1mod assertions;
2mod example;
3mod examples;
4mod explorer;
5mod ids;
6mod instance;
7mod tool_metrics;
8
9use assertions::{AssertionsReport, display_error_row};
10use instance::{ExampleInstance, JudgeOutput, RunOutput, run_git};
11use language_extension::LspAccess;
12pub(crate) use tool_metrics::*;
13
14use ::fs::RealFs;
15use clap::Parser;
16use client::{Client, ProxySettings, UserStore};
17use collections::{HashMap, HashSet};
18use extension::ExtensionHostProxy;
19use futures::future;
20use gpui::http_client::read_proxy_from_env;
21use gpui::{App, AppContext, Application, AsyncApp, Entity, UpdateGlobal};
22use gpui_tokio::Tokio;
23use language::LanguageRegistry;
24use language_model::{ConfiguredModel, LanguageModel, LanguageModelRegistry, SelectedModel};
25use node_runtime::{NodeBinaryOptions, NodeRuntime};
26use project::Project;
27use project::project_settings::ProjectSettings;
28use prompt_store::PromptBuilder;
29use release_channel::AppVersion;
30use reqwest_client::ReqwestClient;
31use settings::{Settings, SettingsStore};
32use std::cell::RefCell;
33use std::collections::VecDeque;
34use std::env;
35use std::path::{Path, PathBuf};
36use std::rc::Rc;
37use std::str::FromStr;
38use std::sync::{Arc, LazyLock};
39use util::ResultExt as _;
40
41static CARGO_MANIFEST_DIR: LazyLock<PathBuf> =
42 LazyLock::new(|| PathBuf::from(env!("CARGO_MANIFEST_DIR")));
43
44#[derive(Parser, Debug)]
45#[command(name = "eval", disable_version_flag = true)]
46struct Args {
47 /// Runs all examples and threads that contain these substrings. If unspecified, all examples and threads are run.
48 #[arg(value_name = "EXAMPLE_SUBSTRING")]
49 filter: Vec<String>,
50 /// provider/model to use for agent
51 #[arg(long, default_value = "anthropic/claude-3-7-sonnet-latest")]
52 model: String,
53 /// provider/model to use for judges
54 #[arg(long, default_value = "anthropic/claude-3-7-sonnet-latest")]
55 judge_model: String,
56 #[arg(long, value_delimiter = ',', default_value = "rs,ts,py")]
57 languages: Vec<String>,
58 /// How many times to run each example.
59 #[arg(long, default_value = "8")]
60 repetitions: usize,
61 /// Maximum number of examples to run concurrently.
62 #[arg(long, default_value = "4")]
63 concurrency: usize,
64}
65
66fn main() {
67 dotenvy::from_filename(CARGO_MANIFEST_DIR.join(".env")).ok();
68
69 env_logger::init();
70
71 let system_id = ids::get_or_create_id(&ids::eval_system_id_path()).ok();
72 let installation_id = ids::get_or_create_id(&ids::eval_installation_id_path()).ok();
73 let session_id = uuid::Uuid::new_v4().to_string();
74 let run_timestamp = chrono::Local::now().format("%Y-%m-%d_%H-%M-%S");
75 let run_id = match env::var("GITHUB_RUN_ID") {
76 Ok(run_id) => format!("github/{}", run_id),
77 Err(_) => format!("local/{}", run_timestamp),
78 };
79
80 let root_dir = Path::new(std::env!("CARGO_MANIFEST_DIR"))
81 .parent()
82 .unwrap()
83 .parent()
84 .unwrap()
85 .canonicalize()
86 .unwrap();
87 let eval_crate_dir = root_dir.join("crates").join("eval");
88 let repos_dir = eval_crate_dir.join("repos");
89 let worktrees_dir = eval_crate_dir.join("worktrees");
90 let examples_dir = eval_crate_dir.join("src").join("examples");
91 let run_dir = eval_crate_dir
92 .join("runs")
93 .join(format!("{}", run_timestamp));
94 std::fs::create_dir_all(&run_dir).unwrap();
95 std::fs::create_dir_all(&repos_dir).unwrap();
