1use crate::{AgentAppState, ToolMetrics};
2use agent::{ThreadEvent, ThreadStore};
3use anyhow::{Context as _, Result, anyhow};
4use assistant_tool::ToolWorkingSet;
5use client::proto::LspWorkProgress;
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::{Buffer, DiagnosticSeverity, OffsetRangeExt};
12use language_model::{
13 LanguageModel, LanguageModelCompletionEvent, LanguageModelRequest, LanguageModelRequestMessage,
14 MessageContent, Role, StopReason, TokenUsage,
15};
16use project::{Project, ProjectPath};
17use serde::{Deserialize, Serialize};
18use std::cell::RefCell;
19use std::fmt::Write as _;
20use std::fs::File;
21use std::io::Write as _;
22use std::rc::Rc;
23use std::sync::{Arc, Mutex};
24use std::time::Duration;
25use std::{
26 fs,
27 path::{Path, PathBuf},
28};
29use unindent::Unindent as _;
30use util::ResultExt as _;
31use util::command::new_smol_command;
32use util::markdown::MarkdownString;
33use util::serde::default_true;
34
35pub const EXAMPLES_DIR: &str = "./crates/eval/examples";
36pub const REPOS_DIR: &str = "./crates/eval/repos";
37pub const WORKTREES_DIR: &str = "./crates/eval/worktrees";
38
39const THREAD_EVENT_TIMEOUT: Duration = Duration::from_secs(60 * 2);
40
41#[derive(Clone, Debug, Deserialize)]
42pub struct ExampleBase {
43 pub url: String,
44 pub revision: String,
45 pub language_extension: Option<String>,
46 pub insert_id: Option<String>,
47 #[serde(default = "default_true")]
48 pub require_lsp: bool,
49 #[serde(default)]
50 pub allow_preexisting_diagnostics: bool,
51}
52
53impl ExampleBase {
54 pub fn repo_name(&self) -> String {
55 self.url
56 .split('/')
57 .next_back()
58 .unwrap_or(&"")
59 .trim_end_matches(".git")
60 .into()
61 }
62}
63
64#[derive(Clone, Debug)]
65pub struct Example {
66 pub name: String,
67 /// Content of `base.toml`
68 pub base: ExampleBase,
69 /// Content of `prompt.md`
70 pub prompt: String,
71 /// Content of `diff_criteria.md`
72 pub diff_criteria: String,
73 /// Content of `thread_criteria.md`, if that file exists (it's optional)
74 pub thread_criteria: Option<String>,
75 /// Path to the directory containing the requests and responses for the agentic loop
76 pub run_directory_path: PathBuf,
77 /// Prefix used for logging that identifies this example
78 pub log_prefix: String,
79}
80
81#[derive(Debug, Serialize, Deserialize, Clone)]
82pub struct RunOutput {
83 pub repository_diff: String,
84 pub ran_diagnostics_check: bool,
85 pub diagnostics_before: Option<String>,
86 pub diagnostics_after: Option<String>,
87 pub response_count: usize,
88 pub token_usage: TokenUsage,
89 pub tool_metrics: ToolMetrics,
90 pub last_request: LanguageModelRequest,
91}
92
93#[derive(Debug, Clone, Serialize, Deserialize)]
94pub struct JudgeDiffInput {
95 pub repository_diff: String,
96 pub ran_diagnostics_check: bool,
97 #[serde(skip_serializing_if = "Option::is_none")]
98 pub diagnostics_before: Option<String>,
99 #[serde(skip_serializing_if = "Option::is_none")]
100 pub diagnostics_after: Option<String>,
101 pub criteria: String,
102}
103
104#[derive(Debug, Clone, Serialize, Deserialize)]
105pub struct JudgeThreadInput {
106 pub messages: String,
107 pub criteria: String,
108}
109
110#[derive(Debug, Clone, Serialize, Deserialize)]
111pub struct JudgeResponse {
112 pub analysis: String,
113 pub score: u32,
114}
115
116#[derive(Debug, Clone, Serialize, Deserialize)]
117pub struct JudgeOutput {
118 pub thread: Option<JudgeResponse>,
119 pub diff: JudgeResponse,
120}
121
122impl Example {
123 /// Load an example from a directory containing base.toml, prompt.md, and criteria.md
124 pub fn load_from_directory(dir_path: &Path, run_dir: &Path) -> Result<Self> {
125 let name = Self::name_from_path(dir_path);
126 let base_path = dir_path.join("base.toml");
127 let prompt_path = dir_path.join("prompt.md");
128 let diff_criteria_path = dir_path.join("diff_criteria.md");
129 let thread_criteria_path = dir_path.join("thread_criteria.md");
