XiuChengWu
f2da0f2ed4
- Introduced a new Python script `check_agents_metadata.py` for validating agent metadata, including YAML parsing, rating ranges, and cross-references. - Added usage instructions and exit codes for the script. - Created a new markdown file `模块目录和功能说明.md` to outline the directory structure and functionality of the modules. - Added a text file `说明此文件不可AI修改.txt` to specify that certain files should not be modified by AI, including important information about the `dstalk` framework and its modules.
788 lines
35 KiB
C++
788 lines
35 KiB
C++
/*
|
||
* @file smoke_test.cpp
|
||
* @brief Basic smoke test: verifies dstalk_init/shutdown cycle, service queries,
|
||
* file_io, session, null-safety, escape boundaries, tool chain, and
|
||
* regression protections R1-R4 (W13.6 qa-xu).
|
||
* 基础冒烟测试：验证 dstalk_init/shutdown 生命周期、服务查询、file_io、session、
|
||
* 空指针安全、转义边界、工具链调用，以及回归保护 R1-R4 (W13.6 qa-xu)。
|
||
* Copyright (c) 2026 dstalk contributors. GPLv3.
|
||
*/
|
||
|
||
#include "dstalk/dstalk_host.h"
|
||
|
||
#include <cstring>
|
||
#include <filesystem>
|
||
#include <fstream>
|
||
#include <iostream>
|
||
#include <string>
|
||
|
||
// ---- 回归测试断言 (W13.6 qa-xu) ----
|
||
// Regression test assertion macro (W13.6 qa-xu): prints [OK]/[FAIL] and tracks failures
|
||
static int g_regression_failures = 0;
|
||
#define REGCHECK(cond, msg) do { \
|
||
if (cond) { \
|
||
std::cout << "[OK] " << (msg) << "\n"; \
|
||
} else { \
|
||
std::cerr << "[FAIL] " << (msg) << "\n"; \
|
||
g_regression_failures++; \
|
||
} \
|
||
} while (0)
|
||
|
||
// W21.5 mock tool handler (qa-xu): increments call counter and returns mock result JSON
|
||
// W21.5 模拟工具处理函数 (qa-xu)：递增调用计数器并返回模拟结果 JSON
|
||
static int g_mock_tool_called = 0;
|
||
static char* mock_tool_handler(const char* /*args_json*/) {
|
||
g_mock_tool_called++;
|
||
return dstalk_strdup("{\"mock_result\":\"ok\"}");
|
||
}
|
||
|
||
// 冒烟测试主流程：init → 服务查询 → file_io → session → ai → config，
|
||
// 然后是扩展测试（空指针安全、转义边界、工具链、session 健壮性），
|
||
// 接着是回归保护 R1-R3、W21.5 工具调用边界和 R4 生命周期循环。
|
||
// Smoke test main: init -> service queries -> file_io -> session -> ai -> config,
|
||
// then extended tests (null-safety, escape, tool chain, session robustness),
|
||
// then regression protections R1-R3, W21.5 tool-call boundaries, and R4 lifecycle cycles.
|
||
int main()
|
||
{
|
||
const auto dir = std::filesystem::temp_directory_path() / "dstalk-smoke-test";
|
||
std::filesystem::create_directories(dir);
|
||
|
||
// 写一个配置文件用于初始化 / Write a config file for initialization
|
||
const auto config_path = dir / "config.toml";
|
||
{
|
||
std::ofstream config(config_path);
|
||
config << "[api]\n"
|
||
<< "provider = \"deepseek\"\n"
|
||
<< "base_url = \"https://api.deepseek.com/v1\"\n"
|
||
<< "api_key = \"test-key\"\n"
|
||
<< "model = \"deepseek-v4-pro\"\n";
|
||
}
|
||
|
||
// 初始化主机（会自动扫描 plugins/ 加载插件）/ Init host (auto-scans plugins/ to load plugins)
|
||
if (dstalk_init(config_path.string().c_str()) != 0) {
|
||
std::cerr << "dstalk_init failed\n";
|
||
return 1;
|
||
}
|
||
std::cout << "[OK] dstalk_init succeeded\n";
|
||
|
||
// 验证插件列表 / Verify plugin list
|
||
{
|
||
char* list_json = nullptr;
|
||
int ret = dstalk_plugin_list(&list_json);
|
||
if (ret == 0 && list_json) {
|
||
std::cout << "[OK] plugins loaded: " << list_json << "\n";
|
||
dstalk_free(list_json);
|
||
} else {
|
||
std::cerr << "[WARN] dstalk_plugin_list returned: " << ret << "\n";
|
||
}
|
||
}
|
||
|
||
// 测试服务查询: file_io / Test service query: file_io
|
||
auto* file_io = static_cast<const dstalk_file_io_service_t*>(
|
||
dstalk_service_query("file_io", 1));
|
||
if (file_io) {
|
||
std::cout << "[OK] file_io service found\n";
|
||
|
||
// 测试写入 / Test write
|
||
const auto file_path = dir / "sample.txt";
|
||
constexpr const char* sample_content = "hello dstalk\nquote=\"yes\" tab=\t slash=\\";
|
||
if (file_io->write(file_path.string().c_str(), sample_content) == 0) {
