diff --git a/.agents/docs/2026-10-04-add-taskflow-spec.md b/.agents/docs/2026-10-04-add-taskflow-spec.md
new file mode 100644
index 00000000..b937559d
--- /dev/null
+++ b/.agents/docs/2026-10-04-add-taskflow-spec.md
@@ -0,0 +1,66 @@
+# Taskflow 4.1.0 模块接入
+
+## 范围与消费方式
+
+以最小必要改动接入上游 CPU 模块，包名为 `taskflow.taskflow@4.1.0`，入口保持 `import tf;`。不提供独立 compat 包、CUDA feature 或自行维护的 API wrapper。
+
+```toml
+[dependencies.taskflow]
+taskflow = "4.1.0"
+```
+
+消费者使用 `import std; import tf;`。模块只提供上游实际导出的 API，预处理宏不随模块导出，版本查询使用 `tf::version()`。模块与文本头文件混用不在本次验收范围内。
+
+## 上游与包结构
+
+- 上游：[taskflow/taskflow v4.1.0](https://github.com/taskflow/taskflow/releases/tag/v4.1.0)，MIT。
+- Tag commit：`45366fe5bc4f2f8ec9aa590b40c504e296886865`。
+- 归档：`https://github.com/taskflow/taskflow/archive/refs/tags/v4.1.0.tar.gz`。
+- SHA-256：`2107f90e315e48a676922010b036357ff2b0c6b9160ce17fa9396e5860b1d715`，独立下载摘要一致。
+- 归档根目录为 `taskflow-4.1.0/`，没有符号链接；Linux、macOS、Windows 使用相同归档及校验值。
+- 描述符为内联 Form B，采用 C++23，`import_std = false`，include 根为归档根目录。
+- 显式编译 `tf.core.cppm`、`tf.algorithm.cppm`、`tf.utility.cppm`、`tf.cppm`；不通配 modules 目录，因为 `tf_with_cuda.cppm` 同样提供主模块 `tf`。
+- 未配置 CN 镜像，使用纯字符串上游 URL；需要镜像时应上传相同归档字节。
+
+该形态参考 `khronos.vulkan-hpp` 的上游模块复用方式，安装期适配参考 `compat.muduo`，Linux 线程链接配置参考 `chriskohlhoff.asio`。
+
+## 必要适配
+
+安装钩子逐项检查目标文件可读且替换位置恰好出现一次，不匹配立即失败。下载归档保持上游原样，适配只发生在解包后的三个模块文件中。
+
+| 文件 | 适配 | 验证依据 |
+|---|---|---|
+| `modules/tf.core.cppm` | 移除 `HasGraph`、`ProfileData` 两条失效导出 | 4.1.0 头文件中不存在这两个名字，原模块在 Clang 下直接编译失败 |
+| `modules/tf.core.cppm`、`modules/tf.cppm` | 将总头文件包含及版本函数导出移至 core，主接口先导入 core 再导入 algorithm | GCC 16.1.0 消费原模块时找不到 `tf::Executor`、`tf::Taskflow`；宿主 GCC 16.2.1 独立复现，组合调整后通过 |
+| `modules/tf.utility.cppm` | 显式包含 `<algorithm>` | libc++ 下 `iterator.hpp` 使用的 `std::max` 缺少声明 |
+
+逐文件摘要比较确认其余上游文件未改变，包括所有头文件及调度实现。没有独立 fork，也未修改 mcpp 引擎或现有 CI 配置。
+
+Linux 最终链接使用 `-pthread`。不单独给依赖模块的编译命令增加该参数，否则 LLVM 消费者与 PCM 的线程配置不一致。实际构建命令确认链接参数到达消费者。
+
+## 测试与索引登记
+
+`tests/examples/taskflow-module` 通过单条 `taskflow` 索引重定向消费当前 checkout。一个测试 main 与一个辅助 TU 共同链接，覆盖：
+
+- `import std; import tf;` 的多 TU 消费。
+- 两个 worker 执行两条 DAG 分支并汇合，检查结果为 42。
+- 1024 个元素的 `reduce`，检查结果为 1024。
+- `next_pow2` 与 `tf::version()`。
+
+断言失败返回非零退出码，运行时间由 mcpp 测试超时约束。成员已登记到根 workspace，中英文描述符目录已有对应条目。
+
+## 验证结果
+
+使用 CI 固定的 mcpp `2026.10.1.2`：
+
+```sh
+mcpp test -p taskflow-module --cache off --timeout 30
+mcpp test -p taskflow-module --toolchain llvm@22.1.8 --cache off --timeout 30
+mcpp xpkg parse --all-os pkgs/t/taskflow.taskflow.lua
+```
+
+- Linux GCC 16.1.0、LLVM 22.1.8/libc++：各 `1 passed; 0 failed`，GCC 增量运行通过。
+- 冷安装、Lua 语法、三平台描述符解析通过。
+- 镜像 URL、包身份、保留 namespace、全仓跨包引用、平台版本一致性、重复版本检查通过。
+- workspace 与描述符变更均能由现有 CI 选择规则命中 `taskflow-module`。
+- macOS、Windows 尚待 PR CI 构建和运行验证；描述符解析不代表运行验收。
diff --git a/docs/descriptor-examples.md b/docs/descriptor-examples.md
index 9cbfe21b..f3ce8543 100644
--- a/docs/descriptor-examples.md
+++ b/docs/descriptor-examples.md
@@ -47,3 +47,4 @@ in the [root README](../README.md#reference-examples).
