1use agent::{RequestKind, ThreadEvent, ThreadStore};
2use anyhow::{Context as _, Result, anyhow};
3use assistant_tool::ToolWorkingSet;
4use client::proto::LspWorkProgress;
5use collections::HashMap;
6use dap::DapRegistry;
7use futures::channel::mpsc;
8use futures::{FutureExt, StreamExt as _, select_biased};
9use gpui::{App, AppContext as _, AsyncApp, Entity, Task};
10use handlebars::Handlebars;
11use language::{DiagnosticSeverity, OffsetRangeExt};
12use language_model::{
13 LanguageModel, LanguageModelRequest, LanguageModelRequestMessage, MessageContent, Role,
14 StopReason, TokenUsage,
15};
16use project::{LspStore, Project, ProjectPath};
17use serde::{Deserialize, Serialize};
18use std::fmt::Write as _;
19use std::fs::File;
20use std::io::Write as _;
21use std::sync::{Arc, Mutex};
22use std::time::Duration;
23use std::{
24 fs,
25 path::{Path, PathBuf},
26};
27use unindent::Unindent as _;
28use util::ResultExt as _;
29use util::command::new_smol_command;
30use util::serde::default_true;
31
32use crate::AgentAppState;
33
34pub const EXAMPLES_DIR: &str = "./crates/eval/examples";
35pub const REPOS_DIR: &str = "./crates/eval/repos";
36pub const WORKTREES_DIR: &str = "./crates/eval/worktrees";
37
38const THREAD_EVENT_TIMEOUT: Duration = Duration::from_secs(60 * 2);
39
40#[derive(Clone, Debug, Deserialize)]
41pub struct ExampleBase {
42 pub url: String,
43 pub revision: String,
44 pub language_extension: Option<String>,
45 pub insert_id: Option<String>,
46 #[serde(default = "default_true")]
47 pub require_lsp: bool,
48}
49
50#[derive(Clone, Debug)]
51pub struct Example {
52 pub name: String,
53 /// Content of `base.toml`
54 pub base: ExampleBase,
55 /// Content of `prompt.md`
56 pub prompt: String,
57 /// Content of `criteria.md`
58 pub criteria: String,
59 /// Markdown output file to append to
60 pub output_file: Option<Arc<Mutex<File>>>,
61 /// Path to the output run directory.
62 pub run_dir: PathBuf,
63 /// Path to markdown output file
64 pub output_file_path: PathBuf,
65 /// Prefix used for logging that identifies this example
66 pub log_prefix: String,
67}
68
69#[derive(Debug, Serialize, Deserialize, Clone)]
70pub struct RunOutput {
71 pub repository_diff: String,
72 pub diagnostics: String,
73 pub response_count: usize,
74 pub token_usage: TokenUsage,
75 pub tool_use_counts: HashMap<Arc<str>, u32>,
76}
77
78#[derive(Debug, Clone, Serialize, Deserialize)]
79pub struct JudgeInput {
80 pub repository_diff: String,
81 pub criteria: String,
82}
83
84#[derive(Debug, Clone, Serialize, Deserialize)]
85pub struct JudgeOutput {
86 pub analysis: String,
87 pub score: u32,
88}
89
90impl Example {
91 /// Load an example from a directory containing base.toml, prompt.md, and criteria.md
92 pub fn load_from_directory(dir_path: &Path, run_dir: &Path) -> Result<Self> {
93 let name = Self::name_from_path(dir_path);
94 let base_path = dir_path.join("base.toml");
95 let prompt_path = dir_path.join("prompt.md");
96 let criteria_path = dir_path.join("criteria.md");
97 let output_file_path = run_dir.join(format!("{}.md", name));
98
99 Ok(Example {
100 name: name.clone(),
101 base: toml::from_str(&fs::read_to_string(&base_path)?)?,
102 prompt: fs::read_to_string(prompt_path.clone())?,
103 criteria: fs::read_to_string(criteria_path.clone())?,
104 run_dir: run_dir.to_path_buf(),
105 output_file: None,
106 output_file_path,
107 log_prefix: name,
108 })
109 }
110
111 pub fn set_repetition_number(&mut self, repetition_number: u32) {
112 if repetition_number > 0 {
