+
 def _prepare_for_execution(
     spec: TaskSpec[Any],
+    context: Mapping[str, Any],
     report: RunReport | None,
     on_event: EventCallback | None,
 ) -> TaskResult[Any] | None:
-    """执行前的统一预检：上游跳过 / 条件跳过。
+    """执行前预检：上游跳过 / 条件跳过。
 
-    Returns
+    返回 SKIPPED TaskResult 或 ``None``（继续执行）。
-    -------
-    TaskResult | None
-        若应跳过，返回已填好的 SKIPPED 结果；否则返回 None 表示继续执行。
     """
-    # 上游跳过检查
     should_skip, skip_reason = _check_upstream_skipped(spec, report)
     if should_skip:
         return _make_skipped_result(spec, skip_reason or "上游任务被跳过", on_event)
 
-    # 条件 / skip_if_missing 检查（单次求值）
+    skip_reason = _evaluate_conditions(spec, context)
-    skip_reason = _evaluate_skip_reason(spec)
     if skip_reason is not None:
         return _make_skipped_result(spec, skip_reason, on_event)
 
     return None
 
 
+def _finalize_failure(
+    result: TaskResult[Any],
+    layer_idx: int | None,
+    on_event: EventCallback | None = None,
+    continue_on_error: bool = False,
+) -> None:
+    """标记任务为 FAILED。若 ``continue_on_error`` 为真则不抛出异常。"""
+    result.status = TaskStatus.FAILED
+    result.finished_at = datetime.now()
+    _emit(on_event, result)
+    if continue_on_error:
+        logger.warning(
+            "task %r failed but continue_on_error=True; continuing.",
+            result.spec.name,
+        )
+        return
+    raise TaskFailedError(
+        task=result.spec.name,
+        cause=result.error if result.error is not None else RuntimeError("unknown"),
+        attempts=result.attempts,
+        layer=layer_idx,
+    )
+
+
+def _sleep_for_retry(spec: TaskSpec[Any], attempt: int) -> None:
+    """重试前的同步等待。"""
+    wait = spec.retry.wait_seconds(attempt)
+    if wait > 0:
+        import time
+
+        time.sleep(wait)
+
+
+async def _async_sleep_for_retry(spec: TaskSpec[Any], attempt: int) -> None:
+    """重试前的异步等待。"""
+    wait = spec.retry.wait_seconds(attempt)
+    if wait > 0:
+        await asyncio.sleep(wait)
+
+
+# ---------------------------------------------------------------------- #
+# 同步执行内核
+# ---------------------------------------------------------------------- #
 def _run_sync_with_retry(
     spec: TaskSpec[Any],
     context: Mapping[str, Any],
@@ -198,44 +282,47 @@ def _run_sync_with_retry(
     report: RunReport | None = None,
 ) -> TaskResult[Any]:
     """执行同步任务并带重试；返回填充好的 TaskResult。"""
-    # 统一预检：上游跳过 / 条件跳过（条件单次求值）
+    skipped = _prepare_for_execution(spec, context, report, on_event)
-    skipped = _prepare_for_execution(spec, report, on_event)
     if skipped is not None:
         return skipped
 
     result: TaskResult[Any] = TaskResult(spec=spec)
     result.started_at = datetime.now()
-    max_attempts = spec.retries + 1
+    max_attempts = spec.retry.max_attempts
     args, kwargs = build_call_args(spec, context)
 
+    _run_hooks(spec.hooks, "pre_run", spec)
+
     while True:
         result.attempts += 1
         try:
-            result.value = spec.effective_fn(*args, **kwargs)
+            with spec.env_context():
+                result.value = spec.effective_fn(*args, **kwargs)
             result.status = TaskStatus.SUCCESS
             result.finished_at = datetime.now()
+            _run_hooks(spec.hooks, "post_run", spec, result.value)
             return result
         except Exception as exc:
             result.error = exc
-            if result.attempts >= max_attempts:
+            if result.attempts >= max_attempts or not spec.retry.should_retry(exc):
-                _finalize_failure(result, layer_idx, on_event)
+                _run_hooks(spec.hooks, "on_failure", spec, exc)
+                _finalize_failure(result, layer_idx, on_event, spec.continue_on_error)
+                return result
             _log_retry(spec, result.attempts, max_attempts, exc)
-    raise AssertionError("unreachable")  # pragma: no cover
+            _sleep_for_retry(spec, result.attempts)
+    # pragma: no cover
 
 
+# ---------------------------------------------------------------------- #
+# 异步执行内核
+# ---------------------------------------------------------------------- #
 async def _execute_async_task(
     spec: TaskSpec[Any],
     args: tuple[Any, ...],
     kwargs: dict[str, Any],
     loop: asyncio.AbstractEventLoop,
 ) -> Any:
-    """执行异步或同步任务（带超时处理）。
+    """执行异步或同步任务（带超时处理）。"""
-
-    Returns
-    -------
-    Any
-        任务返回值
-    """
     if _is_async_fn(spec):
         coro = cast(Awaitable[Any], spec.effective_fn(*args, **kwargs))
         if spec.timeout is not None:
@@ -243,9 +330,10 @@ async def _execute_async_task(
         else:
             return await coro
     else:
-        # 将同步工作卸载到线程，保持事件循环存活。
+
         def fn_call() -> Any:
-            return spec.effective_fn(*args, **kwargs)
+            with spec.env_context():
+                return spec.effective_fn(*args, **kwargs)
 
         if spec.timeout is not None:
             return await asyncio.wait_for(loop.run_in_executor(None, fn_call), timeout=spec.timeout)
@@ -259,76 +347,74 @@ async def _run_async_with_retry(
     layer_idx: int | None,
     on_event: EventCallback | None = None,
     report: RunReport | None = None,
+    semaphore: asyncio.Semaphore | None = None,
 ) -> TaskResult[Any]:
     """在事件循环上执行任务（同步或异步）并带重试。"""
-    # 统一预检：上游跳过 / 条件跳过（条件单次求值）
+    skipped = _prepare_for_execution(spec, context, report, on_event)
-    skipped = _prepare_for_execution(spec, report, on_event)
     if skipped is not None:
         return skipped
 
-    result: TaskResult[Any] = TaskResult[Any](spec=spec)
+    if semaphore is not None:
+        async with semaphore:
+            return await _run_async_inner(spec, context, layer_idx, on_event, report)
+    return await _run_async_inner(spec, context, layer_idx, on_event, report)
+
+
+async def _run_async_inner(
+    spec: TaskSpec[Any],
+    context: Mapping[str, Any],
+    layer_idx: int | None,
+    on_event: EventCallback | None = None,
+    report: RunReport | None = None,  # noqa: ARG001
+) -> TaskResult[Any]:
+    """异步执行内核的内部实现（已获取 semaphore 后）。"""
+    result: TaskResult[Any] = TaskResult(spec=spec)
     result.started_at = datetime.now()
-    max_attempts = spec.retries + 1
+    max_attempts = spec.retry.max_attempts
     args, kwargs = build_call_args(spec, context)
     loop = asyncio.get_event_loop()
 
+    _run_hooks(spec.hooks, "pre_run", spec)
+
     while True:
         result.attempts += 1
         try:
             result.value = await _execute_async_task(spec, args, kwargs, loop)
             result.status = TaskStatus.SUCCESS
             result.finished_at = datetime.now()
+            _run_hooks(spec.hooks, "post_run", spec, result.value)
             return result
         except asyncio.TimeoutError:
-            result.error = TaskTimeoutError(spec.name, spec.timeout or 0.0)
+            exc: BaseException = TaskTimeoutError(spec.name, spec.timeout or 0.0)
-            if result.attempts >= max_attempts:
+            result.error = exc
-                _finalize_failure(result, layer_idx, on_event)
+            if result.attempts >= max_attempts or not spec.retry.should_retry(exc):
+                _run_hooks(spec.hooks, "on_failure", spec, exc)
+                _finalize_failure(result, layer_idx, on_event, spec.continue_on_error)
+                return result
             logger.warning(
                 "task %r timed out (attempt %d/%d); retrying",
                 spec.name,
                 result.attempts,
                 max_attempts,
             )
+            await _async_sleep_for_retry(spec, result.attempts)
         except Exception as exc:
             result.error = exc
-            if result.attempts >= max_attempts:
+            if result.attempts >= max_attempts or not spec.retry.should_retry(exc):
-                _finalize_failure(result, layer_idx, on_event)
+                _run_hooks(spec.hooks, "on_failure", spec, exc)
+                _finalize_failure(result, layer_idx, on_event, spec.continue_on_error)
+                return result
             _log_retry(spec, result.attempts, max_attempts, exc)
-    raise AssertionError("unreachable")  # pragma: no cover
+            await _async_sleep_for_retry(spec, result.attempts)
+    # pragma: no cover
 
 
 # ---------------------------------------------------------------------- #
-# 层驱动器
+# 层执行器
 # ---------------------------------------------------------------------- #
-def _build_context(
+def _sort_by_priority(layer: list[str], graph: Graph) -> list[str]:
-    spec: TaskSpec[Any],
+    """按优先级降序排序（稳定排序）。"""
-    global_context: Mapping[str, Any],
+    return sorted(layer, key=lambda n: -graph.resolved_spec(n).priority)
-) -> Mapping[str, Any]:
-    """将全局上下文限制为本任务的依赖。"""
-    return {dep: global_context[dep] for dep in spec.depends_on if dep in global_context}
-
-
-def _apply_cached(
-    name: str,
-    graph: Graph,
-    context: dict[str, Any],
-    report: RunReport,
-    backend: StateBackend,
-    on_event: EventCallback | None,
-) -> bool:
-    """若 ``name`` 命中缓存，写入 context/report 并返回 True；否则返回 False。
-
-    sequential / thread / async 三种层驱动共用，消除缓存命中分支的重复代码。
-    """
-    if not backend.has(name):
-        return False
-    cached = backend.get(name)
-    context[name] = cached
-    result = TaskResult(spec=graph.spec(name), status=TaskStatus.SKIPPED, value=cached, reason="缓存命中")
-    report.results[name] = result
-    _emit(on_event, result)
-    logger.info("task %r skipped (cached)", name)
-    return True
 
 
 def _execute_layer_sequential(
@@ -340,14 +426,16 @@ def _execute_layer_sequential(
     layer_idx: int,
     on_event: EventCallback | None,
 ) -> None:
-    """逐个运行某层的任务。"""
+    """逐个运行某层的任务（按优先级排序）。"""
-    for name in layer:
+    for name in _sort_by_priority(layer, graph):
-        spec = graph.spec(name)
+        spec = graph.resolved_spec(name)
-        if _apply_cached(name, graph, context, report, backend, on_event):
+        if _apply_cached(name, spec, context, report, backend, on_event):
             continue
-        result = _run_sync_with_retry(spec, _build_context(spec, context), layer_idx, on_event, report)
+        task_ctx = _build_context(spec, context, report)
+        result = _run_sync_with_retry(spec, task_ctx, layer_idx, on_event, report)
         context[name] = result.value
-        backend.save(name, result.value)
+        if result.status == TaskStatus.SUCCESS:
+            backend.save(spec.storage_key(task_ctx), result.value)
         report.results[name] = result
         _emit(on_event, result)
 
@@ -361,42 +449,68 @@ def _execute_layer_threaded(
     layer_idx: int,
     on_event: EventCallback | None,
     max_workers: int,
+    concurrency_limits: Mapping[str, int],
 ) -> None:
     """在线程池中并发运行某层的任务。"""
-    # 先同步满足已缓存任务。
     to_run: list[str] = []
     for name in layer:
-        if _apply_cached(name, graph, context, report, backend, on_event):
+        spec = graph.resolved_spec(name)
+        task_ctx = _build_context(spec, context, report)
+        if _apply_cached(name, spec, context, report, backend, on_event):
             continue
         to_run.append(name)
 
     if not to_run:
         return
 
+    to_run = _sort_by_priority(to_run, graph)
+
+    # 为每个 concurrency_key 创建线程信号量
+    semaphores: dict[str, threading.Semaphore] = {}
+    for name in to_run:
+        spec = graph.resolved_spec(name)
+        key = spec.concurrency_key
+        if key is not None and key not in semaphores:
+            limit = concurrency_limits.get(key, 1)
+            semaphores[key] = threading.Semaphore(limit)
+
+    context_snapshot = dict(context)
+    lock = threading.Lock()
+
+    def _run_threaded_task(name: str) -> TaskResult[Any]:
+        spec = graph.resolved_spec(name)
+        task_ctx = _build_context(spec, context_snapshot, report)
+        sem = semaphores.get(spec.concurrency_key) if spec.concurrency_key else None
+        if sem is not None:
+            sem.acquire()
+        try:
+            return _run_sync_with_retry(spec, task_ctx, layer_idx, on_event, report)
+        finally:
+            if sem is not None:
+                sem.release()
+
     with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as pool:
         future_to_name: dict[concurrent.futures.Future[TaskResult[Any]], str] = {}
         for name in to_run:
-            spec = graph.spec(name)
+            fut = pool.submit(_run_threaded_task, name)
-            # 为本任务快照上下文以避免竞态。
-            task_ctx = _build_context(spec, context)
-            fut = pool.submit(_run_sync_with_retry, spec, task_ctx, layer_idx, on_event, report)
             future_to_name[fut] = name
 
-        # 统一收集后再写 context，与 async 版本行为一致：
-        # 避免边完成边写共享 dict 造成的可见性不一致。
         completed: dict[str, TaskResult[Any]] = {}
         try:
             for fut in concurrent.futures.as_completed(future_to_name):
                 name = future_to_name[fut]
-                result = fut.result()  # 失败时抛出 TaskFailedError
+                result = fut.result()
                 completed[name] = result
         finally:
-            # 无论是否抛出，都先把已完成任务的结果落盘并写回 context/report。
+            with lock:
-            for name, result in completed.items():
+                for name, result in completed.items():
-                context[name] = result.value
+                    context[name] = result.value
-                backend.save(name, result.value)
+                    if result.status == TaskStatus.SUCCESS:
-                report.results[name] = result
+                        spec = graph.resolved_spec(name)
-                _emit(on_event, result)
+                        task_ctx = _build_context(spec, context_snapshot, report)
+                        backend.save(spec.storage_key(task_ctx), result.value)
+                    report.results[name] = result
+                    _emit(on_event, result)
 
 
 async def _execute_layer_async(
@@ -407,52 +521,122 @@ async def _execute_layer_async(
     backend: StateBackend,
     layer_idx: int,
     on_event: EventCallback | None,
+    concurrency_limits: Mapping[str, int],
 ) -> None:
     """在事件循环上并发运行某层的任务。"""
     to_run: list[str] = []
     for name in layer:
-        if _apply_cached(name, graph, context, report, backend, on_event):
+        spec = graph.resolved_spec(name)
+        if _apply_cached(name, spec, context, report, backend, on_event):
             continue
         to_run.append(name)
 
     if not to_run:
         return
 
-    coros = []
+    to_run = _sort_by_priority(to_run, graph)
-    for name in to_run:
-        spec = graph.spec(name)
-        task_ctx = _build_context(spec, context)
-        coros.append(_run_async_with_retry(spec, task_ctx, layer_idx, on_event, report))
 
