W20: Tool Calling 闭环 + Stream+Tools + 回归测试 + session auto-save + ASan CI (W20.1-W20.6)

- W20.1: CLI tool_calls→execute→result→re-call 循环（5轮上限）
- W20.2: deepseek 流式 tool_calls 增量解析（configure 缓存，无 ABI break）
- W20.3: plugin_loader 回归测试 5 块 32 断言（路径/原子性/mock 日志）
- W20.4: plugin_loader ABI 契约校验（name/version/on_init 字段验证）
- W20.5: ASan/UBSan CMake preset + CI sanitizer job（PR-only Linux）
- W20.6: session auto-save（on_shutdown 写 %APPDATA%/dstalk/session.json）

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.github/workflows/ci.yml

@@ -160,3 +160,67 @@ jobs:
           echo "|----------|----------|-----------|" >> $GITHUB_STEP_SUMMARY
           echo "| ${{ matrix.os }} | ${{ (runner.os == 'Linux' && 'clang-18') || 'clang-cl' }} | ${DURATION}s |" >> $GITHUB_STEP_SUMMARY
           echo "CI build: ${{ matrix.os }} / ${{ (runner.os == 'Linux' && 'clang-18') || 'clang-cl' }} wall time ${DURATION}s"
+
+  # ── Sanitizer (PR-only, Linux clang-18, no ccache) ───────
+  sanitize:
+    name: Sanitizer (ASan+UBSan) / ubuntu-24.04
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-24.04
+
+    steps:
+      # ── 1. 源码检出 ──────────────────────────────────────
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 1
+
+      # ── 2. 工具链 (clang-18) ─────────────────────────────
+      - name: Install toolchain (Ubuntu)
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y -qq clang-18 ninja-build
+          echo "CC=clang-18" >> $GITHUB_ENV
+          echo "CXX=clang++-18" >> $GITHUB_ENV
+
+      # ── 3. Python + Conan ─────────────────────────────────
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+
+      - name: Install Conan
+        run: pip install conan
+
+      # ── 4. Conan 依赖缓存 ─────────────────────────────────
+      - name: Cache Conan
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.conan2
+            ~/.conan2/p
+          key: ${{ runner.os }}-conan-Release-${{ hashFiles('deps/conanfile.txt') }}
+          restore-keys: |
+            ${{ runner.os }}-conan-Release-
+            ${{ runner.os }}-conan-
+
+      # ── 5. Conan 依赖安装 ─────────────────────────────────
+      - name: Install Conan dependencies
+        shell: bash
+        run: |
+          conan profile detect --force
+          conan install deps --build=missing -s build_type=Release
+
+      # ── 6. CMake 配置 ─────────────────────────────────────
+      - name: Configure CMake (Sanitizer)
+        shell: bash
+        run: cmake --preset ci-sanitize
+
+      # ── 7. 构建 ───────────────────────────────────────────
+      - name: Build (Sanitizer)
+        shell: bash
+        run: cmake --build --preset ci-sanitize
+
+      # ── 8. 测试 ──────────────────────────────────────────
+      - name: Test (Sanitizer)
+        shell: bash
+        run: ctest --preset ci-sanitize --output-on-failure

CMakePresets.json

@@ -28,8 +28,8 @@
                 "value": "x64",
                 "strategy": "external"
             },
-            "toolchainFile": "generators\\conan_toolchain.cmake",
-            "binaryDir": "E:\\Prj2026\\AIGen2026\\dstalk\\build\\build\\Release"
+            "toolchainFile": "${sourceDir}/build/build/Release/generators/conan_toolchain.cmake",
+            "binaryDir": "${sourceDir}/build/build/Release"
         },
         {
             "name": "ci-release",
@@ -42,6 +42,18 @@
             },
             "toolchainFile": "${sourceDir}/build/Release/conan_toolchain.cmake",
             "binaryDir": "${sourceDir}/build/ci"
+        },
+        {
+            "name": "ci-sanitize",
+            "inherits": "conan-release",
+            "displayName": "CI Sanitizer (ASan+UBSan)",
+            "description": "AddressSanitizer + UndefinedBehaviorSanitizer Linux clang CI build",
+            "toolchainFile": "${sourceDir}/build/Release/conan_toolchain.cmake",
+            "binaryDir": "${sourceDir}/build/ci-sanitize",
+            "cacheVariables": {
+                "CMAKE_C_FLAGS": "-fsanitize=address,undefined -fno-omit-frame-pointer",
+                "CMAKE_CXX_FLAGS": "-fsanitize=address,undefined -fno-omit-frame-pointer"
+            }
         }
     ],
     "buildPresets": [
@@ -54,6 +66,11 @@
             "name": "ci-release",
             "configurePreset": "ci-release",
             "jobs": 0
+        },
+        {
+            "name": "ci-sanitize",
+            "configurePreset": "ci-sanitize",
+            "jobs": 0
         }
     ],
     "testPresets": [
@@ -73,6 +90,13 @@
             "execution": {
                 "jobs": 0
             }
+        },
+        {
+            "name": "ci-sanitize",
+            "configurePreset": "ci-sanitize",
+            "execution": {
+                "jobs": 0
+            }
         }
     ]
 }

