1use crate::{AgentAppState, ToolMetrics};
2use agent::{ThreadEvent, ThreadStore};
3use anyhow::{Context as _, Result, anyhow};
4use assistant_tool::ToolWorkingSet;
5use client::proto::LspWorkProgress;
6use futures::channel::mpsc;
7use futures::{FutureExt, StreamExt as _, select_biased};
8use gpui::{App, AppContext as _, AsyncApp, Entity, Task};
9use handlebars::Handlebars;
10use language::{Buffer, DiagnosticSeverity, OffsetRangeExt};
11use language_model::{
12 LanguageModel, LanguageModelCompletionEvent, LanguageModelRequest, LanguageModelRequestMessage,
13 MessageContent, Role, StopReason, TokenUsage,
14};
15use project::{Project, ProjectPath};
16use serde::{Deserialize, Serialize};
17use std::cell::RefCell;
18use std::fmt::Write as _;
19use std::fs::File;
20use std::io::Write as _;
21use std::rc::Rc;
22use std::sync::{Arc, Mutex};
23use std::time::Duration;
24use std::{
25 fs,
26 path::{Path, PathBuf},
27};
28use unindent::Unindent as _;
29use util::ResultExt as _;
30use util::command::new_smol_command;
31use util::markdown::MarkdownString;
32use util::serde::default_true;
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
40const ZED_REPO_URL: &str = "https://github.com/zed-industries/zed.git";
41
42#[derive(Clone, Debug, Deserialize)]
43pub struct ExampleBase {
44 pub url: String,
45 pub revision: String,
46 pub language_extension: Option<String>,
47 pub insert_id: Option<String>,
48 #[serde(default = "default_true")]
49 pub require_lsp: bool,
50 #[serde(default)]
51 pub allow_preexisting_diagnostics: bool,
52}
53
54impl ExampleBase {
55 pub fn repo_name(&self) -> String {
56 self.url
57 .split('/')
58 .next_back()
59 .unwrap_or(&"")
60 .trim_end_matches(".git")
61 .into()
62 }
63}
64
65#[derive(Clone, Debug)]
66pub struct Example {
67 pub name: String,
68 /// Content of `base.toml`
69 pub base: ExampleBase,
70 /// Content of `prompt.md`
71 pub prompt: String,
72 /// Content of `diff_criteria.md`
73 pub diff_criteria: String,
74 /// Content of `thread_criteria.md`, if that file exists (it's optional)
75 pub thread_criteria: Option<String>,
76 /// Path to the directory containing the requests and responses for the agentic loop
77 pub run_directory_path: PathBuf,
78 /// Prefix used for logging that identifies this example
79 pub log_prefix: String,
80}
81
82#[derive(Debug, Serialize, Deserialize, Clone)]
83pub struct RunOutput {
84 pub repository_diff: String,
85 pub ran_diagnostics_check: bool,
86 pub diagnostics_before: Option<String>,
87 pub diagnostics_after: Option<String>,
88 pub response_count: usize,
89 pub token_usage: TokenUsage,
90 pub tool_metrics: ToolMetrics,
91 pub last_request: LanguageModelRequest,
92}
93
94#[derive(Debug, Clone, Serialize, Deserialize)]
95pub struct JudgeDiffInput {
96 pub repository_diff: String,
97 pub ran_diagnostics_check: bool,
98 #[serde(skip_serializing_if = "Option::is_none")]
99 pub diagnostics_before: Option<String>,
100 #[serde(skip_serializing_if = "Option::is_none")]
101 pub diagnostics_after: Option<String>,
102 pub criteria: String,
103}
104
105#[derive(Debug, Clone, Serialize, Deserialize)]
106pub struct JudgeThreadInput {
107 pub messages: String,
108 pub criteria: String,
109}
110
111#[derive(Debug, Clone, Serialize, Deserialize)]
112pub struct JudgeResponse {
113 pub analysis: String,
114 pub score: u32,
115}
116
117#[derive(Debug, Clone, Serialize, Deserialize)]
118pub struct JudgeOutput {
119 pub thread: Option<JudgeResponse>,
120 pub diff: JudgeResponse,
121}
122
123impl Example {
124 /// Load an example from a directory containing base.toml, prompt.md, and criteria.md
125 pub fn load_from_directory(dir_path: &Path, run_dir: &Path) -> Result<Self> {
126 let name = Self::name_from_path(dir_path);
127 let base_path = dir_path.join("base.toml");
128 let prompt_path = dir_path.join("prompt.md");
129 let diff_criteria_path = dir_path.join("diff_criteria.md");
130 let thread_criteria_path = dir_path.join("thread_criteria.md");