130 let thread_criteria = if thread_criteria_path.exists() {
131 Some(fs::read_to_string(thread_criteria_path.clone())?)
132 } else {
133 None
134 };
135
136 Ok(Example {
137 name: name.clone(),
138 base: toml::from_str(&fs::read_to_string(&base_path)?)?,
139 prompt: fs::read_to_string(prompt_path.clone())?,
140 thread_criteria,
141 diff_criteria: fs::read_to_string(diff_criteria_path.clone())?,
142 run_directory_path: run_dir.to_path_buf(),
143 log_prefix: name,
144 })
145 }
146
147 pub fn set_repetition_number(&mut self, repetition_number: u32) {
148 if repetition_number > 0 {
149 self.name = format!("{}-{}", self.name, repetition_number);
150 }
151 }
152
153 pub fn example_output_directory(&self) -> PathBuf {
154 self.run_directory_path.join(&self.name)
155 }
156
157 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
158 self.log_prefix = format!(
159 "{}{:<width$}\x1b[0m | ",
160 color,
161 self.name,
162 width = name_width
163 );
164 }
165
166 pub fn name_from_path(path: &Path) -> String {
167 path.file_name().unwrap().to_string_lossy().to_string()
168 }
169
170 pub fn worktree_path(&self) -> PathBuf {
171 Path::new(WORKTREES_DIR)
172 .canonicalize()
173 .context(format!("No such directory {WORKTREES_DIR}"))
174 .unwrap()
175 .join(&self.name)
176 .join(self.base.repo_name())
177 }
178
179 /// Set up the example by checking out the specified Git revision
180 pub async fn setup(&mut self) -> Result<()> {
181 let repo_path = repo_path_for_url(&self.base.url);
182
183 let revision_exists = run_git(
184 &repo_path,
185 &["rev-parse", &format!("{}^{{commit}}", self.base.revision)],
186 )
187 .await
188 .is_ok();
189
190 if !revision_exists {
191 println!(
192 "{}Fetching revision {}",
193 self.log_prefix, &self.base.revision
194 );
195 run_git(
196 &repo_path,
197 &["fetch", "--depth", "1", "origin", &self.base.revision],
198 )
199 .await?;
200 }
201
202 let worktree_path = self.worktree_path();
203
204 if worktree_path.is_dir() {
205 println!("{}Resetting existing worktree", self.log_prefix);
206
207 // TODO: consider including "-x" to remove ignored files. The downside of this is that
208 // it will also remove build artifacts, and so prevent incremental reuse there.
209 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
210 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
211 run_git(&worktree_path, &["checkout", &self.base.revision]).await?;
212 } else {
213 println!("{}Creating worktree", self.log_prefix);
214
215 let worktree_path_string = worktree_path.to_string_lossy().to_string();
216
217 run_git(
218 &repo_path,
219 &[
220 "worktree",
221 "add",
222 "-f",
223 &worktree_path_string,
224 &self.base.revision,
225 ],
226 )
227 .await?;
228 }
229
230 std::fs::create_dir_all(self.example_output_directory())?;
231
232 Ok(())
233 }
234
235 pub fn run(
236 &self,
237 model: Arc<dyn LanguageModel>,
238 app_state: Arc<AgentAppState>,
239 cx: &mut App,
240 ) -> Task<Result<RunOutput>> {
241 let project = Project::local(
242 app_state.client.clone(),
243 app_state.node_runtime.clone(),
244 app_state.user_store.clone(),
245 app_state.languages.clone(),
246 Arc::new(DapRegistry::default()),
247 app_state.fs.clone(),
248 None,
249 cx,
250 );
251
252 let worktree_path = self.worktree_path();
253 let worktree = project.update(cx, |project, cx| {
254 project.create_worktree(&worktree_path, true, cx)
255 });
256
257 let tools = cx.new(|_| ToolWorkingSet::default());
258 let thread_store =
259 ThreadStore::load(project.clone(), tools, app_state.prompt_builder.clone(), cx);
260 let this = self.clone();
261
262 cx.spawn(async move |cx| {
263 let worktree = worktree.await?;
264
265 // Wait for worktree scan to finish before choosing a file to open.
266 worktree
267 .update(cx, |worktree, _cx| {
268 worktree.as_local().unwrap().scan_complete()
269 })?
270 .await;
271
272 let lsp = if this.base.require_lsp {
273 let language_extension = this.base.language_extension.as_deref().context(
274 "language_extension field is required in base.toml when `require_lsp == true`",
275 )?;
276
277 // Open a file that matches the language to cause LSP to start.