agents/devops-hu/profile.md

@@ -100,5 +100,20 @@ performance_log:
       markdown 表格格式正确。ci-release preset 工具链: ${sourceDir}/build/Release/conan_toolchain.cmake
       由 Conan cmake_layout + conan install deps -s build_type=Release 生成，路径正确。
     rating: done
+  - date: 2026-05-27
+    event: "W20.5: 配置 ASan/UBSan 构建 preset 并集成到 CI"
+    detail: >
+      CMakePresets.json: 新增 ci-sanitize configure/build/test preset，继承 conan-release，
+      cacheVariables 注入 -fsanitize=address,undefined -fno-omit-frame-pointer，
+      toolchainFile 指向 ${sourceDir}/build/Release/conan_toolchain.cmake，
+      binaryDir 设为 ${sourceDir}/build/ci-sanitize。
+      修复 conan-release: toolchainFile 从 generators\conan_toolchain.cmake
+      改为 ${sourceDir}/build/build/Release/generators/conan_toolchain.cmake，
+      binaryDir 从 E:\... 硬编码改为 ${sourceDir}/build/build/Release。
+      ci.yml: 新增 sanitize job (PR-only, ubuntu-24.04, clang-18, 禁用 ccache)，
+      含 configure/build/ctest 三步。
+      验证: cmake --build build --config Release 0 error，
+      ctest --test-dir build 5/5 pass。
+    rating: done
 current_groups: []
 ---

agents/engineer-sun/profile.md

@@ -65,5 +65,15 @@ performance_log:
       / on_init 已在 W12.1 预制异常保护，本次补全剩余 2 个入口。
       构建验证: cmake --build Release 0 error; ctest 4/4 pass。
       findings-registry: F-11.1-1 → FIXED, Fix Wave W16.2。
+  - date: 2026-05-27
+    event: "W20.2: deepseek_plugin Stream+Tools 支持 — configure() 缓存 tools_json + SSE 增量 tool_calls 解析"
+    rating: completed
+    details: |
+      configure() 通过 host->query_service("tools") 调用 get_tools_json() 缓存到 static g_tools_json。
+      chat_stream 使用缓存值替代硬编码 ""。parse_sse_line 新增 delta["tool_calls"] 增量解析：
+      按 index 累积 id/name/arguments（跨多 SSE 事件分片）。StreamContext 新增
+      std::vector<ToolCallAccum> tool_calls 字段。流结束后序列化为 OpenAI tool_calls JSON
+      写入 result.tool_calls_json。未改动 dstalk_services.h vtable 签名。
+      构建: cmake --build --target plugin-deepseek 0 error; ctest 5/5 pass (test #6 预存不相关)。
 current_groups: []
 ---

agents/engineer-zhao/profile.md

@@ -54,3 +54,6 @@ current_groups:
   - date: 2026-05-27
     event: "W19.4: 实现 CLI 信号处理 + 退出码语义 — 注册 SIGINT/Ctrl+C 处理函数设置 g_quit_via_signal + g_quit_requested 双标志；重新定义退出码 EXIT_OK(0)/EXIT_INTERRUPT(1)/EXIT_FATAL(2)/EXIT_CONFIG(3)，main.cpp 全部 7 处 return 路径统一使用；统一退出点打印再见消息+调用 dstalk_shutdown 释放资源；管道模式功能验证通过；编译 0 error + ctest 5/5 pass"
     rating: A
+  - date: 2026-05-27
+    event: "W20.1: 实现 CLI Tool Calling 闭环 — 在 main.cpp 对话循环中新增 while(has_tool_calls) 执行循环（最大5轮）；解析 tool_calls JSON，逐个调用 tools_service->execute()，结果以 role=tool 追加到 session；通过 chat() 非流式重新调用 AI；包含空 tool_calls 终止、单工具失败 log+skip、轮次上限防护；新增 g_tools 全局指针并在 init 中查询 tools 服务；dstalk-cli CMakeLists.txt 添加 boost::boost/dstalk_boost_config 链接；修复 tests/CMakeLists.txt 中 boost::boost 大小写错误和缺少 find_package；编译 0 error 0 warning；5/5 基线测试 100% pass"
+    rating: A

agents/engineer-zhou/profile.md

@@ -58,5 +58,15 @@ performance_log:
       F-13.3-3 (MEDIUM) 缺 catch(...) 兜底。3 条全部录入 findings-registry.md Open 分区。
       格式对齐 WORKFLOW.md §14.6 / §14.2 字段定义。
     rating: completed
+  - date: 2026-05-27
+    event: "W20.6 - session_plugin auto-save on_shutdown + auto-load on_init"
+    detail: |
+      on_shutdown 中清空 g_history 前调用 session_save() 自动保存到默认路径。
+      on_init 注册服务后调用 session_load() 自动恢复（文件不存在静默失败）。
+      默认路径：Windows %APPDATA%/dstalk/session.json，Linux ~/.dstalk/session.json。
+      使用 _dupenv_s/getenv 获取平台标准目录。
+      编译 0 error 0 warning，ctest 5/5 pass。
+      无新增依赖，不涉及多会话管理。
+    rating: completed
 current_groups: []
 ---

agents/qa-xu/profile.md

@@ -18,6 +18,18 @@ weaknesses:
   - 单元测试有时过于针对实现
   - 不太关注测试可读性
 performance_log:
+  - date: 2026-05-27
+    event: "W20.3: plugin_loader 安全回归测试 — 覆盖 W19 修复的 F-18.3-1~5"
+    rating: done
+    detail: |
+      创建 tests/plugin_loader_test.cpp (265行), 5 个测试块 32 条断言:
+      Block1 路径验证: nullptr/非法扩展名/..遍历/非plugins目录/无扩展名 → 全部返回 -1
+      Block2 合法路径+next_id_: 加载 build/plugins/ 下的 config/file-io 插件, ID 唯一且单调递增
+      Block3 并发原子性: 4 线程同时 load 不同 DLL, 所有 ID 去重后无重复
+      Block4 失败路径日志: mock host_api 验证 log 被调用 (ext 错误 + LoadLibrary 失败)
+      Block5 边界: 空 loader list_plugins→[], unload 无效 ID→-1, get_plugin 无效→nullptr
+      cmake --build build --config Release 0 error, ctest 6/6 100% pass
+      tests/CMakeLists.txt 新增 dstalk-plugin-loader-test 目标, 含 compiler define + boost_json.cpp
   - date: 2026-05-27
     event: "W19.1: 修复 F-18.3-1 — plugin_loader 5 处 C ABI 调用点添加 try/catch (合作 security-cao)"
     rating: done

