fix: skill metadata truncation must fall back to UTF-8 sequence boundary

src/skills/skill_loader.cpp

@@ -34,7 +34,10 @@ std::string read_frontmatter_chunk(const fs::path& path) {
 
 std::string truncate(const std::string& s, size_t n) {
     if (s.size() <= n) return s;
-    return s.substr(0, n > 3 ? n - 3 : n) + "...";
+    // 字节预算截断必须回退到 UTF-8 序列边界,否则切在多字节字符中间会
+    // 留下"残缺引导字节 + '.'"的非法序列(如 0xE9 0x94 + "..." 的 0x2E),
+    // 进入 skills 索引后会被 body.dump() 以 type_error.316 打挂整个请求。
+    return truncate_utf8_prefix(s, n, "...");
 }
 
 std::string first_non_empty_body_line(const std::string& body) {
@@ -162,13 +165,13 @@ std::optional<SkillMetadata> load_skill_from_dir(const fs::path& dir,
 
     std::string desc = get_string(fm, "description");
     if (desc.empty()) desc = first_non_empty_body_line(body);
-    meta.description = truncate(desc, 1024);
+    meta.description = ensure_utf8(truncate(desc, 1024));
 
     // 可选触发条件:写明"什么时候该用这个 skill"。主键 whenToUse 与
     // claude-code 的 frontmatter 约定一致,snake_case 作别名。
     std::string when = get_string(fm, "whenToUse");
     if (when.empty()) when = get_string(fm, "when_to_use");
-    meta.when_to_use = truncate(when, 1024);
+    meta.when_to_use = ensure_utf8(truncate(when, 1024));
 
     meta.category = derive_category(dir, scan_root);
     meta.platforms = get_list(fm, "platforms");

tests/skills/skill_loader_utf8_test.cpp

@@ -0,0 +1,101 @@
+// 覆盖 src/skills/skill_loader.cpp 的元数据截断:description / whenToUse 超预算
+// 时必须回退到 UTF-8 序列边界,不得切在多字节字符中间。
+//
+// 背景 bug:skill_loader 的 truncate() 曾用原始字节 substr + "..." 截断。
+// 当 description 超过 1024 字节且截断点落在多字节字符中间时,会留下
+// "残缺引导字节 + '.'" 的非法序列(如 0xE9 0x94 + "..." 的 0x2E),进入
+// skills 索引后,请求体 body.dump() 以 nlohmann type_error.316
+// ("invalid UTF-8 byte at index …: 0x2E") 打挂整个请求。本测试锁定该回归。
+
+#include <gtest/gtest.h>
+
+#include <filesystem>
+#include <fstream>
+
+#include "skills/skill_loader.hpp"
+#include "utils/encoding.hpp"
+
+namespace {
+
+using acecode::load_skill_from_dir;
+using acecode::is_valid_utf8;
+
+// 构造一个 SKILL.md,其 description 超过 1024 字节,且 1021 字节的
+// 截断预算落在某个 3 字节汉字(如"错"= 0xE9 0x94 0x99)中间——正是线上
+// ~/.agents/skills/lark-apps/SKILL.md 触发崩溃的形态。
+TEST(SkillLoaderTruncation, DescriptionCutInsideMultiByteCharStaysValidUtf8) {
+    const std::string base =
+        "---\n"
+        "name: trunc-utf8-test\n"
+        "description: ";
+    // truncate(..., 1024) 为 "..." 预留 3 字节,所以原文预算是 1021。
+    // 让三字节的"错"从偏移 1020 开始,原实现会只保留它的首字节。
+    std::string description =
+        std::string(1020, 'a') + u8"\u9519" + std::string(300, 'b');
+    ASSERT_GT(description.size(), static_cast<size_t>(1024));
+
+    const std::string content =
+        base + description + "\n"
+        "---\n"
+        "body\n";
+
+    const std::string dir = "trunc-utf8-test";
+    // 写入临时目录
+    const auto tmp_root = std::filesystem::temp_directory_path();
+    const auto skill_dir = tmp_root / "acecode_skill_loader_utf8_test" / dir;
+    std::filesystem::create_directories(skill_dir);
+    const auto skill_md = skill_dir / "SKILL.md";
+    {
+        std::ofstream ofs(skill_md, std::ios::binary);
+        ofs << content;
+    }
+
+    const auto scan_root = tmp_root / "acecode_skill_loader_utf8_test";
+    const auto meta = load_skill_from_dir(skill_dir, scan_root);
+    ASSERT_TRUE(meta.has_value());
+
+    // 核心断言:截断后的 description 必须是合法 UTF-8(回归前会失败)。
+    EXPECT_TRUE(is_valid_utf8(meta->description))
+        << "description after truncation must remain valid UTF-8";
+    // 且带有截断后缀(说明确实触发了截断路径)。
+    // 预算 1021 字节 + "..."(3 字节) = 至多 1024;若 1021 落在多字节
+    // 字符中间则回退到序列边界,总长只会更短。
+    EXPECT_EQ(meta->description, std::string(1020, 'a') + "...");
+
+    std::filesystem::remove_all(tmp_root / "acecode_skill_loader_utf8_test");
+}
+
+// whenToUse 同样走 truncate 路径,验证同一条防御。
+TEST(SkillLoaderTruncation, WhenToUseCutInsideMultiByteCharStaysValidUtf8) {
+    // 这次让"错"从偏移 1019 开始,截断预算会落在它的第三个字节上,
+    // 从而同时覆盖需要回退两个 continuation byte 的情况。
+    const std::string when =
+        std::string(1019, 'x') + u8"\u9519" + std::string(300, 'y');
+    const std::string content =
+        "---\n"
+        "name: trunc-when-test\n"
+        "description: short\n"
+        "whenToUse: " + when + "\n"
+        "---\n"
+        "body\n";
+
+    const auto tmp_root = std::filesystem::temp_directory_path();
+    const auto skill_dir = tmp_root / "acecode_skill_loader_utf8_test2" / "trunc-when-test";
+    std::filesystem::create_directories(skill_dir);
+    const auto skill_md = skill_dir / "SKILL.md";
+    {
+        std::ofstream ofs(skill_md, std::ios::binary);
+        ofs << content;
+    }
+
+    const auto scan_root = tmp_root / "acecode_skill_loader_utf8_test2";
+    const auto meta = load_skill_from_dir(skill_dir, scan_root);
+    ASSERT_TRUE(meta.has_value());
+    EXPECT_TRUE(is_valid_utf8(meta->when_to_use))
+        << "when_to_use after truncation must remain valid UTF-8";
+    EXPECT_EQ(meta->when_to_use, std::string(1019, 'x') + "...");
+
+    std::filesystem::remove_all(tmp_root / "acecode_skill_loader_utf8_test2");
+}
+
+} // namespace