131 let thread_criteria = if thread_criteria_path.exists() {
132 Some(fs::read_to_string(thread_criteria_path.clone())?)
133 } else {
134 None
135 };
136
137 Ok(Example {
138 name: name.clone(),
139 base: toml::from_str(&fs::read_to_string(&base_path)?)?,
140 prompt: fs::read_to_string(prompt_path.clone())?,
141 thread_criteria,
142 diff_criteria: fs::read_to_string(diff_criteria_path.clone())?,
143 run_directory_path: run_dir.to_path_buf(),
144 log_prefix: name,
145 })
146 }
147
148 pub fn set_repetition_number(&mut self, repetition_number: u32) {
149 if repetition_number > 0 {
150 self.name = format!("{}-{}", self.name, repetition_number);
151 }
152 }
153
154 pub fn example_output_directory(&self) -> PathBuf {
155 self.run_directory_path.join(&self.name)
156 }
157
158 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
159 self.log_prefix = format!(
160 "{}{:<width$}\x1b[0m | ",
161 color,
162 self.name,
163 width = name_width
164 );
165 }
166
167 pub fn name_from_path(path: &Path) -> String {
168 path.file_name().unwrap().to_string_lossy().to_string()
169 }
170
171 pub fn worktree_path(&self) -> PathBuf {
172 Path::new(WORKTREES_DIR)
173 .canonicalize()
174 .context(format!("No such directory {WORKTREES_DIR}"))
175 .unwrap()
176 .join(&self.name)
177 .join(self.base.repo_name())
178 }
179
180 /// Set up the example by checking out the specified Git revision
181 pub async fn setup(&mut self) -> Result<()> {
182 let repo_path = repo_path_for_url(&self.base.url);
183
184 let revision_exists = run_git(
185 &repo_path,
186 &["rev-parse", &format!("{}^{{commit}}", self.base.revision)],
187 )
188 .await
189 .is_ok();
190
191 if !revision_exists {
192 println!(
193 "{}Fetching revision {}",
194 self.log_prefix, &self.base.revision
195 );
196 run_git(
197 &repo_path,
198 &["fetch", "--depth", "1", "origin", &self.base.revision],
199 )
200 .await?;
201 }
202
203 let worktree_path = self.worktree_path();
204
205 if worktree_path.is_dir() {
206 println!("{}Resetting existing worktree", self.log_prefix);
207
208 // TODO: consider including "-x" to remove ignored files. The downside of this is that
209 // it will also remove build artifacts, and so prevent incremental reuse there.
210 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
211 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
212 run_git(&worktree_path, &["checkout", &self.base.revision]).await?;
213 } else {
214 println!("{}Creating worktree", self.log_prefix);
215
216 let worktree_path_string = worktree_path.to_string_lossy().to_string();
217
218 run_git(
219 &repo_path,
220 &[
221 "worktree",
222 "add",
223 "-f",
224 &worktree_path_string,
225 &self.base.revision,
226 ],
227 )
228 .await?;
229 }
230
231 if self.base.url == ZED_REPO_URL {
232 std::fs::write(worktree_path.join(".rules"), std::fs::read(".rules")?)?;
233 }
234
235 std::fs::create_dir_all(self.example_output_directory())?;
236
237 Ok(())
238 }
239
240 pub fn run(
241 &self,
242 model: Arc<dyn LanguageModel>,
243 app_state: Arc<AgentAppState>,
244 cx: &mut App,
245 ) -> Task<Result<RunOutput>> {
246 let project = Project::local(
247 app_state.client.clone(),
248 app_state.node_runtime.clone(),
249 app_state.user_store.clone(),
250 app_state.languages.clone(),
251 app_state.fs.clone(),
252 None,
253 cx,
254 );
255
256 let worktree_path = self.worktree_path();
257 let worktree = project.update(cx, |project, cx| {
258 project.create_worktree(&worktree_path, true, cx)
259 });
260
261 let tools = cx.new(|_| ToolWorkingSet::default());
262 let thread_store =
263 ThreadStore::load(project.clone(), tools, app_state.prompt_builder.clone(), cx);
264 let this = self.clone();
265
266 cx.spawn(async move |cx| {
267 let worktree = worktree.await?;
268
269 // Wait for worktree scan to finish before choosing a file to open.
270 worktree
271 .update(cx, |worktree, _cx| {
272 worktree.as_local().unwrap().scan_complete()
273 })?
274 .await;
275
276 let lsp = if this.base.require_lsp {
277 let language_extension = this.base.language_extension.as_deref().context(
278 "language_extension field is required in base.toml when `require_lsp == true`",
279 )?;
280
281 // Open a file that matches the language to cause LSP to start.
