9#include "absl/status/status.h"
10#include "absl/status/statusor.h"
11#include "absl/strings/numbers.h"
12#include "absl/strings/str_cat.h"
13#include "absl/strings/str_format.h"
14#include "absl/time/clock.h"
15#include "absl/time/time.h"
17#include "nlohmann/json.hpp"
34struct RecordingState {
35 std::string recording_id;
36 std::string host =
"localhost";
38 std::string output_path;
41std::filesystem::path RecordingStateFilePath() {
43 std::filesystem::path base = std::filesystem::temp_directory_path(ec);
45 base = std::filesystem::current_path();
47 return base /
"yaze" /
"agent" /
"recording_state.json";
50absl::Status SaveRecordingState(
const RecordingState& state) {
51 auto path = RecordingStateFilePath();
53 std::filesystem::create_directories(path.parent_path(), ec);
55 json[
"recording_id"] = state.recording_id;
56 json[
"host"] = state.host;
57 json[
"port"] = state.port;
58 json[
"output_path"] = state.output_path;
60 std::ofstream out(path, std::ios::out | std::ios::trunc);
62 return absl::InternalError(
63 absl::StrCat(
"Failed to write recording state to ", path.string()));
67 return absl::InternalError(
68 absl::StrCat(
"Failed to flush recording state to ", path.string()));
70 return absl::OkStatus();
73absl::StatusOr<RecordingState> LoadRecordingState() {
74 auto path = RecordingStateFilePath();
75 std::ifstream in(path);
77 return absl::NotFoundError(
78 "No active recording session found. Run 'z3ed agent test record start' "
85 }
catch (
const nlohmann::json::parse_error& error) {
86 return absl::InternalError(
87 absl::StrCat(
"Failed to parse recording state at ", path.string(),
": ",
92 state.recording_id = json.value(
"recording_id",
"");
93 state.host = json.value(
"host",
"localhost");
94 state.port = json.value(
"port", 50052);
95 state.output_path = json.value(
"output_path",
"");
97 if (state.recording_id.empty()) {
98 return absl::InvalidArgumentError(absl::StrCat(
99 "Recording state at ", path.string(),
" is missing a recording_id"));
105absl::Status ClearRecordingState() {
106 auto path = RecordingStateFilePath();
108 std::filesystem::remove(path, ec);
109 if (ec && ec != std::errc::no_such_file_or_directory) {
110 return absl::InternalError(
111 absl::StrCat(
"Failed to clear recording state: ", ec.message()));
113 return absl::OkStatus();
116std::string DefaultRecordingOutputPath() {
117 absl::Time now = absl::Now();
119 "tests/gui/recording-",
120 absl::FormatTime(
"%Y%m%dT%H%M%S", now, absl::LocalTimeZone()),
".json");
126absl::Status HandleTestRunCommand(
const std::vector<std::string>& args);
127absl::Status HandleTestReplayCommand(
const std::vector<std::string>& args);
128absl::Status HandleTestStatusCommand(
const std::vector<std::string>& args);
129absl::Status HandleTestListCommand(
const std::vector<std::string>& args);
130absl::Status HandleTestResultsCommand(
const std::vector<std::string>& args);
131absl::Status HandleTestRecordCommand(
const std::vector<std::string>& args);
133absl::Status HandleTestRunCommand(
const std::vector<std::string>& args) {
134 if (args.empty() || args[0] !=
"--prompt") {
135 return absl::InvalidArgumentError(
136 "Usage: agent test run --prompt <description> [--host <host>] [--port "
138 "Example: agent test run --prompt \"Open the overworld editor and "
139 "verify it loads\"");
142 std::string prompt = args.size() > 1 ? args[1] :
"";
143 std::string host =
"localhost";
147 for (
size_t i = 2; i < args.size(); ++i) {
148 if (args[i] ==
"--host" && i + 1 < args.size()) {
150 }
else if (args[i] ==
"--port" && i + 1 < args.size()) {
151 port = std::stoi(args[++i]);