+    # 为每个 concurrency_key 创建异步信号量
+    semaphores: dict[str, asyncio.Semaphore] = {}
+    for name in to_run:
+        spec = graph.resolved_spec(name)
+        key = spec.concurrency_key
+        if key is not None and key not in semaphores:
+            limit = concurrency_limits.get(key, 1)
+            semaphores[key] = asyncio.Semaphore(limit)
+
+    context_snapshot = dict(context)
+
+    async def _run_async_task_wrapped(name: str) -> TaskResult[Any]:
+        spec = graph.resolved_spec(name)
+        task_ctx = _build_context(spec, context_snapshot, report)
+        sem = semaphores.get(spec.concurrency_key) if spec.concurrency_key else None
+        if sem is not None:
+            async with sem:
+                return await _run_async_with_retry(spec, task_ctx, layer_idx, on_event, report)
+        return await _run_async_with_retry(spec, task_ctx, layer_idx, on_event, report)
+
+    coros = [_run_async_task_wrapped(name) for name in to_run]
     results = await asyncio.gather(*coros)
     for name, result in zip(to_run, results):
         context[name] = result.value
-        backend.save(name, result.value)
+        if result.status == TaskStatus.SUCCESS:
+            spec = graph.resolved_spec(name)
+            task_ctx = _build_context(spec, context_snapshot, report)
+            backend.save(spec.storage_key(task_ctx), result.value)
         report.results[name] = result
         _emit(on_event, result)
 
 
+# ---------------------------------------------------------------------- #
+# 依赖驱动调度
+# ---------------------------------------------------------------------- #
+async def _drive_dependency_async(
+    graph: Graph,
+    context: dict[str, Any],
+    report: RunReport,
+    backend: StateBackend,
+    on_event: EventCallback | None,
+    concurrency_limits: Mapping[str, int],
+) -> None:
+    """依赖驱动调度：任务在硬依赖完成后立即启动，无层屏障。
+
+    所有任务通过 asyncio 并发调度。同步任务卸载到线程池。
+    """
+    all_names = set(graph.all_specs().keys())
+    semaphores: dict[str, asyncio.Semaphore] = {}
+    for name in all_names:
+        spec = graph.resolved_spec(name)
+        key = spec.concurrency_key
+        if key is not None and key not in semaphores:
+            limit = concurrency_limits.get(key, 1)
+            semaphores[key] = asyncio.Semaphore(limit)
+
+    futures: dict[str, asyncio.Future[TaskResult[Any]]] = {}
+
+    async def _run_task(name: str) -> TaskResult[Any]:
+        spec = graph.resolved_spec(name)
+        # 等待所有硬依赖完成
+        for dep in spec.depends_on:
+            if dep in futures:
+                await futures[dep]
+        # 等待所有软依赖完成（但不检查其状态）
+        for dep in spec.soft_depends_on:
+            if dep in futures:
+                await futures[dep]
+
+        task_ctx = _build_context(spec, context, report)
+        if _apply_cached(name, spec, context, report, backend, on_event):
+            return report.results[name]
+
+        sem = semaphores.get(spec.concurrency_key) if spec.concurrency_key else None
+        if sem is not None:
+            async with sem:
+                result = await _run_async_with_retry(spec, task_ctx, None, on_event, report)
+        else:
+            result = await _run_async_with_retry(spec, task_ctx, None, on_event, report)
+
+        context[name] = result.value
+        if result.status == TaskStatus.SUCCESS:
+            backend.save(spec.storage_key(task_ctx), result.value)
+        report.results[name] = result
+        _emit(on_event, result)
+        return result
+
+    loop = asyncio.get_event_loop()
+    for name in all_names:
+        futures[name] = loop.create_task(_run_task(name))
+
+    await asyncio.gather(*futures.values())
+
+
 # ---------------------------------------------------------------------- #
 # 公共 API
 # ---------------------------------------------------------------------- #
-def _make_verbose_callback(
+def _make_verbose_callback(on_event: EventCallback | None) -> EventCallback:
-    on_event: EventCallback | None,
+    """包装 on_event 回调, 在 verbose 模式下打印任务生命周期。"""
-) -> EventCallback | None:
-    """包装 on_event 回调, 在 verbose 模式下打印任务生命周期.
-
-    Parameters
-    ----------
-    on_event : EventCallback | None
-        用户提供的原始回调, 若为 None 则仅打印.
-
-    Returns
-    -------
-    EventCallback | None
-        包装后的回调.
-    """
 
     def _verbose_callback(event: TaskEvent) -> None:
-        # 先打印生命周期信息
         dur = f" ({event.duration:.3f}s)" if event.duration is not None else ""
         if event.status == TaskStatus.RUNNING:  # pragma: no cover
             print(f"[verbose] 任务 {event.task!r} 开始执行...", flush=True)
@@ -464,13 +648,9 @@ def _make_verbose_callback(
                 f"[verbose] 任务 {event.task!r} 失败{dur} (尝试 {event.attempts} 次){err}",
                 flush=True,
             )
-        elif event.status == TaskStatus.SKIPPED:  # pragma: no branch
+        elif event.status == TaskStatus.SKIPPED:
             reason = f" ({event.reason})" if event.reason else ""
             print(f"[verbose] 任务 {event.task!r} 跳过{reason}", flush=True)
-        else:  # pragma: no cover
-            # 不可达: 执行器只发出 RUNNING/SUCCESS/FAILED/SKIPPED 事件
-            pass
-        # 再调用用户回调
         if on_event is not None:
             on_event(event)
 
@@ -486,6 +666,7 @@ def run(
     verbose: bool = False,
     on_event: EventCallback | None = None,
     state: StateBackend | None = None,
+    concurrency_limits: Mapping[str, int] | None = None,
 ) -> RunReport:
     """执行图并返回 :class:`RunReport`。
 
@@ -494,29 +675,28 @@ def run(
     graph:
         待执行的已校验 :class:`Graph`。
     strategy:
-        执行策略, 接受 :class:`Strategy` 枚举成员或字符串
+        执行策略: ``"sequential"`` / ``"thread"`` / ``"async"`` /
-        (``"sequential"`` / ``"thread"`` / ``"async"``). 默认 ``Strategy.SEQUENTIAL``.
+        ``"dependency"``。``"dependency"`` 为依赖驱动调度，无层屏障。
     max_workers:
         ``"thread"`` 的线程池大小。默认 ``min(32, len(layer))``。
     dry_run:
-        若为 ``True``，打印执行计划（层 + 注入）并返回空报告，不执行
+        若为 ``True``，打印执行计划并返回空报告，不执行任务。
-        任何任务。
     verbose:
-        若为 ``True``, 打印任务生命周期 (开始/成功/失败/跳过) 到 stdout.
+        若为 ``True``, 打印任务生命周期到 stdout。
-        注意: subprocess 命令的输出由 :class:`TaskSpec` 的 ``verbose`` 字段控制.
     on_event:
         可选回调，在每次状态转换时调用。
     state:
-        可选 :class:`StateBackend`，用于断点续跑。默认为内存后端
+        可选 :class:`StateBackend`，用于断点续跑。
-        （不跨进程持久化）。
+    concurrency_limits:
+        ``{concurrency_key: max_concurrent}`` 映射。具有相同
+        ``concurrency_key`` 的任务共享信号量，限制同时运行实例数。
 
     抛出
     ----
     ValueError
         ``strategy`` 不被识别时。
     TaskFailedError
-        任何任务耗尽重试后仍失败时。运行在失败层中止；后续层的任务
+        任何任务耗尽重试后仍失败时（除非 ``continue_on_error=True``）。
-        不会被执行。
     """
     graph.validate()
     layers = graph.layers()
@@ -525,20 +705,23 @@ def run(
         _print_dry_run(graph, layers)
         return RunReport(success=True)
 
-    # verbose 模式下包装事件回调
     effective_callback: EventCallback | None = _make_verbose_callback(on_event) if verbose else on_event
-
     backend = resolve_backend(state)
     report = RunReport()
     context: dict[str, Any] = {}
+    limits = concurrency_limits or {}
 
     try:
         if strategy == "sequential":
             _drive_sequential(graph, layers, context, report, backend, effective_callback)
         elif strategy == "thread":
-            _drive_threaded(graph, layers, context, report, backend, effective_callback, max_workers)
+            _drive_threaded(graph, layers, context, report, backend, effective_callback, max_workers, limits)
+        elif strategy == "async":
+            _drive_async(graph, layers, context, report, backend, effective_callback, limits)
+        elif strategy == "dependency":
+            asyncio.run(_drive_dependency_async(graph, context, report, backend, effective_callback, limits))
         else:
-            _drive_async(graph, layers, context, report, backend, effective_callback)
+            raise ValueError(f"Unknown strategy: {strategy!r}")
     except TaskFailedError:
         report.success = False
         raise
@@ -552,7 +735,7 @@ def _print_dry_run(graph: Graph, layers: list[list[str]]) -> None:
     for idx, layer in enumerate(layers, 1):
         print(f"  Layer {idx}: {layer}")
         for name in layer:
-            print(f"    - {describe_injection(graph.spec(name))}")
+            print(f"    - {describe_injection(graph.resolved_spec(name))}")
 
 
 def _drive_sequential(
@@ -575,10 +758,11 @@ def _drive_threaded(
     backend: StateBackend,
     on_event: EventCallback | None,
     max_workers: int | None,
+    concurrency_limits: Mapping[str, int],
 ) -> None:
     for idx, layer in enumerate(layers, 1):
         workers = max_workers or max(1, min(32, len(layer)))
-        _execute_layer_threaded(layer, graph, context, report, backend, idx, on_event, workers)
+        _execute_layer_threaded(layer, graph, context, report, backend, idx, on_event, workers, concurrency_limits)
 
 
 def _drive_async(
@@ -588,8 +772,9 @@ def _drive_async(
     report: RunReport,
     backend: StateBackend,
     on_event: EventCallback | None,
+    concurrency_limits: Mapping[str, int],
 ) -> None:
-    asyncio.run(_async_drive(graph, layers, context, report, backend, on_event))
+    asyncio.run(_async_drive(graph, layers, context, report, backend, on_event, concurrency_limits))
 
 
 async def _async_drive(
@@ -599,6 +784,7 @@ async def _async_drive(
     report: RunReport,
     backend: StateBackend,
     on_event: EventCallback | None,
+    concurrency_limits: Mapping[str, int],
 ) -> None:
     for idx, layer in enumerate(layers, 1):
-        await _execute_layer_async(layer, graph, context, report, backend, idx, on_event)
+        await _execute_layer_async(layer, graph, context, report, backend, idx, on_event, concurrency_limits)

src/pyflowx/graph.py

@@ -2,28 +2,53 @@
 
 使用标准库的 :mod:`graphlib`（3.9+）或 :mod:`graphlib_backport`（3.8）
 进行拓扑排序。图以增量方式构建并即时校验，使配置错误在构建时（而非执行时）快速失败。
+
+支持：
+* 图级默认值 :class:`GraphDefaults`，TaskSpec 字段为 ``None`` 时回退。
+* :meth:`Graph.map` 工厂批量生成 fan-out 任务。
+* 字符串引用与 :func:`compose` 编程式组合多个图。
+* 软依赖：仅用于上下文注入，不参与拓扑分层。
 """
 
 from __future__ import annotations
 
 import sys
 from dataclasses import dataclass, field, replace
-from typing import Any, Iterable, Mapping, Sequence
+from typing import Any, Callable, Iterable, Mapping, Sequence
 
 from .errors import CycleError, DuplicateTaskError, MissingDependencyError
-from .task import TaskSpec
+from .task import RetryPolicy, TaskSpec
 
-# graphlib 自 3.9 起进入标准库；3.8 回退到 backport。
 if sys.version_info >= (3, 9):  # pragma: no cover
     import graphlib  # pyright: ignore[reportUnreachable]
 
     _TopologicalSorter = graphlib.TopologicalSorter
 else:  # pragma: no cover
-    import graphlib  # type: ignore[import-untyped]  # pragma: no cover
+    import graphlib  # type: ignore[import-untyped]
 
     _TopologicalSorter = graphlib.TopologicalSorter  # pragma: no cover
 
 
+@dataclass
+class GraphDefaults:
+    """图级默认值。TaskSpec 对应字段为 ``None`` 时回退到此处。
+
+    仅对可空字段生效（retry/timeout/strategy/env/cwd/tags/priority/
+    continue_on_error/concurrency_key）。非空字段（name/fn/cmd）不回退。
+    """
+
+    retry: RetryPolicy | None = None
+    timeout: float | None = None
+    strategy: str | None = None
+    tags: tuple[str, ...] = ()
+    env: Mapping[str, str] | None = None
+    cwd: Any = None  # Path | None
+    priority: int = 0
+    continue_on_error: bool = False
+    concurrency_key: str | None = None
+    verbose: bool = False
+
+
 @dataclass
 class Graph:
     """校验后的有向无环任务图。
@@ -34,16 +59,11 @@ class Graph:
 
     图仅持有*配置*；运行时状态存于 :class:`~pyflowx.report.RunReport`。
     这使图可安全重复运行并在线程间共享。
-
-    Note
-    -----
-    Graph 不再使用 ``frozen=True``：内部 ``specs``/``deps`` 本就是可变 dict，
-    frozen 既无法真正保证不可变，又迫使 ``_pending_refs`` 等场景用
-    ``object.__setattr__`` 绕过。改为普通 dataclass，让赋值显式且可审计。
     """
 
     specs: dict[str, TaskSpec[Any]] = field(default_factory=dict)
     deps: dict[str, tuple[str, ...]] = field(default_factory=dict)
+    defaults: GraphDefaults = field(default_factory=GraphDefaults)
     # 待解析的字符串引用列表（由 GraphComposer 消费）；为空表示无引用。
     _pending_refs: list[str] = field(default_factory=list)
 
@@ -51,69 +71,47 @@ class Graph:
     # 构建
     # ------------------------------------------------------------------ #
     def add(self, spec: TaskSpec[Any]) -> Graph:
-        """注册一个任务 spec，并即时校验。
+        """注册一个任务 spec，并即时校验。返回 ``self`` 支持链式调用。"""
-
+        self._register(spec)
-        返回 ``self`` 以支持链式调用，但推荐入口是 :meth:`from_specs`，
-        它会整批校验（允许单次调用中的前向引用）。
-        """
-        if spec.name in self.specs:
-            raise DuplicateTaskError(spec.name)
-        self.specs[spec.name] = spec
-        self.deps[spec.name] = spec.depends_on
-        # 为增量 API 即时检查重名与缺失依赖。
         self._validate_references()
         return self
 
+    def _register(self, spec: TaskSpec[Any]) -> None:
+        if spec.name in self.specs:
+            raise DuplicateTaskError(spec.name)
+        self.specs[spec.name] = spec
+        # 拓扑依赖仅含硬依赖；软依赖仅用于注入，不影响分层。
+        self.deps[spec.name] = spec.depends_on
+
     @classmethod
-    def from_specs(cls, specs: Iterable[TaskSpec[Any] | str]) -> Graph:
+    def from_specs(
-        """从可迭代的 task spec 构建图.
+        cls,
+        specs: Iterable[TaskSpec[Any] | str],
+        defaults: GraphDefaults | None = None,
+    ) -> Graph:
+        """从可迭代的 task spec 构建图。
 
-        先收集所有 spec，再统一校验。这意味着任务可以引用*后出现*的
+        先收集所有 spec，再统一校验。允许前向引用。支持字符串引用，
-        依赖——顺序无关，就像声明式配置文件的读取方式。
+        由 :func:`compose` 或 :class:`GraphComposer` 解析展开。
-
-        支持字符串引用，允许引用其他命令图中的任务。
-        字符串引用将在CliRunner中解析展开。
 