278 let language_file = worktree.read_with(cx, |worktree, _cx| {
279 worktree
280 .files(false, 0)
281 .find_map(|e| {
282 if e.path.clone().extension().and_then(|ext| ext.to_str())
283 == Some(language_extension)
284 {
285 Some(ProjectPath {
286 worktree_id: worktree.id(),
287 path: e.path.clone(),
288 })
289 } else {
290 None
291 }
292 })
293 .context("Failed to find a file for example language")
294 })??;
295
296 let open_language_file_buffer_task = project.update(cx, |project, cx| {
297 project.open_buffer(language_file.clone(), cx)
298 })?;
299
300 let language_file_buffer = open_language_file_buffer_task.await?;
301
302 let lsp_open_handle = project.update(cx, |project, cx| {
303 project.register_buffer_with_language_servers(&language_file_buffer, cx)
304 })?;
305
306 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
307
308 Some((lsp_open_handle, language_file_buffer))
309 } else {
310 None
311 };
312
313 let diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
314 if diagnostics_before.is_some() && !this.base.allow_preexisting_diagnostics {
315 return Err(anyhow!("Example has pre-existing diagnostics. If you want to run this example regardless, set `allow_preexisting_diagnostics` to `true` in `base.toml`"));
316 }
317
318 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
319 return Err(anyhow!("Setup only mode"));
320 }
321
322 let thread_store = thread_store.await?;
323 let thread =
324 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
325 let last_request = Rc::new(RefCell::new(None));
326
327 thread.update(cx, |thread, _cx| {
328 let mut request_count = 0;
329 let example_dir_path = this.example_output_directory();
330
331 let last_request = Rc::clone(&last_request);
332 thread.set_request_callback(move |request, response_events| {
333 *last_request.borrow_mut() = Some(request.clone());
334
335 request_count += 1;
336 let messages_file_path = example_dir_path.join(format!("{request_count}.messages.md"));
337 let last_messages_file_path = example_dir_path.join("last.messages.md");
338 let request_markdown = RequestMarkdown::new(request);
339 let response_events_markdown = response_events_to_markdown(response_events);
340
341 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
342 fs::write(messages_file_path, messages.clone()).expect("failed to write messages file");
343 fs::write(last_messages_file_path, messages).expect("failed to write last messages file");
344
345 if request_count == 1 {
346 let tools_file_path = example_dir_path.join("tools.md");
347 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
348 }
349 });
350 })?;
351
352 let tool_metrics = Arc::new(Mutex::new(ToolMetrics::default()));
353
354 let (thread_event_tx, mut thread_event_rx) = mpsc::unbounded();
355
356 let subscription = cx.subscribe(&thread, move |_thread, event: &ThreadEvent, _cx| {
357 thread_event_tx.unbounded_send(event.clone()).log_err();
358 });
359
360 let event_handler_task = cx.spawn({
361 let log_prefix = this.log_prefix.clone();
362 let tool_metrics = tool_metrics.clone();
363 let thread = thread.downgrade();
364 async move |cx| {
365 loop {
366 let event = select_biased! {
367 event = thread_event_rx.next() => event,
368 _ = cx.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
369 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
370 }
371 };
372 let Some(event) = event else {
373 return Err(anyhow!("ThreadEvent channel ended early"));
374 };
375
376 match event {
377 ThreadEvent::Stopped(reason) => match reason {
378 Ok(StopReason::EndTurn) => {
379 return Ok(());
380 }
381 Ok(StopReason::MaxTokens) => {
382 return Err(anyhow!("Exceeded maximum tokens"));
383 }
384 Ok(StopReason::ToolUse) => {
385 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
386 println!("{}StopReason: Tool use", log_prefix);
387 }
388 }
389 Err(error) => {
390 return Err(anyhow!(error.clone()));
391 }
392 },
393 ThreadEvent::ShowError(thread_error) => {
394 break Err(anyhow!(thread_error.clone()));
395 }
396 ThreadEvent::StreamedAssistantText(_, _)| ThreadEvent::StreamedAssistantThinking(_, _) | ThreadEvent::UsePendingTools { .. } => {
397 }
398 ThreadEvent::ToolFinished {
399 tool_use_id,
400 pending_tool_use,
401 ..
