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>

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) {