         Parameters
         ----------
-        specs : Iterable[TaskSpec[Any] | str]
+        specs:
-            TaskSpec对象或字符串引用的列表
+            TaskSpec 对象或字符串引用的列表。
-
+        defaults:
-        Returns
+            图级默认值。``None`` 使用空 :class:`GraphDefaults`。
-        -------
-        Graph
-            构建完成的图
-
-        Note
-        -----
-        字符串引用格式：
-        - "command_name" - 引用整个命令图
-        - "command_name.task_name" - 引用特定任务
-
-        Examples
-        --------
-        >>> graph = Graph.from_specs([
-        ...     TaskSpec("build", cmd=["uv", "build"]),
-        ...     "test",  # 引用test命令图
-        ... ])
         """
-        graph = cls()
+        graph = cls(defaults=defaults or GraphDefaults())
         pending_refs: list[str] = []
 
         for spec in specs:
             if isinstance(spec, str):
-                # 字符串引用，稍后解析
                 pending_refs.append(spec)
             elif isinstance(spec, TaskSpec):
-                if spec.name in graph.specs:
+                graph._register(spec)
-                    raise DuplicateTaskError(spec.name)
-                graph.specs[spec.name] = spec
-                graph.deps[spec.name] = spec.depends_on
             else:
-                raise TypeError(f"from_specs只接受TaskSpec或str，收到: {type(spec)}")
+                raise TypeError(f"from_specs 只接受 TaskSpec 或 str，收到: {type(spec)}")
 
-        # 存储待解析的引用，稍后由 GraphComposer 解析展开。
-        # Graph 不再 frozen，可直接赋值；保留属性名以保持向后兼容。
         if pending_refs:
             graph._pending_refs = pending_refs
 
@@ -125,26 +123,22 @@ class Graph:
     # 校验
     # ------------------------------------------------------------------ #
     def _validate_references(self) -> None:
-        """确保每个依赖名都存在于图中。"""
+        """确保每个依赖名都存在于图中。硬依赖与软依赖都校验。"""
-        for name, deps in self.deps.items():
+        for name, spec in self.specs.items():
-            for dep in deps:
+            for dep in spec.depends_on:
+                if dep not in self.specs:
+                    raise MissingDependencyError(name, dep)
+            for dep in spec.soft_depends_on:
                 if dep not in self.specs:
                     raise MissingDependencyError(name, dep)
 
     def validate(self) -> None:
-        """执行完整 DAG 校验。
+        """执行完整 DAG 校验。存在环时抛出 :class:`CycleError`。"""
-
-        存在环时抛出 :class:`~pyflowx.errors.CycleError`。
-        依赖存在性由 :meth:`_validate_references` 检查。
-        """
         self._validate_references()
         sorter = _TopologicalSorter(self.deps)
         try:
-            # prepare() 在有环时抛出 CycleError；此处不需要
-            # static_order() 的结果，仅利用其校验副作用。
             sorter.prepare()
-        except graphlib.CycleError as exc:
+        except graphlib.CycleError as exc:  # type: ignore[name-defined]
-            # exc.args[1] 是构成环的节点列表。
             cycle: Sequence[str] = exc.args[1] if len(exc.args) > 1 else []
             raise CycleError(list(cycle)) from exc
 
@@ -160,10 +154,49 @@ class Graph:
         """返回 ``name`` 的 spec；不存在则 ``KeyError``。"""
         return self.specs[name]
 
+    def resolved_spec(self, name: str) -> TaskSpec[Any]:
+        """返回应用图级默认值后的 spec（不修改原图）。
+
+        对于 ``retry``/``timeout``/``strategy``/``env``/``cwd`` 等可空
+        字段，若 spec 字段为默认空值且图级默认值非空，则用
+        :func:`dataclasses.replace` 生成带默认值的副本。
+        """
+        spec = self.specs[name]
+        d = self.defaults
+        overrides: dict[str, Any] = {}
+        if spec.retry == RetryPolicy() and d.retry is not None:
+            overrides["retry"] = d.retry
+        if spec.timeout is None and d.timeout is not None:
+            overrides["timeout"] = d.timeout
+        if spec.strategy is None and d.strategy is not None:
+            overrides["strategy"] = d.strategy
+        if spec.env is None and d.env is not None:
+            overrides["env"] = d.env
+        if spec.cwd is None and d.cwd is not None:
+            overrides["cwd"] = d.cwd
+        if spec.priority == 0 and d.priority != 0:
+            overrides["priority"] = d.priority
+        if not spec.continue_on_error and d.continue_on_error:
+            overrides["continue_on_error"] = True
+        if spec.concurrency_key is None and d.concurrency_key is not None:
+            overrides["concurrency_key"] = d.concurrency_key
+        if not spec.verbose and d.verbose:
+            overrides["verbose"] = True
+        if not spec.tags and d.tags:
+            overrides["tags"] = d.tags
+        if not overrides:
+            return spec
+        return replace(spec, **overrides)
+
     def dependencies(self, name: str) -> tuple[str, ...]:
-        """``name`` 的直接前驱。"""
+        """``name`` 的直接硬依赖前驱。"""
         return self.deps[name]
 
+    def all_deps(self, name: str) -> tuple[str, ...]:
+        """``name`` 的硬依赖 + 软依赖。"""
+        spec = self.specs[name]
+        return tuple(spec.depends_on) + tuple(spec.soft_depends_on)
+
     def all_specs(self) -> Mapping[str, TaskSpec[Any]]:
         """name -> spec 的只读视图。"""
         return self.specs
@@ -171,18 +204,15 @@ class Graph:
     def layers(self) -> list[list[str]]:
         """将任务分组为可并行执行的层（Kahn 算法）。
 
-        同层任务无相互依赖，可并发执行。层按执行顺序返回。
+        同层任务无相互硬依赖，可并发执行。软依赖不参与分层。
-
+        层按执行顺序返回。图有环时抛出 :class:`CycleError`。
-        图有环时抛出 :class:`~pyflowx.errors.CycleError`。
         """
         self.validate()
         sorter = _TopologicalSorter(self.deps)
         result: list[list[str]] = []
-        # ``get_ready`` + ``done`` 每次给出一层，正好是并行执行所需的分组。
         sorter.prepare()
         while sorter.is_active():
             ready = list(sorter.get_ready())
-            # 排序以保证确定性、可复现的执行计划。
             ready.sort()
             result.append(ready)
             for node in ready:
@@ -193,12 +223,7 @@ class Graph:
     # 子图 / 标签过滤
     # ------------------------------------------------------------------ #
     def subgraph(self, tags: Iterable[str]) -> Graph:
-        """返回仅包含匹配任意标签的任务的新图。
+        """返回仅包含匹配任意标签的任务的新图。依赖边被修剪。"""
-
-        依赖会被修剪，仅保留被保留任务之间的边；指向被丢弃任务的边
-        会被移除（被保留的任务不再等待它们）。用于调试时运行大型
-        DAG 的切片。
-        """
         wanted: set[str] = set(tags)
         kept: list[TaskSpec[Any]] = []
         for spec in self.specs.values():
@@ -206,10 +231,11 @@ class Graph:
                 pruned_deps = tuple(
                     d for d in spec.depends_on if d in self.specs and (wanted & set(self.specs[d].tags))
                 )
-                # 使用 replace 保留所有字段（verbose/skip_if_missing/allow_upstream_skip 等），
+                pruned_soft = tuple(
-                # 避免手动逐字段重建时遗漏新增字段。
+                    d for d in spec.soft_depends_on if d in self.specs and (wanted & set(self.specs[d].tags))
-                kept.append(replace(spec, depends_on=pruned_deps))
+                )
-        return Graph.from_specs(kept)
+                kept.append(replace(spec, depends_on=pruned_deps, soft_depends_on=pruned_soft))
+        return Graph.from_specs(kept, defaults=self.defaults)
 
     def subgraph_by_names(self, names: Iterable[str]) -> Graph:
         """返回限定于 ``names`` 的新图（边已修剪）。"""
@@ -221,18 +247,71 @@ class Graph:
         for spec in self.specs.values():
             if spec.name in wanted:
                 pruned_deps = tuple(d for d in spec.depends_on if d in wanted)
-                kept.append(replace(spec, depends_on=pruned_deps))
+                pruned_soft = tuple(d for d in spec.soft_depends_on if d in wanted)
-        return Graph.from_specs(kept)
+                kept.append(replace(spec, depends_on=pruned_deps, soft_depends_on=pruned_soft))
+        return Graph.from_specs(kept, defaults=self.defaults)
+
+    # ------------------------------------------------------------------ #
+    # Fan-out / map-reduce
+    # ------------------------------------------------------------------ #
+    def map(
+        self,
+        name_fn: Callable[[int], str],
+        spec: TaskSpec[Any],
+        items: Sequence[Any],
+        arg_factory: Callable[[Any], tuple[Any, ...]] | None = None,
+        depends_on_per: Callable[[int], tuple[str, ...]] | None = None,
+    ) -> list[TaskSpec[Any]]:
+        """为 ``items`` 中每个元素生成一个 TaskSpec 并加入图。
+
+        用于 fan-out / map-reduce 模式。返回生成的 spec 列表，便于
+        后续 reduce 任务依赖。
+
+        Parameters
+        ----------
+        name_fn:
+            接受索引 ``i``，返回任务名。需保证唯一。
+        spec:
+            模板 spec。其 ``name`` 与 ``args`` 会被覆盖。
+        items:
+            待分发的数据序列。
+        arg_factory:
+            接受一个 item，返回位置参数元组，覆盖 spec.args。
+            ``None`` 则将单个 item 作为唯一位置参数。
+        depends_on_per:
+            接受索引 ``i``，返回该任务的额外硬依赖。``None`` 则继承 spec.depends_on。
+
+        Returns
+        -------
+        list[TaskSpec]
+            生成的 spec 列表（已加入图）。
+
+        Examples
+        --------
+        >>> fetch_tmpl = px.TaskSpec("", fn=fetch_user)
+        >>> specs = graph.map(lambda i: f"fetch_{i}", fetch_tmpl, [1, 2, 3])
+        >>> reduce_spec = px.TaskSpec("reduce", fn=reduce_fn, depends_on=tuple(s.name for s in specs))
+        """
+        generated: list[TaskSpec[Any]] = []
+        for i, item in enumerate(items):
+            name = name_fn(i)
+            args = arg_factory(item) if arg_factory is not None else (item,)
+            extra_deps = depends_on_per(i) if depends_on_per is not None else ()
+            new_spec = replace(
+                spec,
+                name=name,
+                args=tuple(args),
+                depends_on=tuple(spec.depends_on) + tuple(extra_deps),
+            )
+            self.add(new_spec)
+            generated.append(new_spec)
+        return generated
 
     # ------------------------------------------------------------------ #
     # 可视化
     # ------------------------------------------------------------------ #
     def to_mermaid(self, orientation: str = "TD") -> str:
-        """将 DAG 渲染为 Mermaid ``graph`` 定义字符串。
+        """将 DAG 渲染为 Mermaid ``graph`` 定义字符串。"""
-
-        无外部依赖；输出可粘贴到 Markdown、由 VS Code 的 Mermaid 预览
-        渲染，或保存为文件。
-        """
         valid = {"TD", "TB", "BT", "LR", "RL"}
         orientation = orientation.upper()
         if orientation not in valid:
@@ -243,6 +322,10 @@ class Graph:
         for name, deps in self.deps.items():
             for dep in deps:
                 lines.append(f"    {dep} --> {name}")
+        # 软依赖用虚线
+        for name, spec in self.specs.items():
+            for dep in spec.soft_depends_on:
+                lines.append(f"    {dep} -.-> {name}")
         return "\n".join(lines) + "\n"
 
     # ------------------------------------------------------------------ #
@@ -268,19 +351,12 @@ class Graph:
 class GraphComposer:
     """将带字符串引用的图展开为纯 :class:`TaskSpec` 图。
 
-    从 ``CliRunner`` 抽出，使 ``Graph``（数据）与引用解析（组合逻辑）
+    引用格式：
-    职责分离。引用按顺序展开，后续引用的任务依赖前面引用的最后一个任务；
-    原始 ``TaskSpec`` 之间也按出现顺序串行依赖。
-
-    引用格式
-    --------
     * ``"command_name"`` —— 引用整个命令图。
     * ``"command_name.task_name"`` —— 引用特定任务。
 
-    Parameters
+    引用按顺序展开，后续引用的任务依赖前面引用的最后一个任务；
-    ----------
+    原始 ``TaskSpec`` 之间也按出现顺序串行依赖。
-    graphs : dict[str, Graph]
-        命令名到图的映射，引用据此解析。
     """
 
     def __init__(self, graphs: dict[str, Graph]) -> None:
@@ -294,18 +370,7 @@ class GraphComposer:
         return resolved
 
     def expand_refs(self, graph: Graph, current_cmd: str) -> Graph:
-        """展开图中的字符串引用。
+        """展开图中的字符串引用。若无 ``_pending_refs``，原样返回。"""
-
-        若图无 ``_pending_refs``，原样返回。
-
-        Note
-        -----
-        引用按顺序展开，后续引用的任务依赖于前面引用的任务完成。
-        例如 ``["c", "tc", bump]`` 展开为：
-        - c 的所有任务（无依赖）
-        - tc 的所有任务（依赖于 c 的最后一个任务）
-        - bump 任务（依赖于 tc 的最后一个任务）
-        """
         pending_refs = graph._pending_refs
         if not pending_refs:
             return graph
@@ -313,23 +378,16 @@ class GraphComposer:
         all_specs: list[TaskSpec[Any]] = []
         previous_ref_last_task: str | None = None
 
-        # 先解析每个引用，并建立依赖链。
         for ref in pending_refs:
             expanded_specs = self.parse_ref(ref, current_cmd)
-
-            # 若有前一个引用，让当前引用的任务依赖其最后一个任务。
             if previous_ref_last_task and expanded_specs:
                 for i, task in enumerate(expanded_specs):
-                    # 只为没有依赖的任务（或第一个任务）添加依赖。
                     if i == 0 or not task.depends_on:
                         expanded_specs[i] = replace(task, depends_on=tuple({*task.depends_on, previous_ref_last_task}))
-
             if expanded_specs:
                 previous_ref_last_task = expanded_specs[-1].name
-
             all_specs.extend(expanded_specs)
 
-        # 然后添加原始 TaskSpec，按出现顺序串行依赖。
         original_specs = list(graph.all_specs().values())
         if original_specs:
             if previous_ref_last_task:
@@ -337,49 +395,53 @@ class GraphComposer:
                 all_specs.append(replace(first, depends_on=tuple({*first.depends_on, previous_ref_last_task})))
             else:
                 all_specs.append(original_specs[0])
-
             for i in range(1, len(original_specs)):
                 current_task = original_specs[i]
                 previous_task_name = original_specs[i - 1].name
                 all_specs.append(
-                    replace(
+                    replace(current_task, depends_on=tuple({*current_task.depends_on, previous_task_name}))
-                        current_task,
-                        depends_on=tuple({*current_task.depends_on, previous_task_name}),
-                    )
                 )
 