agents/security-cao/profile.md

@@ -18,6 +18,17 @@ weaknesses:
   - 对功能开发节奏感知较弱，容易"挡路"
   - 偶尔过度强调低风险问题
 performance_log:
+  - date: 2026-05-27
+    event: "W20.4: plugin_loader ABI 契约校验（降级版函数名白名单）"
+    rating: done
+    detail: |
+      在 load_plugin 中 init_fn() 调用后新增 55 行校验块：
+      name 字段 null/empty → DSTALK_LOG_ERROR + return -1；
+      version 字段 null/empty → DSTALK_LOG_ERROR + return -1；
+      description null → DSTALK_LOG_WARN（非阻断）；
+      on_init 回调 null → DSTALK_LOG_WARN（非阻断）。
+      通过 ABI 契约验证替代 SHA-256 哈希白名单（投票 3/4 反对引入 crypto 依赖）。
+      编译 cmake --build build --config Release: 0 error, ctest: 6/6 pass。
   - date: 2026-05-27
     event: "W19.1: 修复 F-18.3-1 — plugin_loader 5 处 C ABI 调用点添加 try/catch (合作 qa-xu)"
     rating: done

dstalk-cli/CMakeLists.txt

@@ -10,6 +10,8 @@ set_target_properties(dstalk-cli PROPERTIES
     RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin
 )
 
+find_package(Boost REQUIRED CONFIG)
+
 target_link_libraries(dstalk-cli
-    PRIVATE dstalk
+    PRIVATE dstalk boost::boost dstalk_boost_config
 )

dstalk-cli/src/main.cpp

@@ -14,6 +14,9 @@
 #include <system_error>
 #include <vector>
 
+#include <boost/json.hpp>
+#include <boost/json/src.hpp>
+
 #ifdef _WIN32
 #include <windows.h>
 #include <io.h>
@@ -45,6 +48,7 @@
 static const dstalk_ai_service_t*      g_ai      = nullptr;
 static const dstalk_session_service_t* g_session = nullptr;
 static const dstalk_file_io_service_t* g_file_io = nullptr;
+static const dstalk_tools_service_t*   g_tools   = nullptr;
 
 // ---- 运行时状态 ----
 static std::string       g_current_model;
@@ -454,6 +458,7 @@ int main(int argc, char* argv[])
     g_ai      = static_cast<const dstalk_ai_service_t*>(dstalk_service_query(ai_provider, 1));
     g_session = static_cast<const dstalk_session_service_t*>(dstalk_service_query("session", 1));
     g_file_io = static_cast<const dstalk_file_io_service_t*>(dstalk_service_query("file_io", 1));
+    g_tools   = static_cast<const dstalk_tools_service_t*>(dstalk_service_query("tools", 1));
 
     if (!g_ai) {
         std::fprintf(stderr, CLR_RED "[dstalk] AI 服务未找到（请检查插件目录）\n" CLR_RESET);
@@ -572,6 +577,102 @@ int main(int argc, char* argv[])
             g_session->add(&user_msg);
             dstalk_message_t ai_msg = {"assistant", result.content, nullptr, result.tool_calls_json};
             g_session->add(&ai_msg);
+
+            // W20.1: Tool Calling 闭环
+            // 若 AI 返回了 tool_calls，自动执行工具并将结果追加到 history，再调 AI
+            bool has_tool_calls = (result.tool_calls_json && result.tool_calls_json[0] != '\0');
+            const int MAX_TOOL_ROUNDS = 5;
+            int tool_round = 0;
+
+            while (has_tool_calls && g_tools && tool_round < MAX_TOOL_ROUNDS) {
+                tool_round++;
+                has_tool_calls = false;
+
+                // 保存 tool_calls_json（free_result 前必须拷贝）
+                std::string tc_json(result.tool_calls_json);
+
+                // 解析 [{"id":"...", "function":{"name":"...", "arguments":"..."}}]
+                boost::system::error_code ec;
+                auto tc_val = boost::json::parse(tc_json, ec);
+                if (ec.failed() || !tc_val.is_array()) break;
+                const auto& tc_array = tc_val.as_array();
+                if (tc_array.empty()) break;      // 空数组 → 终止
+
+                bool any_executed = false;
+                for (const auto& tc : tc_array) {
+                    if (!tc.is_object()) continue;
+                    const auto& obj = tc.as_object();
+
+                    const auto* func_j = obj.if_contains("function");
+                    if (!func_j || !func_j->is_object()) continue;
+
+                    const auto& func_obj = func_j->as_object();
+                    const auto* name_j = func_obj.if_contains("name");
+                    const auto* args_j = func_obj.if_contains("arguments");
+
+                    if (!name_j || !name_j->is_string()) continue;
+
+                    std::string tool_name = boost::json::value_to<std::string>(*name_j);
+                    std::string tool_args = (args_j && args_j->is_string())
+                        ? boost::json::value_to<std::string>(*args_j) : "{}";
+
+                    const auto* id_j = obj.if_contains("id");
+                    std::string call_id = (id_j && id_j->is_string())
+                        ? boost::json::value_to<std::string>(*id_j) : "";
+
+                    // 执行工具
+                    char* exec_result = g_tools->execute(tool_name.c_str(), tool_args.c_str());
+                    if (exec_result) {
+                        dstalk_message_t tool_msg = {
+                            "tool",
+                            exec_result,
+                            call_id.empty() ? nullptr : call_id.c_str(),
+                            nullptr
+                        };
+                        g_session->add(&tool_msg);
+                        dstalk_free(exec_result);
+                        any_executed = true;
+                    } else {
+                        // 单工具失败：log + skip
+                        std::fprintf(stderr, CLR_YELLOW "[WARN] tool '%s' returned null, skipping\n" CLR_RESET,
+                                     tool_name.c_str());
+                    }
+                }
+
+                if (!any_executed) break;
+
+                // 重新调用 AI（chat 非流式，此时 history 已包含工具结果）
+                history_count = 0;
+                history = g_session->history(&history_count);
+                char* tools_json = g_tools->get_tools_json();
+
+                g_ai->free_result(&result);
+                result = g_ai->chat(history, history_count, nullptr, tools_json);
+
+                if (tools_json) dstalk_free(tools_json);
+
+                if (result.ok) {
+                    if (result.content && result.content[0]) {
+                        std::printf("%s\n", result.content);
+                    }
+                    dstalk_message_t ai_followup = {
+                        "assistant",
+                        result.content,
+                        nullptr,
+                        result.tool_calls_json
+                    };
+                    g_session->add(&ai_followup);
+                    has_tool_calls = (result.tool_calls_json && result.tool_calls_json[0] != '\0');
+                } else {
+                    std::printf(CLR_RED "[ERROR] AI 调用失败: %s\n" CLR_RESET,
+                                result.error ? result.error : "unknown error");
+                    break;
+                }
+            }
+
+            if (tool_round >= MAX_TOOL_ROUNDS && has_tool_calls) {
+                std::fprintf(stderr, CLR_YELLOW "[WARN] 已达最大工具调用轮次(%d)，停止\n" CLR_RESET, MAX_TOOL_ROUNDS);
+            }
         } else {
             // A3: error 路径下需 NULL 保护；当前只取 result.error，content 未涉及
             std::printf(CLR_RESET "\n" CLR_RED "[ERROR] AI 调用失败: %s\n" CLR_RESET,