155 std::cout <<
"\n=== GUI Automation Test ===\n";
156 std::cout <<
"Prompt: " << prompt <<
"\n";
157 std::cout <<
"Server: " << host <<
":" << port <<
"\n\n";
160 TestWorkflowGenerator generator;
161 auto workflow_or = generator.GenerateWorkflow(prompt);
162 if (!workflow_or.ok()) {
163 std::cerr <<
"Failed to generate workflow: "
164 << workflow_or.status().message() << std::endl;
165 return workflow_or.status();
168 auto workflow = workflow_or.value();
169 std::cout <<
"Generated " << workflow.steps.size() <<
" test steps\n\n";
172 GuiAutomationClient client(absl::StrCat(host,
":", port));
173 auto status = client.Connect();
175 std::cerr <<
"Failed to connect to test harness: " << status.message()
181 for (
size_t i = 0; i < workflow.steps.size(); ++i) {
182 const auto& step = workflow.steps[i];
183 std::cout <<
"Step " << (i + 1) <<
": " << step.ToString() <<
"... ";
186 absl::StatusOr<AutomationResult> result(
187 absl::InternalError(
"Unknown step type"));
191 result = client.Click(step.target);
194 result = client.Type(step.target, step.text, step.clear_first);
197 result = client.Wait(step.condition, step.timeout_ms);
200 result = client.Assert(step.condition);
203 std::cout <<
"✗ SKIPPED (unknown type)\n";
208 std::cout <<
"✗ FAILED\n";
209 std::cerr <<
" Error: " << result.status().message() <<
"\n";
210 return result.status();
213 if (!result.value().success) {
214 std::cout <<
"✗ FAILED\n";
215 std::cerr <<
" Error: " << result.value().message <<
"\n";
216 return absl::InternalError(result.value().message);
222 std::cout <<
"\n✅ Test passed!\n";
223 return absl::OkStatus();
226absl::Status HandleTestReplayCommand(
const std::vector<std::string>& args) {
228 return absl::InvalidArgumentError(
229 "Usage: agent test replay <script.json> [--host <host>] [--port "
230 "<port>] [--ci] [--set key=value ...]\n"
231 "Example: agent test replay tests/overworld_load.json --set "
235 std::string script_path = args[0];
236 std::string host =
"localhost";
238 bool ci_mode =
false;
239 std::map<std::string, std::string> parameter_overrides;
241 for (
size_t i = 1; i < args.size(); ++i) {
242 const std::string& token = args[i];
243 if (token ==
"--host" && i + 1 < args.size()) {
245 }
else if (token ==
"--port" && i + 1 < args.size()) {
246 const std::string& port_value = args[++i];
248 if (!absl::SimpleAtoi(port_value, &parsed_port)) {
249 return absl::InvalidArgumentError(
250 absl::StrCat(
"Invalid --port value: ", port_value));
253 }
else if (token ==
"--set" && i + 1 < args.size()) {
254 const std::string& assignment = args[++i];
255 size_t separator = assignment.find(
'=');
256 if (separator == std::string::npos || separator == 0 ||
257 separator == assignment.size() - 1) {
258 return absl::InvalidArgumentError(absl::StrCat(
259 "Invalid --set value (expected key=value): ", assignment));
261 parameter_overrides[assignment.substr(0, separator)] =
262 assignment.substr(separator + 1);
263 }
else if (token ==
"--ci") {
266 return absl::InvalidArgumentError(
267 absl::StrCat(
"Unknown replay option: ", token));
271 std::cout <<
"\n=== Replay Test ===\n";
272 std::cout <<
"Script: " << script_path <<
"\n";
273 std::cout <<
"Server: " << host <<
":" << port <<
"\n\n";
275 std::cout <<
"CI mode: enabled\n";
277 if (!parameter_overrides.empty()) {
278 std::cout <<
"Overrides:\n";
279 for (
const auto& [key, value] : parameter_overrides) {
280 std::cout <<
" - " << key <<
"=" << value <<
"\n";
285 GuiAutomationClient client(absl::StrCat(host,
":", port));
286 auto status = client.Connect();
288 std::cerr <<
"Failed to connect: " << status.message() << std::endl;
