fix: resolve LLVM toolchain install failure on Linux

.agents/docs/llvm-install-failure-analysis.md

@@ -0,0 +1,129 @@
+# LLVM 工具链安装失败分析
+
+## 现象
+
+`mcpp toolchain install llvm` 依赖包（libxml2, zlib, glibc 等）安装成功，但 LLVM 本体（800MB）缺失：
+
+```
+~/.mcpp/registry/data/xpkgs/
+├── xim-x-libxml2/     ✓ 安装成功
+├── xim-x-zlib/         ✓ 安装成功
+├── xim-x-glibc/        ✓ 安装成功
+├── xim-x-llvm/         ✗ 不存在
+```
+
+## 根因分析
+
+### 问题 1：`</dev/null` 关闭 stdin 可能破坏 xlings 子进程通信
+
+`platform/process.cppm:79-84` 的 `seal_stdin()` 对所有 POSIX 命令追加 `</dev/null`。
+
+这个修复解决了 macOS 首次运行卡住的问题，但副作用是：xlings 内部的子进程（如解压 800MB LLVM 的 tar 进程）可能依赖 stdin 进行进程间通信或信号传递。小包（libxml2 等）不受影响，大包（LLVM）因为解压时间长，子进程链更复杂，可能被 broken stdin 导致静默失败。
+
+### 问题 2：`2>/dev/null` 吞掉所有错误信息
+
+`xlings.cppm:432-434` 构建的命令：
+
+```bash
+cd ~/.mcpp && ... xlings interface install_packages --args '...' 2>/dev/null </dev/null
+```
+
+stderr 被完全丢弃。如果 xlings 安装 LLVM 时输出了错误信息到 stderr，我们完全看不到。
+
+### 问题 3：NDJSON handler 只处理 download_progress 事件
+
+`xlings.cppm:645-692` 的 `handle_line` 回调：
+
+```cpp
+if (kind != "data") return;                    // 忽略非 data 事件
+if (ls.find_str("dataKind") != "download_progress") return;  // 只关心下载进度
+```
+
+如果 xlings 发出了 error 事件或 log 事件报告安装失败，全部被静默丢弃。
+
+### 问题 4：Windows 有 fallback 但 Linux 没有
+
+`package_fetcher.cppm:608-638` 有一个 Windows-only 的 workaround：
+
+```cpp
+#if defined(_WIN32)
+    // 如果 verdir 不存在，检查全局 xlings 目录 ~/.xlings/data/xpkgs/ 并复制过来
+    if (!std::filesystem::exists(verdir)) {
+        // ... copy from ~/.xlings/ to ~/.mcpp/
+    }
+#endif
+```
+
+这个 workaround 处理了 "xlings 把包装到全局目录而非 XLINGS_HOME 指定目录" 的情况。**Linux 没有这个 fallback**。
+
+### 为什么 CI 没有这个问题
+
+CI 设置了 `MCPP_VENDORED_XLINGS="$XLINGS_BIN"`：
+
+```yaml
+export MCPP_VENDORED_XLINGS="$XLINGS_BIN"
+"$MCPP" build --target x86_64-linux-musl
+```
+
+`MCPP_VENDORED_XLINGS` 触发 `make_xlings_env()` 中的特殊路径，使用全局 xlings 二进制。而且 CI 中的工具链安装走的是 xlings 全局 sandbox（因为 MCPP_HOME 显式设置），与用户本地的嵌套沙箱场景完全不同。
+
+实际上 **CI 也没有测试 `mcpp toolchain install llvm` 这个用户流程**——CI 只测试 `mcpp build`（使用预装的工具链）。
+
+## 修复方案
+
+### 修复 1：`install_with_progress()` Linux 路径改为直接命令（对齐 Windows）
+
+Windows 已经用直接 `xlings install ... -y` 命令而非 interface 模式。Linux 也应该如此：
+
+```cpp
+int install_with_progress(const Env& env, std::string_view target,
+                          const BootstrapProgressCallback& cb)
+{
+    // 所有平台统一：先用直接命令安装
+    auto directCmd = build_command_prefix(env) + std::format(" install {} -y", target);
+    int directRc = mcpp::platform::process::run_silent(directCmd);
+    if (directRc == 0) return 0;
+
+    // 直接命令失败则 fallback 到 interface 模式（保留进度回调能力）
+    // ...
+}
+```
+
+### 修复 2：Linux 增加与 Windows 相同的 fallback 检查
+
+在 `resolve_xpkg_path()` 中，将 Windows 的全局目录 fallback 扩展到所有平台：
+
+```cpp
+// 移除 #if defined(_WIN32)，改为所有平台通用
+if (!std::filesystem::exists(verdir)) {
+    // 检查全局 xlings 目录
+    auto homeDir = std::getenv("HOME");
+    if (homeDir) {
+        std::filesystem::path globalXpkgs =
+            std::filesystem::path(homeDir) / ".xlings" / "data" / "xpkgs";
+        auto globalVerdir = globalXpkgs / verdir.filename().parent_path().filename() / verdir.filename();
+        if (std::filesystem::exists(globalVerdir)) {
+            // 复制或软链接到 sandbox
+        }
+    }
+}
+```
+
+### 修复 3：不对 xlings install 命令关闭 stdin
+
+为 `install_with_progress()` 添加不关闭 stdin 的选项，或让直接 install 命令走 `std::system()` 而非 `platform::process`：
+
+```cpp
+// 直接命令不通过 platform::process（不追加 </dev/null）
+int directRc = std::system(directCmd.c_str());
+```
+
+### 修复 4：CI 增加工具链安装测试
+
+在 `ci.yml` 中增加专门测试 `mcpp toolchain install llvm` 的步骤，确保这个用户核心流程被覆盖。
+
+## 推荐实施顺序
+
+1. **修复 1 + 修复 3**：Linux 改用直接命令 + 不关闭 stdin（最可能解决问题）
+2. **修复 2**：增加全局目录 fallback（兜底）
+3. **修复 4**：增加 CI 测试（防止回归）

