Files
dstalk/tests/deepseek_plugin_test.cpp
XiuChengWu b2b381b9b3
Some checks failed
CI / Determine matrix (push) Has been cancelled
CI / ${{ matrix.os }} / ${{ matrix.build_type }} (push) Has been cancelled
CI / Sanitizer (ASan+UBSan) / ubuntu-24.04 (push) Has been cancelled
W21: anthropic Stream+Tools + --prompt batch + sanitizer fix + plugin unit tests (W21.1-W21.6)
- W21.1: ci-sanitize preset 独立 Linux-clang + ci-threadsan (TSan)
- W21.2: anthropic tool_use content_block 解析 + configure 缓存 tools_json
- W21.3: --prompt 非交互批处理模式
- W21.4: session auto-save 失败告警 + 当前目录 fallback
- W21.5: smoke 补 tool_calls 边界用例 4 块 12 断言
- W21.6: anthropic 11 块 78 CHECK + deepseek 12 块 78 CHECK

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-27 20:40:58 +08:00

670 lines
26 KiB
C++

// ============================================================================
// deepseek_plugin_test.cpp — DeepSeek AI 插件单元测试
// W21.6 (qa-wang): 覆盖 SSE 解析 / [DONE] 匹配 / JSON 请求构建 / tool_calls
// 通过 #include plugin source 访问 file-scope static 函数
// ============================================================================
#define BOOST_JSON_HEADER_ONLY
#define BOOST_ALL_NO_LIB
#include "../plugins/deepseek/src/deepseek_plugin.cpp"
#include <cstring>
#include <iostream>
#include <string>
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)
// Test helper: populate g_cfg for build functions
static void setup_config() {
g_cfg.provider = "deepseek";
g_cfg.base_url = "https://api.deepseek.com/v1";
g_cfg.api_key = "sk-ds-test-key-67890";
g_cfg.model = "deepseek-v4-pro";
g_cfg.max_tokens = 4096;
g_cfg.temperature = 0.7;
}
int main()
{
// ================================================================
// Test Block 1: parse_sse_line — invalid/malformed inputs
// ================================================================
std::cout << "\n--- Block 1: parse_sse_line invalid/malformed ---\n";
{
std::string token;
bool ret = parse_sse_line("", token, nullptr);
CHECK(!ret, "T1.1: empty line returns false");
}
{
std::string token;
bool ret = parse_sse_line("not a data line", token, nullptr);
CHECK(!ret, "T1.2: non-'data:' prefix returns false");
}
{
std::string token;
bool ret = parse_sse_line("event: message", token, nullptr);
CHECK(!ret, "T1.3: 'event:' line returns false");
}
{
// "data:" without space — rfind("data: ", 0) should fail
std::string token;
bool ret = parse_sse_line("data:{\"x\":1}", token, nullptr);
CHECK(!ret, "T1.4: 'data:' without trailing space returns false (rfind mismatch)");
}
{
// "data: " followed by invalid JSON
std::string token;
bool ret = parse_sse_line("data: not valid json!!!", token, nullptr);
CHECK(!ret, "T1.5: 'data: ' + invalid JSON returns false (no crash)");
}
{
// "data: " followed by binary garbage
std::string token;
bool ret = parse_sse_line("data: \x00\x01\xFF\xFE", token, nullptr);
CHECK(!ret, "T1.6: 'data: ' + binary garbage returns false (no crash)");
}
{
// Empty data after "data: "
std::string token;
bool ret = parse_sse_line("data: ", token, nullptr);
CHECK(!ret, "T1.7: 'data: ' with empty payload returns false");
}
// ================================================================
// Test Block 2: parse_sse_line — [DONE] sentinel
// ================================================================
std::cout << "\n--- Block 2: parse_sse_line [DONE] sentinel ---\n";
{
std::string token = "SHOULD_BE_CLEARED";
bool ret = parse_sse_line("data: [DONE]", token, nullptr);
CHECK(ret, "T2.1: 'data: [DONE]' returns true (stream end)");
CHECK(token.empty(), "T2.2: [DONE] clears token");