96 std::fs::create_dir_all(&worktrees_dir).unwrap();
97 std::fs::create_dir_all(&examples_dir).unwrap();
98 std::fs::create_dir_all(&paths::config_dir()).unwrap();
99
100 let zed_commit_sha = commit_sha_for_path(&root_dir);
101 let zed_branch_name = git_branch_for_path(&root_dir);
102 let args = Args::parse();
103 let languages: HashSet<String> = args.languages.into_iter().collect();
104
105 let http_client = Arc::new(ReqwestClient::new());
106 let app = Application::headless().with_http_client(http_client);
107 let all_threads = examples::all(&examples_dir);
108
109 app.run(move |cx| {
110 let app_state = init(cx);
111
112 let telemetry = app_state.client.telemetry();
113 telemetry.start(system_id, installation_id, session_id, cx);
114
115 let enable_telemetry = env::var("ZED_EVAL_TELEMETRY").is_ok_and(|value| value == "1")
116 && telemetry.has_checksum_seed();
117 if enable_telemetry {
118 println!("Telemetry enabled");
119 telemetry::event!(
120 "Agent Eval Started",
121 zed_commit_sha = zed_commit_sha,
122 zed_branch_name = zed_branch_name,
123 run_id = run_id,
124 );
125 }
126
127 let mut cumulative_tool_metrics = ToolMetrics::default();
128
129 let agent_model = load_model(&args.model, cx).unwrap();
130 let judge_model = load_model(&args.judge_model, cx).unwrap();
131
132 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
133 registry.set_default_model(Some(agent_model.clone()), cx);
134 });
135
136 let auth1 = agent_model.provider.authenticate(cx);
137 let auth2 = judge_model.provider.authenticate(cx);
138
139 cx.spawn(async move |cx| {
140 auth1.await?;
141 auth2.await?;
142
143 let mut examples = Vec::new();
144
145 const COLORS: [&str; 12] = [
146 "\x1b[31m", // Red
147 "\x1b[32m", // Green
148 "\x1b[33m", // Yellow
149 "\x1b[34m", // Blue
150 "\x1b[35m", // Magenta
151 "\x1b[36m", // Cyan
152 "\x1b[91m", // Bright Red
153 "\x1b[92m", // Bright Green
154 "\x1b[93m", // Bright Yellow
155 "\x1b[94m", // Bright Blue
156 "\x1b[95m", // Bright Magenta
157 "\x1b[96m", // Bright Cyan
158 ];
159
160 let mut skipped = Vec::new();
161
162 for thread in all_threads {
163 let meta = thread.meta();
164 if !args.filter.is_empty() && !args.filter.iter().any(|sub| meta.name.contains(sub))
165 {
166 skipped.push(meta.name);
167 continue;
168 }
169
170 if let Some(language) = meta.language_server
171 && !languages.contains(&language.file_extension) {
172 panic!(
173 "Eval for {:?} could not be run because no language server was found for extension {:?}",
174 meta.name,
175 language.file_extension
176 );
177 }
178
179 // TODO: This creates a worktree per repetition. Ideally these examples should
180 // either be run sequentially on the same worktree, or reuse worktrees when there
181 // are more examples to run than the concurrency limit.
182 for repetition_number in 0..args.repetitions {
183 let example_instance = ExampleInstance::new(
184 thread.clone(),
185 &repos_dir,
186 &run_dir,
187 &worktrees_dir,
188 repetition_number,
189 );
190
191 examples.push(example_instance);
192 }
193 }
194
195 if !skipped.is_empty() {
196 println!("Skipped threads: {}", skipped.join(", "));
197 }
198
199 if examples.is_empty() {
200 eprintln!("Filter matched no examples");
201 return cx.update(|cx| cx.quit());
202 }
203
204 let mut repo_urls = HashSet::default();
205 let mut clone_tasks = Vec::new();
206
207 let max_name_width = examples