.github/workflows/ci.yml

@@ -134,6 +134,23 @@ jobs:
           "$MCPP" build
           "$MCPP" test
 
+      - name: Toolchain install smoke test (mcpp toolchain install llvm)
+        run: |
+          # Test the core user flow: install a toolchain, create a project,
+          # build with it. Uses the freshly-built mcpp (not bootstrap).
+          MCPP=$(realpath "$(find target -type f -name mcpp -printf '%T@ %p\n' | sort -rn | head -1 | cut -d' ' -f2)")
+          # Install LLVM toolchain into mcpp's sandbox
+          "$MCPP" toolchain install llvm 20.1.7
+          # Set as default so the build picks it up
+          "$MCPP" toolchain default llvm@20.1.7
+          # Build a hello-world project with the installed toolchain
+          TMP=$(mktemp -d)
+          cd "$TMP"
+          "$MCPP" new hello
+          cd hello
+          "$MCPP" build
+          "$MCPP" run
+
       - name: Fresh user experience (xlings install mcpp → new → run)
         continue-on-error: true
         run: |

src/build/flags.cppm

@@ -88,9 +88,24 @@ CompileFlags compute_flags(const BuildPlan& plan) {
         include_flags += " -I" + escape_path(abs);
     }
 
-    // Sysroot
+    // Sysroot + config override for macOS.
+    // On macOS, xlings LLVM's clang++.cfg contains hardcoded --sysroot and
+    // -isystem paths from the original install location. When the package is
+    // copied to mcpp's sandbox, these paths become stale. We pass
+    // --no-default-config to ignore the cfg and provide correct paths.
     std::string sysroot_flag;
-    if (!plan.toolchain.sysroot.empty()) {
+    bool is_macos_clang = mcpp::toolchain::is_clang(plan.toolchain)
+        && (plan.toolchain.targetTriple.find("apple") != std::string::npos
+         || plan.toolchain.targetTriple.find("darwin") != std::string::npos);
+    if (is_macos_clang) {
+        auto llvmRoot = plan.toolchain.binaryPath.parent_path().parent_path();
+        auto libcxxInclude = llvmRoot / "include" / "c++" / "v1";
+        sysroot_flag = " --no-default-config";
+        sysroot_flag += " -isystem" + escape_path(libcxxInclude);
+        if (auto sdk = mcpp::platform::macos::sdk_path())
+            sysroot_flag += " --sysroot=" + escape_path(*sdk);
+        f.sysroot = sysroot_flag;
+    } else if (!plan.toolchain.sysroot.empty()) {
         sysroot_flag = " --sysroot=" + escape_path(plan.toolchain.sysroot);
         f.sysroot = sysroot_flag;
     }

src/pm/package_fetcher.cppm

@@ -605,14 +605,17 @@ Fetcher::resolve_xpkg_path(std::string_view target,
     };
 
     auto resolve = [&]() -> std::expected<XpkgPayload, CallError> {
-#if defined(_WIN32)
-        // Workaround: xlings on Windows may extract large packages (e.g. LLVM)
-        // into its global data dir instead of the mcpp sandbox, because the
-        // extraction subprocess doesn't inherit XLINGS_HOME. Detect this and
-        // copy the payload into the sandbox so mcpp remains self-contained.
+        // Workaround: xlings may extract large packages (e.g. LLVM) into its
+        // global data dir instead of the mcpp sandbox, because the extraction
+        // subprocess doesn't always inherit XLINGS_HOME. Detect this and copy
+        // the payload into the sandbox so mcpp remains self-contained.
+        // Originally Windows-only; extended to all platforms for the same
+        // reason (xlings subprocess XLINGS_HOME propagation is unreliable).
         if (!std::filesystem::exists(verdir)) {
-            // Try xlings' own data dir (where `xlings self install` placed it)
-            auto xhome = std::getenv("USERPROFILE");
+            const char* xhome = nullptr;
+#if defined(_WIN32)
+            xhome = std::getenv("USERPROFILE");
+#endif
             if (!xhome) xhome = std::getenv("HOME");
             if (xhome) {
                 // xlings stores xpkgs at <home>/.xlings/data/xpkgs/ or
@@ -635,7 +638,6 @@ Fetcher::resolve_xpkg_path(std::string_view target,
                 }
             }
         }
-#endif
         if (!std::filesystem::exists(verdir)) {
             return std::unexpected(CallError{
                 std::format("xpkg payload missing: {}", verdir.string())});