}
{
// [DONE] with leading whitespace
std::string token;
bool ret = parse_sse_line("data: [DONE]", token, nullptr);
CHECK(ret, "T2.3: 'data: [DONE]' (leading spaces) returns true");
CHECK(token.empty(), "T2.4: whitespace-trimmed [DONE] clears token");
}
{
// [DONE] with trailing whitespace
std::string token;
bool ret = parse_sse_line("data: [DONE] ", token, nullptr);
CHECK(ret, "T2.5: 'data: [DONE] ' (trailing spaces) returns true");
CHECK(token.empty(), "T2.6: trailing-whitespace [DONE] clears token");
}
{
// [DONE] with tabs and newlines around it
std::string token;
bool ret = parse_sse_line("data: \t [DONE] \t\r\n", token, nullptr);
CHECK(ret, "T2.7: '[DONE]' with mixed whitespace returns true");
CHECK(token.empty(), "T2.8: mixed-whitespace [DONE] clears token");
}
{
// [DONE] without spaces — exact match
std::string token;
bool ret = parse_sse_line("data: [DONE]", token, nullptr);
CHECK(ret, "T2.9: '[DONE]' exact match returns true");
}
{
// "[done]" lowercase — should NOT match (case-sensitive)
std::string token;
bool ret = parse_sse_line("data: [done]", token, nullptr);
CHECK(!ret, "T2.10: '[done]' lowercase NOT treated as DONE (case-sensitive)");
}
{
// "[DONE" without closing bracket
std::string token;
bool ret = parse_sse_line("data: [DONE", token, nullptr);
CHECK(!ret, "T2.11: '[DONE' (no closing bracket) not treated as DONE");
}
// ================================================================
// Test Block 3: parse_sse_line — content delta
// ================================================================
std::cout << "\n--- Block 3: parse_sse_line content delta ---\n";
{
std::string token;
const char* json =
"data: {\"choices\":[{\"delta\":{\"content\":\"Hello\"},"
"\"index\":0}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(ret, "T3.1: delta with content 'Hello' returns true");
CHECK(token == "Hello", "T3.2: token equals 'Hello'");
}
{
std::string token;
const char* json =
"data: {\"choices\":[{\"delta\":{\"content\":\"\"},\"index\":0}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(ret, "T3.3: delta with empty content returns true");
CHECK(token.empty(), "T3.4: empty content token is empty");
}
{
// Delta with no content field
std::string token;
const char* json =
"data: {\"choices\":[{\"delta\":{},\"index\":0}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(!ret, "T3.5: delta without 'content' field returns false");
}
{
// Empty choices array
std::string token;
const char* json =
"data: {\"choices\":[]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(!ret, "T3.6: empty choices array returns false");
}
{
// Single character token (typical streaming)
std::string token;
const char* json =
"data: {\"choices\":[{\"delta\":{\"content\":\"H\"},\"index\":0}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(ret, "T3.7: single-char delta returns true");
CHECK(token == "H", "T3.8: single-char token correct");
}
{
// Multi-byte UTF-8 content (emoji) in delta
std::string token;
const char* json =
"data: {\"choices\":[{\"delta\":{\"content\":\"\\uD83D\\uDE00\"},"
"\"index\":0}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(ret, "T3.9: emoji delta returns true");
// U+1F600 in UTF-8: F0 9F 98 80
std::string emoji = "\xF0\x9F\x98\x80";
CHECK(token == emoji, "T3.10: emoji token decoded correctly (U+1F600)");
}
{
// Malformed JSON structure — no "delta" key
std::string token;
const char* json =
"data: {\"choices\":[{\"no_delta\":{},\"index\":0}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(!ret, "T3.11: choice without 'delta' key returns false");