113 self.name = format!("{}-{}", self.name, repetition_number);
114 self.output_file_path = self.run_dir.join(format!("{}.md", self.name));
115 }
116 }
117
118 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
119 self.log_prefix = format!(
120 "{}{:<width$}\x1b[0m | ",
121 color,
122 self.name,
123 width = name_width
124 );
125 }
126
127 pub fn name_from_path(path: &Path) -> String {
128 path.file_name().unwrap().to_string_lossy().to_string()
129 }
130
131 pub fn worktree_path(&self) -> PathBuf {
132 Path::new(WORKTREES_DIR)
133 .canonicalize()
134 .context(format!("No such directory {WORKTREES_DIR}"))
135 .unwrap()
136 .join(&self.name)
137 }
138
139 /// Set up the example by checking out the specified Git revision
140 pub async fn setup(&mut self) -> Result<()> {
141 let repo_path = repo_path_for_url(&self.base.url);
142
143 println!("{}Fetching", self.log_prefix);
144
145 run_git(
146 &repo_path,
147 &["fetch", "--depth", "1", "origin", &self.base.revision],
148 )
149 .await?;
150
151 let worktree_path = self.worktree_path();
152
153 if worktree_path.is_dir() {
154 println!("{}Resetting existing worktree", self.log_prefix);
155
156 // TODO: consider including "-x" to remove ignored files. The downside of this is that
157 // it will also remove build artifacts, and so prevent incremental reuse there.
158 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
159 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
160 run_git(&worktree_path, &["checkout", &self.base.revision]).await?;
161 } else {
162 println!("{}Creating worktree", self.log_prefix);
163
164 let worktree_path_string = worktree_path.to_string_lossy().to_string();
165
166 run_git(
167 &repo_path,
168 &[
169 "worktree",
170 "add",
171 "-f",
172 &worktree_path_string,
173 &self.base.revision,
174 ],
175 )
176 .await?;
177 }
178
179 // Create the output file
180 let output_file = Arc::new(Mutex::new(File::create(&self.output_file_path)?));
181 self.output_file = Some(output_file);
182
183 Ok(())
184 }
185
186 /// Returns the output file, panicking if it's not set
187 fn output_file(&self) -> Arc<Mutex<File>> {
188 self.output_file
189 .clone()
190 .expect("Output file not created. Call setup() first.")
191 }
192
193 pub fn run(
194 &self,
195 model: Arc<dyn LanguageModel>,
196 app_state: Arc<AgentAppState>,
197 cx: &mut App,
198 ) -> Task<Result<RunOutput>> {
199 let project = Project::local(
200 app_state.client.clone(),
201 app_state.node_runtime.clone(),
202 app_state.user_store.clone(),
203 app_state.languages.clone(),
204 Arc::new(DapRegistry::default()),
205 app_state.fs.clone(),
206 None,
207 cx,
208 );
209
210 let worktree_path = self.worktree_path();
211 let worktree = project.update(cx, |project, cx| {
212 project.create_worktree(&worktree_path, true, cx)
213 });
214
215 let tools = cx.new(|_| ToolWorkingSet::default());
216 let thread_store =
217 ThreadStore::load(project.clone(), tools, app_state.prompt_builder.clone(), cx);
218 let this = self.clone();
219
220 cx.spawn(async move |cx| {
221 let worktree = worktree.await?;
222
223 // Wait for worktree scan to finish before choosing a file to open.
224 worktree
225 .update(cx, |worktree, _cx| {
226 worktree.as_local().unwrap().scan_complete()
227 })?
228 .await;
229
230 let lsp_open_handle_and_store = if this.base.require_lsp {
231 let language_extension = this.base.language_extension.as_deref().context(
232 "language_extension field is required in base.toml when `require_lsp == true`",
233 )?;
234
235 // Open a file that matches the language to cause LSP to start.