402 } => {
403 thread.update(cx, |thread, _cx| {
404 if let Some(tool_use) = pending_tool_use {
405 let mut tool_metrics = tool_metrics.lock().unwrap();
406 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
407 let message = if tool_result.is_error {
408 format!("TOOL FAILED: {}", tool_use.name)
409 } else {
410 format!("TOOL FINISHED: {}", tool_use.name)
411 };
412 println!("{log_prefix}{message}");
413 tool_metrics.insert(tool_result.tool_name.clone(), !tool_result.is_error);
414 } else {
415 let message = format!("TOOL FINISHED WITHOUT RESULT: {}", tool_use.name);
416 println!("{log_prefix}{message}");
417 tool_metrics.insert(tool_use.name.clone(), true);
418 }
419 }
420 })?;
421 }
422 ThreadEvent::ToolConfirmationNeeded => {
423 panic!("{}Bug: Tool confirmation should not be required in eval", log_prefix);
424 },
425 ThreadEvent::StreamedCompletion |
426 ThreadEvent::MessageAdded(_) |
427 ThreadEvent::MessageEdited(_) |
428 ThreadEvent::MessageDeleted(_) |
429 ThreadEvent::SummaryChanged |
430 ThreadEvent::SummaryGenerated |
431 ThreadEvent::CheckpointChanged |
432 ThreadEvent::UsageUpdated(_) => {
433 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
434 println!("{}Event: {:#?}", log_prefix, event);
435 }
436 }
437 }
438 }
439 }
440 });
441
442 thread.update(cx, |thread, cx| {
443 let context = vec![];
444 thread.insert_user_message(this.prompt.clone(), context, None, cx);
445 thread.send_to_model(model, cx);
446 })?;
447
448 event_handler_task.await?;
449
450 println!("{}Stopped", this.log_prefix);
451
452 if let Some((_, language_file_buffer)) = lsp.as_ref() {
453 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
454 }
455
456 println!("{}Getting repository diff", this.log_prefix);
457 let repository_diff = this.repository_diff().await?;
458
459 let example_output_dir = this.example_output_directory();
460 let repository_diff_path = example_output_dir.join("patch.diff");
461 let mut repository_diff_output_file = File::create(&repository_diff_path)?;
462 writeln!(&mut repository_diff_output_file, "{}", &repository_diff).log_err();
463
464 println!("{}Getting diagnostics", this.log_prefix);
465 let diagnostics_after = cx
466 .update(move |cx| {
467 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
468 })?
469 .await?;
470 println!("{}Got diagnostics", this.log_prefix);
471
472 let Some(last_request) = last_request.borrow_mut().take() else {
473 return Err(anyhow!("No requests ran."));
474 };
475
476 drop(subscription);
477 drop(lsp);
478
479 if let Some(diagnostics_before) = &diagnostics_before {
480 fs::write(example_output_dir.join("diagnostics_before.txt"), diagnostics_before)?;
481 }
482
483 if let Some(diagnostics_after) = &diagnostics_after {
484 fs::write(example_output_dir.join("diagnostics_after.txt"), diagnostics_after)?;
485 }
486
487
488 thread.update(cx, |thread, _cx| {
489 let response_count = thread
490 .messages()
491 .filter(|message| message.role == language_model::Role::Assistant)
492 .count();
493 RunOutput {
494 repository_diff,
495 ran_diagnostics_check: this.base.require_lsp,
496 diagnostics_before,
497 diagnostics_after,
498 response_count,
499 token_usage: thread.cumulative_token_usage(),
500 tool_metrics: tool_metrics.lock().unwrap().clone(),
501 last_request,
502 }
503 })
504 })
505 }
506
507 async fn judge_diff(
508 &self,
509 model: Arc<dyn LanguageModel>,
510 run_output: &RunOutput,
511 judge_number: u32,
512 cx: &AsyncApp,
513 ) -> Result<(String, JudgeResponse)> {
514 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
515 let judge_diff_prompt_name = "judge_diff_prompt";
516 let mut hbs = Handlebars::new();
517 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)?;
518
519 let diff_prompt = hbs.render(
520 judge_diff_prompt_name,
521 &JudgeDiffInput {
522 repository_diff: run_output.repository_diff.clone(),
523 ran_diagnostics_check: run_output.ran_diagnostics_check,
524 diagnostics_before: run_output.diagnostics_before.clone(),
525 diagnostics_after: run_output.diagnostics_after.clone(),
526 criteria: self.diff_criteria.clone(),
527 },
528 )?;
529
530 let request = LanguageModelRequest {
531 thread_id: None,
532 prompt_id: None,
533 messages: vec![LanguageModelRequestMessage {
534 role: Role::User,