tests/web/web_server_smoke_test.cpp

@@ -2355,13 +2355,26 @@ TEST(WebServerHttp, GlobalSessionSearchIncludesHiddenAndMarkerlessProjects) {
         EXPECT_NE(workspace.value("hash", std::string{}), markerless_hash);
     }
 
-    auto content = cpr::Get(
-        cpr::Url{fx.url("/api/session-search/user-messages")},
-        cpr::Parameters{{"q", "global-hidden-content-needle"}, {"limit", "10"},
-                        {"request_id", "hidden-content-smoke"}});
-    ASSERT_EQ(content.status_code, 200) << content.text;
-    const auto matches = json::parse(content.text)["matches"];
-    ASSERT_EQ(matches.size(), 1u);
+    cpr::Response content;
+    json content_body;
+    const auto content_deadline = std::chrono::steady_clock::now() + 3s;
+    do {
+        content = cpr::Get(
+            cpr::Url{fx.url("/api/session-search/user-messages")},
+            cpr::Parameters{{"q", "global-hidden-content-needle"}, {"limit", "10"},
+                            {"request_id", "hidden-content-smoke"}});
+        ASSERT_EQ(content.status_code, 200) << content.text;
+        content_body = json::parse(content.text);
+        if (content_body["progress"].value("complete", false)) {
+            break;
+        }
+        std::this_thread::sleep_for(5ms);
+    } while (std::chrono::steady_clock::now() < content_deadline);
+
+    ASSERT_TRUE(content_body["progress"].value("complete", false)) << content.text;
+    ASSERT_TRUE(content_body["matches"].is_array()) << content.text;
+    const auto& matches = content_body["matches"];
+    ASSERT_EQ(matches.size(), 1u) << content.text;
     EXPECT_EQ(matches[0]["id"], hidden_id);
     EXPECT_EQ(matches[0]["workspace_hash"], hidden_hash);
     EXPECT_FALSE(matches[0]["workspace_visible"].get<bool>());