236 let language_file = worktree.read_with(cx, |worktree, _cx| {
237 worktree
238 .files(false, 0)
239 .find_map(|e| {
240 if e.path.clone().extension().and_then(|ext| ext.to_str())
241 == Some(language_extension)
242 {
243 Some(ProjectPath {
244 worktree_id: worktree.id(),
245 path: e.path.clone(),
246 })
247 } else {
248 None
249 }
250 })
251 .context("Failed to find a file for example language")
252 })??;
253
254 let open_language_file_buffer_task = project.update(cx, |project, cx| {
255 project.open_buffer(language_file.clone(), cx)
256 })?;
257
258 let language_file_buffer = open_language_file_buffer_task.await?;
259
260 let (lsp_open_handle, lsp_store) = project.update(cx, |project, cx| {
261 (
262 project.register_buffer_with_language_servers(&language_file_buffer, cx),
263 project.lsp_store().clone(),
264 )
265 })?;
266
267 // TODO: remove this once the diagnostics tool waits for new diagnostics
268 cx.background_executor().timer(Duration::new(5, 0)).await;
269 wait_for_lang_server(&lsp_store, this.log_prefix.clone(), cx).await?;
270
271 lsp_store.update(cx, |lsp_store, cx| {
272 lsp_open_handle.update(cx, |buffer, cx| {
273 buffer.update(cx, |buffer, cx| {
274 let has_language_server = lsp_store
275 .language_servers_for_local_buffer(buffer, cx)
276 .next()
277 .is_some();
278 if has_language_server {
279 Ok(())
280 } else {
281 Err(anyhow!(
282 "`{:?}` was opened to cause the language server to start, \
283 but no language servers are registered for its buffer. \
284 Set `require_lsp = false` in `base.toml` to skip this.",
285 language_file
286 ))
287 }
288 })
289 })
290 })??;
291
292 Some((lsp_open_handle, lsp_store))
293 } else {
294 None
295 };
296
297 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
298 return Err(anyhow!("Setup only mode"));
299 }
300
301 let thread_store = thread_store.await;
302 let thread =
303 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
304
305 {
306 let output_file_ref = this.output_file();
307 let mut output_file = output_file_ref.lock().unwrap();
308 writeln!(&mut output_file, "👤 USER:").log_err();
309 writeln!(&mut output_file, "{}", this.prompt).log_err();
310 writeln!(&mut output_file, "🤖 ASSISTANT:").log_err();
311 output_file.flush().log_err();
312 }
313
314 let tool_use_counts: Arc<Mutex<HashMap<Arc<str>, u32>>> =
315 Mutex::new(HashMap::default()).into();
316
317 let (thread_event_tx, mut thread_event_rx) = mpsc::unbounded();
318
319 let subscription = cx.subscribe(&thread, move |_thread, event: &ThreadEvent, _cx| {
320 thread_event_tx.unbounded_send(event.clone()).log_err();
321 });
322
323 let event_handler_task = cx.spawn({
324 // Need to clone the Arc here because the reference from output_file() won't live long enough
325 let output_file = this.output_file.clone().unwrap();
326 let log_prefix = this.log_prefix.clone();
327 let tool_use_counts = tool_use_counts.clone();
328 let thread = thread.downgrade();
329 async move |cx| {
330 loop {
331 let event = select_biased! {
332 event = thread_event_rx.next() => event,
333 _ = cx.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
334 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
335 }
336 };
337 let Some(event) = event else {
338 return Err(anyhow!("ThreadEvent channel ended early"));
339 };
340
341 let mut output_file = output_file.lock().unwrap();
342
343 match event {
344 ThreadEvent::Stopped(reason) => match reason {
345 Ok(StopReason::EndTurn) => {
346 return Ok(());
347 }
348 Ok(StopReason::MaxTokens) => {
349 return Err(anyhow!("Exceeded maximum tokens"));
350 }
351 Ok(StopReason::ToolUse) => {
352 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
353 println!("{}StopReason: Tool use", log_prefix);
354 }
355 }
356 Err(error) => {
357 return Err(anyhow!(error.clone()));
358 }
359 },
360 ThreadEvent::ShowError(thread_error) => {
361 break Err(anyhow!(thread_error.clone()));
362 }
363 ThreadEvent::StreamedAssistantText(_, chunk) => {
364 write!(&mut output_file, "{}", chunk).log_err();
365 }
366 ThreadEvent::StreamedAssistantThinking(_, chunk) => {
367 write!(&mut output_file, "{}", chunk).log_err();
368 }
369 ThreadEvent::UsePendingTools { tool_uses } => {
370 writeln!(&mut output_file, "\n\nUSING TOOLS:").log_err();
371 for tool_use in tool_uses {
372 writeln!(&mut output_file, "{}: {}", tool_use.name, tool_use.input)
373 .log_err();
374 }
375 }
376 ThreadEvent::ToolFinished {
377 tool_use_id,
378 pending_tool_use,
379 ..