-        return Graph.from_specs(all_specs)
+        return Graph.from_specs(all_specs, defaults=graph.defaults)
 
     def parse_ref(self, ref: str, current_cmd: str) -> list[TaskSpec[Any]]:
-        """解析单个字符串引用，返回对应的 TaskSpec 列表。
+        """解析单个字符串引用，返回对应的 TaskSpec 列表。"""
-
-        Raises
-        ------
-        ValueError
-            引用无效、目标命令/任务不存在，或检测到循环引用。
-        """
-        # 避免循环引用。
         if ref == current_cmd:
             raise ValueError(f"循环引用: 命令 '{current_cmd}' 引用了自己")
 
         if "." in ref:
-            # 特定任务引用: "command_name.task_name"
             cmd_name, task_name = ref.split(".", 1)
             if cmd_name not in self.graphs:
                 raise ValueError(f"引用的命令 '{cmd_name}' 不存在")
-
             ref_graph = self.graphs[cmd_name]
             if task_name not in ref_graph.all_specs():
                 raise ValueError(f"任务 '{task_name}' 不存在于命令 '{cmd_name}' 中")
-
             return [ref_graph.all_specs()[task_name]]
         else:
-            # 整个命令图引用: "command_name"
             cmd_name = ref
             if cmd_name not in self.graphs:
                 raise ValueError(f"引用的命令 '{cmd_name}' 不存在")
-
             ref_graph = self.graphs[cmd_name]
-            # 递归展开（若引用的图自身也含引用）。
             ref_graph = self.expand_refs(ref_graph, cmd_name)
             return list(ref_graph.all_specs().values())
+
+
+def compose(
+    graphs: dict[str, Graph],
+) -> dict[str, Graph]:
+    """编程式解析多图的字符串引用，返回展开后的新图映射。
+
+    与 :class:`GraphComposer` 等价，但作为独立函数暴露，供不使用
+    :class:`~pyflowx.runner.CliRunner` 的编程式用户调用。
+
+    Examples
+    --------
+    >>> graphs = {
+    ...     "build": px.Graph.from_specs([px.TaskSpec("b", cmd=["echo", "b"])]),
+    ...     "all": px.Graph.from_specs(["build", px.TaskSpec("t", cmd=["echo", "t"])]),
+    ... }
+    >>> resolved = px.compose(graphs)
+    >>> "b" in resolved["all"].all_specs()
+    True
+    """
+    return GraphComposer(graphs).resolve_all()

src/pyflowx/storage.py

@@ -4,20 +4,18 @@
 执行器向后端查询某任务是否已有存储结果；若有则跳过该任务，并将其
 存储值注入下游任务。
 
-本模块刻意保持最小化：仅持久化*成功*结果（失败任务会重跑），存储
+存储键由 :meth:`TaskSpec.storage_key` 计算，默认为任务名；若任务配置
-形态为扁平的 ``{task_name: result}`` 映射。内置两个后端：
+了 ``cache_key``，则键为 ``"name:cache_key_value"``，使不同输入产生
+独立缓存条目。
 
-* :class:`MemoryBackend` —— 快速、进程内、无 I/O。默认。
+支持 TTL：``has`` 在条目过期时返回 ``False``。
-* :class:`JSONBackend` —— 持久化到 JSON 文件，支持跨进程续跑。
-
-两者均零依赖（``json`` 为标准库）。用户可子类化
-:class:`StateBackend` 接入 SQLite、Redis 等。
 """
 
 from __future__ import annotations
 
 import json
 import sys
+import time
 from abc import ABC, abstractmethod
 from pathlib import Path
 from typing import Any, Mapping
@@ -31,23 +29,26 @@ from .errors import StorageError
 
 
 class StateBackend(ABC):
-    """可续跑状态存储的抽象基类。"""
+    """可续跑状态存储的抽象基类。
+
+    所有方法以 ``key`` 为参数（通常为任务名或 ``name:cache_key``）。
+    """
 
     @abstractmethod
     def load(self) -> Mapping[str, Any]:
         """返回完整的存储映射（可能为空）。"""
 
     @abstractmethod
-    def save(self, name: str, value: Any) -> None:
+    def save(self, key: str, value: Any) -> None:
         """持久化单个任务的成功结果。"""
 
     @abstractmethod
-    def has(self, name: str) -> bool:
+    def has(self, key: str) -> bool:
-        """``name`` 是否已有存储结果。"""
+        """``key`` 是否已有未过期的存储结果。"""
 
     @abstractmethod
-    def get(self, name: str) -> Any:
+    def get(self, key: str) -> Any:
-        """返回 ``name`` 的存储结果（不存在则抛 ``KeyError``）。"""
+        """返回 ``key`` 的存储结果（不存在则抛 ``KeyError``）。"""
 
     @abstractmethod
     def clear(self) -> None:
@@ -55,43 +56,66 @@ class StateBackend(ABC):
 
 
 class MemoryBackend(StateBackend):
-    """进程内 dict 后端。进程退出即丢失。"""
+    """进程内 dict 后端。进程退出即丢失。
 
-    def __init__(self) -> None:
+    Parameters
-        self._store: dict[str, Any] = {}
+    ----------
+    ttl:
+        条目存活秒数。``None`` 表示永不过期。``has`` 在条目超过 ttl 后
+        返回 ``False``（但不主动删除，下次 ``save`` 覆盖）。
+    """
+
+    def __init__(self, ttl: float | None = None) -> None:
+        self._store: dict[str, tuple[Any, float]] = {}
+        self._ttl = ttl
 
     @override
     def load(self) -> Mapping[str, Any]:
-        return dict(self._store)
+        return {k: v for k, (v, _ts) in self._store.items() if not self._expired(k)}
 
     @override
-    def save(self, name: str, value: Any) -> None:
+    def save(self, key: str, value: Any) -> None:
-        self._store[name] = value
+        self._store[key] = (value, time.monotonic())
 
     @override
-    def has(self, name: str) -> bool:
+    def has(self, key: str) -> bool:
-        return name in self._store
+        return key in self._store and not self._expired(key)
 
     @override
-    def get(self, name: str) -> Any:
+    def get(self, key: str) -> Any:
-        return self._store[name]
+        if key not in self._store or self._expired(key):
+            raise KeyError(key)
+        return self._store[key][0]
 
     @override
     def clear(self) -> None:
         self._store.clear()
 
+    def _expired(self, key: str) -> bool:
+        if self._ttl is None or key not in self._store:
+            return False
+        _value, ts = self._store[key]
+        return (time.monotonic() - ts) > self._ttl
+
 
 class JSONBackend(StateBackend):
     """基于文件的 JSON 存储，用于跨进程续跑。
 
-    结果必须可 JSON 序列化。不可序列化的值会抛出
+    存储格式：``{key: {"value": v, "ts": epoch_seconds}}``。
-    :class:`~pyflowx.errors.StorageError`（运行本身不会中止；仅该条
+    ``ts`` 用于 TTL 判断。结果必须可 JSON 序列化。
-    结果的持久化失败）。
+
+    Parameters
+    ----------
+    path:
+        JSON 文件路径。
+    ttl:
+        条目存活秒数。``None`` 表示永不过期。
     """
 
-    def __init__(self, path: str) -> None:
+    def __init__(self, path: str, ttl: float | None = None) -> None:
         self._path: str = path
-        self._store: dict[str, Any] = {}
+        self._ttl = ttl
+        self._store: dict[str, dict[str, Any]] = {}
         self._load()
 
     def _load(self) -> None:
@@ -101,7 +125,14 @@ class JSONBackend(StateBackend):
             with open(self._path, encoding="utf-8") as fh:
                 data: Any = json.load(fh)
             if isinstance(data, dict):
-                self._store = data
+                # 兼容纯值格式与带元数据格式
+                self._store = {}
+                for k, v in data.items():
+                    if isinstance(v, dict) and "value" in v and "ts" in v:
+                        self._store[k] = v
+                    else:
+                        # 旧格式：纯值
+                        self._store[k] = {"value": v, "ts": time.time()}
         except (OSError, json.JSONDecodeError) as exc:
             raise StorageError(f"cannot read state file {self._path!r}", exc) from exc
 
@@ -110,32 +141,40 @@ class JSONBackend(StateBackend):
         try:
             with open(tmp, "w", encoding="utf-8") as fh:
                 json.dump(self._store, fh, ensure_ascii=False, indent=2)
-
             _ = Path(tmp).replace(Path(self._path))
         except (OSError, TypeError) as exc:
             raise StorageError(f"cannot write state file {self._path!r}", exc) from exc
 
-    @override
+    def _now(self) -> float:
-    def load(self) -> Mapping[str, Any]:
+        return time.time()
-        return dict(self._store)
+
+    def _expired(self, entry: dict[str, Any]) -> bool:
+        if self._ttl is None:
+            return False
+        return (self._now() - float(entry.get("ts", 0))) > self._ttl
 
     @override
-    def save(self, name: str, value: Any) -> None:
+    def load(self) -> Mapping[str, Any]:
-        # 在修改内存状态前先校验可序列化性。
+        return {k: v["value"] for k, v in self._store.items() if not self._expired(v)}
+
+    @override
+    def save(self, key: str, value: Any) -> None:
         try:
             _ = json.dumps(value)
         except (TypeError, ValueError) as exc:
-            raise StorageError(f"result of task {name!r} is not JSON-serialisable", exc) from exc
+            raise StorageError(f"result of key {key!r} is not JSON-serialisable", exc) from exc
-        self._store[name] = value
+        self._store[key] = {"value": value, "ts": self._now()}
         self._flush()
 
     @override
-    def has(self, name: str) -> bool:
+    def has(self, key: str) -> bool:
-        return name in self._store
+        return key in self._store and not self._expired(self._store[key])
 
     @override
-    def get(self, name: str) -> Any:
+    def get(self, key: str) -> Any:
-        return self._store[name]
+        if key not in self._store or self._expired(self._store[key]):
+            raise KeyError(key)
+        return self._store[key]["value"]
 
     @override
     def clear(self) -> None:

src/pyflowx/task.py

@@ -15,9 +15,11 @@
 * ``TaskStatus`` 是封闭枚举；执行器绝不发明临时字符串。
 """
 
+import os
 import shutil
 import subprocess
 import sys
+from contextlib import contextmanager
 from dataclasses import dataclass, field
 from datetime import datetime
 from enum import Enum
@@ -25,8 +27,10 @@ from pathlib import Path
 from typing import (
     Any,
     Callable,
+    ContextManager,
     Coroutine,
     Generic,
+    Iterator,
     List,
     Mapping,
     Optional,
@@ -59,8 +63,95 @@ TaskCmd = Union[
     Callable[..., Any],  # Python 函数
 ]
 
-# 条件判断函数类型
+# 执行策略：sequential/thread/async 为层屏障模型，dependency 为依赖驱动模型。
-Condition = Callable[[], bool]
+Strategy = Union[str, "StrategyKind"]
+StrategyKind = Any  # 占位，避免循环；executors 模块用 Literal 约束
+
+# 条件判断函数类型：接收依赖上下文（可能为空映射），返回是否应执行。
+Condition = Callable[[Context], bool]
+
+# 缓存键计算函数：基于依赖上下文计算稳定字符串键。
+CacheKeyFn = Callable[[Context], str]
+
+
+# ---------------------------------------------------------------------- #
+# 重试策略
+# ---------------------------------------------------------------------- #
+@dataclass(frozen=True)
+class RetryPolicy:
+    """任务失败重试策略。
+
+    参数
+    ----
+    max_attempts:
+        最大尝试次数（含首次）。``1`` 表示仅尝试一次，不重试。
+    delay:
+        两次尝试之间的初始等待秒数。
+    backoff:
+        退避倍率。第 n 次重试等待 ``delay * backoff ** (n-1)``。
+    jitter:
+        抖动上限秒数。每次等待加上 ``[0, jitter)`` 的随机量，避免惊群。
+    retry_on:
+        仅对这些异常类型重试。默认 ``(Exception,)`` 重试所有异常。
+        传入空元组等价于不重试。
+
+    Note
+    -----
+    替代旧版 ``retries: int``。``retries=2`` 等价于
+    ``RetryPolicy(max_attempts=3)``。
+    """
+
+    max_attempts: int = 1
+    delay: float = 0.0
+    backoff: float = 1.0
+    jitter: float = 0.0
+    retry_on: Tuple[type[BaseException], ...] = (Exception,)
+
+    def __post_init__(self) -> None:
+        if self.max_attempts < 1:
+            raise ValueError(f"RetryPolicy.max_attempts must be >= 1, got {self.max_attempts}.")
+        if self.delay < 0:
+            raise ValueError(f"RetryPolicy.delay must be >= 0, got {self.delay}.")
+        if self.backoff < 0:
+            raise ValueError(f"RetryPolicy.backoff must be >= 0, got {self.backoff}.")
+        if self.jitter < 0:
+            raise ValueError(f"RetryPolicy.jitter must be >= 0, got {self.jitter}.")
+
+    @property
+    def retries(self) -> int:
+        """重试次数（不含首次），等价于 ``max_attempts - 1``。"""
+        return self.max_attempts - 1
+
+    def should_retry(self, exc: BaseException) -> bool:
+        """异常是否属于可重试类型。"""
+        return isinstance(exc, self.retry_on)
+
+    def wait_seconds(self, attempt: int) -> float:
+        """第 ``attempt`` 次失败后应等待的秒数（attempt 从 1 开始）。"""
+        if attempt < 1:
+            return 0.0
+        import random
+
+        base = self.delay * (self.backoff ** max(0, attempt - 1))
+        jitter = random.uniform(0, self.jitter) if self.jitter > 0 else 0.0
+        return base + jitter
+
+
+# ---------------------------------------------------------------------- #
+# 任务钩子
+# ---------------------------------------------------------------------- #
+@dataclass(frozen=True)
+class TaskHooks:
+    """任务生命周期钩子。
+
+    所有钩子均为可选。``pre_run`` 在任务实际执行前调用；``post_run``
+    在成功后调用并接收返回值；``on_failure`` 在最终失败后调用并接收异常。
+    钩子异常不会影响任务状态，仅记录日志。
+    """
+
+    pre_run: Optional[Callable[["TaskSpec[Any]"], None]] = None
+    post_run: Optional[Callable[["TaskSpec[Any]", Any], None]] = None
+    on_failure: Optional[Callable[["TaskSpec[Any]", BaseException], None]] = None
 