|
||
std::cout << "[OK] file_io->write succeeded\n";
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->write failed\n";
|
||
dstalk_shutdown();
|
||
return 1;
|
||
}
|
||
|
||
// 测试读取 / Test read
|
||
char* content = nullptr;
|
||
if (file_io->read(file_path.string().c_str(), &content) == 0 && content) {
|
||
bool ok = std::strcmp(content, sample_content) == 0;
|
||
dstalk_free(content);
|
||
if (ok) {
|
||
std::cout << "[OK] file_io->read content matches\n";
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->read content mismatch\n";
|
||
dstalk_shutdown();
|
||
return 1;
|
||
}
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->read failed\n";
|
||
dstalk_shutdown();
|
||
return 1;
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] file_io service not found (plugin may not be in plugins/ dir)\n";
|
||
}
|
||
|
||
// 测试服务查询: session / Test service query: session
|
||
auto* session = static_cast<const dstalk_session_service_t*>(
|
||
dstalk_service_query("session", 1));
|
||
if (session) {
|
||
std::cout << "[OK] session service found\n";
|
||
|
||
// 测试 session save/load / Test session save/load
|
||
const auto session_path = dir / "session.jsonl";
|
||
const auto saved_path = dir / "session-saved.jsonl";
|
||
constexpr const char* session_content =
|
||
"{\"role\":\"user\",\"content\":\"line\\n\\\"quote\\\"\\\\slash\"}\n"
|
||
"{\"role\":\"assistant\",\"content\":\"ok\\tready\"}\n";
|
||
|
||
if (file_io) {
|
||
file_io->write(session_path.string().c_str(), session_content);
|
||
}
|
||
|
||
if (session->load(session_path.string().c_str()) == 0) {
|
||
std::cout << "[OK] session->load succeeded\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session->load failed\n";
|
||
dstalk_shutdown();
|
||
return 1;
|
||
}
|
||
|
||
if (session->save(saved_path.string().c_str()) == 0) {
|
||
std::cout << "[OK] session->save succeeded\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session->save failed\n";
|
||
dstalk_shutdown();
|
||
return 1;
|
||
}
|
||
|
||
// 验证保存的内容 / Verify saved content
|
||
if (file_io) {
|
||
char* saved = nullptr;
|
||
if (file_io->read(saved_path.string().c_str(), &saved) == 0 && saved) {
|
||
bool session_ok = std::strcmp(saved, session_content) == 0;
|
||
dstalk_free(saved);
|
||
if (session_ok) {
|
||
std::cout << "[OK] session content matches after save/load\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session content mismatch after save/load\n";
|
||
dstalk_shutdown();
|
||
return 1;
|
||
}
|
||
}
|
||
}
|
||
|
||
// 测试 token 计数 / Test token count
|
||
int tokens = session->token_count();
|
||
std::cout << "[OK] session->token_count: " << tokens << "\n";
|
||
|
||
// 测试 history / Test history
|
||
int count = 0;
|
||
session->history(&count);
|
||
std::cout << "[OK] session->history count: " << count << "\n";
|
||
|
||
// 测试 clear / Test clear
|
||
session->clear();
|
||
session->history(&count);
|
||
if (count == 0) {
|
||
std::cout << "[OK] session->clear succeeded\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] session service not found\n";
|
||
}
|
||
|
||
// 测试服务查询: ai（可能因为没有真实 API key 而失败，但服务应存在）
|
||
// Test service query: ai (may fail without real API key, but service should exist)
|
||
const char* ai_provider = dstalk_config_get("ai.provider");
|
||
if (!ai_provider) ai_provider = "ai.deepseek";
|
||
auto* ai = static_cast<const dstalk_ai_service_t*>(
|
||
dstalk_service_query(ai_provider, 1));
|
||
if (ai) {
|
||
std::cout << "[OK] ai service found\n";
|
||
} else {
|
||
std::cerr << "[WARN] ai service not found\n";
|
||
}
|
||
|
||
// 测试服务查询: config / Test service query: config
|
||
auto* config_svc = static_cast<const dstalk_config_service_t*>(
|
||
dstalk_service_query("config", 1));
|
||
if (config_svc) {
|
||
std::cout << "[OK] config service found\n";
|
||
const char* val = config_svc->get("api.model");
|
||
if (val) {
|
||