 | Module layer over a compat source build (external Form-A repo) | [`godotengine.godot-cpp-m`](../pkgs/g/godotengine.godot-cpp-m.lua) (two versions tracking upstream: `10.0.0-rc1` = Godot 4.6, `4.5.0` = Godot 4.5. `import godot_cpp;` re-exports the whole `godot` namespace, ~1800 names GENERATED from the headers rather than curated; the 1022-TU build stays in `compat.godot-cpp`, so the index carries only this descriptor. Macros — `GDCLASS`, `GDREGISTER_CLASS`, `memnew`, `ERR_*` — are the one thing a named module cannot export, so the package ships a side header to include next to the import. It also ships a generated `hashfuncs.hpp` shim — upstream's header minus `static` on two functions whose bodies declare an unnamed union — without which GCC refuses the module interface outright, a hard error no `-W` flag reaches) |
 | C++23 module wrapper | [`nlohmann.json`](../pkgs/n/nlohmann.json.lua) · [`marzer.tomlplusplus`](../pkgs/m/marzer.tomlplusplus.lua) · [`neargye.magic_enum`](../pkgs/n/neargye.magic_enum.lua) · [`boost-ext.ut`](../pkgs/b/boost-ext.ut.lua) (upstream's own `include/boost/ut.cppm` reproduced verbatim but for one `__argc`/`__argv` shim that Clang-on-MSVC needs; namespace `boost-ext` since it is NOT an official Boost library) |
 | C++23 module, upstream's own unit | [`khronos.vulkan-hpp`](../pkgs/k/khronos.vulkan-hpp.lua) (Vulkan-Hpp 1.4.357.0 — Khronos generates `vulkan.cppm` / `vulkan_video.cppm` into every Vulkan-Headers release, so `sources` names the two units and NOTHING is authored here; the payload is the same tarball, URL and sha256 as `compat.vulkan-headers`, which is what makes the module and the headers it includes impossible to skew. `import_std = true` is forced by the unit's own unconditional `export import std;`. No `include_dirs`: the headers arrive with the `compat.vulkan` dependency, which is also what satisfies the STATIC dispatcher's direct calls at link — depending on headers alone gives a package that compiles and then fails at every consumer's link. Module names stay upstream's `vulkan` / `vulkan_video`, never `khronos.vulkan`. The second unit `import`s the first and mcpp orders the pair from the scan, which `tests/examples/vulkan-hpp-module/tests/video.cpp` is the regression for) |
+| C++23 module, upstream's CPU partitions | [`taskflow.taskflow`](../pkgs/t/taskflow.taskflow.lua) (Taskflow 4.1.0, `import tf;`; reuses the four upstream CPU module units and omits the competing CUDA entry point. The checked install hook removes two nonexistent exports, moves the umbrella include/version export into core and imports core first for GCC, and supplies `<algorithm>` to utility for libc++. Only three module files are patched; headers and scheduler implementation stay upstream. No fork; consumers use `tf::version()` because macros are not exported) |
diff --git a/docs/zh/descriptor-examples.md b/docs/zh/descriptor-examples.md
index 5f976ef8..cf3d08c7 100644
--- a/docs/zh/descriptor-examples.md
+++ b/docs/zh/descriptor-examples.md
@@ -45,3 +45,4 @@
 | 模块层叠在 compat 源码构建之上(外部 Form-A 仓) | [`godotengine.godot-cpp-m`](../../pkgs/g/godotengine.godot-cpp-m.lua)(两个版本与上游对齐:`10.0.0-rc1` 对应 Godot 4.6,`4.5.0` 对应 Godot 4.5。`import godot_cpp;` 重导出整个 `godot` 命名空间,约 1800 个名字由头文件**生成**而非手工罗列;1022 个 TU 的构建留在 `compat.godot-cpp`,索引侧只留这一个描述符。宏 —— `GDCLASS`、`GDREGISTER_CLASS`、`memnew`、`ERR_*` —— 是具名模块唯一带不走的东西,故包内附一个与 import 并排包含的侧头文件。另外还带一份生成的 `hashfuncs.hpp` 遮蔽头 —— 上游那个头去掉两个函数的 `static`(它们体内声明了匿名 union)—— 否则 GCC 直接拒绝该模块接口,且是任何 `-W` 开关都够不到的硬错误) |