282 let language_file = worktree.read_with(cx, |worktree, _cx| {
283 worktree
284 .files(false, 0)
285 .find_map(|e| {
286 if e.path.clone().extension().and_then(|ext| ext.to_str())
287 == Some(language_extension)
288 {
289 Some(ProjectPath {
290 worktree_id: worktree.id(),
291 path: e.path.clone(),
292 })
293 } else {
294 None
295 }
296 })
297 .context("Failed to find a file for example language")
298 })??;
299
300 let open_language_file_buffer_task = project.update(cx, |project, cx| {
301 project.open_buffer(language_file.clone(), cx)
302 })?;
303
304 let language_file_buffer = open_language_file_buffer_task.await?;
305
306 let lsp_open_handle = project.update(cx, |project, cx| {
307 project.register_buffer_with_language_servers(&language_file_buffer, cx)
308 })?;
309
310 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
311
312 Some((lsp_open_handle, language_file_buffer))
313 } else {
314 None
315 };
316
317 let diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
318 if diagnostics_before.is_some() && !this.base.allow_preexisting_diagnostics {
319 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`"));
320 }
321
322 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
323 return Err(anyhow!("Setup only mode"));
324 }
325
326 let thread_store = thread_store.await?;
327 let thread =
328 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
329 let last_request = Rc::new(RefCell::new(None));
330
331 thread.update(cx, |thread, _cx| {
332 let mut request_count = 0;
333 let example_dir_path = this.example_output_directory();
334
335 let last_request = Rc::clone(&last_request);
336 thread.set_request_callback(move |request, response_events| {
337 *last_request.borrow_mut() = Some(request.clone());
338
339 request_count += 1;
340 let messages_file_path = example_dir_path.join(format!("{request_count}.messages.md"));
341 let last_messages_file_path = example_dir_path.join("last.messages.md");
342 let request_markdown = RequestMarkdown::new(request);
343 let response_events_markdown = response_events_to_markdown(response_events);
344
345 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
346 fs::write(messages_file_path, messages.clone()).expect("failed to write messages file");
347 fs::write(last_messages_file_path, messages).expect("failed to write last messages file");
348
349 if request_count == 1 {
350 let tools_file_path = example_dir_path.join("tools.md");
351 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
352 }
353 });
354 })?;
355
356 let tool_metrics = Arc::new(Mutex::new(ToolMetrics::default()));
357
358 let (thread_event_tx, mut thread_event_rx) = mpsc::unbounded();
359
360 let subscription = cx.subscribe(&thread, move |_thread, event: &ThreadEvent, _cx| {
361 thread_event_tx.unbounded_send(event.clone()).log_err();
362 });
363
364 let event_handler_task = cx.spawn({
365 let log_prefix = this.log_prefix.clone();
366 let tool_metrics = tool_metrics.clone();
367 let thread = thread.downgrade();
368 async move |cx| {
369 loop {
370 let event = select_biased! {
371 event = thread_event_rx.next() => event,
372 _ = cx.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
373 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
374 }
375 };
376 let Some(event) = event else {
377 return Err(anyhow!("ThreadEvent channel ended early"));
378 };
379
380 match event {
381 ThreadEvent::Stopped(reason) => match reason {
382 Ok(StopReason::EndTurn) => {
383 return Ok(());
384 }
385 Ok(StopReason::MaxTokens) => {
386 return Err(anyhow!("Exceeded maximum tokens"));
387 }
388 Ok(StopReason::ToolUse) => {
389 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
390 println!("{}StopReason: Tool use", log_prefix);
391 }
392 }
393 Err(error) => {
394 return Err(anyhow!(error.clone()));
395 }
396 },
397 ThreadEvent::ShowError(thread_error) => {
398 break Err(anyhow!(thread_error.clone()));
399 }
400 ThreadEvent::StreamedAssistantText(_, _)| ThreadEvent::StreamedAssistantThinking(_, _) | ThreadEvent::UsePendingTools { .. } => {
401 }
402 ThreadEvent::ToolFinished {
403 tool_use_id,
404 pending_tool_use,
405 ..