std::cout << "[OK] config->get(\"api.model\"): " << val << "\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] config service not found\n";
|
||
}
|
||
|
||
// 测试 dstalk_config_get（主机级配置 API）/ Test dstalk_config_get (host-level config API)
|
||
const char* model = dstalk_config_get("api.model");
|
||
if (model) {
|
||
std::cout << "[OK] dstalk_config_get(\"api.model\"): " << model << "\n";
|
||
}
|
||
|
||
// 测试 dstalk_log / Test dstalk_log
|
||
dstalk_log(DSTALK_LOG_INFO, "Smoke test completed successfully");
|
||
|
||
// ========================================================================
|
||
// 扩展测试块 C2: null-safety / 转义边界 / tools 调用链 / session 健壮性
|
||
// Extended test block C2: null-safety / escape boundaries / tools chain / session robustness
|
||
// ========================================================================
|
||
std::cout << "\n--- Extended Smoke Tests (C2) ---\n";
|
||
|
||
// 提前查询 tools 服务，供后续测试块使用 / Pre-query tools service for subsequent test blocks
|
||
auto* tools = static_cast<const dstalk_tools_service_t*>(
|
||
dstalk_service_query("tools", 1));
|
||
|
||
// ---- 1. Null-safety 测试 ----
|
||
// 对所有服务 API 传 null 参数，验证不崩溃且返回错误
|
||
std::cout << "\n[Block] Null-safety tests\n";
|
||
|
||
if (file_io) {
|
||
char* dummy = nullptr;
|
||
int ret = file_io->read(nullptr, &dummy);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] file_io->read(nullptr, ...) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->read(nullptr, ...) should return error\n";
|
||
}
|
||
|
||
ret = file_io->write(nullptr, "test_content");
|
||
if (ret != 0) {
|
||
std::cout << "[OK] file_io->write(nullptr, ...) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->write(nullptr, ...) should return error\n";
|
||
}
|
||
|
||
// read 的 content 参数也为 null / read's content param also null
|
||
ret = file_io->read("dummy_path", nullptr);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] file_io->read(path, nullptr) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->read(path, nullptr) should return error\n";
|
||
}
|
||
|
||
// write 的 content 参数为 null / write's content param is null
|
||
ret = file_io->write("dummy_path", nullptr);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] file_io->write(path, nullptr) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] file_io->write(path, nullptr) should return error\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] file_io service not available for null-safety tests\n";
|
||
}
|
||
|
||
if (session) {
|
||
session->add(nullptr);
|
||
std::cout << "[OK] session->add(nullptr) did not crash\n";
|
||
|
||
int ret = session->save(nullptr);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] session->save(nullptr) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session->save(nullptr) should return error\n";
|
||
}
|
||
|
||
ret = session->load(nullptr);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] session->load(nullptr) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session->load(nullptr) should return error\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] session service not available for null-safety tests\n";
|
||
}
|
||
|
||
if (tools) {
|
||
char* result = tools->execute(nullptr, nullptr);
|
||
if (result) {
|
||
// 实现返回了错误字符串（如 {"error":"tool name is null"}），未崩溃
|
||
// Implementation returned error string (e.g. {"error":"tool name is null"}), no crash
|
||
std::cout << "[OK] tools->execute(nullptr, nullptr) did not crash"
|
||
<< " (returned: " << result << ")\n";
|
||
dstalk_free(result);
|
||
} else {
|
||
std::cout << "[OK] tools->execute(nullptr, nullptr) returned null without crash\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] tools service not available for null-safety tests\n";
|
||
}
|
||
|
||
if (config_svc) {
|
||
const char* val = config_svc->get(nullptr);
|
||
if (val == nullptr) {
|
||
std::cout << "[OK] config->get(nullptr) returned nullptr\n";