380 } => {
381 thread.update(cx, |thread, _cx| {
382 if let Some(tool_use) = pending_tool_use {
383 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
384 let message = if tool_result.is_error {
385 format!("TOOL FAILED: {}", tool_use.name)
386 } else {
387 format!("TOOL FINISHED: {}", tool_use.name)
388 };
389 println!("{log_prefix}{message}");
390 writeln!(&mut output_file, "\n{}", message).log_err();
391 writeln!(&mut output_file, "\n{}\n", tool_result.content).log_err();
392 let mut tool_use_counts = tool_use_counts.lock().unwrap();
393 *tool_use_counts
394 .entry(tool_result.tool_name.clone())
395 .or_insert(0) += 1;
396 } else {
397 let message = format!("TOOL FINISHED WITHOUT RESULT: {}", tool_use.name);
398 println!("{log_prefix}{message}");
399 writeln!(&mut output_file, "\n{}", message).log_err();
400 }
401 }
402 })?;
403 }
404 ThreadEvent::ToolConfirmationNeeded => {
405 panic!("{}Bug: Tool confirmation should not be required in eval", log_prefix);
406 },
407 ThreadEvent::StreamedCompletion |
408 ThreadEvent::MessageAdded(_) |
409 ThreadEvent::MessageEdited(_) |
410 ThreadEvent::MessageDeleted(_) |
411 ThreadEvent::SummaryChanged |
412 ThreadEvent::SummaryGenerated |
413 ThreadEvent::CheckpointChanged => {
414 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
415 println!("{}Event: {:#?}", log_prefix, event);
416 }
417 }
418 }
419
420 output_file.flush().log_err();
421 }
422 }
423 });
424
425 thread.update(cx, |thread, cx| {
426 let context = vec![];
427 thread.insert_user_message(this.prompt.clone(), context, None, cx);
428 thread.send_to_model(model, RequestKind::Chat, cx);
429 })?;
430
431 event_handler_task.await?;
432
433 println!("{}Stopped", this.log_prefix);
434
435 if let Some((_, lsp_store)) = lsp_open_handle_and_store.as_ref() {
436 wait_for_lang_server(lsp_store, this.log_prefix.clone(), cx).await?;
437 }
438
439 println!("{}Getting repository diff", this.log_prefix);
440 let repository_diff = this.repository_diff().await?;
441
442 let repository_diff_path = this.run_dir.join(format!("{}.diff", this.name));
443 let mut repository_diff_output_file = File::create(&repository_diff_path)?;
444 writeln!(&mut repository_diff_output_file, "{}", &repository_diff).log_err();
445
446 println!("{}Getting diagnostics", this.log_prefix);
447 let diagnostics = cx
448 .update(move |cx| {
449 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
450 })?