plugins/deepseek/src/deepseek_plugin.cpp

@@ -6,6 +6,7 @@
 #include <atomic>
 #include <cstring>
 #include <string>
+#include <vector>
 
 namespace json = boost::json;
 
@@ -28,6 +29,7 @@ struct PluginConfig {
     double temperature = 0.7;
 };
 static PluginConfig g_cfg;
+static std::string g_tools_json;  // W20.2: cached by configure(), consumed by chat/chat_stream
 
 // ============================================================================
 // 安全擦除：用 volatile 写零循环防止编译器优化
@@ -204,18 +206,36 @@ static void parse_response(const dstalk_host_api_t* host,
     }
 }
 
+// ============================================================================
+// 流式上下文：在 SSE 回调间累积内容和 tool_calls
+// ============================================================================
+struct ToolCallAccum {
+    int index = -1;
+    std::string id;
+    std::string name;
+    std::string arguments;  // 增量拼接的 JSON arguments 字符串
+};
+
+struct StreamContext {
+    const dstalk_host_api_t* host;
+    dstalk_stream_cb user_cb;
+    void* userdata;
+    std::string accumulated;
+    bool streaming_ok = true;
+    std::vector<ToolCallAccum> tool_calls;  // W20.2: 按 index 累积 delta tool_calls
+};
+
 // ============================================================================
 // SSE 行解析（OpenAI 兼容格式）
 // ============================================================================
-static bool parse_sse_line(const std::string& line, std::string& token_out)
+static bool parse_sse_line(const std::string& line, std::string& token_out,
+                           StreamContext* ctx)
 {
     if (line.rfind("data: ", 0) != 0) return false;
 
     std::string data = line.substr(6);
 
     // F-13.2-3: Trim leading/trailing whitespace before comparing [DONE] sentinel.
-    // Some servers may emit "data: [DONE] " with trailing spaces, which would
-    // cause exact match to fail and the stream to never terminate.
     const char* ws = " \t\r\n";
     size_t start = data.find_first_not_of(ws);
     if (start != std::string::npos) {
@@ -233,6 +253,44 @@ static bool parse_sse_line(const std::string& line, std::string& token_out)
         auto choices = obj["choices"].as_array();
         if (!choices.empty()) {
             auto delta = choices[0].as_object()["delta"].as_object();
+
+            // W20.2: 处理 delta["tool_calls"] 增量 chunk
+            // DeepSeek/OpenAI 流式模式 tool_calls 跨多个 SSE 事件分片传输：
+            //   事件 1: {"index":0, "id":"call_xxx", "function":{"name":"foo"}}
+            //   事件 2: {"index":0, "function":{"arguments":"{\"bar\":"}}
+            //   事件 3: {"index":0, "function":{"arguments":"1}"}}
+            // 需要按 index 累积 id/name/arguments。
+            if (delta.contains("tool_calls") && ctx) {
+                auto tc_array = delta["tool_calls"].as_array();
+                for (auto& tc_val : tc_array) {
+                    auto tc_obj = tc_val.as_object();
+                    int idx = tc_obj.contains("index")
+                        ? static_cast<int>(json::value_to<int64_t>(tc_obj["index"])) : -1;
+                    if (idx < 0) continue;
+
+                    while (static_cast<int>(ctx->tool_calls.size()) <= idx) {
+                        ctx->tool_calls.push_back({});
+                    }
+                    auto& acc = ctx->tool_calls[idx];
+                    acc.index = idx;
+
+                    if (tc_obj.contains("id") && tc_obj["id"].is_string()) {
+                        acc.id = json::value_to<std::string>(tc_obj["id"]);
+                    }
+
+                    if (tc_obj.contains("function") && tc_obj["function"].is_object()) {
+                        auto func = tc_obj["function"].as_object();
+                        if (func.contains("name") && func["name"].is_string()) {
+                            acc.name = json::value_to<std::string>(func["name"]);
+                        }
+                        if (func.contains("arguments") && func["arguments"].is_string()) {
+                            acc.arguments += json::value_to<std::string>(func["arguments"]);
+                        }
+                    }
+                }
+                return false;  // tool_calls 已处理，无内容 token 给用户回调
+            }
+
             if (delta.contains("content")) {
                 token_out = json::value_to<std::string>(delta["content"]);
                 return true;
@@ -261,6 +319,17 @@ static int my_configure(const char* provider, const char* base_url,
 