}
{
// Realistic DeepSeek streaming chunk (with finish_reason)
std::string token;
const char* json =
"data: {\"id\":\"chatcmpl-xxx\","
"\"object\":\"chat.completion.chunk\","
"\"created\":1712345678,"
"\"model\":\"deepseek-v4-pro\","
"\"choices\":[{\"index\":0,"
"\"delta\":{\"content\":\" World\"},"
"\"finish_reason\":null}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(ret, "T3.12: realistic DeepSeek chunk returns true");
CHECK(token == " World", "T3.13: realistic chunk token correct");
}
// ================================================================
// Test Block 4: parse_sse_line — tool_calls delta
// ================================================================
std::cout << "\n--- Block 4: parse_sse_line tool_calls delta ---\n";
{
// tool_calls chunk with id + function name (first chunk)
StreamContext ctx = {};
std::string token;
const char* json =
"data: {\"choices\":[{\"index\":0,"
"\"delta\":{\"tool_calls\":[{\"index\":0,"
"\"id\":\"call_abc123\","
"\"type\":\"function\","
"\"function\":{\"name\":\"get_weather\",\"arguments\":\"\"}}]}}]}";
bool ret = parse_sse_line(json, token, &ctx);
CHECK(!ret, "T4.1: tool_calls first chunk returns false (no content token)");
CHECK(ctx.tool_calls.size() >= 1, "T4.2: tool_calls accumulated in ctx");
if (ctx.tool_calls.size() >= 1) {
CHECK(ctx.tool_calls[0].index == 0, "T4.3: tool_call index=0");
CHECK(ctx.tool_calls[0].id == "call_abc123", "T4.4: tool_call id stored");
CHECK(ctx.tool_calls[0].name == "get_weather", "T4.5: tool_call name stored");
}
}
{
// tool_calls arguments chunk (second chunk, same index)
StreamContext ctx;
// First, set up the initial state
ctx.tool_calls.push_back({0, "call_abc123", "get_weather", ""});
std::string token;
const char* json =
"data: {\"choices\":[{\"index\":0,"
"\"delta\":{\"tool_calls\":[{\"index\":0,"
"\"function\":{\"arguments\":\"{\\\"city\\\":\\\"\"}}]}}]}";
bool ret = parse_sse_line(json, token, &ctx);
CHECK(!ret, "T4.6: tool_calls arguments chunk returns false");
if (ctx.tool_calls.size() >= 1) {
CHECK(ctx.tool_calls[0].arguments == "{\"city\":\"",
"T4.7: arguments accumulated (first fragment)");
}
}
{
// tool_calls final arguments chunk
StreamContext ctx;
ctx.tool_calls.push_back({0, "call_abc123", "get_weather", "{\"city\":\""});
std::string token;
const char* json =
"data: {\"choices\":[{\"index\":0,"
"\"delta\":{\"tool_calls\":[{\"index\":0,"
"\"function\":{\"arguments\":\"Beijing\\\"}\"}}]}}]}";
bool ret = parse_sse_line(json, token, &ctx);
CHECK(!ret, "T4.8: tool_calls final chunk returns false");
if (ctx.tool_calls.size() >= 1) {
CHECK(ctx.tool_calls[0].arguments == "{\"city\":\"Beijing\"}",
"T4.9: full arguments accumulated correctly");
}
}
{
// tool_calls with null ctx — should skip tool_calls processing
std::string token;
const char* json =
"data: {\"choices\":[{\"index\":0,"
"\"delta\":{\"tool_calls\":[{\"index\":0,"
"\"function\":{\"arguments\":\"{}\"}}]}}]}";
bool ret = parse_sse_line(json, token, nullptr);
CHECK(!ret, "T4.10: tool_calls with null ctx returns false (no crash)");
}
{
// Multiple tool_calls in single chunk (unusual but valid)
StreamContext ctx;
std::string token;
const char* json =
"data: {\"choices\":[{\"index\":0,"
"\"delta\":{\"tool_calls\":["
"{\"index\":0,\"function\":{\"arguments\":\"a\"}},"
"{\"index\":1,\"function\":{\"arguments\":\"b\"}}"
"]}}]}";
bool ret = parse_sse_line(json, token, &ctx);
CHECK(!ret, "T4.11: multi-tool_call chunk returns false");
CHECK(ctx.tool_calls.size() >= 2, "T4.12: both tool_calls accumulated");