535 content: vec![MessageContent::Text(diff_prompt)],
536 cache: false,
537 }],
538 temperature: None,
539 tools: Vec::new(),
540 stop: Vec::new(),
541 };
542
543 let diff_response = send_language_model_request(model, request, cx).await?;
544 let diff_output = JudgeResponse::parse(&diff_response)?;
545
546 println!(
547 "{}Judge #{judge_number} - Diff score: {}",
548 self.log_prefix, diff_output.score
549 );
550
551 Ok((diff_response, diff_output))
552 }
553
554 async fn judge_thread(
555 &self,
556 model: Arc<dyn LanguageModel>,
557 run_output: &RunOutput,
558 judge_number: u32,
559 cx: &AsyncApp,
560 ) -> Result<(String, Option<JudgeResponse>)> {
561 if let Some(criteria) = self.thread_criteria.clone() {
562 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
563 let judge_thread_prompt_name = "judge_thread_prompt";
564 let mut hbs = Handlebars::new();
565 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)?;
566
567 let request_markdown = RequestMarkdown::new(&run_output.last_request);
568 let thread_prompt = hbs.render(
569 judge_thread_prompt_name,
570 &JudgeThreadInput {
571 messages: request_markdown.messages,
572 criteria,
573 },
574 )?;
575
576 let request = LanguageModelRequest {
577 thread_id: None,
578 prompt_id: None,
579 messages: vec![LanguageModelRequestMessage {
580 role: Role::User,
581 content: vec![MessageContent::Text(thread_prompt)],
582 cache: false,
583 }],
584 temperature: None,
585 tools: Vec::new(),
586 stop: Vec::new(),
587 };
588
589 let thread_response = send_language_model_request(model, request, cx).await?;
590 let thread_output = JudgeResponse::parse(&thread_response)?;
591
592 println!(
593 "{}Judge #{judge_number} - Thread score: {}",
594 self.log_prefix, thread_output.score
595 );
596
597 Ok((thread_response, Some(thread_output)))
598 } else {
599 let msg = "There were no criteria specified for this thread, so this example was not judged on its thread.".to_string();
600 Ok((msg, None))
601 }
602 }
603
604 pub async fn judge(
605 &self,
606 model: Arc<dyn LanguageModel>,
607 run_output: &RunOutput,
608 judge_number: u32,
609 cx: &AsyncApp,
610 ) -> Result<JudgeOutput> {
611 let mut output_file = File::create(
612 self.example_output_directory()
613 .join(format!("judge_{}.md", judge_number)),
614 )
615 .expect("failed to create judge.md");
616
617 println!("{}Running judge #{judge_number}", self.log_prefix);
618
619 let diff_task = self.judge_diff(model.clone(), &run_output, judge_number, cx);
620 let thread_task = self.judge_thread(model.clone(), &run_output, judge_number, cx);
621
622 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
623
624 let (diff_response, diff_output) = diff_result?;
625 let (thread_response, thread_output) = thread_result?;
626
627 writeln!(
628 &mut output_file,
629 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
630 )
631 .log_err();
632
633 Ok(JudgeOutput {
634 thread: thread_output,
635 diff: diff_output,
636 })
637 }
638
639 async fn repository_diff(&self) -> Result<String> {
640 let worktree_path = self.worktree_path();
641 run_git(&worktree_path, &["add", "."]).await?;
642 run_git(&worktree_path, &["diff", "--staged"]).await
643 }
644}
645
646fn wait_for_lang_server(
647 project: &Entity<Project>,
648 buffer: &Entity<Buffer>,
649 log_prefix: String,
650 cx: &mut AsyncApp,
651) -> Task<Result<()>> {
652 println!("{}⏵ Waiting for language server", log_prefix);
653
654 let (mut tx, mut rx) = mpsc::channel(1);
655
656 let lsp_store = project
657 .update(cx, |project, _| project.lsp_store())
658 .unwrap();
659
660 let has_lang_server = buffer
661 .update(cx, |buffer, cx| {
662 lsp_store.update(cx, |lsp_store, cx| {
663 lsp_store
664 .language_servers_for_local_buffer(&buffer, cx)
665 .next()
666 .is_some()
667 })
668 })
669 .unwrap_or(false);
670
671 if has_lang_server {
672 project
673 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
674 .unwrap()
675 .detach();
676 }
677
678 let subscriptions =
679 [
680 cx.subscribe(&lsp_store, {
681 let log_prefix = log_prefix.clone();
682 move |_, event, _| match event {
683 project::LspStoreEvent::LanguageServerUpdate {
684 message:
685 client::proto::update_language_server::Variant::WorkProgress(
686 LspWorkProgress {
687 message: Some(message),
688 ..
689 },
690 ),
691 ..