 
 class TaskStatus(Enum):
@@ -90,181 +181,239 @@ class TaskSpec(Generic[T]):
         - ``list[str]``: 命令及参数列表，如 ``["ls", "-la"]``
         - ``str``: shell 命令字符串，如 ``"pip freeze > requirements.txt"``
         - ``Callable``: Python 函数，与 ``fn`` 参数等效
-        若提供此参数，会自动包装为执行函数，覆盖 ``fn`` 参数。
     depends_on:
-        必须先完成才能运行本任务的任务名列表。顺序无关；框架会做
+        硬依赖任务名。必须全部成功完成才会运行本任务。
-        拓扑排序。
+        上游被 SKIPPED 时，本任务也会被 SKIPPED（除非
+        ``allow_upstream_skip=True``）。
+    soft_depends_on:
+        软依赖任务名。会等待其完成，但其结果不影响本任务是否执行：
+        - 上游成功：注入其返回值
+        - 上游 SKIPPED 或失败：注入 :attr:`defaults` 中提供的默认值
+        适用于"可选输入"场景。
+    defaults:
+        软依赖的默认值映射 ``{dep_name: default_value}``。
+        软依赖未提供结果时使用。未在 defaults 中出现的软依赖默认为 ``None``。
     args:
-        静态位置参数，追加在注入参数*之后*。适用于参数化任务
+        静态位置参数，追加在注入参数*之后*。
-        （如 ``fetch_user(uid)``）。
     kwargs:
         静态关键字参数。若与注入名冲突则抛出
         :class:`~pyflowx.errors.InjectionError`。
-    retries:
+    retry:
-        失败后的重试次数。``0`` 表示仅尝试一次。
+        :class:`RetryPolicy` 重试策略。默认仅尝试一次。
     timeout:
         最大执行时长（秒）。``None`` 表示不限制。异步任务使用
-        :func:`asyncio.wait_for`；线程/异步执行器中的同步任务会
+        :func:`asyncio.wait_for`；同步任务通过线程 future 取消。
-        取消 worker future。
     tags:
-        自由标签，供 :meth:`Graph.subgraph` 做选择性执行与调试。
+        自由标签，供 :meth:`Graph.subgraph` 做选择性执行与调试，
+        也可用于并发限制分组。
     conditions:
-        条件判断函数列表，只有所有条件都返回 ``True`` 时才执行任务。
+        条件判断函数列表，接收依赖上下文，全部返回 ``True`` 时才执行任务。
-        若任一条件返回 ``False``，任务会被标记为 SKIPPED。
+        任一返回 ``False`` 则任务被标记为 SKIPPED。
-        用于平台判断、环境变量检查等场景。
     cwd:
-        命令执行的工作目录，仅在使用 ``cmd`` 参数时有效。
+        工作目录。对 ``cmd`` 任务作为子进程工作目录；对 ``fn`` 任务
-        ``None`` 表示当前目录。
+        通过临时切换当前目录生效。
+    env:
+        环境变量覆盖映射。对 ``cmd`` 任务合并到子进程环境；对 ``fn``
+        任务在执行期间临时设置。
     verbose:
-        是否在命令执行时显示详细输出。``True`` 时会打印执行的命令
+        是否打印详细输出。``True`` 时打印执行的命令、返回码与输出
-        及其标准输出/标准错误。仅在使用 ``cmd`` 参数时有效。
+        （仅 ``cmd``），以及任务生命周期。
-        ``False`` 时静默捕获输出（失败时仍会包含在错误信息中）。
     skip_if_missing:
-        仅对 ``cmd`` 为 ``list[str]`` 的任务有效。``True`` 时自动检查
+        仅对 ``cmd`` 为 ``list[str]`` 有效。``True`` 时通过
-        命令是否存在（通过 :func:`shutil.which`），不存在则跳过任务
+        :func:`shutil.which` 检查命令是否存在，不存在则跳过。
-        （标记为 SKIPPED）而非失败。适用于构建工具场景，避免因未安装
-        某些工具（如 maturin、tox）而导致整个图执行失败。
-        对于 ``str`` (shell) 和 ``Callable`` 类型的 ``cmd``，此参数无效。
     allow_upstream_skip:
-        若为 ``True``，当上游任务因条件不满足被跳过时，本任务仍会执行
+        若为 ``True``，硬依赖被 SKIPPED 时本任务仍执行（软依赖不影响）。
-        （而非跟随跳过）。适用于清理类任务：即使某些删除操作因目标不存在
+        适用于清理类任务。
-        而跳过，后续操作（如重启服务）仍应执行。默认为 ``False``。
+    strategy:
+        单任务执行策略覆盖。``None`` 表示继承图级策略。
+        ``"sequential"`` 同步直接调用；``"thread"``/``"async"`` 将同步
+        任务卸载到线程池，异步任务跑在事件循环上。
+    priority:
+        同层任务调度优先级。数值越大越先启动。仅影响同层内启动顺序，
+        不打破层屏障。默认 ``0``。
+    concurrency_key:
+        并发限制分组键。具有相同键的任务共享一个信号量，限制同时
+        运行的实例数。具体限额由 :func:`run` 的 ``concurrency_limits``
+        参数提供 ``{key: limit}`` 映射。``None`` 表示不限制。
+    continue_on_error:
+        若为 ``True``，任务最终失败时不中止整图，仅标记本任务 FAILED，
+        其硬依赖下游被 SKIPPED，其余任务继续。默认 ``False``。
+    cache_key:
+        缓存键计算函数。若提供，则用其基于依赖上下文计算的字符串键
+        存取状态后端，使不同输入产生独立缓存条目。``None`` 表示用任务名。
+    hooks:
+        :class:`TaskHooks` 生命周期钩子。
     """
 
     name: str
     fn: Optional[TaskFn[T]] = None
     cmd: Optional[TaskCmd] = None
     depends_on: Tuple[str, ...] = ()
+    soft_depends_on: Tuple[str, ...] = ()
+    defaults: Mapping[str, Any] = field(default_factory=dict)
     args: Tuple[Any, ...] = ()
     kwargs: Mapping[str, Any] = field(default_factory=dict)
-    retries: int = 0
+    retry: RetryPolicy = field(default_factory=RetryPolicy)
     timeout: Optional[float] = None
     tags: Tuple[str, ...] = ()
     conditions: Tuple[Condition, ...] = ()
     cwd: Optional[Path] = None
+    env: Optional[Mapping[str, str]] = None
     verbose: bool = False
     skip_if_missing: bool = False
     allow_upstream_skip: bool = False
+    strategy: Optional[str] = None
+    priority: int = 0
+    concurrency_key: Optional[str] = None
+    continue_on_error: bool = False
+    cache_key: Optional[CacheKeyFn] = None
+    hooks: TaskHooks = field(default_factory=TaskHooks)
 
     def __post_init__(self) -> None:
         if not self.name:
             raise ValueError("TaskSpec.name must be a non-empty string.")
-        if self.retries < 0:
+        if self.retry.max_attempts < 1:
-            raise ValueError(f"TaskSpec '{self.name}': retries must be >= 0.")
+            raise ValueError(f"TaskSpec '{self.name}': retry.max_attempts must be >= 1.")
         if self.timeout is not None and self.timeout <= 0:
             raise ValueError(f"TaskSpec '{self.name}': timeout must be > 0.")
-        if self.name in self.depends_on:
+        if self.name in self.depends_on or self.name in self.soft_depends_on:
             raise ValueError(f"TaskSpec '{self.name}' cannot depend on itself.")
+        overlap = set(self.depends_on) & set(self.soft_depends_on)
+        if overlap:
+            raise ValueError(f"TaskSpec '{self.name}': depends_on 与 soft_depends_on 不能重叠: {sorted(overlap)}")
         if self.fn is None and self.cmd is None:
             raise ValueError(f"TaskSpec '{self.name}': 必须提供 fn 或 cmd 参数。")
 
     @property
     def effective_fn(self) -> TaskFn[T]:
-        """获取有效的执行函数.
+        """获取有效的执行函数。
 
-        若提供了 ``cmd`` 参数，则返回包装后的命令执行函数；
+        若提供 ``cmd``，返回包装后的命令执行函数；否则返回 ``fn``。
-        否则返回 ``fn`` 参数。
+        包装函数在每次调用时从 ``self`` 读取 ``verbose``/``cwd``/``env``/
-
+        ``timeout``，避免闭包捕获运行期参数，使翻转字段无需重建 spec。
-        Note
-        -----
-        命令执行逻辑已抽到模块级 :func:`_run_command`，此处仅返回轻量
-        转发闭包。``verbose`` / ``cwd`` / ``timeout`` 不再在创建时闭包
-        捕获，而是在每次调用时从 ``self`` 读取——这使得翻转 ``verbose``
-        无需重建 spec（见 :func:`pyflowx.runner._apply_verbose_to_graph`）。
         """
         if self.cmd is not None:
             return self._wrap_cmd()
         if self.fn is not None:
             return self.fn
-
         raise ValueError(f"TaskSpec '{self.name}': 没有可执行的函数或命令。")  # pragma: no cover
 
     def _wrap_cmd(self) -> TaskFn[Any]:
-        """将 cmd 包装为可执行函数.
+        """将 cmd 包装为可执行函数。"""
-
-        返回的闭包仅持有 ``self`` 引用，每次调用时从 spec 读取
-        ``verbose``/``cwd``/``timeout``，避免闭包捕获运行期参数。
-
-        Returns
-        -------
-        TaskFn[Any]
-            包装后的执行函数.
-        """
         spec = self
 
-        if isinstance(spec.cmd, list):
+        def _run() -> T:
+            return cast(T, _run_command(spec))
 
-            def _run_list() -> T:
+        _run.__name__ = spec.name
-                return cast(T, _run_command(spec))
+        return _run  # type: ignore[return-value]
 
-            _run_list.__name__ = spec.name
+    def should_execute(self, context: Context) -> Tuple[bool, Optional[str]]:
-            return _run_list  # type: ignore[return-value]
+        """检查任务是否应执行。
-
-        if isinstance(spec.cmd, str):
-
-            def _run_shell() -> T:
-                return cast(T, _run_command(spec))
-
-            _run_shell.__name__ = spec.name
-            return _run_shell  # type: ignore[return-value]
-
-        if callable(spec.cmd):
-            return spec.cmd  # type: ignore[return-value]
-
-        raise TypeError(f"TaskSpec '{spec.name}': 不支持的 cmd 类型 {type(spec.cmd).__name__}")  # pragma: no cover
-
-    def should_execute(self) -> bool:
-        """检查任务是否应该执行.
 
         Returns
         -------
-        bool
+        (should_run, skip_reason)
-            若所有条件都返回 ``True``，且 ``skip_if_missing`` 检查通过，
+            ``should_run`` 为 False 时 ``skip_reason`` 描述跳过原因。
-            则返回 ``True``；否则返回 ``False``。
         """
-        if not all(condition() for condition in self.conditions):
+        # 逐个求值条件，记录失败项。
-            return False
+        failed_conditions: list[str] = []
+        for condition in self.conditions:
+            try:
+                ok = condition(context)
+            except Exception:
+                ok = False
+                name = getattr(condition, "__name__", None) or "匿名条件(执行错误)"
+                failed_conditions.append(name)
+                continue
+            if not ok:
+                failed_conditions.append(getattr(condition, "__name__", None) or "匿名条件")
 
-        return not (self.skip_if_missing and not self._is_cmd_available())
+        if failed_conditions:
+            return False, f"条件不满足: {', '.join(failed_conditions)}"
+
+        if self.skip_if_missing and not self._is_cmd_available():
+            cmd_name = self.cmd[0] if isinstance(self.cmd, list) and self.cmd else "unknown"
+            return False, f"命令不存在: {cmd_name}"
+
+        return True, None
 
     def _is_cmd_available(self) -> bool:
-        """检查 ``cmd`` 是否可用.
+        """检查 ``cmd`` 是否可用（仅 list[str]）。"""
-
-        仅对 ``list[str]`` 类型的 ``cmd`` 进行检查（通过 :func:`shutil.which`）。
-        对于 ``str`` (shell) 和 ``Callable`` 类型，始终返回 ``True``。
-
-        Returns
-        -------
-        bool
-            命令可用返回 ``True``，否则返回 ``False``。
-        """
-
         cmd = self.cmd
         if isinstance(cmd, list) and cmd:
-            first_arg = cmd[0]
+            return shutil.which(cmd[0]) is not None
-            return shutil.which(first_arg) is not None
         return True
 
+    def env_context(self) -> ContextManager[None]:
+        """返回临时应用 ``env`` 与 ``cwd`` 的上下文管理器。
 
-def _run_command(spec: "TaskSpec[Any]") -> Any:
+        对 ``fn`` 任务生效。``cmd`` 任务在 :func:`_run_command` 中直接
-    """执行 ``spec.cmd`` 指定的命令（list 或 shell 字符串）。
+        传给子进程。
+        """
+        return _env_and_cwd(self.env, self.cwd)
 
-    list 与 shell 两条路径的异常处理、输出捕获、返回码判断完全一致，
+    def storage_key(self, context: Context) -> str:
-    合并于此消除重复。``verbose``/``cwd``/``timeout`` 在调用时从
+        """计算状态后端存储键。"""
-    ``spec`` 读取，而非闭包捕获——这是 ``_wrap_cmd`` 不再捕获运行期
+        if self.cache_key is not None:
-    参数的关键。
+            try:
+                return f"{self.name}:{self.cache_key(context)}"
+            except Exception:
+                return self.name
+        return self.name
 
-    成功返回 ``None``；失败抛 ``RuntimeError``，错误信息包含命令、
+
-    返回码与（非 verbose 模式下的）stderr。
+@contextmanager
-    """
+def _env_and_cwd(
+    env: Optional[Mapping[str, str]],
+    cwd: Optional[Path],
+) -> Iterator[None]:
+    """临时设置环境变量与工作目录。"""
+    saved_env: dict[str, str] = {}
+    saved_cwd: Optional[str] = None
+    if env:
+        for k, v in env.items():
+            if k in os.environ:
+                saved_env[k] = os.environ[k]
+            os.environ[k] = v
+    if cwd is not None:
+        saved_cwd = str(Path.cwd())
+        os.chdir(cwd)
+    try:
+        yield
+    finally:
+        if saved_cwd is not None:
+            os.chdir(saved_cwd)
+        # 恢复环境变量
+        if env:
+            for k in env:
+                if k in saved_env:
+                    os.environ[k] = saved_env[k]
+                else:
+                    os.environ.pop(k, None)
+
+
+def _run_command(spec: "TaskSpec[Any]") -> Any:  # noqa: PLR0912
+    """执行 ``spec.cmd`` 指定的命令（list / shell 字符串 / 可调用对象）。"""
     cmd = spec.cmd
-    is_list = isinstance(cmd, list)
     verbose = spec.verbose
     cwd = spec.cwd
     timeout = spec.timeout
+    env_override = spec.env
 
-    # 统一展示用的命令字符串与标签。保持 "执行命令" / "执行 Shell" 连续，
+    # 可调用对象：直接调用，返回其结果。
-    # 以兼容既有输出格式与测试断言。
+    if callable(cmd) and not isinstance(cmd, (list, str)):
+        name = getattr(cmd, "__name__", "callable")
+        if verbose:
+            print(f"[verbose] 执行可调用命令: {name}", flush=True)
+            if cwd is not None:
+                print(f"[verbose] 工作目录: {cwd}", flush=True)
+        try:
+            return cmd()
+        except Exception as e:
+            raise RuntimeError(f"可调用命令执行异常: {name}: {e}") from e
+
+    is_list = isinstance(cmd, list)
     if is_list:
         cmd_str = " ".join(arg for arg in cmd)  # type: ignore[union-attr]
         verb = "执行命令"
@@ -279,14 +428,18 @@ def _run_command(spec: "TaskSpec[Any]") -> Any:
         if cwd is not None:
             print(f"[verbose] 工作目录: {cwd}", flush=True)
 
+    # 合并环境变量
+    run_env: Optional[dict[str, str]] = None
+    if env_override:
+        run_env = dict(os.environ)
+        run_env.update(env_override)
+
     try:
-        # cmd 此处必为 list[str] 或 str（_wrap_cmd 的 isinstance 守卫已排除
-        # None 与 Callable），但类型检查器无法跨函数推断，故 cast 收窄到
-        # subprocess.run 接受的 Union[str, Sequence[str]]。
         result = subprocess.run(
             cast(Union[str, List[str]], cmd),
             shell=not is_list,
             cwd=cwd,
+            env=run_env,
             timeout=timeout,
             capture_output=not verbose,
             text=True,
@@ -311,13 +464,42 @@ def _run_command(spec: "TaskSpec[Any]") -> Any:
     raise RuntimeError(err_msg)
 
 
+# ---------------------------------------------------------------------- #
+# 任务模板：批量生成相似 TaskSpec 的工厂
+# ---------------------------------------------------------------------- #
+def task_template(
+    fn: Optional[TaskFn[Any]] = None,
+    cmd: Optional[TaskCmd] = None,
+    **defaults: Any,
+) -> Callable[..., TaskSpec[Any]]:
+    """创建任务模板工厂。
+
+    返回的工厂接受 ``name`` 与任意覆盖字段，生成 :class:`TaskSpec`。
+    适用于批量创建相似任务（如 fan-out）。
+
+    Examples
+    --------
+    >>> Fetch = px.task_template(fn=fetch_user, retry=px.RetryPolicy(max_attempts=3))
+    >>> specs = [Fetch(f"fetch_{uid}", args=(uid,)) for uid in range(5)]
+    """
+    base = dict(defaults)
+    if fn is not None:
+        base["fn"] = fn
+    if cmd is not None:
+        base["cmd"] = cmd
+
+    def _factory(name: str, **overrides: Any) -> TaskSpec[Any]:
+        merged = dict(base)
+        merged.update(overrides)
+        return TaskSpec(name, **merged)
+
+    _factory.__name__ = "task_template_factory"
+    return _factory
+
+
 @dataclass
 class TaskResult(Generic[T]):
-    """运行期间产生的可变单任务记录。
+    """运行期间产生的可变单任务记录。"""
-
-    每次运行都会创建全新的 :class:`TaskResult`；spec 本身保持不可变。
-    这让同一个图可以安全地重复运行。
-    """
 
     spec: TaskSpec[T]
     status: TaskStatus = TaskStatus.PENDING
@@ -338,15 +520,11 @@ class TaskResult(Generic[T]):
 
 @dataclass(frozen=True)
 class TaskEvent:
-    """执行期间向观察者发出的不可变事件。
+    """执行期间向观察者发出的不可变事件。"""
-
-    传递给 :func:`pyflowx.run` 的 ``on_event`` 回调，让调用者无需耦合
-    执行器内部即可构建进度条、指标或结构化日志。
-    """
 
     task: str
     status: TaskStatus
     attempts: int = 0
     error: Optional[str] = None
     duration: Optional[float] = None
-    reason: Optional[str] = None  # 跳过原因，如 "条件不满足"、"上游任务被跳过"、"缓存"
+    reason: Optional[str] = None

tests/test_conditions.py

@@ -1,142 +1,218 @@
 """Tests for conditions module."""
 