208 .iter()
209 .map(|e| e.worktree_name().len())
210 .max()
211 .unwrap_or(0);
212
213 for (i, example_instance) in examples.iter_mut().enumerate() {
214 let color = COLORS[i % COLORS.len()].to_string();
215 example_instance.set_log_prefix_style(&color, max_name_width);
216
217 println!(
218 "{}Logging to: {}",
219 example_instance.log_prefix,
220 example_instance.run_directory.display()
221 );
222
223 let repo_url = example_instance.repo_url();
224 if repo_urls.insert(repo_url.clone()) {
225 let repo_path = example_instance.repo_path.clone();
226
227 if !repo_path.join(".git").is_dir() {
228 println!(
229 "{:<width$} < {}",
230 "↓ Cloning",
231 repo_url,
232 width = max_name_width
233 );
234
235 let git_task = cx.spawn(async move |_cx| {
236 std::fs::create_dir_all(&repo_path)?;
237 run_git(&repo_path, &["init"]).await?;
238 run_git(&repo_path, &["remote", "add", "origin", &repo_url]).await
239 });
240
241 clone_tasks.push(git_task);
242 } else {
243 println!(
244 "{:<width$} < {}",
245 "✔︎ Already cloned",
246 repo_url,
247 width = max_name_width
248 );
249
250 let actual_origin =
251 run_git(&repo_path, &["remote", "get-url", "origin"]).await?;
252 anyhow::ensure!(
253 actual_origin == repo_url,
254 "remote origin {actual_origin} does not match expected origin {repo_url}"
255 );
256 }
257 }
258 }
259
260 future::join_all(clone_tasks).await;
261
262 for example_instance in examples.iter_mut() {
263 example_instance.fetch().await?;
264 }
265
266 let examples = Rc::new(RefCell::new(VecDeque::from(examples)));
267 let results_by_example_name = Rc::new(RefCell::new(HashMap::default()));
268
269 future::join_all((0..args.concurrency).map(|_| {
270 let app_state = app_state.clone();
271 let model = agent_model.model.clone();
272 let judge_model = judge_model.model.clone();
273 let zed_commit_sha = zed_commit_sha.clone();
274 let zed_branch_name = zed_branch_name.clone();
275 let run_id = run_id.clone();
276 let examples = examples.clone();
277 let results = results_by_example_name.clone();
278 cx.spawn(async move |cx| {
279 loop {
280 let Some(mut example) = examples.borrow_mut().pop_front() else {
281 break;
282 };
283 let result = async {
284 example.setup().await?;
285 let run_output = cx
286 .update(|cx| example.run(model.clone(), app_state.clone(), cx))?
287 .await?;
288 let judge_output = judge_example(
289 example.clone(),
290 judge_model.clone(),
291 &zed_commit_sha,
292 &zed_branch_name,
293 &run_id,
294 &run_output,
295 enable_telemetry,
296 cx,
297 )
298 .await;
299 anyhow::Ok((run_output, judge_output))
300 }
301 .await;
302 results
303 .borrow_mut()
304 .entry(example.name.clone())
305 .or_insert(Vec::new())
306 .push((example.clone(), result));
307 }
308 })
309 }))
310 .await;
311
312 print_report(
313 &mut results_by_example_name.borrow_mut(),
314 &mut cumulative_tool_metrics,
315 &run_dir,
316 )?;
317
318 app_state.client.telemetry().flush_events().await;
319
320 cx.update(|cx| cx.quit())
321 })
322 .detach_and_log_err(cx);
323 });
324}
325
326/// Subset of `workspace::AppState` needed by `HeadlessAssistant`, with additional fields.
327pub struct AgentAppState {
328 pub languages: Arc<LanguageRegistry>,
329 pub client: Arc<Client>,
330 pub user_store: Entity<UserStore>,
331 pub fs: Arc<dyn fs::Fs>,
332 pub node_runtime: NodeRuntime,
333
334 // Additional fields not present in `workspace::AppState`.