451 .await?;
452 println!("{}Got diagnostics", this.log_prefix);
453
454 drop(subscription);
455 drop(lsp_open_handle_and_store);
456
457 thread.update(cx, |thread, _cx| {
458 let response_count = thread
459 .messages()
460 .filter(|message| message.role == language_model::Role::Assistant)
461 .count();
462 RunOutput {
463 repository_diff,
464 diagnostics,
465 response_count,
466 token_usage: thread.cumulative_token_usage(),
467 tool_use_counts: tool_use_counts.lock().unwrap().clone(),
468 }
469 })
470 })
471 }
472
473 pub async fn judge(
474 &self,
475 model: Arc<dyn LanguageModel>,
476 repository_diff: String,
477 judge_repetitions: u32,
478 cx: &AsyncApp,
479 ) -> Result<JudgeOutput> {
480 let judge_prompt = include_str!("judge_prompt.hbs");
481 let judge_prompt_name = "judge_prompt";
482 let mut handlebars = Handlebars::new();
483 handlebars.register_template_string(judge_prompt_name, judge_prompt)?;
484 let prompt = handlebars.render(
485 judge_prompt_name,
486 &JudgeInput {
487 repository_diff,
488 criteria: self.criteria.clone(),
489 },
490 )?;
491
492 let request = LanguageModelRequest {
493 messages: vec![LanguageModelRequestMessage {
494 role: Role::User,
495 content: vec![MessageContent::Text(prompt)],
496 cache: false,
497 }],
498 temperature: None,
499 tools: Vec::new(),
500 stop: Vec::new(),
501 };
502
503 let response = send_language_model_request(model, request, cx).await?;
504
505 let judge_file_path = self.run_dir.join(format!(
506 "{}_judge_{}.md",
507 self.name, // This is the eval_name
508 judge_repetitions
509 ));
510
511 let mut judge_output_file = File::create(&judge_file_path)?;
512 writeln!(&mut judge_output_file, "{}", &response).log_err();
513
514 parse_judge_output(&response)
515 }
516
517 pub async fn repository_diff(&self) -> Result<String> {
518 let worktree_path = self.worktree_path();
519 run_git(&worktree_path, &["add", "-N"]).await?;
520 run_git(&worktree_path, &["diff"]).await
521 }
522}
523
524fn wait_for_lang_server(
525 lsp_store: &Entity<LspStore>,
526 log_prefix: String,
527 cx: &mut AsyncApp,
528) -> Task<Result<()>> {
529 if cx
530 .update(|cx| !has_pending_lang_server_work(lsp_store, cx))
531 .unwrap()
532 || std::env::var("ZED_EVAL_SKIP_LS_WAIT").is_ok()
533 {
534 return Task::ready(anyhow::Ok(()));
535 }
536
537 println!("{}⏵ Waiting for language server", log_prefix);
538
539 let (mut tx, mut rx) = mpsc::channel(1);
540
541 let subscription =
542 cx.subscribe(&lsp_store, {
543 let log_prefix = log_prefix.clone();
544 move |lsp_store, event, cx| {
545 match event {
546 project::LspStoreEvent::LanguageServerUpdate {
547 message:
548 client::proto::update_language_server::Variant::WorkProgress(
549 LspWorkProgress {
550 message: Some(message),
551 ..
552 },
553 ),
554 ..
555 } => println!("{}⟲ {message}", log_prefix),
556 _ => {}
557 }
558
559 if !has_pending_lang_server_work(&lsp_store, cx) {
560 tx.try_send(()).ok();
561 }
562 }
563 });
564
565 cx.spawn(async move |cx| {
566 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
567 let result = futures::select! {
568 _ = rx.next() => {
569 println!("{}⚑ Language server idle", log_prefix);
570 anyhow::Ok(())
571 },
572 _ = timeout.fuse() => {
573 Err(anyhow!("LSP wait timed out after 5 minutes"))
574 }
575 };
576 drop(subscription);
577 result
578 })
579}
580
581fn has_pending_lang_server_work(lsp_store: &Entity<LspStore>, cx: &App) -> bool {
582 lsp_store
583 .read(cx)
584 .language_server_statuses()
585 .any(|(_, status)| !status.pending_work.is_empty())
586}
587
588async fn query_lsp_diagnostics(project: Entity<Project>, cx: &mut AsyncApp) -> Result<String> {
589 let paths_with_diagnostics = project.update(cx, |project, cx| {
590 project
591 .diagnostic_summaries(true, cx)
592 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
593 .map(|(project_path, _, _)| project_path)
594 .collect::<Vec<_>>()
595 })?;
596
597 let mut output = String::new();
598 for project_path in paths_with_diagnostics {
599 let buffer = project
600 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
601 .await?;
602 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
603
604 for (_, group) in snapshot.diagnostic_groups(None) {
605 let entry = &group.entries[group.primary_ix];
606 let range = entry.range.to_point(&snapshot);
607 let severity = match entry.diagnostic.severity {
608 DiagnosticSeverity::ERROR => "error",
609 DiagnosticSeverity::WARNING => "warning",
610 _ => continue,
611 };
612
613 writeln!(
614 output,
615 "{} at line {}: {}",
616 severity,
617 range.start.row + 1,
618 entry.diagnostic.message
619 )?;
620 }
621 }
622 anyhow::Ok(output)
623}
624
625fn parse_judge_output(response: &str) -> Result<JudgeOutput> {
626 let analysis = get_tag("analysis", response)?.to_string();
627 let score = get_tag("score", response)?