src/toolchain/probe.cppm

@@ -262,7 +262,10 @@ probe_sysroot(const std::filesystem::path& compilerBin,
         auto s = trim_line(*r);
         if (!s.empty() && std::filesystem::exists(s)) return s;
     }
-    // macOS fallback: use xcrun to discover the SDK path
+    // macOS fallback: use xcrun to discover the SDK path.
+    // The sysroot is used for regular compilation flags (flags.cppm) but
+    // skipped for std module precompilation on macOS (stdmod.cppm) to
+    // avoid breaking SDK internal header dependencies.
     if (auto sdk = mcpp::platform::macos::sdk_path())
         return *sdk;
     return {};

src/toolchain/stdmod.cppm

@@ -92,8 +92,26 @@ std::expected<StdModule, StdModError> ensure_built(
                   : mcpp::toolchain::gcc::std_bmi_path(sm.cacheDir);
     sm.objectPath = sm.cacheDir / "std.o";
 
+    // Build sysroot + include flags for std module precompilation.
+    // On macOS, xlings LLVM's clang++.cfg contains hardcoded --sysroot and
+    // -isystem paths from the original install location. When the LLVM package
+    // is copied to mcpp's sandbox, these cfg paths become stale (still point
+    // to the original xlings directory). We override both:
+    //   --sysroot   → current active SDK (from xcrun)
+    //   --no-default-config  → ignore stale cfg entirely
+    //   -isystem    → correct libc++ headers in the sandbox copy
     std::string sysroot_flag;
-    if (!tc.sysroot.empty()) {
+    bool is_macos = tc.targetTriple.find("apple") != std::string::npos
+                 || tc.targetTriple.find("darwin") != std::string::npos;
+    if (is_macos && is_clang(tc)) {
+        // Ignore the stale clang++.cfg and provide correct flags directly.
+        auto llvmRoot = tc.binaryPath.parent_path().parent_path();
+        auto libcxxInclude = llvmRoot / "include" / "c++" / "v1";
+        sysroot_flag = " --no-default-config";
+        sysroot_flag += std::format(" -isystem'{}'", libcxxInclude.string());
+        if (auto sdk = mcpp::platform::macos::sdk_path())
+            sysroot_flag += std::format(" --sysroot='{}'", sdk->string());
+    } else if (!tc.sysroot.empty()) {
         sysroot_flag = std::format(" --sysroot='{}'", tc.sysroot.string());
     }
 

src/xlings.cppm

@@ -609,24 +609,29 @@ int install_with_progress(const Env& env, std::string_view target,
     auto argsJson = std::format(
         R"({{"targets":["{}"],"yes":true}})", target);
 
-    if constexpr (mcpp::platform::is_windows) {
-        mcpp::platform::env::set("XLINGS_HOME", env.home.string());
-        mcpp::platform::env::set("XLINGS_PROJECT_DIR", "");
-        std::error_code ec_mkdir;
-        std::filesystem::create_directories(env.home, ec_mkdir);
-        // Use direct `install` command instead of `interface install_packages`
-        // on Windows. The NDJSON interface may have issues with large packages
-        // where the extraction subprocess doesn't respect XLINGS_HOME.
-        auto directCmd = std::format("{} install {} -y",
-            env.binary.string(), target);
-        int directRc = mcpp::platform::process::run_silent(directCmd);
+    // All platforms: try direct `xlings install ... -y` first.
+    // The direct command is more reliable for large packages (e.g. LLVM
+    // ~800MB) because:
+    //   - it doesn't pipe through NDJSON interface (simpler subprocess chain)
+    //   - xlings manages its own stdin/stdout/stderr
+    //   - extraction subprocess coordination works normally
+    // The NDJSON interface path is kept as a fallback for progress reporting.
+    {
+        auto directCmd = build_command_prefix(env) +
+            std::format(" install {} -y {}", target, mcpp::platform::shell::silent_redirect);
+        // Use std::system() directly — do NOT redirect stdin via </dev/null
+        // because xlings may need stdin for subprocess coordination during
+        // large package extraction.
+        int directRc = mcpp::platform::process::extract_exit_code(
+            std::system(directCmd.c_str()));
         if (directRc == 0) return 0;
     }
+
+    // Fallback: NDJSON interface path (provides progress callbacks).
     auto cmd = [&]() -> std::string {
         if constexpr (mcpp::platform::is_windows) {
-            // Fallback to interface path if direct install fails
             return std::format("{} interface install_packages --args {} {}",
-                env.binary.string(),
+                build_command_prefix(env),
                 shq(argsJson),
                 mcpp::platform::null_redirect);
         } else {