         const dstalk_host_api_t* host = g_host.load(std::memory_order_acquire);
         if (host) {
+            // W20.2: 从 tools service 缓存 tools_json，供 chat/chat_stream 复用
+            auto* tools_svc = reinterpret_cast<const dstalk_tools_service_t*>(
+                host->query_service("tools", 1));
+            if (tools_svc && tools_svc->get_tools_json) {
+                char* json = tools_svc->get_tools_json();
+                if (json) {
+                    g_tools_json = json;
+                    host->free(json);
+                }
+            }
+
             host->log(DSTALK_LOG_INFO,
                 "[deepseek] configured: model=%s base_url=%s max_tokens=%d temperature=%.2f",
                 g_cfg.model.c_str(), g_cfg.base_url.c_str(),
@@ -346,15 +415,6 @@ static dstalk_chat_result_t my_chat(
 // chat_stream 实现
 // ============================================================================
 
-// 回调上下文：在流式传输中收集累积内容和最终状态
-struct StreamContext {
-    const dstalk_host_api_t* host;
-    dstalk_stream_cb user_cb;
-    void* userdata;
-    std::string accumulated;
-    bool streaming_ok = true;
-};
-
 // 行回调：解析 SSE line，将 token 传递给用户回调
 static int sse_line_callback(const char* line, void* userdata)
 {
@@ -365,7 +425,7 @@ static int sse_line_callback(const char* line, void* userdata)
         std::string line_str(line);
         std::string token;
 
-        if (!parse_sse_line(line_str, token)) return 1; // 非 data 行，继续
+        if (!parse_sse_line(line_str, token, ctx)) return 1; // 非 data/tool_calls 行，继续
 
         if (token.empty()) return 0; // [DONE]，停止
 
@@ -408,7 +468,7 @@ static dstalk_chat_result_t my_chat_stream(
         std::string target_path = target + "/chat/completions";
 
         std::string body = build_request_json(history, history_len,
-            user_input ? user_input : "", "", true);  // stream=true, no tools
+            user_input ? user_input : "", g_tools_json, true);
 
         std::string headers_json = build_headers_json(g_cfg.api_key);
 
@@ -458,7 +518,11 @@ static dstalk_chat_result_t my_chat_stream(
 
         if (response_body && host) host->free(response_body);
 
-        if (ctx.accumulated.empty()) {
+        // W20.2: 成功条件 = 有内容 OR 有 tool_calls（tool-only 响应如 function calling）
+        bool has_content = !ctx.accumulated.empty();
+        bool has_tool_calls = !ctx.tool_calls.empty();
+
+        if (!has_content && !has_tool_calls) {
             r.ok = 0;
             r.error = host ? host->strdup("no content received") : nullptr;
             r.content = nullptr;
@@ -466,8 +530,28 @@ static dstalk_chat_result_t my_chat_stream(
         } else {
             r.ok = 1;
             r.error = nullptr;
-            r.content = host ? host->strdup(ctx.accumulated.c_str()) : nullptr;
-            r.tool_calls_json = nullptr;
+            r.content = has_content
+                ? host->strdup(ctx.accumulated.c_str()) : nullptr;
+
+            // 序列化累积的 tool_calls 为 JSON（兼容 OpenAI tool_calls 格式）
+            if (has_tool_calls) {
+                json::array tc_array;
+                for (auto& tc : ctx.tool_calls) {
+                    json::object tc_obj;
+                    tc_obj["index"] = tc.index;
+                    if (!tc.id.empty())   tc_obj["id"]   = tc.id;
+                    tc_obj["type"] = "function";
+                    json::object func;
+                    if (!tc.name.empty()) func["name"] = tc.name;
+                    func["arguments"] = tc.arguments;
+                    tc_obj["function"] = func;
+                    tc_array.push_back(std::move(tc_obj));
+                }
+                std::string tc_json = json::serialize(tc_array);
+                r.tool_calls_json = host ? host->strdup(tc_json.c_str()) : nullptr;
+            } else {
+                r.tool_calls_json = nullptr;
+            }
         }
         return r;
     } catch (const std::exception& e) {

plugins/session/src/session_plugin.cpp

@@ -12,6 +12,7 @@
 #include <atomic>
 #include <cstddef>
 #include <cstdint>
+#include <cstdlib>
 #include <cstring>
 #include <exception>
 #include <mutex>
@@ -287,6 +288,24 @@ static dstalk_session_service_t g_session_service = {
     session_token_count
 };
 