335 pub prompt_builder: Arc<PromptBuilder>,
336}
337
338pub fn init(cx: &mut App) -> Arc<AgentAppState> {
339 let app_version = AppVersion::load(env!("ZED_PKG_VERSION"));
340 release_channel::init(app_version, cx);
341 gpui_tokio::init(cx);
342
343 let mut settings_store = SettingsStore::new(cx);
344 settings_store
345 .set_default_settings(settings::default_settings().as_ref(), cx)
346 .unwrap();
347 cx.set_global(settings_store);
348 client::init_settings(cx);
349
350 // Set User-Agent so we can download language servers from GitHub
351 let user_agent = format!(
352 "Zed Agent Eval/{} ({}; {})",
353 app_version,
354 std::env::consts::OS,
355 std::env::consts::ARCH
356 );
357 let proxy_str = ProxySettings::get_global(cx).proxy.to_owned();
358 let proxy_url = proxy_str
359 .as_ref()
360 .and_then(|input| input.parse().ok())
361 .or_else(read_proxy_from_env);
362 let http = {
363 let _guard = Tokio::handle(cx).enter();
364
365 ReqwestClient::proxy_and_user_agent(proxy_url, &user_agent)
366 .expect("could not start HTTP client")
367 };
368 cx.set_http_client(Arc::new(http));
369
370 Project::init_settings(cx);
371
372 let client = Client::production(cx);
373 cx.set_http_client(client.http_client());
374
375 let git_binary_path = None;
376 let fs = Arc::new(RealFs::new(
377 git_binary_path,
378 cx.background_executor().clone(),
379 ));
380
381 let mut languages = LanguageRegistry::new(cx.background_executor().clone());
382 languages.set_language_server_download_dir(paths::languages_dir().clone());
383 let languages = Arc::new(languages);
384
385 let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
386
387 extension::init(cx);
388
389 let (mut tx, rx) = watch::channel(None);
390 cx.observe_global::<SettingsStore>(move |cx| {
391 let settings = &ProjectSettings::get_global(cx).node;
392 let options = NodeBinaryOptions {
393 allow_path_lookup: !settings.ignore_system_version,
394 allow_binary_download: true,
395 use_paths: settings.path.as_ref().map(|node_path| {
396 let node_path = PathBuf::from(shellexpand::tilde(node_path).as_ref());
397 let npm_path = settings
398 .npm_path
399 .as_ref()
400 .map(|path| PathBuf::from(shellexpand::tilde(&path).as_ref()));
401 (
402 node_path.clone(),
403 npm_path.unwrap_or_else(|| {
404 let base_path = PathBuf::new();
405 node_path.parent().unwrap_or(&base_path).join("npm")
406 }),
407 )
408 }),
409 };
410 tx.send(Some(options)).log_err();
411 })
412 .detach();
413 let node_runtime = NodeRuntime::new(client.http_client(), None, rx);
414
415 let extension_host_proxy = ExtensionHostProxy::global(cx);
416
417 language::init(cx);
418 debug_adapter_extension::init(extension_host_proxy.clone(), cx);
419 language_extension::init(LspAccess::Noop, extension_host_proxy, languages.clone());
420 language_model::init(client.clone(), cx);
421 language_models::init(user_store.clone(), client.clone(), cx);
422 languages::init(languages.clone(), fs.clone(), node_runtime.clone(), cx);
423 prompt_store::init(cx);
424 terminal_view::init(cx);
425 let stdout_is_a_pty = false;
426 let prompt_builder = PromptBuilder::load(fs.clone(), stdout_is_a_pty, cx);
427 agent_ui::init(
428 fs.clone(),
429 client.clone(),
430 prompt_builder.clone(),
431 languages.clone(),
432 true,
433 cx,
434 );
435 assistant_tools::init(client.http_client(), cx);
436
437 SettingsStore::update_global(cx, |store, cx| {
438 store.set_user_settings(include_str!("../runner_settings.json"), cx)
439 })
440 .unwrap();
441
442 Arc::new(AgentAppState {
443 languages,
444 client,
445 user_store,
446 fs,
447 node_runtime,
448 prompt_builder,
449 })
450}
451
452pub fn find_model(
453 model_name: &str,
454 model_registry: &LanguageModelRegistry,
455 cx: &App,
456) -> anyhow::Result<Arc<dyn LanguageModel>> {
457 let selected = SelectedModel::from_str(model_name).map_err(|e| anyhow::anyhow!(e))?;
458 model_registry
459 .available_models(cx)
460 .find(|model| model.id() == selected.model && model.provider_id() == selected.provider)
461 .ok_or_else(|| {
462 anyhow::anyhow!(
463 "No language model with ID {}/{} was available. Available models: {}",
464 selected.model.0,
465 selected.provider.0,
466 model_registry
467 .available_models(cx)
468 .map(|model| format!("{}/{}", model.provider_id().0, model.id().0))
469 .collect::<Vec<_>>()
470 .join(", ")
471 )
472 })
473}
474
475pub fn load_model(model_name: &str, cx: &mut App) -> anyhow::Result<ConfiguredModel> {
476 let model = {
477 let model_registry = LanguageModelRegistry::read_global(cx);
478 find_model(model_name, model_registry, cx)?