+from __future__ import annotations
+
 import os
 import sys
 from unittest.mock import patch
 
 from pyflowx.conditions import (
+    IS_LINUX,
+    IS_MACOS,
+    IS_POSIX,
+    IS_WINDOWS,
     BuiltinConditions,
     Constants,
 )
 
+_CTX: dict[str, object] = {}
+
 
 def test_constants_is_windows():
-    """Test Constants.IS_WINDOWS is correct."""
     assert (sys.platform == "win32") == Constants.IS_WINDOWS
 
 
 def test_constants_is_linux():
-    """Test Constants.IS_LINUX is correct."""
     assert (sys.platform == "linux") == Constants.IS_LINUX
 
 
 def test_constants_is_macos():
-    """Test Constants.IS_MACOS is correct."""
     assert (sys.platform == "darwin") == Constants.IS_MACOS
 
 
 def test_constants_is_posix():
-    """Test Constants.IS_POSIX is correct."""
     assert (sys.platform != "win32") == Constants.IS_POSIX
 
 
+def test_module_level_static_conditions():
+    assert IS_WINDOWS(_CTX) == Constants.IS_WINDOWS
+    assert IS_LINUX(_CTX) == Constants.IS_LINUX
+    assert IS_MACOS(_CTX) == Constants.IS_MACOS
+    assert IS_POSIX(_CTX) == Constants.IS_POSIX
 
-def test_builtin_conditions_python_version_major_only():
+
-    """Test BuiltinConditions.PYTHON_VERSION with major only."""
+def test_python_version_major_only():
-    # Test with current Python version
     current_major = sys.version_info.major
-    assert BuiltinConditions.PYTHON_VERSION(current_major) is True
+    assert BuiltinConditions.PYTHON_VERSION(current_major)(_CTX) is True
-    assert BuiltinConditions.PYTHON_VERSION(current_major + 1) is False
+    assert BuiltinConditions.PYTHON_VERSION(current_major + 1)(_CTX) is False
 
 
-def test_builtin_conditions_python_version_with_minor():
+def test_python_version_with_minor():
-    """Test BuiltinConditions.PYTHON_VERSION with major and minor."""
     current_major = sys.version_info.major
     current_minor = sys.version_info.minor
-    assert BuiltinConditions.PYTHON_VERSION(current_major, current_minor) is True
+    assert BuiltinConditions.PYTHON_VERSION(current_major, current_minor)(_CTX) is True
-    assert BuiltinConditions.PYTHON_VERSION(current_major, current_minor + 1) is False
+    assert BuiltinConditions.PYTHON_VERSION(current_major, current_minor + 1)(_CTX) is False
 
 
-def test_builtin_conditions_python_version_at_least():
+def test_python_version_at_least():
-    """Test BuiltinConditions.PYTHON_VERSION_AT_LEAST."""
     current_major = sys.version_info.major
     current_minor = sys.version_info.minor
-    # Current version should be at least itself
+    assert BuiltinConditions.PYTHON_VERSION_AT_LEAST(current_major, current_minor)(_CTX) is True
-    assert BuiltinConditions.PYTHON_VERSION_AT_LEAST(current_major, current_minor) is True
+    assert BuiltinConditions.PYTHON_VERSION_AT_LEAST(current_major - 1, 0)(_CTX) is True
-    # Current version should be at least an older version
+    assert BuiltinConditions.PYTHON_VERSION_AT_LEAST(current_major + 1, 0)(_CTX) is False
-    assert BuiltinConditions.PYTHON_VERSION_AT_LEAST(current_major - 1, 0) is True
-    # Current version should NOT be at least a newer version
-    assert BuiltinConditions.PYTHON_VERSION_AT_LEAST(current_major + 1, 0) is False
 
 
-def test_builtin_conditions_HAS_INSTALLED_true():
+def test_has_installed_true():
-    """Test BuiltinConditions.HAS_INSTALLED when app exists."""
+    condition = BuiltinConditions.HAS_INSTALLED("python3")
-    # Python should always be available
+    assert condition(_CTX) is True
-    condition = BuiltinConditions.HAS_INSTALLED("python")
-    assert condition() is True
 
 
-def test_builtin_conditions_HAS_INSTALLED_false():
+def test_has_installed_false():
-    """Test BuiltinConditions.HAS_INSTALLED when app doesn't exist."""
     condition = BuiltinConditions.HAS_INSTALLED("nonexistent_app_12345")
-    assert condition() is False
+    assert condition(_CTX) is False
 
 
-def test_builtin_conditions_env_var_exists_true():
+def test_env_var_exists_true():
-    """Test BuiltinConditions.ENV_VAR_EXISTS when variable exists."""
     with patch.dict(os.environ, {"TEST_VAR": "value"}):
         condition = BuiltinConditions.ENV_VAR_EXISTS("TEST_VAR")
-        assert condition() is True
+        assert condition(_CTX) is True
 
 
-def test_builtin_conditions_env_var_exists_false():
+def test_env_var_exists_false():
-    """Test BuiltinConditions.ENV_VAR_EXISTS when variable doesn't exist."""
     condition = BuiltinConditions.ENV_VAR_EXISTS("NONEXISTENT_VAR_12345")
-    assert condition() is False
+    assert condition(_CTX) is False
 
 
-def test_builtin_conditions_env_var_equals_true():
+def test_env_var_equals_true():
-    """Test BuiltinConditions.ENV_VAR_EQUALS when value matches."""
     with patch.dict(os.environ, {"TEST_VAR": "expected_value"}):
         condition = BuiltinConditions.ENV_VAR_EQUALS("TEST_VAR", "expected_value")
-        assert condition() is True
+        assert condition(_CTX) is True
 
 
-def test_builtin_conditions_env_var_equals_false():
+def test_env_var_equals_false():
-    """Test BuiltinConditions.ENV_VAR_EQUALS when value doesn't match."""
     with patch.dict(os.environ, {"TEST_VAR": "different_value"}):
         condition = BuiltinConditions.ENV_VAR_EQUALS("TEST_VAR", "expected_value")
-        assert condition() is False
+        assert condition(_CTX) is False
 
 
-def test_builtin_conditions_not():
+def test_not():
-    """Test BuiltinConditions.NOT."""
+    true_cond = BuiltinConditions.HAS_INSTALLED("python3")
-    true_condition = lambda: True  # noqa: E731
+    false_cond = BuiltinConditions.HAS_INSTALLED("nonexistent_app_12345")
-    false_condition = lambda: False  # noqa: E731
 
-    not_true = BuiltinConditions.NOT(true_condition)
+    assert BuiltinConditions.NOT(true_cond)(_CTX) is False
-    assert not_true() is False
+    assert BuiltinConditions.NOT(false_cond)(_CTX) is True
-
-    not_false = BuiltinConditions.NOT(false_condition)
-    assert not_false() is True
 
 
-def test_builtin_conditions_and_all_true():
+def test_and_all_true():
-    """Test BuiltinConditions.AND when all conditions are true."""
+    cond = BuiltinConditions.AND(
-    true_condition = lambda: True  # noqa: E731
+        BuiltinConditions.HAS_INSTALLED("python3"),
-    condition = BuiltinConditions.AND(true_condition, true_condition, true_condition)
+        BuiltinConditions.HAS_INSTALLED("python3"),
-    assert condition() is True
+    )
+    assert cond(_CTX) is True
 
 
-def test_builtin_conditions_and_one_false():
+def test_and_one_false():
-    """Test BuiltinConditions.AND when one condition is false."""
+    cond = BuiltinConditions.AND(
-    true_condition = lambda: True  # noqa: E731
+        BuiltinConditions.HAS_INSTALLED("python3"),
-    false_condition = lambda: False  # noqa: E731
+        BuiltinConditions.HAS_INSTALLED("nonexistent_app"),
-    condition = BuiltinConditions.AND(true_condition, false_condition, true_condition)
+    )
-    assert condition() is False
+    assert cond(_CTX) is False
 
 
-def test_builtin_conditions_or_all_false():
+def test_or_all_false():
-    """Test BuiltinConditions.OR when all conditions are false."""
+    cond = BuiltinConditions.OR(
-    false_condition = lambda: False  # noqa: E731
+        BuiltinConditions.HAS_INSTALLED("nonexistent1"),
-    condition = BuiltinConditions.OR(false_condition, false_condition, false_condition)
+        BuiltinConditions.HAS_INSTALLED("nonexistent2"),
-    assert condition() is False
+    )
+    assert cond(_CTX) is False
 
 
-def test_builtin_conditions_or_one_true():
+def test_or_one_true():
-    """Test BuiltinConditions.OR when one condition is true."""
+    cond = BuiltinConditions.OR(
-    true_condition = lambda: True  # noqa: E731
+        BuiltinConditions.HAS_INSTALLED("nonexistent1"),
-    false_condition = lambda: False  # noqa: E731
+        BuiltinConditions.HAS_INSTALLED("python3"),
-    condition = BuiltinConditions.OR(false_condition, true_condition, false_condition)
+    )
-    assert condition() is True
+    assert cond(_CTX) is True
 
+
+# ---------------------------------------------------------------------- #
+# 上下文条件：基于上游依赖结果
+# ---------------------------------------------------------------------- #
+def test_dep_equals_true():
+    ctx = {"upstream": 42}
+    cond = BuiltinConditions.DEP_EQUALS("upstream", 42)
+    assert cond(ctx) is True
+
+
+def test_dep_equals_false():
+    ctx = {"upstream": 99}
+    cond = BuiltinConditions.DEP_EQUALS("upstream", 42)
+    assert cond(ctx) is False
+
+
+def test_dep_equals_missing_dep():
+    cond = BuiltinConditions.DEP_EQUALS("missing", 42)
+    assert cond({}) is False
+
+
+def test_dep_matches_true():
+    ctx = {"upstream": [1, 2, 3]}
+    cond = BuiltinConditions.DEP_MATCHES("upstream", lambda v: len(v) == 3)
+    assert cond(ctx) is True
+
+
+def test_dep_matches_false():
+    ctx = {"upstream": [1, 2]}
+    cond = BuiltinConditions.DEP_MATCHES("upstream", lambda v: len(v) == 3)
+    assert cond(ctx) is False
+
+
+def test_dep_matches_exception_returns_false():
+    ctx = {"upstream": ""}
+    cond = BuiltinConditions.DEP_MATCHES("upstream", lambda v: v[0])
+    assert cond(ctx) is False
+
+
+def test_dep_present_true():
+    ctx = {"upstream": "value"}
+    cond = BuiltinConditions.DEP_PRESENT("upstream")
+    assert cond(ctx) is True
+
+
+def test_dep_present_false_none():
+    # pyrefly: ignore [implicit-any-empty-container]
+    ctx = {"upstream": None}
+    cond = BuiltinConditions.DEP_PRESENT("upstream")
+    assert cond(ctx) is False
+
+
+def test_dep_present_false_missing():
+    cond = BuiltinConditions.DEP_PRESENT("missing")
+    assert cond({}) is False
+
+
+def test_dep_truthy_true():
+    ctx = {"upstream": [1]}
+    cond = BuiltinConditions.DEP_TRUTHY("upstream")
+    assert cond(ctx) is True
+
+
+def test_dep_truthy_false():
+    # pyrefly: ignore [implicit-any-empty-container]
+    ctx = {"upstream": []}
+    cond = BuiltinConditions.DEP_TRUTHY("upstream")
+    assert cond(ctx) is False
+
+
+def test_dep_truthy_missing():
+    cond = BuiltinConditions.DEP_TRUTHY("missing")
+    assert cond({}) is False
+
+
+def test_logical_combination_with_dep_conditions():
+    ctx = {"a": 1, "b": 0}
+    cond = BuiltinConditions.AND(
+        BuiltinConditions.DEP_EQUALS("a", 1),
+        BuiltinConditions.NOT(BuiltinConditions.DEP_TRUTHY("b")),
+    )
+    assert cond(ctx) is True

tests/test_context.py

@@ -141,7 +141,7 @@ class TestDescribeInjection:
 
         spec = px.TaskSpec("t", fn, depends_on=("a",))
         desc = describe_injection(spec)
-        assert "a=<result:a>" in desc
+        assert "a=<dep:a>" in desc
         assert "ctx=<Context>" in desc
         assert "flag=<default>" in desc
 

tests/test_executors.py

@@ -84,7 +84,9 @@ def test_retries_then_succeeds() -> None:
             raise RuntimeError("not yet")
         return "ok"
 
-    graph = px.Graph.from_specs([px.TaskSpec("flaky", flaky, retries=2)])
+    graph = px.Graph.from_specs([
+        px.TaskSpec("flaky", flaky, retry=px.RetryPolicy(max_attempts=3)),
+    ])
     report = px.run(graph, strategy="sequential")
     assert report.success
     assert report["flaky"] == "ok"
@@ -95,7 +97,9 @@ def test_retries_exhausted() -> None:
     def always_fail() -> None:
         raise RuntimeError("nope")
 
-    graph = px.Graph.from_specs([px.TaskSpec("f", always_fail, retries=2)])
+    graph = px.Graph.from_specs([
+        px.TaskSpec("f", always_fail, retry=px.RetryPolicy(max_attempts=3)),
+    ])
     with pytest.raises(TaskFailedError) as exc_info:
         _ = px.run(graph, strategy="sequential")
     assert exc_info.value.attempts == 3
@@ -332,7 +336,9 @@ def test_async_timeout_retry_then_succeed() -> None:
             await asyncio.sleep(10)  # 触发超时
         return "ok"
 