 | C++23 module wrapper | [`nlohmann.json`](../../pkgs/n/nlohmann.json.lua) · [`marzer.tomlplusplus`](../../pkgs/m/marzer.tomlplusplus.lua) · [`neargye.magic_enum`](../../pkgs/n/neargye.magic_enum.lua) · [`boost-ext.ut`](../../pkgs/b/boost-ext.ut.lua)(逐字复用上游自带的 `include/boost/ut.cppm`,仅加一处 Clang-on-MSVC 需要的 `__argc`/`__argv` shim;命名空间取 `boost-ext`,因其并非 boost 官方库) |
 | C++23 module,上游自带单元 | [`khronos.vulkan-hpp`](../../pkgs/k/khronos.vulkan-hpp.lua)(Vulkan-Hpp 1.4.357.0 —— Khronos 把 `vulkan.cppm` / `vulkan_video.cppm` 生成进每个 Vulkan-Headers release,所以 `sources` 点名这两个单元即可,本仓**不写一行**包装体;载荷与 `compat.vulkan-headers` 是同一份 tarball、同一个 URL 与 sha256,这让模块与它 include 的头不可能错配。`import_std = true` 由单元自身无条件的 `export import std;` 决定。不声明 `include_dirs`:头随 `compat.vulkan` 依赖到达,而该依赖同时满足静态 dispatcher 在链接期的直接调用 —— 只依赖头会得到一个「能编译、每个消费者都链接失败」的包。模块名保持上游的 `vulkan` / `vulkan_video`,绝不写成 `khronos.vulkan`。第二个单元 `import` 第一个,顺序由 mcpp 扫描决定,`tests/examples/vulkan-hpp-module/tests/video.cpp` 就是这条的回归)|
+| C++23 module，上游 CPU 分区 | [`taskflow.taskflow`](../../pkgs/t/taskflow.taskflow.lua)（Taskflow 4.1.0，`import tf;`；复用上游四个 CPU 模块单元，排除提供同名主模块的 CUDA 入口。安装钩子逐项校验匹配次数：移除两个不存在的导出，将总头文件及版本导出移至 core 并优先导入 core 以兼容 GCC，为 utility 补 `<algorithm>` 以兼容 libc++。仅适配三个模块文件，头文件和调度实现保持上游原样，无独立 fork；宏不随模块导出，版本查询使用 `tf::version()`） |
diff --git a/mcpp.toml b/mcpp.toml
index 9c63f694..785f7f0d 100644
--- a/mcpp.toml
+++ b/mcpp.toml
@@ -145,6 +145,7 @@ members = [
     "tests/examples/spdlog-compiled",
     "tests/examples/sqlite3",
     "tests/examples/sqlitecpp",
+    "tests/examples/taskflow-module",
     "tests/examples/tinyhttps",
     "tests/examples/usockets",
     "tests/examples/opencl",
diff --git a/pkgs/t/taskflow.taskflow.lua b/pkgs/t/taskflow.taskflow.lua
new file mode 100644
index 00000000..fe5fd17e
--- /dev/null
+++ b/pkgs/t/taskflow.taskflow.lua
@@ -0,0 +1,86 @@
+package = {
+    spec        = "1",
+    namespace   = "taskflow",
+    name        = "taskflow",
+    description = "Taskflow CPU task graphs through the upstream tf C++ module",
+    licenses    = { "MIT" },
+    repo        = "https://github.com/taskflow/taskflow",
+    type        = "package",
+
+    xpm = {
+        linux = {
+            ["4.1.0"] = {
+                url = "https://github.com/taskflow/taskflow/archive/refs/tags/v4.1.0.tar.gz",
+                sha256 = "2107f90e315e48a676922010b036357ff2b0c6b9160ce17fa9396e5860b1d715",
+            },
+        },
+        macosx = {
+            ["4.1.0"] = {
+                url = "https://github.com/taskflow/taskflow/archive/refs/tags/v4.1.0.tar.gz",
+                sha256 = "2107f90e315e48a676922010b036357ff2b0c6b9160ce17fa9396e5860b1d715",
+            },
+        },
+        windows = {
+            ["4.1.0"] = {
+                url = "https://github.com/taskflow/taskflow/archive/refs/tags/v4.1.0.tar.gz",
+                sha256 = "2107f90e315e48a676922010b036357ff2b0c6b9160ce17fa9396e5860b1d715",
+            },
+        },
+    },
+
+    mcpp = {
+        language     = "c++23",
+        import_std   = false,
+        modules      = { "tf" },
+        include_dirs = { "*" },
+        -- CUDA 入口同样提供 tf 主模块，与 CPU 入口不能同时编译
+        sources = {
+            "*/modules/tf.core.cppm",
+            "*/modules/tf.algorithm.cppm",
+            "*/modules/tf.utility.cppm",
+            "*/modules/tf.cppm",
+        },
+        targets = { ["taskflow"] = { kind = "lib" } },