628 .parse()
629 .context("error parsing score")?;
630
631 Ok(JudgeOutput { analysis, score })
632}
633
634fn get_tag(name: &'static str, response: &str) -> Result<String> {
635 let start_tag = format!("<{}>", name);
636 let end_tag = format!("</{}>", name);
637
638 let start_ix = response
639 .find(&start_tag)
640 .context(format!("{} start tag not found", name))?;
641 let content_start_ix = start_ix + start_tag.len();
642
643 let end_ix = content_start_ix
644 + response[content_start_ix..]
645 .find(&end_tag)
646 .context(format!("{} end tag not found", name))?;
647
648 let content = response[content_start_ix..end_ix].trim().unindent();
649
650 anyhow::Ok(content)
651}
652
653pub fn repo_path_for_url(repo_url: &str) -> PathBuf {
654 let repo_name = repo_url
655 .trim_start_matches("https://")
656 .replace(|c: char| !c.is_alphanumeric(), "-");
657 Path::new(REPOS_DIR)
658 .canonicalize()
659 .context(format!("No such directory {REPOS_DIR}"))
660 .unwrap()
661 .join(repo_name)
662}
663
664pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
665 let output = new_smol_command("git")
666 .current_dir(repo_path)
667 .args(args)
668 .output()
669 .await?;
670
671 if output.status.success() {
672 Ok(String::from_utf8(output.stdout)?.trim().to_string())
673 } else {
674 Err(anyhow!(
675 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
676 args.join(" "),
677 repo_path.display(),
678 output.status,
679 String::from_utf8_lossy(&output.stderr),
680 String::from_utf8_lossy(&output.stdout),
681 ))
682 }
683}
684
685pub async fn send_language_model_request(
686 model: Arc<dyn LanguageModel>,
687 request: LanguageModelRequest,
688 cx: &AsyncApp,
689) -> anyhow::Result<String> {
690 match model.stream_completion_text(request, &cx).await {
691 Ok(mut stream) => {
692 let mut full_response = String::new();
693 while let Some(chunk_result) = stream.stream.next().await {
694 match chunk_result {
695 Ok(chunk_str) => {
696 full_response.push_str(&chunk_str);
697 }
698 Err(err) => {
699 return Err(anyhow!(
700 "Error receiving response from language model: {err}"
701 ));
702 }
703 }
704 }
705 Ok(full_response)
706 }
707 Err(err) => Err(anyhow!(
708 "Failed to get response from language model. Error was: {err}"
709 )),
710 }
711}
712
713#[cfg(test)]
714mod test {
715 use super::*;
716
717 #[test]
718 fn test_parse_judge_output() {
719 let response = r#"
720 <analysis>The model did a good job but there were still compilations errors.</analysis>
721 <score>3</score>
722 "#
723 .unindent();
724
725 let output = parse_judge_output(&response).unwrap();
726 assert_eq!(
727 output.analysis,
728 "The model did a good job but there were still compilations errors."
729 );
730 assert_eq!(output.score, 3);
731
732 let response = r#"
733 Text around ignored
734
735 <analysis>
736 Failed to compile:
737 - Error 1
738 - Error 2
739 </analysis>
740
741 <score>1</score>
742 "#
743 .unindent();
744
745 let output = parse_judge_output(&response).unwrap();
746 assert_eq!(output.analysis, "Failed to compile:\n- Error 1\n- Error 2");
747 assert_eq!(output.score, 1);
748 }
749}