|
||
} else {
|
||
std::cerr << "[FAIL] config->get(nullptr) should return nullptr\n";
|
||
}
|
||
|
||
int ret = config_svc->set(nullptr, nullptr);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] config->set(nullptr, nullptr) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] config->set(nullptr, nullptr) should return error\n";
|
||
}
|
||
|
||
// set 的 value 为 null / set's value is null
|
||
ret = config_svc->set("some.key", nullptr);
|
||
if (ret != 0) {
|
||
std::cout << "[OK] config->set(key, nullptr) returned error (" << ret << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] config->set(key, nullptr) should return error\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] config service not available for null-safety tests\n";
|
||
}
|
||
|
||
// ---- 2. 转义边界测试 ----
|
||
// 写入含特殊字符的内容，读回后验证内容一致
|
||
// ---- Escape boundary tests ----
|
||
// Write content with special chars, verify round-trip integrity
|
||
std::cout << "\n[Block] Escape boundary tests\n";
|
||
|
||
if (file_io) {
|
||
// 构造包含各种特殊字节的内容：
|
||
// - 实际换行符 (0x0A)
|
||
// - 实际双引号 (0x22)
|
||
// - 实际反斜杠 (0x5C)
|
||
// - 实际制表符 (0x09)
|
||
// - 以及字面上的 \n \" \\ \t 转义序列文本
|
||
// Build content with various special bytes:
|
||
// - literal newline (0x0A)
|
||
// - literal double-quote (0x22)
|
||
// - literal backslash (0x5C)
|
||
// - literal tab (0x09)
|
||
// - plus textual \n \" \\ \t escape sequences
|
||
constexpr const char* escape_content =
|
||
"line1\nline2\n"
|
||
"quote=\"yes\"\n"
|
||
"backslash=\\path\n"
|
||
"tab=\there\n"
|
||
"literal-escapes: newline=\\n quote=\\\" backslash=\\\\ tab=\\t\n"
|
||
"endswithbackslash\\\\\n"
|
||
"mixed\\t\\\"quoted\\\"\\\\path\n";
|
||
|
||
const auto escape_path = dir / "escape_test.txt";
|
||
|
||
if (file_io->write(escape_path.string().c_str(), escape_content) == 0) {
|
||
std::cout << "[OK] escape content write succeeded\n";
|
||
|
||
char* read_back = nullptr;
|
||
if (file_io->read(escape_path.string().c_str(), &read_back) == 0 && read_back) {
|
||
bool match = (std::strcmp(read_back, escape_content) == 0);
|
||
if (match) {
|
||
std::cout << "[OK] escape content round-trip matches"
|
||
<< " (length=" << std::strlen(escape_content) << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] escape content round-trip mismatch\n"
|
||
<< " expected length: " << std::strlen(escape_content) << "\n"
|
||
<< " got length: " << std::strlen(read_back) << "\n";
|
||
}
|
||
dstalk_free(read_back);
|
||
} else {
|
||
std::cerr << "[FAIL] escape content read-back failed\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[FAIL] escape content write failed\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] file_io service not available for escape tests\n";
|
||
}
|
||
|
||
// ---- 3. Tools 调用链测试 ----
|
||
// 通过 tools->execute("file_read", ...) 验证内置工具可正确调用 file_io
|
||
// ---- Tools call chain tests ----
|
||
// Verify built-in tools correctly call file_io via tools->execute("file_read", ...)
|
||
std::cout << "\n[Block] Tools call chain tests\n";
|
||
|
||
if (tools && file_io) {
|
||
// 准备测试文件 / Prepare test file
|
||
const auto chain_path = dir / "tool_chain_test.txt";
|
||
constexpr const char* chain_content = "tools-chain-ok\n";
|
||
file_io->write(chain_path.string().c_str(), chain_content);
|
||
|
||
// 用 generic_string() 获取正斜杠路径，避免 JSON 中反斜杠转义问题
|
||
// Use generic_string() for forward-slash paths to avoid backslash escaping in JSON
|
||
std::string generic_path = chain_path.generic_string();
|
||
std::string args_json = "{\"path\":\"" + generic_path + "\"}";
|
||
|
||
char* result = tools->execute("file_read", args_json.c_str());
|
||
if (result) {
|
||
std::cout << "[OK] tools->execute(\"file_read\", ...) returned result\n";
|
||
// 验证返回的 JSON 中包含原始文件内容 / Verify returned JSON contains original file content
|
||
if (std::strstr(result, "tools-chain-ok")) {
|
||
std::cout << "[OK] tools->execute chain correctly called file_io\n";