if (ctx.tool_calls.size() >= 2) {
CHECK(ctx.tool_calls[0].index == 0, "T4.13: first tool_call index=0");
CHECK(ctx.tool_calls[1].index == 1, "T4.14: second tool_call index=1");
}
}
// ================================================================
// Test Block 5: build_request_json — basic cases
// ================================================================
setup_config();
std::cout << "\n--- Block 5: build_request_json basic ---\n";
{
std::string json = build_request_json(
nullptr, 0, "Hello", "", false);
CHECK(!json.empty(), "T5.1: non-empty JSON produced");
CHECK(json.find("\"messages\"") != std::string::npos,
"T5.2: contains 'messages' array");
CHECK(json.find("\"user\"") != std::string::npos,
"T5.3: contains 'user' role");
CHECK(json.find("\"Hello\"") != std::string::npos,
"T5.4: contains user input");
CHECK(json.find("\"stream\":false") != std::string::npos,
"T5.5: stream=false present");
CHECK(json.find("\"model\":\"deepseek-v4-pro\"")
!= std::string::npos,
"T5.6: model field present");
CHECK(json.find("\"max_tokens\":4096") != std::string::npos,
"T5.7: max_tokens present");
CHECK(json.find("\"temperature\"") != std::string::npos,
"T5.8: temperature field present (always included in DeepSeek)");
}
{
// With user+assistant history
dstalk_message_t msgs[2] = {
{"user", "What is 2+2?", nullptr, nullptr},
{"assistant", "It is 4.", nullptr, nullptr}
};
std::string json = build_request_json(
msgs, 2, "Thanks!", "", false);
CHECK(json.find("\"role\":\"user\"") != std::string::npos,
"T5.9: user role present");
CHECK(json.find("\"role\":\"assistant\"") != std::string::npos,
"T5.10: assistant role present");
CHECK(json.find("Thanks!") != std::string::npos,
"T5.11: current user input present");
}
{
// Stream=true
std::string json = build_request_json(
nullptr, 0, "Hi", "", true);
CHECK(json.find("\"stream\":true") != std::string::npos,
"T5.12: stream=true present");
}
{
// Empty user input — no user message appended
std::string json = build_request_json(
nullptr, 0, "", "", false);
CHECK(!json.empty(), "T5.13: empty user input produces valid JSON");
// DeepSeek's build_request_json checks `if (!user_input.empty())` before adding
// So there should be no user message for empty input
CHECK(json.find("\"role\":\"user\"") == std::string::npos,
"T5.14: empty user input NOT added to messages (DeepSeek guard)");
}
// ================================================================
// Test Block 6: build_request_json — tools / edge cases
// ================================================================
std::cout << "\n--- Block 6: build_request_json tools / edges ---\n";
{
// With tools_json
std::string tools = "[{\"type\":\"function\","
"\"function\":{\"name\":\"get_weather\","
"\"description\":\"Get current weather\","
"\"parameters\":{\"type\":\"object\","
"\"properties\":{\"city\":{\"type\":\"string\"}},"
"\"required\":[\"city\"]}}}]";
std::string json = build_request_json(
nullptr, 0, "Weather in Beijing?", tools, false);
CHECK(json.find("\"tools\"") != std::string::npos,
"T6.1: 'tools' field present when tools_json provided");
CHECK(json.find("get_weather") != std::string::npos,
"T6.2: tool name present in serialized JSON");
}
{
// Empty tools_json — no tools field
std::string json = build_request_json(
nullptr, 0, "Hello", "", false);
CHECK(json.find("\"tools\"") == std::string::npos,
"T6.3: no 'tools' field when tools_json is empty");
}
{
// Malformed tools_json — build_request_json calls json::parse()
// without try/catch, so it will throw std::exception.