-    graph = px.Graph.from_specs([px.TaskSpec("a", flaky, retries=2, timeout=0.05)])
+    graph = px.Graph.from_specs([
+        px.TaskSpec("a", flaky, retry=px.RetryPolicy(max_attempts=3), timeout=0.05),
+    ])
     report = px.run(graph, strategy="async")
     assert report.success
     assert report["a"] == "ok"
@@ -349,7 +355,9 @@ def test_async_failure_retry_branch(caplog: pytest.LogCaptureFixture) -> None:
             raise RuntimeError("not yet")
         return "ok"
 
-    graph = px.Graph.from_specs([px.TaskSpec("a", flaky, retries=2)])
+    graph = px.Graph.from_specs([
+        px.TaskSpec("a", flaky, retry=px.RetryPolicy(max_attempts=3)),
+    ])
     with caplog.at_level("WARNING", logger="pyflowx"):
         report = px.run(graph, strategy="async")
     assert report.success
@@ -489,7 +497,7 @@ def test_run_empty_graph() -> None:
 # ---------------------------------------------------------------------- #
 def test_downstream_skipped_when_upstream_skipped_sequential() -> None:
     """上游任务被 SKIPPED 后，下游任务也应被 SKIPPED（sequential 策略）."""
-    never_true = lambda: False  # noqa: E731
+    never_true = lambda _ctx: False  # noqa: E731
 
     def downstream(upstream: str) -> str:
         return upstream + "_processed"
@@ -506,7 +514,7 @@ def test_downstream_skipped_when_upstream_skipped_sequential() -> None:
 
 def test_downstream_skipped_when_upstream_skipped_thread() -> None:
     """上游任务被 SKIPPED 后，下游任务也应被 SKIPPED（thread 策略）."""
-    never_true = lambda: False  # noqa: E731
+    never_true = lambda _ctx: False  # noqa: E731
 
     def downstream(upstream: str) -> str:
         return upstream + "_processed"
@@ -530,7 +538,7 @@ def test_downstream_skipped_when_upstream_skipped_async() -> None:
     async def downstream(upstream: str) -> str:
         return upstream + "_processed"
 
-    never_true = lambda: False  # noqa: E731
+    never_true = lambda _ctx: False  # noqa: E731
 
     graph = px.Graph.from_specs([
         px.TaskSpec("upstream", upstream, conditions=(never_true,)),
@@ -544,7 +552,7 @@ def test_downstream_skipped_when_upstream_skipped_async() -> None:
 
 def test_downstream_executes_when_upstream_succeeds() -> None:
     """上游任务成功时，下游任务应正常执行."""
-    always_true = lambda: True  # noqa: E731
+    always_true = lambda _ctx: True  # noqa: E731
 
     def upstream() -> str:
         return "hello"

tests/test_executors_edge_cases.py

@@ -85,7 +85,7 @@ def test_verbose_run_with_skipped_lifecycle(capsys: pytest.CaptureFixture[str]):
     spec = px.TaskSpec(
         "test",
         fn=lambda: "result",
-        conditions=(lambda: False,),
+        conditions=(lambda _ctx: False,),
     )
     graph = px.Graph.from_specs([spec])
     report = px.run(graph, strategy="sequential", verbose=True)
@@ -140,7 +140,7 @@ def test_verbose_event_callback_skipped():
     spec = px.TaskSpec(
         "test",
         fn=lambda: "result",
-        conditions=(lambda: False,),
+        conditions=(lambda _ctx: False,),
         verbose=True,
     )
     graph = px.Graph.from_specs([spec])
@@ -161,7 +161,11 @@ def test_execute_sync_with_retries():
             raise ValueError("temporary error")
         return "success"
 
-    spec = px.TaskSpec("retry_test", fn=failing_function, retries=3)
+    spec = px.TaskSpec(
+        "retry_test",
+        fn=failing_function,
+        retry=px.RetryPolicy(max_attempts=3),
+    )
     graph = px.Graph.from_specs([spec])
 
     # Should succeed after retries
@@ -182,7 +186,11 @@ def test_execute_async_with_retries():
             raise ValueError("temporary error")
         return "success"
 
-    spec = px.TaskSpec("retry_async_test", fn=failing_async_function, retries=3)
+    spec = px.TaskSpec(
+        "retry_async_test",
+        fn=failing_async_function,
+        retry=px.RetryPolicy(max_attempts=3),
+    )
     graph = px.Graph.from_specs([spec])
 
     # Should succeed after retries
@@ -196,7 +204,7 @@ def test_execute_sync_skip_on_condition():
     spec = px.TaskSpec(
         "skip_test",
         fn=lambda: "result",
-        conditions=(lambda: False,),
+        conditions=(lambda _ctx: False,),
     )
     graph = px.Graph.from_specs([spec])
 
@@ -210,7 +218,7 @@ def test_execute_async_skip_on_condition():
     spec = px.TaskSpec(
         "skip_async_test",
         fn=lambda: "result",
-        conditions=(lambda: False,),
+        conditions=(lambda _ctx: False,),
     )
     graph = px.Graph.from_specs([spec])
 

tests/test_graph.py

@@ -13,13 +13,11 @@ def _fn() -> None:
 
 
 def test_from_specs_builds_graph() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+        px.TaskSpec("b", _fn, depends_on=("a",)),
-            px.TaskSpec("b", _fn, depends_on=("a",)),
+        px.TaskSpec("c", _fn, depends_on=("a", "b")),
-            px.TaskSpec("c", _fn, depends_on=("a", "b")),
+    ])
-        ]
-    )
     assert set(graph.names) == {"a", "b", "c"}
     assert graph.dependencies("c") == ("a", "b")
     assert len(graph) == 3
@@ -28,23 +26,19 @@ def test_from_specs_builds_graph() -> None:
 
 def test_from_specs_allows_forward_references() -> None:
     # b depends on a, but a is declared after b — order should not matter.
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("b", _fn, depends_on=("a",)),
-            px.TaskSpec("b", _fn, depends_on=("a",)),
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+    ])
-        ]
-    )
     assert graph.layers() == [["a"], ["b"]]
 
 
 def test_duplicate_task_raises() -> None:
     with pytest.raises(DuplicateTaskError):
-        _ = px.Graph.from_specs(
+        _ = px.Graph.from_specs([
-            [
+            px.TaskSpec("a", _fn),
-                px.TaskSpec("a", _fn),
+            px.TaskSpec("a", _fn),
-                px.TaskSpec("a", _fn),
+        ])
-            ]
-        )
 
 
 def test_missing_dependency_raises() -> None:
@@ -57,24 +51,20 @@ def test_missing_dependency_raises() -> None:
 
 def test_cycle_detection() -> None:
     with pytest.raises(CycleError):
-        _ = px.Graph.from_specs(
+        _ = px.Graph.from_specs([
-            [
+            px.TaskSpec("a", _fn, depends_on=("c",)),
-                px.TaskSpec("a", _fn, depends_on=("c",)),
+            px.TaskSpec("b", _fn, depends_on=("a",)),
-                px.TaskSpec("b", _fn, depends_on=("a",)),
+            px.TaskSpec("c", _fn, depends_on=("b",)),
-                px.TaskSpec("c", _fn, depends_on=("b",)),
+        ])
-            ]
-        )
 
 
 def test_layers_grouping() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+        px.TaskSpec("b", _fn),
-            px.TaskSpec("b", _fn),
+        px.TaskSpec("c", _fn, depends_on=("a", "b")),
-            px.TaskSpec("c", _fn, depends_on=("a", "b")),
+        px.TaskSpec("d", _fn, depends_on=("c",)),
-            px.TaskSpec("d", _fn, depends_on=("c",)),
+    ])
-        ]
-    )
     layers = graph.layers()
     assert layers == [["a", "b"], ["c"], ["d"]]
 
@@ -85,12 +75,10 @@ def test_self_dependency_rejected() -> None:
 
 
 def test_to_mermaid() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+        px.TaskSpec("b", _fn, depends_on=("a",)),
-            px.TaskSpec("b", _fn, depends_on=("a",)),
+    ])
-        ]
-    )
     mermaid = graph.to_mermaid()
     assert mermaid.startswith("graph TD")
     assert 'a["a"]' in mermaid
@@ -104,13 +92,11 @@ def test_to_mermaid_invalid_orientation() -> None:
 
 
 def test_subgraph_by_tags() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn, tags=("ingest",)),
-            px.TaskSpec("a", _fn, tags=("ingest",)),
+        px.TaskSpec("b", _fn, depends_on=("a",), tags=("ingest",)),
-            px.TaskSpec("b", _fn, depends_on=("a",), tags=("ingest",)),
+        px.TaskSpec("c", _fn, depends_on=("b",), tags=("report",)),
-            px.TaskSpec("c", _fn, depends_on=("b",), tags=("report",)),
+    ])
-        ]
-    )
     sub = graph.subgraph(["ingest"])
     assert set(sub.names) == {"a", "b"}
     # Edge to dropped task c is removed; b no longer waits for anything
@@ -119,13 +105,11 @@ def test_subgraph_by_tags() -> None:
 
 
 def test_subgraph_by_names() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+        px.TaskSpec("b", _fn, depends_on=("a",)),
-            px.TaskSpec("b", _fn, depends_on=("a",)),
+        px.TaskSpec("c", _fn, depends_on=("b",)),
-            px.TaskSpec("c", _fn, depends_on=("b",)),
+    ])
-        ]
-    )
     sub = graph.subgraph_by_names(["a", "b"])
     assert set(sub.names) == {"a", "b"}
     # c is dropped, so b's dep on c (none here) — but a->b edge preserved.
@@ -139,12 +123,10 @@ def test_subgraph_by_names_unknown() -> None:
 
 
 def test_describe() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+        px.TaskSpec("b", _fn, depends_on=("a",)),
-            px.TaskSpec("b", _fn, depends_on=("a",)),
+    ])
-        ]
-    )
     desc = graph.describe()
     assert "Layer 1" in desc
     assert "Layer 2" in desc
@@ -187,12 +169,10 @@ def test_spec_accessor() -> None:
 
 
 def test_dependencies_accessor() -> None:
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec("a", _fn),
-            px.TaskSpec("a", _fn),
+        px.TaskSpec("b", _fn, depends_on=("a",)),
-            px.TaskSpec("b", _fn, depends_on=("a",)),
+    ])
-        ]
-    )
     assert graph.dependencies("a") == ()
     assert graph.dependencies("b") == ("a",)
 
@@ -210,16 +190,20 @@ def test_empty_graph_layers() -> None:
 
 
 def test_subgraph_preserves_metadata() -> None:
-    """子图应保留原任务的 retries/timeout/tags 等元数据。"""
+    """子图应保留原任务的 retry/timeout/tags 等元数据。"""
-    graph = px.Graph.from_specs(
+    graph = px.Graph.from_specs([
-        [
+        px.TaskSpec(
-            px.TaskSpec("a", _fn, tags=("x",), retries=3, timeout=5.0),
+            "a",
-            px.TaskSpec("b", _fn, depends_on=("a",), tags=("y",)),
+            _fn,
-        ]
+            tags=("x",),
-    )
+            retry=px.RetryPolicy(max_attempts=3),
+            timeout=5.0,
+        ),
+        px.TaskSpec("b", _fn, depends_on=("a",), tags=("y",)),
+    ])
     sub = graph.subgraph(["x"])
     spec = sub.spec("a")
-    assert spec.retries == 3
+    assert spec.retry.max_attempts == 3
     assert spec.timeout == 5.0
     assert spec.tags == ("x",)
 

tests/test_runner.py

@@ -29,24 +29,20 @@ def _echo_graph(name: str = "echo_task", msg: str = "hello") -> px.Graph:
 
 def _failing_graph() -> px.Graph:
     """构造一个必定失败的单任务图."""
-    return px.Graph.from_specs(
+    return px.Graph.from_specs([
-        [
+        px.TaskSpec(
-            px.TaskSpec(
+            "fail",
-                "fail",
+            cmd=["python", "-c", "import sys; sys.exit(1)"],
-                cmd=["python", "-c", "import sys; sys.exit(1)"],
+        )
-            )
+    ])
-        ]
-    )
 
 
 def _multi_task_graph() -> px.Graph:
     """构造一个带依赖的多任务图."""
-    return px.Graph.from_specs(
+    return px.Graph.from_specs([
-        [
+        px.TaskSpec("a", cmd=[*ECHO_CMD, "a"]),
-            px.TaskSpec("a", cmd=[*ECHO_CMD, "a"]),
+        px.TaskSpec("b", cmd=[*ECHO_CMD, "b"], depends_on=("a",)),
-            px.TaskSpec("b", cmd=[*ECHO_CMD, "b"], depends_on=("a",)),
+    ])
-        ]
-    )
 
 
 # ---------------------------------------------------------------------- #
@@ -240,12 +236,10 @@ class TestCliRunnerRunSuccess:
         def track_b() -> None:
             executed.append("b")
 
-        runner = px.CliRunner(
+        runner = px.CliRunner({
-            {
+            "a": px.Graph.from_specs([px.TaskSpec("a", track_a)]),
-                "a": px.Graph.from_specs([px.TaskSpec("a", track_a)]),
+            "b": px.Graph.from_specs([px.TaskSpec("b", track_b)]),
-                "b": px.Graph.from_specs([px.TaskSpec("b", track_b)]),
+        })
-            }
-        )
         _ = runner.run(["b"])
         assert executed == ["b"]
 