|
||
} else {
|
||
std::cout << "[WARN] tools->execute result does not contain expected content: "
|
||
<< result << "\n";
|
||
}
|
||
dstalk_free(result);
|
||
} else {
|
||
std::cout << "[WARN] tools->execute(\"file_read\", ...) returned null"
|
||
<< " (tool may not be registered)\n";
|
||
}
|
||
|
||
// 额外测试：查询 tools 返回的工具列表 / Additional test: query tools list
|
||
char* tools_json = tools->get_tools_json();
|
||
if (tools_json) {
|
||
std::cout << "[OK] tools->get_tools_json() returned: " << tools_json << "\n";
|
||
dstalk_free(tools_json);
|
||
} else {
|
||
std::cout << "[WARN] tools->get_tools_json() returned null\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] tools or file_io service not available for chain tests\n";
|
||
}
|
||
|
||
// ---- 4. Session 健壮性测试 ----
|
||
// session->add(nullptr) 后验证 history 不变
|
||
// session->clear 后验证 token_count 为 0
|
||
// ---- Session robustness tests ----
|
||
// Verify history unchanged after session->add(nullptr)
|
||
// Verify token_count == 0 after session->clear
|
||
std::cout << "\n[Block] Session robustness tests\n";
|
||
|
||
if (session) {
|
||
// 记录 add(nullptr) 前的 history 计数 / Record history count before add(nullptr)
|
||
int count_before = 0;
|
||
session->history(&count_before);
|
||
|
||
// 传 null 不应改变 history / Passing null should not change history
|
||
session->add(nullptr);
|
||
|
||
int count_after = 0;
|
||
session->history(&count_after);
|
||
|
||
if (count_before == count_after) {
|
||
std::cout << "[OK] session->add(nullptr) did not change history count"
|
||
<< " (before=" << count_before << ", after=" << count_after << ")\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session->add(nullptr) changed history count: "
|
||
<< count_before << " -> " << count_after << "\n";
|
||
}
|
||
|
||
// clear 后 token_count 应为 0 / token_count should be 0 after clear
|
||
session->clear();
|
||
int tokens = session->token_count();
|
||
if (tokens == 0) {
|
||
std::cout << "[OK] session->token_count() == 0 after clear\n";
|
||
} else {
|
||
std::cerr << "[FAIL] session->token_count() == " << tokens
|
||
<< " after clear, expected 0\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] session service not available for robustness tests\n";
|
||
}
|
||
|
||
// ========================================================================
|
||
// W13.6 回归保护点 R1-R3 (qa-xu 徐磊)
|
||
// 覆盖: W11.7 BUG-2/3/4 + W11.1 Discovery 2/3 + W12.2/W12.3 修复
|
||
// W13.6 regression protections R1-R3 (qa-xu)
|
||
// Covers: W11.7 BUG-2/3/4 + W11.1 Discovery 2/3 + W12.2/W12.3 fixes
|
||
// ========================================================================
|
||
std::cout << "\n--- Regression Tests (R1-R3: W11.7/W12 bug protection) ---\n";
|
||
|
||
// ---- R1: context max_tokens 生效 ----
|
||
// 回归: W11.1 Discovery 3 (g_max_tokens 死变量 — W12.3 已修, W18.1 彻底移除)
|
||
// W11.7 BUG-3 (/context 静默 — W12.3 已修)
|
||
// 验证: trim 能正确裁剪消息数，调用链完整不崩溃
|
||
// ---- R1: context max_tokens takes effect ----
|
||
// Regression: W11.1 Discovery 3 (g_max_tokens dead var — fixed W12.3, removed W18.1)
|
||
// W11.7 BUG-3 (/context silent — fixed W12.3)
|
||
// Verify: trim reduces message count correctly, full call chain without crash
|
||
{
|
||
auto* ctx = static_cast<const dstalk_context_service_t*>(
|
||
dstalk_service_query("context", 1));
|
||
if (ctx) {
|
||
std::cout << "[OK] R1: context service found\n";
|
||
|
||
// 构造 5 条消息，每条 ~50 字符 / ~15 token，总计 ~75 token > 50 max
|
||
// Build 5 messages, each ~50 chars / ~15 tokens, total ~75 tokens > 50 max
|
||
dstalk_message_t msgs[5];
|
||
msgs[0] = {"user", "Hello this is message one with enough text to count tokens", nullptr, nullptr};
|
||
msgs[1] = {"assistant", "Message two also has sufficient length for token counting", nullptr, nullptr};
|
||
msgs[2] = {"user", "Message three continues the conversation with more text", nullptr, nullptr};
|
||