// This test verifies that the exception is thrown (rather than crashing).
bool threw = false;
try {
build_request_json(nullptr, 0, "Hello", "NOT JSON", false);
} catch (const std::exception&) {
threw = true;
} catch (...) {
threw = true;
}
CHECK(threw, "T6.4: malformed tools_json throws (expected, not a crash)");
}
{
// History with null role
dstalk_message_t msgs[1] = {
{nullptr, "some content", nullptr, nullptr}
};
std::string json = build_request_json(msgs, 1, "Hi", "", false);
CHECK(!json.empty(), "T6.5: null role produces valid JSON (no crash)");
}
{
// History with null content
dstalk_message_t msgs[1] = {
{"user", nullptr, nullptr, nullptr}
};
std::string json = build_request_json(msgs, 1, "Hi", "", false);
CHECK(!json.empty(), "T6.6: null content produces valid JSON (no crash)");
}
{
// Very long message
std::string long_input(5000, 'A');
std::string json = build_request_json(
nullptr, 0, long_input, "", false);
CHECK(!json.empty(), "T6.7: 5000-char input produces valid JSON");
CHECK(json.length() > 5000, "T6.8: JSON longer than input (wraps content)");
}
// ================================================================
// Test Block 7: build_headers_json
// ================================================================
std::cout << "\n--- Block 7: build_headers_json ---\n";
{
std::string headers = build_headers_json("sk-test-123");
CHECK(headers.find("Authorization") != std::string::npos,
"T7.1: contains Authorization header");
CHECK(headers.find("Bearer sk-test-123") != std::string::npos,
"T7.2: contains 'Bearer sk-test-123'");
}
{
// Empty API key
std::string headers = build_headers_json("");
CHECK(headers.find("Authorization") != std::string::npos,
"T7.3: Authorization header present with empty key");
CHECK(headers.find("Bearer ") != std::string::npos,
"T7.4: 'Bearer ' prefix present even with empty key");
}
// ================================================================
// Test Block 8: extract_host_port (same logic as anthropic)
// ================================================================
std::cout << "\n--- Block 8: extract_host_port ---\n";
{
std::string scheme, host, port, target;
bool ret = extract_host_port(
"https://api.deepseek.com/v1/chat/completions",
scheme, host, port, target);
CHECK(ret, "T8.1: valid HTTPS URL returns true");
CHECK(scheme == "https", "T8.2: scheme is 'https'");
CHECK(host == "api.deepseek.com", "T8.3: host extracted");
CHECK(port == "443", "T8.4: default HTTPS port 443");
CHECK(target == "/v1/chat/completions", "T8.5: target path extracted");
}
{
std::string scheme, host, port, target;
bool ret = extract_host_port(
"http://localhost:11434/api/generate",
scheme, host, port, target);
CHECK(ret, "T8.6: HTTP URL with port returns true");
CHECK(scheme == "http", "T8.7: scheme is 'http'");
CHECK(host == "localhost", "T8.8: host is 'localhost'");
CHECK(port == "11434", "T8.9: explicit port 11434");
}
{
std::string scheme, host, port, target;
bool ret = extract_host_port(
"no-scheme-url.com",
scheme, host, port, target);
CHECK(!ret, "T8.10: URL without scheme returns false");
}
{
std::string scheme, host, port, target;
bool ret = extract_host_port(
"https://api.deepseek.com",
scheme, host, port, target);
CHECK(ret, "T8.11: URL without path returns true");
CHECK(target == "/", "T8.12: target defaults to '/'");
}
// ================================================================
// Test Block 9: secure_zero
// ================================================================
std::cout << "\n--- Block 9: secure_zero ---\n";
{
char buf[16];
memset(buf, 0xFF, sizeof(buf));
secure_zero(buf, sizeof(buf));
bool all_zero = true;
for (int i = 0; i < 16; ++i) {
if (buf[i] != 0) { all_zero = false; break; }
}
CHECK(all_zero, "T9.1: secure_zero zeros entire buffer");
}
{
secure_zero(nullptr, 0);
CHECK(true, "T9.2: secure_zero(nullptr, 0) does not crash");
}
// ================================================================
// Test Block 10: append_history
// ================================================================