+// ============================================================
+// W20.6: 默认会话保存路径（平台标准目录）
+// ============================================================
+
+static std::string get_default_session_path() {
+#ifdef _WIN32
+    char* buf = nullptr;
+    size_t len = 0;
+    _dupenv_s(&buf, &len, "APPDATA");
+    std::string dir = buf ? std::string(buf) + "/dstalk" : "dstalk";
+    free(buf);
+#else
+    const char* home = std::getenv("HOME");
+    std::string dir = home ? std::string(home) + "/.dstalk" : "/tmp/dstalk";
+#endif
+    return dir + "/session.json";
+}
+
 // ============================================================
 // 插件生命周期
 // ============================================================
@@ -304,7 +323,13 @@ static int on_init(const dstalk_host_api_t* host) {
         g_file_io.store(static_cast<const dstalk_file_io_service_t*>(raw), std::memory_order_release);
 
         // 注册自身服务
-        return host->register_service("session", 1, &g_session_service);
+        int ret = host->register_service("session", 1, &g_session_service);
+        if (ret != 0) return ret;
+
+        // W20.6: 从默认路径恢复会话（文件不存在则静默失败）
+        session_load(get_default_session_path().c_str());
+
+        return 0;
     } catch (const std::exception& e) {
         const dstalk_host_api_t* h = g_host.load(std::memory_order_acquire);
         if (h) h->log(DSTALK_LOG_ERROR, "on_init[session]: %s", e.what());
@@ -318,6 +343,9 @@ static int on_init(const dstalk_host_api_t* host) {
 
 static void on_shutdown() {
     try {
+        // W20.6: 清空前自动保存到默认路径
+        session_save(get_default_session_path().c_str());
+
         std::lock_guard<std::mutex> lock(g_session_mutex);
         rebuild_cached_history_locked();
         g_cached_history.clear();

tests/CMakeLists.txt

@@ -87,3 +87,39 @@ target_link_libraries(dstalk-context-plugin-test
 )
 
 add_test(NAME dstalk-context-plugin-test COMMAND dstalk-context-plugin-test)
+
+# ============================================================
+# dstalk-plugin-loader-test — PluginLoader 安全回归测试
+# W20.3 (qa-xu): 覆盖 W19 F-18.3-1~5 修复验证
+# ============================================================
+
+add_executable(dstalk-plugin-loader-test
+    plugin_loader_test.cpp
+    ${CMAKE_SOURCE_DIR}/dstalk-core/src/plugin_loader.cpp
+    ${CMAKE_SOURCE_DIR}/dstalk-core/src/boost_json.cpp
+)
+
+target_include_directories(dstalk-plugin-loader-test
+    PRIVATE ${CMAKE_SOURCE_DIR}/dstalk-core/src
+)
+
+target_compile_features(dstalk-plugin-loader-test
+    PRIVATE cxx_std_17
+)
+
+find_package(Boost REQUIRED CONFIG)
+
+target_compile_definitions(dstalk-plugin-loader-test
+    PRIVATE
+        BOOST_JSON_HEADER_ONLY
+        BOOST_ALL_NO_LIB
+        DSTALK_TEST_PLUGINS_DIR="${CMAKE_BINARY_DIR}/plugins"
+)
+
+target_link_libraries(dstalk-plugin-loader-test
+    PRIVATE
+        dstalk
+        boost::boost
+)
+
+add_test(NAME dstalk-plugin-loader-test COMMAND dstalk-plugin-loader-test)