479 };
480
481 let provider = {
482 let model_registry = LanguageModelRegistry::read_global(cx);
483 model_registry
484 .provider(&model.provider_id())
485 .ok_or_else(|| anyhow::anyhow!("Provider not found: {}", model.provider_id()))?
486 };
487
488 Ok(ConfiguredModel {
489 provider: provider.clone(),
490 model: model.clone(),
491 })
492}
493
494pub fn commit_sha_for_path(repo_path: &Path) -> String {
495 futures::executor::block_on(run_git(repo_path, &["rev-parse", "HEAD"])).unwrap()
496}
497
498pub fn git_branch_for_path(repo_path: &Path) -> String {
499 match std::env::var("GITHUB_REF_NAME") {
500 Ok(branch) => branch,
501 Err(_) => {
502 futures::executor::block_on(run_git(repo_path, &["rev-parse", "--abbrev-ref", "HEAD"]))
503 .unwrap_or_else(|_| "unknown".to_string())
504 }
505 }
506}
507
508async fn judge_example(
509 example: ExampleInstance,
510 model: Arc<dyn LanguageModel>,
511 zed_commit_sha: &str,
512 zed_branch_name: &str,
513 run_id: &str,
514 run_output: &RunOutput,
515 enable_telemetry: bool,
516 cx: &AsyncApp,
517) -> JudgeOutput {
518 let judge_output = example.judge(model.clone(), run_output, cx).await;
519
520 if enable_telemetry {
521 telemetry::event!(
522 "Agent Example Evaluated",
523 zed_commit_sha = zed_commit_sha,
524 zed_branch_name = zed_branch_name,
525 run_id = run_id,
526 example_name = example.name.clone(),
527 example_repetition = example.repetition,
528 diff_evaluation = judge_output.diff.clone(),
529 thread_evaluation = judge_output.thread,
530 tool_metrics = run_output.tool_metrics,
531 response_count = run_output.response_count,
532 token_usage = run_output.token_usage,
533 model = model.telemetry_id(),
534 model_provider = model.provider_id().to_string(),
535 repository_url = example.repo_url(),
536 repository_revision = example.revision(),
537 diagnostic_summary_before = run_output.diagnostic_summary_before,
538 diagnostic_summary_after = run_output.diagnostic_summary_after,
539 diagnostics_before = run_output.diagnostics_before,
540 diagnostics_after = run_output.diagnostics_after,
541 );
542 }
543
544 judge_output
545}
546
547const HEADER_WIDTH: usize = 65;
548
549fn print_h1(header: &str) {
550 println!("\n\n{:=^HEADER_WIDTH$}", "");
551 println!("{:^HEADER_WIDTH$}", header);
552 println!("{:=^HEADER_WIDTH$}\n", "");
553}
554
555fn print_h2(header: &str) {
556 println!("\n{:-^HEADER_WIDTH$}", "");
557 println!("{:^HEADER_WIDTH$}", header);
558 println!("{:-^HEADER_WIDTH$}\n", "");
559}
560
561fn print_report(
562 results_by_example_name: &mut HashMap<
563 String,
564 Vec<(ExampleInstance, anyhow::Result<(RunOutput, JudgeOutput)>)>,
565 >,
566 cumulative_tool_metrics: &mut ToolMetrics,
567 run_dir: &Path,
568) -> anyhow::Result<()> {
569 print_h1("EVAL RESULTS");
570
571 let mut diff_scores = Vec::new();
572 let mut thread_scores = Vec::new();
573 let mut programmatic_scores = Vec::new();
574 let mut error_count = 0;
575
576 for (example_name, results) in results_by_example_name.iter_mut() {
577 print_h2(example_name);
578
579 results.sort_unstable_by_key(|(example, _)| example.repetition);
580 let mut example_cumulative_tool_metrics = ToolMetrics::default();
581
582 let mut table_rows = String::new();
583
584 for (example, result) in results.iter() {
585 match result {
586 Err(err) => {
587 display_error_row(&mut table_rows, example.repetition, err.to_string())?;
588 error_count += 1;
589 programmatic_scores.push(0.0);
590 diff_scores.push(0.0);
591 thread_scores.push(0.0);
592 }
593 Ok((run_output, judge_output)) => {
594 cumulative_tool_metrics.merge(&run_output.tool_metrics);
595 example_cumulative_tool_metrics.merge(&run_output.tool_metrics);
596
597 if run_output.programmatic_assertions.total_count() > 0 {
598 for assertion in &run_output.programmatic_assertions.ran {