std::cout << "\n--- Block 10: append_history ---\n";
{
json::array msgs;
dstalk_message_t m = {"user", "Hello", nullptr, nullptr};
append_history(msgs, &m, 1);
CHECK(msgs.size() == 1, "T10.1: one message appended");
CHECK(msgs[0].as_object()["role"].as_string() == "user",
"T10.2: role preserved");
CHECK(msgs[0].as_object()["content"].as_string() == "Hello",
"T10.3: content preserved");
}
{
// Tool message (should include tool_call_id)
json::array msgs;
dstalk_message_t m = {"tool", "result data", "call_xyz", nullptr};
append_history(msgs, &m, 1);
CHECK(msgs.size() == 1, "T10.4: tool message appended");
auto obj = msgs[0].as_object();
CHECK(obj["role"].as_string() == "tool", "T10.5: tool role preserved");
CHECK(obj["tool_call_id"].as_string() == "call_xyz",
"T10.6: tool_call_id preserved");
CHECK(obj["content"].as_string() == "result data",
"T10.7: tool message content preserved");
}
{
// Assistant with tool_calls_json
json::array msgs;
const char* tc_json = "[{\"id\":\"call_1\",\"type\":\"function\","
"\"function\":{\"name\":\"get_weather\",\"arguments\":\"{}\"}}]";
dstalk_message_t m = {"assistant", "Let me check", nullptr, tc_json};
append_history(msgs, &m, 1);
CHECK(msgs.size() == 1, "T10.8: assistant with tool_calls appended");
auto obj = msgs[0].as_object();
CHECK(obj["role"].as_string() == "assistant", "T10.9: assistant role");
CHECK(obj["content"].as_string() == "Let me check", "T10.10: content");
CHECK(obj.if_contains("tool_calls") != nullptr, "T10.11: tool_calls present");
}
{
// Empty history (0 messages)
json::array msgs;
append_history(msgs, nullptr, 0);
CHECK(msgs.size() == 0, "T10.12: empty history produces empty array");
}
{
// Multiple messages
json::array msgs;
dstalk_message_t ms[2] = {
{"user", "Q1", nullptr, nullptr},
{"assistant", "A1", nullptr, nullptr}
};
append_history(msgs, ms, 2);
CHECK(msgs.size() == 2, "T10.13: two messages appended");
}
{
// Null role and null content — default to empty strings
json::array msgs;
dstalk_message_t m = {nullptr, nullptr, nullptr, nullptr};
append_history(msgs, &m, 1);
CHECK(msgs.size() == 1, "T10.14: null fields produce valid JSON (no crash)");
auto obj = msgs[0].as_object();
CHECK(obj["role"].as_string() == "", "T10.15: null role → empty string");
CHECK(obj["content"].as_string() == "", "T10.16: null content → empty string");
}
// ================================================================
// Test Block 11: my_free_result — null safety
// ================================================================
std::cout << "\n--- Block 11: my_free_result null safety ---\n";
{
// g_host is nullptr, so free_result should early-return
my_free_result(nullptr);
CHECK(true, "T11.1: free_result(nullptr) does not crash (null host)");
}
{
dstalk_chat_result_t r = {};
r.ok = 1;
my_free_result(&r);
CHECK(true, "T11.2: free_result with zeroed fields does not crash");
}
// ================================================================
// Test Block 12: my_configure — null host safety
// ================================================================
std::cout << "\n--- Block 12: my_configure null host safety ---\n";
{
int ret = my_configure(
"deepseek", "https://api.deepseek.com/v1",
"sk-key", "deepseek-v4", 2048, 0.5);
CHECK(ret == 0, "T12.1: my_configure returns 0 with null host");
CHECK(g_cfg.provider == "deepseek", "T12.2: provider stored");
CHECK(g_cfg.max_tokens == 2048, "T12.3: max_tokens stored");
CHECK(g_cfg.temperature == 0.5, "T12.4: temperature stored");
}
{
int ret = my_configure(nullptr, nullptr, nullptr, nullptr, 4096, 1.0);
CHECK(ret == 0, "T12.5: my_configure with all-null strings returns 0");
}
// ================================================================
// Summary
// ================================================================
std::cout << "\n";
if (g_failures == 0) {
std::cout << "=== All deepseek plugin tests passed ===\n";
return 0;
} else {
std::cerr << "=== " << g_failures << " test(s) FAILED ===\n";
return 1;
}
}