292 auto result = client.ReplayTest(script_path, ci_mode, parameter_overrides);
294 std::cerr <<
"Replay failed: " << result.status().message() << std::endl;
295 return result.status();
299 if (!result->success) {
300 return absl::InternalError(result.value().message);
303 return absl::OkStatus();
306absl::Status HandleTestStatusCommand(
const std::vector<std::string>& args) {
308 std::string host =
"localhost";
311 for (
size_t i = 0; i < args.size(); ++i) {
312 if (args[i] ==
"--test-id" && i + 1 < args.size()) {
314 }
else if (args[i] ==
"--host" && i + 1 < args.size()) {
316 }
else if (args[i] ==
"--port" && i + 1 < args.size()) {
317 port = std::stoi(args[++i]);
321 if (test_id.empty()) {
322 return absl::InvalidArgumentError(
323 "Usage: agent test status --test-id <id> [--host <host>] [--port "
327 GuiAutomationClient client(absl::StrCat(host,
":", port));
328 auto status = client.Connect();
333 auto details = client.GetTestStatus(test_id);
335 return details.status();
338 std::cout <<
"\n=== Test Status ===\n";
339 std::cout <<
"Test ID: " << test_id <<
"\n";
340 std::cout <<
"Status: " << TestRunStatusToString(details.value().status)
347 if (!details.value().error_message.empty()) {
348 std::cout <<
"Error: " << details.value().error_message <<
"\n";
351 return absl::OkStatus();
354absl::Status HandleTestListCommand(
const std::vector<std::string>& args) {
355 std::string host =
"localhost";
358 for (
size_t i = 0; i < args.size(); ++i) {
359 if (args[i] ==
"--host" && i + 1 < args.size()) {
361 }
else if (args[i] ==
"--port" && i + 1 < args.size()) {
362 port = std::stoi(args[++i]);
366 GuiAutomationClient client(absl::StrCat(host,
":", port));
367 auto status = client.Connect();
372 auto batch = client.ListTests(
"", 100,
"");
374 return batch.status();
377 std::cout <<
"\n=== Available Tests ===\n";
378 std::cout <<
"Total: " << batch.value().total_count <<
"\n\n";
380 for (
const auto& test : batch.value().tests) {
381 std::cout <<
"• " << test.name <<
"\n";
382 std::cout <<
" ID: " << test.test_id <<
"\n";
383 std::cout <<
" Category: " << test.category <<
"\n";
384 std::cout <<
" Runs: " << test.total_runs <<
" (" << test.pass_count
385 <<
" passed, " << test.fail_count <<
" failed)\n\n";
388 return absl::OkStatus();
391absl::Status HandleTestResultsCommand(
const std::vector<std::string>& args) {
393 std::string host =
"localhost";
395 bool include_logs =
false;
397 for (
size_t i = 0; i < args.size(); ++i) {
398 if (args[i] ==
"--test-id" && i + 1 < args.size()) {
400 }
else if (args[i] ==
"--host" && i + 1 < args.size()) {
402 }
else if (args[i] ==
"--port" && i + 1 < args.size()) {
403 port = std::stoi(args[++i]);
404 }
else if (args[i] ==
"--include-logs") {
409 if (test_id.empty()) {
410 return absl::InvalidArgumentError(
411 "Usage: agent test results --test-id <id> [--include-logs] [--host "
412 "<host>] [--port <port>]");
415 GuiAutomationClient client(absl::StrCat(host,
":", port));
416 auto status = client.Connect();
421 auto details = client.GetTestResults(test_id, include_logs);
423 return details.status();
426 std::cout <<
"\n=== Test Results ===\n";
427 std::cout <<
"Test ID: " << details.value().test_id <<
"\n";
428 std::cout <<
"Name: " << details.value().test_name <<
"\n";
429 std::cout <<
"Success: " << (details.value().success ?
"✓" :
"✗") <<
"\n";
430 std::cout <<
"Duration: " << details.value().duration_ms <<
"ms\n\n";
432 if (!details.value().assertions.empty()) {
433 std::cout <<
"Assertions:\n";
434 for (
const auto& assertion : details.value().assertions) {
435 std::cout <<
" " << (assertion.passed ?