692 } => println!("{}⟲ {message}", log_prefix),
693 _ => {}
694 }
695 }),
696 cx.subscribe(&project, {
697 let buffer = buffer.clone();
698 move |project, event, cx| match event {
699 project::Event::LanguageServerAdded(_, _, _) => {
700 let buffer = buffer.clone();
701 project
702 .update(cx, |project, cx| project.save_buffer(buffer, cx))
703 .detach();
704 }
705 project::Event::DiskBasedDiagnosticsFinished { .. } => {
706 tx.try_send(()).ok();
707 }
708 _ => {}
709 }
710 }),
711 ];
712
713 cx.spawn(async move |cx| {
714 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
715 let result = futures::select! {
716 _ = rx.next() => {
717 println!("{}⚑ Language server idle", log_prefix);
718 anyhow::Ok(())
719 },
720 _ = timeout.fuse() => {
721 Err(anyhow!("LSP wait timed out after 5 minutes"))
722 }
723 };
724 drop(subscriptions);
725 result
726 })
727}
728
729async fn query_lsp_diagnostics(
730 project: Entity<Project>,
731 cx: &mut AsyncApp,
732) -> Result<Option<String>> {
733 let paths_with_diagnostics = project.update(cx, |project, cx| {
734 project
735 .diagnostic_summaries(true, cx)
736 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
737 .map(|(project_path, _, _)| project_path)
738 .collect::<Vec<_>>()
739 })?;
740
741 if paths_with_diagnostics.is_empty() {
742 return Ok(None);
743 }
744
745 let mut output = String::new();
746 for project_path in paths_with_diagnostics {
747 let buffer = project
748 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
749 .await?;
750 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
751
752 for (_, group) in snapshot.diagnostic_groups(None) {
753 let entry = &group.entries[group.primary_ix];
754 let range = entry.range.to_point(&snapshot);
755 let severity = match entry.diagnostic.severity {
756 DiagnosticSeverity::ERROR => "error",
757 DiagnosticSeverity::WARNING => "warning",
758 _ => continue,
759 };
760
761 writeln!(
762 output,
763 "{} at line {}: {}",
764 severity,
765 range.start.row + 1,
766 entry.diagnostic.message
767 )?;
768 }
769 }
770 anyhow::Ok(Some(output))
771}
772
773impl JudgeResponse {
774 fn parse(response: &str) -> Result<Self> {
775 let analysis = get_tag("analysis", response)?.to_string();
776 let score = get_tag("score", response)?
777 .parse()
778 .context("error parsing score")?;
779
780 Ok(Self { analysis, score })
781 }
782}
783
784fn get_tag(name: &'static str, response: &str) -> Result<String> {
785 let start_tag = format!("<{}>", name);
786 let end_tag = format!("</{}>", name);
787
788 let start_ix = response
789 .find(&start_tag)
790 .context(format!("{} start tag not found", name))?;
791 let content_start_ix = start_ix + start_tag.len();
792
793 let end_ix = content_start_ix
794 + response[content_start_ix..]
795 .find(&end_tag)
796 .context(format!("{} end tag not found", name))?;
797
798 let content = response[content_start_ix..end_ix].trim().unindent();
799
800 anyhow::Ok(content)
801}
802
803pub fn repo_path_for_url(repo_url: &str) -> PathBuf {
804 let repo_name = repo_url
805 .trim_start_matches("https://")
806 .replace(|c: char| !c.is_alphanumeric(), "-");
807 Path::new(REPOS_DIR)
808 .canonicalize()
809 .context(format!("No such directory {REPOS_DIR}"))
810 .unwrap()
811 .join(repo_name)
812}
813
814pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
815 let output = new_smol_command("git")
816 .current_dir(repo_path)
817 .args(args)
818 .output()
819 .await?;
820
821 if output.status.success() {
822 Ok(String::from_utf8(output.stdout)?.trim().to_string())
823 } else {
824 Err(anyhow!(
825 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
826 args.join(" "),
827 repo_path.display(),
828 output.status,
829 String::from_utf8_lossy(&output.stderr),
830 String::from_utf8_lossy(&output.stdout),
831 ))
832 }
833}
834
835pub async fn send_language_model_request(
836 model: Arc<dyn LanguageModel>,
837 request: LanguageModelRequest,
838 cx: &AsyncApp,
839) -> anyhow::Result<String> {
840 match model.stream_completion_text(request, &cx).await {
841 Ok(mut stream) => {
842 let mut full_response = String::new();
843 while let Some(chunk_result) = stream.stream.next().await {
844 match chunk_result {
845 Ok(chunk_str) => {
846 full_response.push_str(&chunk_str);
847 }
848 Err(err) => {
849 return Err(anyhow!(
850 "Error receiving response from language model: {err}"
851 ));
852 }
853 }
854 }
855 Ok(full_response)
856 }
857 Err(err) => Err(anyhow!(
858 "Failed to get response from language model. Error was: {err}"
859 )),
860 }
861}
862
863struct RequestMarkdown {
864 tools: String,