599 assertions::display_table_row(
600 &mut table_rows,
601 example.repetition,
602 assertion,
603 )?;
604 }
605
606 programmatic_scores
607 .push(run_output.programmatic_assertions.passed_percentage())
608 }
609
610 if !judge_output.diff.is_empty() {
611 diff_scores.push(judge_output.diff.passed_percentage());
612
613 for assertion in &judge_output.diff.ran {
614 assertions::display_table_row(
615 &mut table_rows,
616 example.repetition,
617 assertion,
618 )?;
619 }
620 }
621
622 if !judge_output.thread.is_empty() {
623 thread_scores.push(judge_output.thread.passed_percentage());
624
625 for assertion in &judge_output.thread.ran {
626 assertions::display_table_row(
627 &mut table_rows,
628 example.repetition,
629 assertion,
630 )?;
631 }
632 }
633 }
634 }
635 }
636
637 let mut all_asserts = Vec::new();
638
639 if !table_rows.is_empty() {
640 assertions::print_table_header();
641 print!("{}", table_rows);
642
643 assertions::print_table_divider();
644
645 for (example, result) in results.iter() {
646 if let Ok((run_output, judge_output)) = result {
647 let asserts = [
648 run_output.programmatic_assertions.clone(),
649 judge_output.diff.clone(),
650 judge_output.thread.clone(),
651 ];
652 all_asserts.extend_from_slice(&asserts);
653 assertions::print_table_round_summary(
654 &example.repetition.to_string(),
655 asserts.iter(),
656 )
657 } else if let Err(err) = result {
658 let assert = AssertionsReport::error(err.to_string());
659 all_asserts.push(assert.clone());
660 assertions::print_table_round_summary(
661 &example.repetition.to_string(),
662 [assert].iter(),
663 )
664 }
665 }
666
667 assertions::print_table_divider();
668
669 assertions::print_table_round_summary("avg", all_asserts.iter());
670
671 assertions::print_table_footer();
672 }
673
674 if !example_cumulative_tool_metrics.is_empty() {
675 println!("{}", &example_cumulative_tool_metrics);
676 }
677 }
678
679 if results_by_example_name.len() > 1 {
680 print_h1("AGGREGATE");
681
682 if error_count > 0 {
683 println!("\n{error_count} examples failed to run!");
684 }
685
686 let programmatic_score_count = programmatic_scores.len();
687 if programmatic_score_count > 0 {
688 let average_programmatic_score = (programmatic_scores.into_iter().sum::<f32>()
689 / (programmatic_score_count as f32))
690 .floor();
691 println!("Average programmatic score: {average_programmatic_score}%");
692 }
693
694 let diff_score_count = diff_scores.len();
695 if diff_score_count > 0 {
696 let average_diff_score =
697 (diff_scores.into_iter().sum::<f32>() / (diff_score_count as f32)).floor();
698 println!("Average diff score: {average_diff_score}%");
699 }
700
701 let thread_score_count = thread_scores.len();
702
703 if thread_score_count > 0 {
704 let average_thread_score =
705 (thread_scores.into_iter().sum::<f32>() / (thread_score_count as f32)).floor();
706 println!("Average thread score: {average_thread_score}%");
707 }
708
709 println!();
710
711 print_h2("CUMULATIVE TOOL METRICS");
712 println!("{}", cumulative_tool_metrics);
713 }
714
715 let explorer_output_path = run_dir.join("overview.html");
716 let mut json_paths: Vec<PathBuf> = results_by_example_name
717 .values()
718 .flat_map(|results| {
719 results.iter().map(|(example, _)| {
720 let absolute_path = run_dir.join(example.run_directory.join("last.messages.json"));
721 let cwd = std::env::current_dir().expect("Can't get current dir");
722 pathdiff::diff_paths(&absolute_path, cwd).unwrap_or_else(|| absolute_path.clone())
723 })
724 })
725 .collect::<Vec<_>>();
726 json_paths.sort();
727 if let Err(err) = explorer::generate_explorer_html(&json_paths, &explorer_output_path) {
728 eprintln!("Failed to generate explorer HTML: {}", err);
729 }
730
731 Ok(())
732}