"✓" :
"✗") <<
" "
436 << assertion.description <<
"\n";
437 if (!assertion.error_message.empty()) {
438 std::cout <<
" Error: " << assertion.error_message <<
"\n";
443 if (include_logs && !details.value().logs.empty()) {
444 std::cout <<
"\nLogs:\n";
445 for (
const auto& log : details.value().logs) {
446 std::cout <<
" " << log <<
"\n";
450 return absl::OkStatus();
453absl::Status HandleTestRecordCommand(
const std::vector<std::string>& args) {
455 return absl::InvalidArgumentError(
456 "Usage: agent test record <start|stop> [options]\n"
457 " start [--output <file>] [--description <text>] [--session <id>]\n"
458 " [--host <host>] [--port <port>]\n"
459 " stop [--validate] [--discard] [--host <host>] [--port <port>]");
462 std::string action = args[0];
463 if (action !=
"start" && action !=
"stop") {
464 return absl::InvalidArgumentError(
465 "Record action must be 'start' or 'stop'");
468 if (action ==
"start") {
469 std::string host =
"localhost";
472 std::string session_name;
473 std::string output_path;
475 for (
size_t i = 1; i < args.size(); ++i) {
476 const std::string& token = args[i];
477 if (token ==
"--output" && i + 1 < args.size()) {
478 output_path = args[++i];
479 }
else if (token ==
"--description" && i + 1 < args.size()) {
481 }
else if (token ==
"--session" && i + 1 < args.size()) {
482 session_name = args[++i];
483 }
else if (token ==
"--host" && i + 1 < args.size()) {
485 }
else if (token ==
"--port" && i + 1 < args.size()) {
486 std::string port_value = args[++i];
488 if (!absl::SimpleAtoi(port_value, &parsed_port)) {
489 return absl::InvalidArgumentError(
490 absl::StrCat(
"Invalid --port value: ", port_value));
496 if (output_path.empty()) {
497 output_path = DefaultRecordingOutputPath();
500 std::filesystem::path absolute_output =
501 std::filesystem::absolute(output_path);
503 std::filesystem::create_directories(absolute_output.parent_path(), ec);
505 GuiAutomationClient client(absl::StrCat(host,
":", port));
508 if (session_name.empty()) {
509 session_name = std::filesystem::path(output_path).stem().string();
513 client.StartRecording(absolute_output.string(),
515 if (!start_result.success) {
516 return absl::InternalError(absl::StrCat(
517 "Harness rejected start-recording request: ", start_result.message));
520 RecordingState state;
521 state.recording_id = start_result.recording_id;
524 state.output_path = absolute_output.string();
527 std::cout <<
"\n=== Recording Session Started ===\n";
528 std::cout <<
"Recording ID: " << start_result.recording_id <<
"\n";
529 std::cout <<
"Server: " << host <<
":" << port <<
"\n";
530 std::cout <<
"Output: " << absolute_output <<
"\n";
532 std::cout <<
"Description: " <<
description <<
"\n";
534 if (start_result.started_at.has_value()) {
535 std::cout <<
"Started: "
536 << absl::FormatTime(
"%Y-%m-%d %H:%M:%S",
537 *start_result.started_at,
538 absl::LocalTimeZone())
541 std::cout <<
"\nPress Ctrl+C to abort the recording session.\n";
543 return absl::OkStatus();
547 bool validate =
false;
548 bool discard =
false;
549 std::optional<std::string> host_override;
550 std::optional<int> port_override;
552 for (
size_t i = 1; i < args.size(); ++i) {
553 const std::string& token = args[i];
554 if (token ==
"--validate") {
556 }
else if (token ==
"--discard") {
558 }
else if (token ==
"--host" && i + 1 < args.size()) {
559 host_override = args[++i];
560 }
else if (token ==
"--port" && i + 1 < args.size()) {
561 std::string port_value = args[++i];
563 if (!absl::SimpleAtoi(port_value, &parsed_port)) {
564 return absl::InvalidArgumentError(