msgs[3] = {"assistant", "Message four adds further content beyond the max budget", nullptr, nullptr};
|
||
msgs[4] = {"user", "Message five with extra text to ensure we overflow limit", nullptr, nullptr};
|
||
|
||
dstalk_message_t* out = nullptr;
|
||
int out_count = 0;
|
||
int ret = ctx->trim(msgs, 5, &out, &out_count, 50);
|
||
|
||
REGCHECK(ret >= 0, "R1: context->trim returned non-negative (no crash)");
|
||
if (out) {
|
||
REGCHECK(out_count < 5, "R1: trim reduced message count (out < in=5)");
|
||
std::cout << "[OK] R1: trim output count = " << out_count
|
||
<< " (in=5, max_tokens=50)\n";
|
||
dstalk_free(out);
|
||
} else if (ret >= 0) {
|
||
// 首条消息即超 max_tokens 时 trim 可能返回空，这也是合法路径
|
||
// When first message exceeds max_tokens, trim may return empty; also valid
|
||
std::cout << "[WARN] R1: trim returned null output (single msg exceeds max?)\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] R1: context service not found, skipping\n";
|
||
}
|
||
}
|
||
|
||
// ---- R2: config 双 store 一致性 ----
|
||
// 回归: W11.2 Discovery 2 (双 ConfigStore 数据孤岛 — W12.2 已修)
|
||
// W11.2 Discovery 3 (c_str() 悬垂 — W12.2 已修)
|
||
// 验证: dstalk_config_set 写入后，dstalk_config_get 和 config_service->get 返回一致值
|
||
// ---- R2: config dual-store consistency ----
|
||
// Regression: W11.2 Discovery 2 (dual ConfigStore islands — fixed W12.2)
|
||
// W11.2 Discovery 3 (c_str() dangling — fixed W12.2)
|
||
// Verify: after dstalk_config_set write, dstalk_config_get and config_service->get return same value
|
||
{
|
||
constexpr const char* k = "__regr_w13_6_dual";
|
||
constexpr const char* v = "dual_ok_42";
|
||
|
||
// 通过 host API 写入 / Write via host API
|
||
int set_ret = dstalk_config_set(k, v);
|
||
REGCHECK(set_ret == 0, "R2: dstalk_config_set returned 0");
|
||
|
||
// 通过 host API 读回 / Read back via host API
|
||
const char* host_val = dstalk_config_get(k);
|
||
REGCHECK(host_val && std::strcmp(host_val, v) == 0,
|
||
"R2: dstalk_config_get matches written value");
|
||
|
||
// 通过 plugin config 服务读回 — 验证双 store 整合后数据可见性一致
|
||
// 注: W12.2 双 store 整合尚未部署，跨 store 可见性当前为已知 gap；
|
||
// 本检查用 WARN 记录现状，待 W12.2 fix 落地后改为 REGCHECK
|
||
// Read back via plugin config service — verify visibility after dual-store merge
|
||
// Note: W12.2 dual-store merge not yet deployed; cross-store visibility is a known gap;
|
||
// this check uses WARN to record status, upgrade to REGCHECK after W12.2 lands
|
||
auto* cfg_svc = static_cast<const dstalk_config_service_t*>(
|
||
dstalk_service_query("config", 1));
|
||
if (cfg_svc) {
|
||
const char* plugin_val = cfg_svc->get(k);
|
||
if (plugin_val && std::strcmp(plugin_val, v) == 0) {
|
||
std::cout << "[OK] R2: config_service->get matches dstalk_config_set value\n";
|
||
} else {
|
||
std::cout << "[WARN] R2: cross-store visibility gap: "
|
||
<< "config_service->get returned '"
|
||
<< (plugin_val ? plugin_val : "(null)")
|
||
<< "' for host-set key '" << k
|
||
<< "' (W12.2 dual-store merge pending)\n";
|
||
}
|
||
} else {
|
||
std::cerr << "[WARN] R2: config service not found, partial skip\n";
|
||
}
|
||
|
||
// 清理测试 key / Clean up test key
|
||
dstalk_config_set(k, "");
|
||
}
|
||
|
||
// ---- R3: HTTP / AI 服务错误路径不崩溃 ----
|
||
// 回归: W12.1 removed TLS/http_client 代码 (移除重写的网络层)
|
||
// W11.7 BUG-4 (/file write 落空) 同类的错误路径静默问题
|
||
// 验证: http post_json 到不可达目标返回错误而不崩溃;
|
||
// 若 http 服务不可用，回退测 ai 服务错误路径
|
||
// ---- R3: HTTP / AI service error paths do not crash ----
|
||
// Regression: W12.1 removed TLS/http_client code (removed rewritten network layer)
|
||
// W11.7 BUG-4 (/file write miss) similar error-path silent issues
|
||
// Verify: http post_json to unreachable target returns error without crash;
|
||
// fall back to ai service error path if http unavailable
|
||
{
|
||
auto* http = static_cast<const dstalk_http_service_t*>(
|
||
dstalk_service_query("http", 1));
|
||
if (http) {
|
||
std::cout << "[OK] R3: http service found\n";