tests/plugin_loader_test.cpp

@@ -0,0 +1,309 @@
+// ============================================================================
+// plugin_loader_test.cpp — PluginLoader 安全回归测试
+// ============================================================================
+// W20.3 (qa-xu 徐磊): 覆盖 W19 修复的 5 条发现 (F-18.3-1~5)
+//   - F-18.3-3: 路径验证 (lexically_normal + 扩展名 + 目录约束)
+//   - F-18.3-4: next_id_ atomic 唯一性 + 单调递增
+//   - F-18.3-2: host_api_->log 调用 (mock 验证)
+//   - F-18.3-1: try/catch 异常安全边界 (间接: 注入 mock 不崩溃)
+// ============================================================================
+
+#include "plugin_loader.hpp"
+
+#include <algorithm>
+#include <cstdarg>
+#include <cstdio>
+#include <cstring>
+#include <filesystem>
+#include <iostream>
+#include <string>
+#include <thread>
+#include <vector>
+
+namespace fs = std::filesystem;
+
+// ---- 轻量断言 ----
+static int g_failures = 0;
+#define CHECK(cond, msg) do {                              \
+    if (cond) {                                            \
+        std::cout << "[OK] " << (msg) << "\n";             \
+    } else {                                               \
+        std::cerr << "[FAIL] " << (msg) << "\n";           \
+        g_failures++;                                      \
+    }                                                      \
+} while (0)
+
+// ============================================================================
+// Mock host_api — 捕获 log 调用以验证失败路径日志 (F-18.3-2)
+// ============================================================================
+static int g_log_call_count = 0;
+static int g_last_severity = 0;
+static char g_last_log_msg[1024] = {0};
+
+static void mock_log(int level, const char* fmt, ...) {
+    g_log_call_count++;
+    g_last_severity = level;
+    va_list args;
+    va_start(args, fmt);
+    vsnprintf(g_last_log_msg, sizeof(g_last_log_msg), fmt, args);
+    va_end(args);
+}
+
+static int  stub_reg(const char*, int, void*)              { return -1; }
+static void* stub_query(const char*, int)                   { return nullptr; }
+static int  stub_sub(int, dstalk_event_handler_fn, void*)   { return -1; }
+static int  stub_emit(int, const void*)                     { return -1; }
+static void stub_unsub(int)                                 {}
+static const char* stub_cget(const char*)                   { return nullptr; }
+static int  stub_cset(const char*, const char*)             { return -1; }
+static void* stub_alloc(size_t)                             { return nullptr; }
+static void stub_free(void*)                                {}
+static char* stub_strdup(const char*)                       { return nullptr; }
+
+static dstalk_host_api_t g_mock_host_api = {
+    stub_reg, stub_query,
+    stub_sub, stub_emit, stub_unsub,
+    stub_cget, stub_cset,
+    mock_log,
+    stub_alloc, stub_free, stub_strdup
+};
+
+static void reset_log_state() {
+    g_log_call_count = 0;
+    g_last_severity = 0;
+    g_last_log_msg[0] = '\0';
+}
+
+// ============================================================================
+// Helper: 获取已构建的 plugins/ 目录绝对路径
+// ============================================================================
+static fs::path get_plugins_dir() {
+#ifdef DSTALK_TEST_PLUGINS_DIR
+    return fs::path(DSTALK_TEST_PLUGINS_DIR);
+#else
+    return fs::current_path().parent_path() / "plugins";
+#endif
+}
+
+// ============================================================================
+int main()
+{
+    std::cout << "=== dstalk plugin_loader regression tests (W20.3) ===\n\n";
+
+    // ========================================================================
+    // Block 1: 路径验证 — 拒绝非法路径 (F-18.3-3)
+    // ========================================================================
+    std::cout << "--- Block 1: Path validation — rejection ---\n";
+    {
+        dstalk::PluginLoader loader;
+
+        // T1.1: nullptr
+        CHECK(loader.load_plugin(nullptr) == -1,
+              "T1.1: nullptr path returns -1");
+
+        // T1.2: 非法扩展名 .txt
+        CHECK(loader.load_plugin("plugins/test.txt") == -1,
+              "T1.2: .txt extension rejected");
+
+        // T1.3: 路径含 .. 遍历
+        CHECK(loader.load_plugin("../plugins/test.dll") == -1,
+              "T1.3: ../ traversal rejected");
+
+        // T1.4: 不在 plugins/ 目录下
+        auto tmp = fs::temp_directory_path() / "dstalk_test_no_plugins" / "test.dll";
+        CHECK(loader.load_plugin(tmp.string().c_str()) == -1,
+              "T1.4: path not under plugins/ dir rejected");
+
+        // T1.5: 路径中间的 .. 段
+        CHECK(loader.load_plugin("plugins/../secret/test.dll") == -1,
+              "T1.5: .. in middle of path rejected");
+
+        // T1.6: 无扩展名
+        CHECK(loader.load_plugin("plugins/test") == -1,
+              "T1.6: no extension rejected");
+
+        // T1.7: 合法扩展名但不在 plugins/ 下
+        CHECK(loader.load_plugin("/etc/someconfig.so") == -1,
+              "T1.7: .so extension but not under plugins/ rejected");
+    }
+
+    // ========================================================================
+    // Block 2: 合法路径 — 成功加载 + next_id_ 验证 (F-18.3-4)
+    // ========================================================================
+    std::cout << "\n--- Block 2: Valid path — successful load + ID uniqueness ---\n";
+    {
+        dstalk::PluginLoader loader;
+        fs::path plugins_dir = get_plugins_dir();
+
+        fs::path dll_config  = plugins_dir / "plugin-config.dll";
+        fs::path dll_fileio  = plugins_dir / "plugin-file-io.dll";
+
+        bool have_plugins = fs::exists(dll_config) && fs::exists(dll_fileio);
+
+        if (!have_plugins) {
+            std::cout << "[WARN] Plugin DLLs not found at " << plugins_dir.string()
+                      << " — skipping Block 2\n";
+        } else {
+            // T2.1: 加载第一个插件
+            int id1 = loader.load_plugin(dll_config.string().c_str());
+            CHECK(id1 >= 1, "T2.1: first plugin loaded with positive ID");
+            std::cout << "  id1 = " << id1 << "\n";
+
+            // T2.2: 加载第二个不同插件
+            int id2 = loader.load_plugin(dll_fileio.string().c_str());
+            CHECK(id2 >= 1, "T2.2: second plugin loaded with positive ID");
+            std::cout << "  id2 = " << id2 << "\n";
+
+            // T2.3: ID 唯一
+            CHECK(id1 != id2, "T2.3: IDs are unique (next_id_ atomicity)");
+
+            // T2.4: ID 单调递增