865 messages: String,
866}
867
868impl RequestMarkdown {
869 fn new(request: &LanguageModelRequest) -> Self {
870 let mut tools = String::new();
871 let mut messages = String::new();
872 let mut assistant_message_number: u32 = 1;
873
874 // Print the tools
875 if !request.tools.is_empty() {
876 for tool in &request.tools {
877 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
878 write!(&mut tools, "{}\n\n", tool.description).unwrap();
879 write!(
880 &mut tools,
881 "{}\n",
882 MarkdownString::code_block("json", &format!("{:#}", tool.input_schema))
883 )
884 .unwrap();
885 }
886 }
887
888 // Print the messages
889 for message in &request.messages {
890 match message.role {
891 Role::System => messages.push_str("# ⚙️ SYSTEM\n\n"),
892 Role::User => messages.push_str("# 👤 USER\n\n"),
893 Role::Assistant => {
894 messages.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
895 assistant_message_number += 1;
896 }
897 };
898
899 for content in &message.content {
900 match content {
901 MessageContent::Text(text) => {
902 messages.push_str(text);
903 messages.push_str("\n\n");
904 }
905 MessageContent::Image(_) => {
906 messages.push_str("[IMAGE DATA]\n\n");
907 }
908 MessageContent::Thinking { text, signature } => {
909 messages.push_str("**Thinking**:\n\n");
910 if let Some(sig) = signature {
911 messages.push_str(&format!("Signature: {}\n\n", sig));
912 }
913 messages.push_str(text);
914 messages.push_str("\n");
915 }
916 MessageContent::RedactedThinking(items) => {
917 messages.push_str(&format!(
918 "**Redacted Thinking**: {} item(s)\n\n",
919 items.len()
920 ));
921 }
922 MessageContent::ToolUse(tool_use) => {
923 messages.push_str(&format!(
924 "**Tool Use**: {} (ID: {})\n",
925 tool_use.name, tool_use.id
926 ));
927 messages.push_str(&format!(
928 "{}\n",
929 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
930 ));
931 }
932 MessageContent::ToolResult(tool_result) => {
933 messages.push_str(&format!(
934 "**Tool Result**: {} (ID: {})\n\n",
935 tool_result.tool_name, tool_result.tool_use_id
936 ));
937 if tool_result.is_error {
938 messages.push_str("**ERROR:**\n");
939 }
940 messages.push_str(&format!("{}\n\n", tool_result.content));
941 }
942 }
943 }
944 }
945
946 Self { tools, messages }
947 }
948}
949
950fn response_events_to_markdown(
951 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
952) -> String {
953 let mut response = String::new();
954 // Print the response events if any
955 response.push_str("# Response\n\n");
956 let mut text_buffer = String::new();
957 let mut thinking_buffer = String::new();
958
959 let flush_buffers =
960 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
961 if !text_buffer.is_empty() {
962 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
963 text_buffer.clear();
964 }
965 if !thinking_buffer.is_empty() {
966 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
967 thinking_buffer.clear();
968 }
969 };
970
971 for event in response_events {
972 match event {
973 Ok(LanguageModelCompletionEvent::Text(text)) => {
974 text_buffer.push_str(text);
975 }
976 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
977 thinking_buffer.push_str(text);
978 }
979 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
980 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
981 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
982 }
983 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
984 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
985 response.push_str(&format!(
986 "**Tool Use**: {} (ID: {})\n",
987 tool_use.name, tool_use.id
988 ));
989 response.push_str(&format!(
990 "{}\n",
991 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
992 ));
993 }
994 Ok(
995 LanguageModelCompletionEvent::UsageUpdate(_)
996 | LanguageModelCompletionEvent::StartMessage { .. },
997 ) => {}
998 Err(error) => {
999 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1000 response.push_str(&format!("**Error**: {}\n\n", error));
1001 }
1002 }
1003 }
1004
1005 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1006
1007 response
1008}
1009
1010#[cfg(test)]
1011mod test {
1012 use super::*;
1013 use handlebars::Handlebars;
1014
1015 #[test]
1016 fn test_parse_judge_output() {
1017 let response = r#"
1018 <analysis>The model did a good job but there were still compilations errors.</analysis>
1019 <score>3</score>
1020 "#
1021 .unindent();
1022
1023 let output = JudgeResponse::parse(&response).unwrap();
1024 assert_eq!(
1025 output.analysis,
1026 "The model did a good job but there were still compilations errors."