+        deps = {},
+        linux = {
+            -- 模块编译选项须与消费者一致，线程依赖在最终链接时满足
+            ldflags = { "-pthread" },
+        },
+    },
+}
+
+import("xim.libxpkg.pkginfo")
+
+function install()
+    local wrap = "taskflow-" .. pkginfo.version()
+    local function patch(relative, before, after)
+        local source = path.join(wrap, relative)
+        local content = assert(io.readfile(source), "taskflow.taskflow: cannot read " .. source)
+        local pattern = before:gsub("(%W)", "%%%1")
+        local patched, count = content:gsub(pattern, function() return after end)
+        assert(count == 1, "taskflow.taskflow: expected exactly one match in " .. source)
+        io.writefile(source, patched)
+    end
+
+    -- 4.1.0 的模块导出表包含两个在该版本头文件中不存在的名字
+    for _, name in ipairs({ "HasGraph", "ProfileData" }) do
+        patch("modules/tf.core.cppm", "    using tf::" .. name .. ";\n", "")
+    end
+
+    -- 主接口的重复头声明会遮蔽 GCC 从分区导出的类型，版本声明随 core 导出
+    patch("modules/tf.core.cppm", "module;\n", "module;\n#include <taskflow/taskflow.hpp>\n")
+    patch("modules/tf.core.cppm", "export namespace tf {\n",
+        "export namespace tf {\n    using tf::version;\n")
+    patch("modules/tf.cppm", "#include <taskflow/taskflow.hpp>\n", "")
+    patch("modules/tf.cppm", "export import :algorithm;\nexport import :core;",
+        "export import :core;\nexport import :algorithm;")
+
+    -- iterator.hpp 使用 std::max，但 libc++ 的间接包含不提供它
+    patch("modules/tf.utility.cppm", "module;\n", "module;\n#include <algorithm>\n")
+
+    local prefix = pkginfo.install_dir()
+    os.tryrm(prefix)
+    os.mkdir(prefix)
+    os.mv(wrap, path.join(prefix, wrap))
+    return true
+end
diff --git a/tests/examples/taskflow-module/mcpp.toml b/tests/examples/taskflow-module/mcpp.toml
new file mode 100644
index 00000000..0817b41f
--- /dev/null
+++ b/tests/examples/taskflow-module/mcpp.toml
@@ -0,0 +1,10 @@
+# 本成员通过本地索引验证 taskflow.taskflow 的模块消费
+[package]
+name = "taskflow-module-tests"
+version = "0.1.0"
+
+[indices]
+taskflow = { path = "../../.." }
+
+[dependencies.taskflow]
+taskflow = "4.1.0"
diff --git a/tests/examples/taskflow-module/src/utility.cpp b/tests/examples/taskflow-module/src/utility.cpp
new file mode 100644
index 00000000..cbfd6020
--- /dev/null
+++ b/tests/examples/taskflow-module/src/utility.cpp
@@ -0,0 +1,7 @@
+import std;
+import tf;
+
+bool taskflow_utility_ok() {
+    return std::string_view(tf::version()) == "4.1.0"
+        && tf::next_pow2(std::uint64_t{17}) == 32;
+}
diff --git a/tests/examples/taskflow-module/tests/taskflow.cpp b/tests/examples/taskflow-module/tests/taskflow.cpp
new file mode 100644
index 00000000..22853231
--- /dev/null
+++ b/tests/examples/taskflow-module/tests/taskflow.cpp
@@ -0,0 +1,26 @@
+import std;
+import tf;
+
+bool taskflow_utility_ok();
+
+int main() {
+    tf::Executor executor(2);
+    tf::Taskflow flow;
+    int left = 0;
+    int right = 0;
+    int joined = 0;
+    auto a = flow.emplace([&] { left = 20; });
+    auto b = flow.emplace([&] { right = 22; });
+    auto c = flow.emplace([&] { joined = left + right; });
+    a.precede(c);
+    b.precede(c);
+    executor.run(flow).get();
+    if (joined != 42) return 1;
+
+    std::vector<int> values(1024, 1);
+    int sum = 0;
+    tf::Taskflow reduction;
+    reduction.reduce(values.begin(), values.end(), sum, std::plus<int>{});
+    executor.run(reduction).get();
+    return sum == 1024 && taskflow_utility_ok() ? 0 : 2;
+}