|
||
// 向 127.0.0.1:1 发请求 — 端口 1 在 Windows 上几乎肯定无服务监听
|
||
// 连接拒绝应立即返回错误而非崩溃
|
||
// Send request to 127.0.0.1:1 — port 1 on Windows almost certainly has no listener
|
||
// Connection refused should return error immediately, not crash
|
||
char* body = nullptr;
|
||
int status = 0;
|
||
int ret = http->post_json("127.0.0.1", "1", "/",
|
||
"{}", "{}", &body, &status);
|
||
REGCHECK(ret != 0, "R3: http->post_json to closed port returned error (no crash)");
|
||
if (body) {
|
||
std::cout << "[OK] R3: http error response body present (status=" << status << ")\n";
|
||
dstalk_free(body);
|
||
} else {
|
||
std::cout << "[OK] R3: http error path, no response body (connection refused)\n";
|
||
}
|
||
} else {
|
||
// 回退：测 AI 服务 (ai.deepseek) 错误路径
|
||
// Fallback: test AI service (ai.deepseek) error path
|
||
auto* ai_svc = static_cast<const dstalk_ai_service_t*>(
|
||
dstalk_service_query("ai.deepseek", 1));
|
||
if (ai_svc) {
|
||
std::cout << "[OK] R3: ai.deepseek service found (http fallback)\n";
|
||
dstalk_message_t msg = {"user", "hi", nullptr, nullptr};
|
||
dstalk_chat_result_t r = ai_svc->chat(&msg, 1, "", nullptr);
|
||
// api_key="test-key" 为无效 key，应返回 error result 而非崩溃
|
||
// api_key="test-key" is invalid, should return error result, not crash
|
||
REGCHECK(r.ok == 0 || r.error != nullptr,
|
||
"R3: ai->chat with invalid key returned error result (no crash)");
|
||
if (r.content) dstalk_free((void*)r.content);
|
||
if (r.error) dstalk_free((void*)r.error);
|
||
if (r.tool_calls_json) dstalk_free((void*)r.tool_calls_json);
|
||
} else {
|
||
std::cerr << "[WARN] R3: neither http nor ai service found, skipping\n";
|
||
}
|
||
}
|
||
}
|
||
|
||
// ========================================================================
|
||
// W21.5 Tool Calls 边界测试 (qa-xu 徐磊)
|
||
// 覆盖: null tool_calls_json / 空数组 "[]" / 有效 tool_calls mock 验证
|
||
// W21.5 Tool Calls boundary tests (qa-xu)
|
||
// Covers: null tool_calls_json / empty array "[]" / valid tool_calls mock verification
|
||
// ========================================================================
|
||
std::cout << "\n--- Tool Calls Boundary Tests (W21.5) ---\n";
|
||
|
||
if (tools && session) {
|
||
// ---- W21.5-1: null tool_calls_json → 正常处理（不崩溃）----
|
||
// ---- W21.5-1: null tool_calls_json → handle normally (no crash) ----
|
||
{
|
||
int before = 0;
|
||
session->history(&before);
|
||
|
||
dstalk_message_t msg_null_tc = {
|
||
"assistant", "content with null tool_calls_json", nullptr, nullptr
|
||
};
|
||
session->add(&msg_null_tc);
|
||
|
||
int after = 0;
|
||
const dstalk_message_t* hist = session->history(&after);
|
||
REGCHECK(after == before + 1,
|
||
"W21.5-1a: null tool_calls_json add — count +1 (no crash)");
|
||
if (hist && after > 0) {
|
||
REGCHECK(hist[after - 1].tool_calls_json == nullptr,
|
||
"W21.5-1b: null tool_calls_json — retrieved field is nullptr");
|
||
}
|
||
}
|
||
|
||
// ---- W21.5-2: 空 JSON 数组 "[]" → 正常处理（不崩溃）----
|
||
// ---- W21.5-2: empty JSON array "[]" → handle normally (no crash) ----
|
||
{
|
||
int before = 0;
|
||
session->history(&before);
|
||
|
||
dstalk_message_t msg_empty = {
|
||
"assistant", "content with empty [] tool_calls", nullptr, "[]"
|
||
};
|
||
session->add(&msg_empty);
|
||
|
||
int after = 0;
|
||
const dstalk_message_t* hist = session->history(&after);
|
||
REGCHECK(after == before + 1,
|
||
"W21.5-2a: empty [] tool_calls_json add — count +1 (no crash)");
|
||
if (hist && after > 0) {
|
||
REGCHECK(hist[after - 1].tool_calls_json != nullptr &&
|
||
std::strcmp(hist[after - 1].tool_calls_json, "[]") == 0,
|
||
"W21.5-2b: empty [] tool_calls_json — retrieved as \"[]\"");
|
||
}
|
||
}
|
||
|
||
// ---- W21.5-3: 有效 tool_calls JSON → 验证 execute 被调用 (mock) ----
|
||
// ---- W21.5-3: valid tool_calls JSON → verify execute is called (mock) ----
|
||
{
|
||
g_mock_tool_called = 0;
|
||
int reg = tools->register_tool(
|
||
"__w21_5_mock",