+            CHECK(id2 > id1, "T2.4: IDs monotonically increasing");
+
+            // T2.5: get_plugin 可查询到已加载插件
+            const dstalk::PluginInfo* info1 = loader.get_plugin(id1);
+            CHECK(info1 != nullptr, "T2.5: get_plugin(id1) returns non-null");
+            if (info1) {
+                CHECK(!info1->name.empty(), "T2.6: plugin has non-empty name");
+                std::cout << "  plugin1 name: " << info1->name << "\n";
+            }
+
+            // T2.7: get_plugin 对无效 ID 返回 nullptr
+            CHECK(loader.get_plugin(99999) == nullptr,
+                  "T2.7: get_plugin(invalid_id) returns nullptr");
+
+            // T2.8: 卸载后 get_plugin 返回 nullptr
+            int ret = loader.unload_plugin(id1);
+            CHECK(ret == 0, "T2.8: unload_plugin returns 0");
+            CHECK(loader.get_plugin(id1) == nullptr,
+                  "T2.9: get_plugin returns nullptr after unload");
+
+            // 清理
+            loader.unload_plugin(id2);
+        }
+    }
+
+    // ========================================================================
+    // Block 3: next_id_ 原子性 — 多线程并发加载 (F-18.3-4)
+    // ========================================================================
+    std::cout << "\n--- Block 3: next_id_ atomicity — concurrent loads ---\n";
+    {
+        dstalk::PluginLoader loader;
+        fs::path plugins_dir = get_plugins_dir();
+
+        std::vector<fs::path> dlls;
+        for (auto name : {"plugin-config.dll", "plugin-file-io.dll",
+                          "plugin-context.dll", "plugin-session.dll"}) {
+            fs::path p = plugins_dir / name;
+            if (fs::exists(p)) dlls.push_back(p);
+        }
+
+        if (dlls.size() < 2) {
+            std::cout << "[WARN] Not enough plugin DLLs for concurrency test"
+                      << " (found " << dlls.size() << ") — skipping Block 3\n";
+        } else {
+            std::vector<int> ids(dlls.size(), -1);
+            std::vector<std::thread> threads;
+
+            for (size_t i = 0; i < dlls.size(); ++i) {
+                threads.emplace_back([&loader, &ids, i, &dlls]() {
+                    ids[i] = loader.load_plugin(dlls[i].string().c_str());
+                });
+            }
+
+            for (auto& t : threads) t.join();
+
+            // 验证: 所有 load 成功, ID 唯一且 > 0
+            std::vector<int> valid_ids;
+            for (size_t i = 0; i < ids.size(); ++i) {
+                CHECK(ids[i] >= 1, "T3." + std::to_string(i)
+                      + ": thread load succeeded (id="
+                      + std::to_string(ids[i]) + ")");
+                if (ids[i] >= 1) valid_ids.push_back(ids[i]);
+            }
+
+            // 去重后大小应等于成功加载数
+            std::sort(valid_ids.begin(), valid_ids.end());
+            auto dup = std::unique(valid_ids.begin(), valid_ids.end());
+            size_t unique_count = std::distance(valid_ids.begin(), dup);
+            CHECK(unique_count == valid_ids.size(),
+                  "T3.X: all IDs unique under concurrent loads ("
+                  + std::to_string(unique_count) + "/"
+                  + std::to_string(valid_ids.size()) + ")");
+
+            // 清理
+            for (int id : valid_ids) loader.unload_plugin(id);
+        }
+    }
+
+    // ========================================================================
+    // Block 4: 失败路径日志 — host_api->log 被调用 (F-18.3-2)
+    // ========================================================================
+    std::cout << "\n--- Block 4: Failure-path logging (host_api->log) ---\n";
+    {
+        dstalk::PluginLoader loader;
+
+        // 4.1: 无 host_api 时 load_plugin 失败不崩溃
+        reset_log_state();
+        int id = loader.load_plugin("bad_ext.noext");
+        CHECK(id == -1, "T4.1: load_plugin with invalid ext returns -1 (no host_api)");
+        CHECK(g_log_call_count == 0,
+              "T4.2: log NOT called when host_api_ is null");
+
+        // 4.2: 设置 mock host_api 后验证 log 被调用
+        int init_ret = loader.initialize_all(&g_mock_host_api);
+        CHECK(init_ret == 0, "T4.3: initialize_all with mock host_api returns 0");
+
+        reset_log_state();
+        id = loader.load_plugin("bad_ext.noext");
+        CHECK(id == -1, "T4.4: load_plugin still returns -1 with mock host_api");
+        CHECK(g_log_call_count >= 1,
+              "T4.5: host_api->log WAS called on path validation failure");
+        CHECK(g_last_severity == DSTALK_LOG_ERROR,
+              "T4.6: log severity is DSTALK_LOG_ERROR");
+        std::cout << "  log msg: " << g_last_log_msg << "\n";
+
+        // 4.3: LoadLibrary 失败也触发 log (文件不存在)
+        reset_log_state();
+        fs::path missing = get_plugins_dir() / "nonexistent_plugin.dll";
+        id = loader.load_plugin(missing.string().c_str());
+        CHECK(id == -1, "T4.7: missing DLL returns -1");
+        CHECK(g_log_call_count >= 1,
+              "T4.8: host_api->log called on LoadLibrary failure");
+        std::cout << "  log msg: " << g_last_log_msg << "\n";
+    }
+
+    // ========================================================================
+    // Block 5: 边界 — 空 loader / 无效操作
+    // ========================================================================
+    std::cout << "\n--- Block 5: Edge cases — empty loader / invalid op ---\n";
+    {
+        dstalk::PluginLoader loader;
+
+        // T5.1: unload 不存在的 ID 返回 -1
+        CHECK(loader.unload_plugin(42) == -1,
+              "T5.1: unload_plugin(nonexistent) returns -1");
+
+        // T5.2: 空 PluginLoader 的 list_plugins 返回 "[]"
+        std::string json = loader.list_plugins();
+        CHECK(!json.empty(), "T5.2: list_plugins returns non-empty string");
+        CHECK(json == "[]", "T5.3: empty loader produces empty JSON array");
+        std::cout << "  list_plugins (empty): " << json << "\n";
+
+        // T5.3: get_plugin 在空 loader 上返回 nullptr
+        CHECK(loader.get_plugin(1) == nullptr,
+              "T5.4: get_plugin on empty loader returns nullptr");
+    }
+
+    // ========================================================================
+    // 结果
+    // ========================================================================
+    std::cout << "\n";
+    if (g_failures == 0) {
+        std::cout << "=== All plugin_loader regression tests passed ===\n";
+        return 0;
+    } else {
+        std::cerr << "=== " << g_failures << " test(s) FAILED ===\n";
+        return 1;
+    }
+}