1027 );
1028 assert_eq!(output.score, 3);
1029
1030 let response = r#"
1031 Text around ignored
1032
1033 <analysis>
1034 Failed to compile:
1035 - Error 1
1036 - Error 2
1037 </analysis>
1038
1039 <score>1</score>
1040 "#
1041 .unindent();
1042
1043 let output = JudgeResponse::parse(&response).unwrap();
1044 assert_eq!(output.analysis, "Failed to compile:\n- Error 1\n- Error 2");
1045 assert_eq!(output.score, 1);
1046 }
1047
1048 #[test]
1049 fn test_judge_prompt_with_diagnostics() {
1050 // Case 1: Both diagnostics before and after are present
1051 let input = JudgeDiffInput {
1052 repository_diff: "diff content goes here".to_string(),
1053 ran_diagnostics_check: true,
1054 diagnostics_before: Some("Error at line 10: variable not found".to_string()),
1055 diagnostics_after: Some("Error at line 15: missing semicolon".to_string()),
1056 criteria: "Fix all bugs".to_string(),
1057 };
1058
1059 let rendered = templates().render(JUDGE_PROMPT_NAME, &input).unwrap();
1060
1061 let expected_diagnostics_section = r#"
1062 Take into account the diagnostics before and after applying the change:
1063
1064 <diagnostics_before>
1065 Error at line 10: variable not found
1066 </diagnostics_before>
1067
1068 <diagnostics_after>
1069 Error at line 15: missing semicolon
1070 </diagnostics_after>
1071 "#
1072 .unindent();
1073
1074 assert!(rendered.contains(&expected_diagnostics_section));
1075 }
1076
1077 #[test]
1078 fn test_judge_prompt_with_empty_diagnostics() {
1079 // Case 2: Diagnostics check run but no diagnostics found
1080 let input = JudgeDiffInput {
1081 repository_diff: "diff content goes here".to_string(),
1082 ran_diagnostics_check: true,
1083 diagnostics_before: None,
1084 diagnostics_after: None,
1085 criteria: "Fix all bugs".to_string(),
1086 };
1087
1088 let rendered = templates().render(JUDGE_PROMPT_NAME, &input).unwrap();
1089
1090 let expected_diagnostics_section = r#"
1091 Take into account the diagnostics before and after applying the change:
1092
1093 <diagnostics_before>
1094 No diagnostics before applying the edits.
1095 </diagnostics_before>
1096
1097 <diagnostics_after>
1098 No diagnostics after applying the edits.
1099 </diagnostics_after>
1100 "#
1101 .unindent();
1102
1103 assert!(rendered.contains(&expected_diagnostics_section));
1104 }
1105
1106 #[test]
1107 fn test_judge_prompt_with_mixed_diagnostics() {
1108 let templates = templates();
1109
1110 // Case 3: Before diagnostics present, after diagnostics absent
1111 let input = JudgeDiffInput {
1112 repository_diff: "diff content goes here".to_string(),
1113 ran_diagnostics_check: true,
1114 diagnostics_before: Some("Error at line 10: variable not found".to_string()),
1115 diagnostics_after: None,
1116 criteria: "Fix all bugs".to_string(),
1117 };
1118
1119 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1120
1121 let expected_diagnostics_section = r#"
1122 Take into account the diagnostics before and after applying the change:
1123
1124 <diagnostics_before>
1125 Error at line 10: variable not found
1126 </diagnostics_before>
1127
1128 <diagnostics_after>
1129 No diagnostics after applying the edits.
1130 </diagnostics_after>
1131 "#
1132 .unindent();
1133
1134 assert!(rendered.contains(&expected_diagnostics_section));
1135
1136 // Case 4: Before diagnostics absent, after diagnostics present
1137 let input = JudgeDiffInput {
1138 repository_diff: "diff content goes here".to_string(),
1139 ran_diagnostics_check: true,
1140 diagnostics_before: None,
1141 diagnostics_after: Some("Error at line 15: missing semicolon".to_string()),
1142 criteria: "Fix all bugs".to_string(),
1143 };
1144
1145 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1146
1147 let expected_diagnostics_section = r#"
1148 Take into account the diagnostics before and after applying the change:
1149
1150 <diagnostics_before>
1151 No diagnostics before applying the edits.
1152 </diagnostics_before>
1153
1154 <diagnostics_after>
1155 Error at line 15: missing semicolon
1156 </diagnostics_after>
1157 "#
1158 .unindent();
1159
1160 assert!(rendered.contains(&expected_diagnostics_section));
1161 }
1162
1163 #[test]
1164 fn test_judge_prompt_without_diagnostics() {
1165 let templates = templates();
1166
1167 // Case 5: No diagnostics check run
1168 let input = JudgeDiffInput {
1169 repository_diff: "diff content goes here".to_string(),
1170 ran_diagnostics_check: false,
1171 diagnostics_before: None,
1172 diagnostics_after: None,
1173 criteria: "Fix all bugs".to_string(),
1174 };
1175
1176 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1177
1178 // Check for the message when no diagnostics were performed
1179 let diagnostics_message = "No diagnostic checks were performed.";
1180
1181 assert!(rendered.contains(diagnostics_message));
1182 assert!(!rendered.contains("<diagnostics_before>"));
1183 assert!(!rendered.contains("<diagnostics_after>"));
1184 }
1185
1186 const JUDGE_PROMPT_NAME: &str = "judge_prompt";
1187
1188 fn templates() -> Handlebars<'static> {
1189 let mut judge_prompt = include_str!("judge_diff_prompt.hbs").to_string();
1190 language::LineEnding::normalize(&mut judge_prompt);
1191 let mut handlebars = Handlebars::new();
1192 handlebars
1193 .register_template_string(JUDGE_PROMPT_NAME, judge_prompt)
1194 .unwrap();
1195 handlebars
1196 }
1197}