|
||
"W21.5 mock tool for boundary test",
|
||
"{}",
|
||
mock_tool_handler
|
||
);
|
||
REGCHECK(reg == 0, "W21.5-3a: register mock tool ok");
|
||
|
||
char* result = tools->execute("__w21_5_mock", "{}");
|
||
REGCHECK(g_mock_tool_called == 1,
|
||
"W21.5-3b: mock tool handler was called (execute works)");
|
||
if (result) {
|
||
REGCHECK(std::strstr(result, "mock_result") != nullptr,
|
||
"W21.5-3c: execute returned mock result json");
|
||
dstalk_free(result);
|
||
}
|
||
|
||
tools->unregister_tool("__w21_5_mock");
|
||
|
||
// 验证已注销的工具返回 error 而非崩溃
|
||
// Verify unregistered tool returns error, not crash
|
||
char* err_result = tools->execute("__w21_5_mock", "{}");
|
||
REGCHECK(err_result && std::strstr(err_result, "error") != nullptr,
|
||
"W21.5-3d: unregistered tool returns error (not crash)");
|
||
if (err_result) dstalk_free(err_result);
|
||
}
|
||
|
||
// ---- W21.5-4: save/load 往返保留 tool_calls_json ----
|
||
// ---- W21.5-4: save/load round-trip preserves tool_calls_json ----
|
||
if (file_io) {
|
||
const auto rtt_path = dir / "w21_5_tc_rtt.jsonl";
|
||
int ret = session->save(rtt_path.string().c_str());
|
||
REGCHECK(ret == 0, "W21.5-4a: session save with tool_calls msgs ok");
|
||
|
||
session->clear();
|
||
ret = session->load(rtt_path.string().c_str());
|
||
REGCHECK(ret == 0, "W21.5-4b: session load round-trip ok");
|
||
|
||
int count = 0;
|
||
session->history(&count);
|
||
REGCHECK(count > 0, "W21.5-4c: history non-empty after load round-trip");
|
||
}
|
||
|
||
session->clear();
|
||
} else {
|
||
std::cerr << "[WARN] W21.5: tools or session service not available\n";
|
||
}
|
||
|
||
// 清理 / Cleanup
|
||
dstalk_shutdown();
|
||
std::cout << "[OK] dstalk_shutdown succeeded\n";
|
||
|
||
// ========================================================================
|
||
// W13.6 回归保护点 R4 (qa-xu 徐磊)
|
||
// W13.6 regression protection R4 (qa-xu)
|
||
// ========================================================================
|
||
|
||
// ---- R4: 重复 init / shutdown 生命周期 ----
|
||
// 回归: W9.8 initialize_all 容错 (插件生命周期健壮性)
|
||
// W11.7 BUG-1 [CRITICAL] build/bin/ 损坏副本 (stale state 残留)
|
||
// 验证: 多次 dstalk_init/dstalk_shutdown 循环不崩溃，每次 reload 正常
|
||
// ---- R4: repeat init/shutdown lifecycle ----
|
||
// Regression: W9.8 initialize_all fault tolerance (plugin lifecycle robustness)
|
||
// W11.7 BUG-1 [CRITICAL] build/bin/ corrupt copy (stale state residue)
|
||
// Verify: multiple dstalk_init/dstalk_shutdown cycles without crash, each reload ok
|
||
{
|
||
std::cout << "\n[Block] R4: Repeat init/shutdown lifecycle\n";
|
||
constexpr int cycles = 3;
|
||
for (int i = 0; i < cycles; i++) {
|
||
// 每轮重写配置（模拟独立启动）/ Rewrite config each cycle (simulate independent start)
|
||
{
|
||
std::ofstream c(config_path);
|
||
c << "[api]\n"
|
||
<< "provider = \"deepseek\"\n"
|
||
<< "base_url = \"https://api.deepseek.com/v1\"\n"
|
||
<< "api_key = \"test-key\"\n"
|
||
<< "model = \"deepseek-v4-pro\"\n";
|
||
}
|
||
|
||
std::cout << "[R4] cycle " << (i + 1) << "/" << cycles << "\n";
|
||
|
||
int r = dstalk_init(config_path.string().c_str());
|
||
REGCHECK(r == 0, "R4: dstalk_init returned 0");
|
||
if (r != 0) {
|
||
std::cerr << "[WARN] R4: cycle " << (i + 1)
|
||
<< " init failed, stopping remaining cycles\n";
|
||
break;
|
||
}
|
||
|
||
// 快速验证服务可用 / Quick verify service is available
|
||
void* q = dstalk_service_query("config", 1);
|
||
REGCHECK(q != nullptr, "R4: service query ok after init");
|
||
|
||
dstalk_shutdown();
|
||
std::cout << "[OK] R4: cycle " << (i + 1) << "/" << cycles
|
||
<< " shutdown ok\n";
|
||
}
|
||
}
|
||
|
||
// ---- 最终结果 / Final result ----
|
||
std::cout << "\n";
|
||
if (g_regression_failures == 0) {
|
||
std::cout << "=== All smoke tests passed ===\n";
|
||
return 0;
|
||
} else {
|
||
std::cerr << "=== " << g_regression_failures
|
||
<< " regression test(s) FAILED ===\n";
|
||
return 1;
|
||
}
|
||
}
|