565 absl::StrCat(
"Invalid --port value: ", port_value));
567 port_override = parsed_port;
571 if (discard && validate) {
572 return absl::InvalidArgumentError(
573 "Cannot use --validate and --discard together");
577 if (host_override.has_value()) {
578 state.host = *host_override;
580 if (port_override.has_value()) {
581 state.port = *port_override;
584 GuiAutomationClient client(absl::StrCat(state.host,
":", state.port));
588 client.StopRecording(state.recording_id, discard));
589 if (!stop_result.success) {
590 return absl::InternalError(
591 absl::StrCat(
"Stop recording failed: ", stop_result.message));
595 std::cout <<
"\n=== Recording Session Completed ===\n";
596 std::cout <<
"Recording ID: " << state.recording_id <<
"\n";
597 std::cout <<
"Server: " << state.host <<
":" << state.port <<
"\n";
598 std::cout <<
"Steps captured: " << stop_result.step_count <<
"\n";
599 std::cout <<
"Duration: " << stop_result.duration.count() <<
" ms\n";
600 if (!stop_result.message.empty()) {
601 std::cout <<
"Message: " << stop_result.message <<
"\n";
603 if (!discard && !stop_result.output_path.empty()) {
604 std::cout <<
"Output saved to: " << stop_result.output_path <<
"\n";
608 std::cout <<
"Recording discarded; no script file was produced."
610 return absl::OkStatus();
613 if (!validate || stop_result.output_path.empty()) {
614 std::cout << std::endl;
615 return absl::OkStatus();
618 std::cout <<
"\nReplaying recorded script to validate...\n";
620 client.ReplayTest(stop_result.output_path,
false, {}));
621 if (!replay_result.success) {
622 return absl::InternalError(
623 absl::StrCat(
"Replay failed: ", replay_result.message));
626 std::cout <<
"Replay succeeded. Steps executed: "
627 << replay_result.steps_executed <<
"\n";
628 return absl::OkStatus();
635 return absl::InvalidArgumentError(
636 "Usage: agent test <subcommand>\n"
638 " run --prompt <text> - Generate and run a GUI automation test\n"
639 " replay <script> - Replay a script [--set key=value] [--ci]\n"
640 " status --test-id <id> - Query test execution status\n"
641 " list - List available tests\n"
642 " results --test-id <id> - Get detailed test results\n"
643 " record start/stop - Record test interactions\n"
644 "\nNote: Test commands require YAZE_WITH_GRPC=ON at build time.");
647#ifndef YAZE_WITH_GRPC
648 return absl::UnimplementedError(
649 "GUI automation test commands require YAZE_WITH_GRPC=ON at build time.\n"
650 "Rebuild with: cmake -B build-grpc-test -DYAZE_WITH_GRPC=ON\n"
651 "Then: cmake --build build-grpc-test --target z3ed");
653 std::string subcommand = args[0];
654 std::vector<std::string> tail(args.begin() + 1, args.end());
656 if (subcommand ==
"run") {
657 return HandleTestRunCommand(tail);
658 }
else if (subcommand ==
"replay") {
659 return HandleTestReplayCommand(tail);
660 }
else if (subcommand ==
"status") {
661 return HandleTestStatusCommand(tail);
662 }
else if (subcommand ==
"list") {
663 return HandleTestListCommand(tail);
664 }
else if (subcommand ==
"results") {
665 return HandleTestResultsCommand(tail);
666 }
else if (subcommand ==
"record") {
667 return HandleTestRecordCommand(tail);
669 return absl::InvalidArgumentError(
670 absl::StrCat(
"Unknown test subcommand: ", subcommand,
671 "\nRun 'z3ed agent test' for usage."));
#define ASSIGN_OR_RETURN(type_variable_name, expression)
std::string FormatOptionalTime(const std::optional< absl::Time > &time)
absl::Status HandleTestCommand(const std::vector< std::string > &args)
void PrintReplayResult(const ReplayTestResult &result, std::ostream &output)
#define RETURN_IF_ERROR(expr)