@@ -318,15 +312,13 @@ class TestCliRunnerVerbose:
 
     def test_verbose_prints_skip_lifecycle(self, capsys: pytest.CaptureFixture[str]) -> None:
         """verbose 模式下跳过的任务应打印跳过信息."""
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [
+            px.TaskSpec(
-                px.TaskSpec(
+                "skip_me",
-                    "skip_me",
+                cmd=[*ECHO_CMD, "skip"],
-                    cmd=[*ECHO_CMD, "skip"],
+                conditions=(lambda _ctx: False,),
-                    conditions=(lambda: False,),
+            ),
-                ),
+        ])
-            ]
-        )
         runner = px.CliRunner({"skip": graph})
         _ = runner.run(["skip"])
         captured = capsys.readouterr()
@@ -394,13 +386,11 @@ class TestCliRunnerList:
 
     def test_list_prints_all_commands(self, capsys: pytest.CaptureFixture[str]) -> None:
         """--list 应打印所有命令."""
-        runner = px.CliRunner(
+        runner = px.CliRunner({
-            {
+            "clean": _echo_graph("c", "clean"),
-                "clean": _echo_graph("c", "clean"),
+            "build": _echo_graph("b", "build"),
-                "build": _echo_graph("b", "build"),
+            "test": _echo_graph("t", "test"),
-                "test": _echo_graph("t", "test"),
+        })
-            }
-        )
         _ = runner.run(["--list"])
         captured = capsys.readouterr()
         assert "clean" in captured.out
@@ -523,30 +513,26 @@ class TestCliRunnerIntegration:
 
     def test_condition_skipped_command_succeeds(self) -> None:
         """条件不满足时任务跳过, 整体仍成功."""
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [
+            px.TaskSpec(
-                px.TaskSpec(
+                "skip_me",
-                    "skip_me",
+                cmd=[*ECHO_CMD, "should not run"],
-                    cmd=[*ECHO_CMD, "should not run"],
+                conditions=(lambda _ctx: False,),
-                    conditions=(lambda: False,),
+            ),
-                ),
+        ])
-            ]
-        )
         runner = px.CliRunner({"skip": graph})
         exit_code = runner.run(["skip"])
         assert exit_code == CliExitCode.SUCCESS.value
 
     def test_condition_met_command_succeeds(self) -> None:
         """条件满足时任务执行, 整体成功."""
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [
+            px.TaskSpec(
-                px.TaskSpec(
+                "run_me",
-                    "run_me",
+                cmd=[*ECHO_CMD, "should run"],
-                    cmd=[*ECHO_CMD, "should run"],
+                conditions=(lambda _ctx: True,),
-                    conditions=(lambda: True,),
+            ),
-                ),
+        ])
-            ]
-        )
         runner = px.CliRunner({"run": graph})
         exit_code = runner.run(["run"])
         assert exit_code == CliExitCode.SUCCESS.value
@@ -562,14 +548,12 @@ class TestCliRunnerIntegration:
 
             return fn
 
-        graph = px.Graph.from_specs(
+        graph = px.Graph.from_specs([
-            [
+            px.TaskSpec("a", make("a")),
-                px.TaskSpec("a", make("a")),
+            px.TaskSpec("b", make("b"), depends_on=("a",)),
-                px.TaskSpec("b", make("b"), depends_on=("a",)),
+            px.TaskSpec("c", make("c"), depends_on=("a",)),
-                px.TaskSpec("c", make("c"), depends_on=("a",)),
+            px.TaskSpec("d", make("d"), depends_on=("b", "c")),
-                px.TaskSpec("d", make("d"), depends_on=("b", "c")),
+        ])
-            ]
-        )
         runner = px.CliRunner({"diamond": graph})
         exit_code = runner.run(["diamond"])
         assert exit_code == CliExitCode.SUCCESS.value
@@ -577,12 +561,10 @@ class TestCliRunnerIntegration:
 
     def test_mixed_fn_and_cmd_commands(self) -> None:
         """混合 fn 和 cmd 的命令应都能执行."""
-        runner = px.CliRunner(
+        runner = px.CliRunner({
-            {
+            "fn_cmd": px.Graph.from_specs([px.TaskSpec("fn", fn=lambda: "fn-result")]),
-                "fn_cmd": px.Graph.from_specs([px.TaskSpec("fn", fn=lambda: "fn-result")]),
+            "cmd_cmd": px.Graph.from_specs([px.TaskSpec("cmd", cmd=[*ECHO_CMD, "cmd-result"])]),
-                "cmd_cmd": px.Graph.from_specs([px.TaskSpec("cmd", cmd=[*ECHO_CMD, "cmd-result"])]),
+        })
-            }
-        )
         assert runner.run(["fn_cmd"]) == CliExitCode.SUCCESS.value
         assert runner.run(["cmd_cmd"]) == CliExitCode.SUCCESS.value
 

tests/test_task.py

@@ -6,7 +6,7 @@ from datetime import datetime
 
 import pytest
 
-from pyflowx.task import TaskResult, TaskSpec, TaskStatus
+from pyflowx.task import RetryPolicy, TaskResult, TaskSpec, TaskStatus
 
 
 def _fn() -> None:
@@ -18,9 +18,9 @@ def test_spec_empty_name_rejected() -> None:
         TaskSpec("", _fn)
 
 
-def test_spec_negative_retries_rejected() -> None:
+def test_spec_negative_max_attempts_rejected() -> None:
-    with pytest.raises(ValueError, match="retries"):
+    with pytest.raises(ValueError, match="max_attempts"):
-        TaskSpec("a", _fn, retries=-1)
+        TaskSpec("a", _fn, retry=RetryPolicy(max_attempts=0))
 
 
 def test_spec_zero_timeout_rejected() -> None:

tests/test_task_edge_cases.py

@@ -67,7 +67,9 @@ def test_taskspec_wrap_cmd_verbose():
 
 def test_taskspec_wrap_cmd_error():
     """Test TaskSpec._wrap_cmd handles command error."""
-    spec = TaskSpec("test", cmd=["python", "-c", "import sys; sys.exit(1)"])
+    import sys
+
+    spec = TaskSpec("test", cmd=[sys.executable, "-c", "import sys; sys.exit(1)"])
     wrapped_fn = spec.effective_fn
 
     with pytest.raises(RuntimeError, match="命令执行失败"):
@@ -105,10 +107,10 @@ def test_taskspec_conditions_check():
     spec = px.TaskSpec(
         "test",
         fn=lambda: "result",
-        conditions=(lambda: True,),
+        conditions=(lambda _ctx: True,),
     )
 
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
 
 def test_taskspec_conditions_false():
@@ -116,10 +118,10 @@ def test_taskspec_conditions_false():
     spec = px.TaskSpec(
         "test",
         fn=lambda: "result",
-        conditions=(lambda: False,),
+        conditions=(lambda _ctx: False,),
     )
 
-    assert spec.should_execute() is False
+    assert spec.should_execute({})[0] is False
 
 
 def test_taskspec_conditions_multiple():
@@ -127,10 +129,10 @@ def test_taskspec_conditions_multiple():
     spec = px.TaskSpec(
         "test",
         fn=lambda: "result",
-        conditions=(lambda: True, lambda: True, lambda: True),
+        conditions=(lambda _ctx: True, lambda _ctx: True, lambda _ctx: True),
     )
 
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
 
 def test_taskspec_conditions_multiple_one_false():
@@ -138,10 +140,10 @@ def test_taskspec_conditions_multiple_one_false():
     spec = px.TaskSpec(
         "test",
         fn=lambda: "result",
-        conditions=(lambda: True, lambda: False, lambda: True),
+        conditions=(lambda _ctx: True, lambda _ctx: False, lambda _ctx: True),
     )
 
-    assert spec.should_execute() is False
+    assert spec.should_execute({})[0] is False
 
 
 def test_taskspec_list_cmd_timeout_mocked():
@@ -218,27 +220,28 @@ def test_taskspec_shell_cmd_os_error_mocked():
 # ---------------------------------------------------------------------- #
 def test_skip_if_missing_with_available_command():
     """skip_if_missing=True 时，命令存在应返回 True."""
-    # python 命令在测试环境中一定存在
+    import sys
-    spec = TaskSpec("test", cmd=["python", "--version"], skip_if_missing=True)
+
-    assert spec.should_execute() is True
+    spec = TaskSpec("test", cmd=[sys.executable, "--version"], skip_if_missing=True)
+    assert spec.should_execute({})[0] is True
 
 
 def test_skip_if_missing_with_missing_command():
     """skip_if_missing=True 时，命令不存在应返回 False."""
     spec = TaskSpec("test", cmd=["definitely_not_installed_app_xyz"], skip_if_missing=True)
-    assert spec.should_execute() is False
+    assert spec.should_execute({})[0] is False
 
 
 def test_skip_if_missing_false_with_missing_command():
     """skip_if_missing=False 时，命令不存在也应返回 True（不检查）."""
     spec = TaskSpec("test", cmd=["definitely_not_installed_app_xyz"], skip_if_missing=False)
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
 
 def test_skip_if_missing_with_shell_cmd_not_checked():
     """skip_if_missing=True 时，shell 命令（str）不检查，应返回 True."""
     spec = TaskSpec("test", cmd="definitely_not_installed_app_xyz", skip_if_missing=True)
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
 
 def test_skip_if_missing_with_callable_cmd_not_checked():
@@ -248,7 +251,7 @@ def test_skip_if_missing_with_callable_cmd_not_checked():
         return 0
 
     spec = TaskSpec("test", cmd=custom_cmd, skip_if_missing=True)
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
 
 def test_skip_if_missing_with_fn_not_checked():
@@ -258,7 +261,7 @@ def test_skip_if_missing_with_fn_not_checked():
         return 0
 
     spec = TaskSpec("test", fn=my_fn, skip_if_missing=True)
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
 
 def test_skip_if_missing_with_empty_cmd_list():
@@ -266,37 +269,39 @@ def test_skip_if_missing_with_empty_cmd_list():
     spec = TaskSpec("test", cmd=[""], skip_if_missing=True)
     # 空字符串命令，shutil.which 返回 None
     # 但 cmd[0] 是空字符串，shutil.which("") 返回 None
-    assert spec.should_execute() is False
+    assert spec.should_execute({})[0] is False
 
 
 def test_skip_if_missing_combined_with_conditions():
     """skip_if_missing=True 与 conditions 组合使用."""
+    import sys
+
     # conditions 返回 False，应跳过
     spec = TaskSpec(
         "test",
-        cmd=["python", "--version"],
+        cmd=[sys.executable, "--version"],
         skip_if_missing=True,
-        conditions=(lambda: False,),
+        conditions=(lambda _ctx: False,),
     )
-    assert spec.should_execute() is False
+    assert spec.should_execute({})[0] is False
 
     # conditions 返回 True，命令存在，应执行
     spec = TaskSpec(
         "test",
-        cmd=["python", "--version"],
+        cmd=[sys.executable, "--version"],
         skip_if_missing=True,
-        conditions=(lambda: True,),
+        conditions=(lambda _ctx: True,),
     )
-    assert spec.should_execute() is True
+    assert spec.should_execute({})[0] is True
 
     # conditions 返回 True，命令不存在，应跳过
     spec = TaskSpec(
         "test",
         cmd=["definitely_not_installed_app_xyz"],
         skip_if_missing=True,
-        conditions=(lambda: True,),
+        conditions=(lambda _ctx: True,),
     )
-    assert spec.should_execute() is False
+    assert spec.should_execute({})[0] is False
 
 
 def test_skip_if_missing_skips_task_in_run():

tests/test_taskspec_commands.py

@@ -52,7 +52,7 @@ def test_taskspec_with_conditions_skip():
     """测试条件不满足时任务被跳过."""
 
     # 创建一个永远不会满足的条件
-    def never_true():
+    def never_true(_ctx):
         return False
 
     graph = px.Graph.from_specs([
@@ -73,7 +73,7 @@ def test_taskspec_with_conditions_execute():
     """测试条件满足时任务正常执行."""
 
     # 创建一个总是满足的条件
-    def always_true():
+    def always_true(_ctx):
         return True
 
     graph = px.Graph.from_specs([
@@ -103,17 +103,17 @@ def test_platform_conditions():
         px.TaskSpec(
             "win_task",
             cmd=win_cmd,
-            conditions=(lambda: Constants.IS_WINDOWS,),
+            conditions=(lambda _ctx: Constants.IS_WINDOWS,),
         ),
         px.TaskSpec(
             "linux_task",
             cmd=posix_cmd,
-            conditions=(lambda: Constants.IS_LINUX,),
+            conditions=(lambda _ctx: Constants.IS_LINUX,),
         ),
         px.TaskSpec(
             "macos_task",
             cmd=posix_cmd,
-            conditions=(lambda: Constants.IS_MACOS,),
+            conditions=(lambda _ctx: Constants.IS_MACOS,),
         ),
     ])
 
@@ -137,17 +137,15 @@ def test_platform_conditions():
 
 def test_app_installed_conditions():
     """测试应用安装条件."""
-    # 测试 python 应该总是安装的
+    # 使用 sys.executable 保证可移植
-    if sys.platform == "win32":
+    python_cmd = [sys.executable, "--version"]
-        python_cmd = ["python", "--version"]
+    py_name = "python" if sys.platform == "win32" else "python3"
-    else:
-        python_cmd = ["python3", "--version"]
 
     graph = px.Graph.from_specs([
         px.TaskSpec(
             "python_check",
             cmd=python_cmd,
-            conditions=(BuiltinConditions.HAS_INSTALLED("python"),),
+            conditions=(BuiltinConditions.HAS_INSTALLED(py_name),),
         ),
     ])
 
@@ -162,18 +160,18 @@ def test_combined_conditions():
     """测试组合条件."""
     # AND 条件
     and_condition = BuiltinConditions.AND(
-        lambda: True,
+        lambda _ctx: True,
-        lambda: True,
+        lambda _ctx: True,
     )
 
     # OR 条件
     or_condition = BuiltinConditions.OR(
-        lambda: True,
+        lambda _ctx: True,
-        lambda: False,
+        lambda _ctx: False,
     )
 
     # NOT 条件
-    not_condition = BuiltinConditions.NOT(lambda: False)
+    not_condition = BuiltinConditions.NOT(lambda _ctx: False)
 
     graph = px.Graph.from_specs([
         px.TaskSpec(
@@ -228,7 +226,7 @@ def test_taskspec_with_timeout():
         # 短时间任务应该成功
         px.TaskSpec(
             "short_task",
-            cmd=["python", "-c", "import time; time.sleep(0.1)"],
+            cmd=[sys.executable, "-c", "import time; time.sleep(0.1)"],
             timeout=1.0,
         ),
     ])
@@ -245,13 +243,13 @@ def test_taskspec_dependency_with_conditions():
         px.TaskSpec(
             "first",
             cmd=[*ECHO_CMD, "first"],
-            conditions=(lambda: True,),
+            conditions=(lambda _ctx: True,),
         ),
         px.TaskSpec(
             "second",
             cmd=[*ECHO_CMD, "second"],
             depends_on=("first",),
-            conditions=(lambda: True,),
+            conditions=(lambda _ctx: True,),
         ),
         px.TaskSpec(
             "third",
@@ -378,7 +376,7 @@ class TestTaskSpecVerbose:
         graph = px.Graph.from_specs([
             px.TaskSpec(
                 "fail",
-                cmd=["python", "-c", "import sys; sys.exit(1)"],
+                cmd=[sys.executable, "-c", "import sys; sys.exit(1)"],
                 verbose=True,
             )
         ])
@@ -414,7 +412,7 @@ class TestTaskSpecCmdErrors:
             px.TaskSpec(
                 "fail",
                 cmd=[
-                    "python",
+                    sys.executable,
                     "-c",
                     "import sys; sys.stderr.write('error-msg'); sys.exit(1)",
                 ],
@@ -437,7 +435,9 @@ class TestTaskSpecCmdErrors:
         """shell 命令失败时应抛出 RuntimeError."""
         from pyflowx.errors import TaskFailedError
 
-        graph = px.Graph.from_specs([px.TaskSpec("fail", cmd='python -c "import sys; sys.exit(1)"')])
+        graph = px.Graph.from_specs([
+            px.TaskSpec("fail", cmd=f'{sys.executable} -c "import sys; sys.exit(1)"'),
+        ])
         with pytest.raises(TaskFailedError) as exc_info:
             _ = px.run(graph, strategy="sequential")
         assert "Shell 命令执行失败" in str(exc_info.value.cause)
@@ -450,7 +450,7 @@ class TestTaskSpecCmdErrors:
         graph = px.Graph.from_specs([
             px.TaskSpec(
                 "slow",
-                cmd=["python", "-c", "import time; time.sleep(5)"],
+                cmd=[sys.executable, "-c", "import time; time.sleep(5)"],
                 timeout=0.1,
             )
         ])
@@ -463,7 +463,13 @@ class TestTaskSpecCmdErrors:
         """shell 命令超时应抛出 RuntimeError."""
         from pyflowx.errors import TaskFailedError
 
-        graph = px.Graph.from_specs([px.TaskSpec("slow", cmd='python -c "import time; time.sleep(5)"', timeout=0.1)])
+        graph = px.Graph.from_specs([
+            px.TaskSpec(
+                "slow",
+                cmd=f'{sys.executable} -c "import time; time.sleep(5)"',
+                timeout=0.1,
+            ),
+        ])
         with pytest.raises(TaskFailedError) as exc_info:
             _ = px.run(graph, strategy="sequential")
         assert "超时" in str(exc_info.value.cause)