406 } => {
407 thread.update(cx, |thread, _cx| {
408 if let Some(tool_use) = pending_tool_use {
409 let mut tool_metrics = tool_metrics.lock().unwrap();
410 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
411 let message = if tool_result.is_error {
412 format!("TOOL FAILED: {}", tool_use.name)
413 } else {
414 format!("TOOL FINISHED: {}", tool_use.name)
415 };
416 println!("{log_prefix}{message}");
417 tool_metrics.insert(tool_result.tool_name.clone(), !tool_result.is_error);
418 } else {
419 let message = format!("TOOL FINISHED WITHOUT RESULT: {}", tool_use.name);
420 println!("{log_prefix}{message}");
421 tool_metrics.insert(tool_use.name.clone(), true);
422 }
423 }
424 })?;
425 }
426 ThreadEvent::ToolConfirmationNeeded => {
427 panic!("{}Bug: Tool confirmation should not be required in eval", log_prefix);
428 },
429 ThreadEvent::StreamedCompletion |
430 ThreadEvent::MessageAdded(_) |
431 ThreadEvent::MessageEdited(_) |
432 ThreadEvent::MessageDeleted(_) |
433 ThreadEvent::SummaryChanged |
434 ThreadEvent::SummaryGenerated |
435 ThreadEvent::CheckpointChanged |
436 ThreadEvent::UsageUpdated(_) => {
437 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
438 println!("{}Event: {:#?}", log_prefix, event);
439 }
440 }
441 }
442 }
443 }
444 });
445
446 thread.update(cx, |thread, cx| {
447 let context = vec![];
448 thread.insert_user_message(this.prompt.clone(), context, None, cx);
449 thread.send_to_model(model, cx);
450 })?;
451
452 event_handler_task.await?;
453
454 println!("{}Stopped", this.log_prefix);
455
456 if let Some((_, language_file_buffer)) = lsp.as_ref() {
457 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
458 }
459
460 println!("{}Getting repository diff", this.log_prefix);
461 let repository_diff = this.repository_diff().await?;
462
463 let example_output_dir = this.example_output_directory();
464 let repository_diff_path = example_output_dir.join("patch.diff");
465 let mut repository_diff_output_file = File::create(&repository_diff_path)?;
466 writeln!(&mut repository_diff_output_file, "{}", &repository_diff).log_err();
467
468 println!("{}Getting diagnostics", this.log_prefix);
469 let diagnostics_after = cx
470 .update(move |cx| {
471 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
472 })?
473 .await?;
474 println!("{}Got diagnostics", this.log_prefix);
475
476 let Some(last_request) = last_request.borrow_mut().take() else {
477 return Err(anyhow!("No requests ran."));
478 };
479
480 drop(subscription);
481 drop(lsp);
482
483 if let Some(diagnostics_before) = &diagnostics_before {
484 fs::write(example_output_dir.join("diagnostics_before.txt"), diagnostics_before)?;
485 }
486
487 if let Some(diagnostics_after) = &diagnostics_after {
488 fs::write(example_output_dir.join("diagnostics_after.txt"), diagnostics_after)?;
489 }
490
491
492 thread.update(cx, |thread, _cx| {
493 let response_count = thread
494 .messages()
495 .filter(|message| message.role == language_model::Role::Assistant)
496 .count();
497 RunOutput {
498 repository_diff,
499 ran_diagnostics_check: this.base.require_lsp,
500 diagnostics_before,
501 diagnostics_after,
502 response_count,
503 token_usage: thread.cumulative_token_usage(),
504 tool_metrics: tool_metrics.lock().unwrap().clone(),
505 last_request,
506 }
507 })
508 })
509 }
510
511 async fn judge_diff(
512 &self,
513 model: Arc<dyn LanguageModel>,
514 run_output: &RunOutput,
515 judge_number: u32,
516 cx: &AsyncApp,
517 ) -> Result<(String, JudgeResponse)> {
518 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
519 let judge_diff_prompt_name = "judge_diff_prompt";
520 let mut hbs = Handlebars::new();
521 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)?;
522
523 let diff_prompt = hbs.render(
524 judge_diff_prompt_name,
525 &JudgeDiffInput {
526 repository_diff: run_output.repository_diff.clone(),
527 ran_diagnostics_check: run_output.ran_diagnostics_check,
528 diagnostics_before: run_output.diagnostics_before.clone(),
529 diagnostics_after: run_output.diagnostics_after.clone(),
530 criteria: self.diff_criteria.clone(),
531 },
532 )?;
533
534 let request = LanguageModelRequest {
535 thread_id: None,
536 prompt_id: None,
537 messages: vec![LanguageModelRequestMessage {
538 role: Role::User,
539 content: vec![MessageContent::Text(diff_prompt)],
540 cache: false,
541 }],
542 temperature: None,
543 tools: Vec::new(),
544 stop: Vec::new(),
545 };
546
547 let diff_response = send_language_model_request(model, request, cx).await?;
548 let diff_output = JudgeResponse::parse(&diff_response)?;
549
550 println!(
551 "{}Judge #{judge_number} - Diff score: {}",
552 self.log_prefix, diff_output.score
553 );
554
555 Ok((diff_response, diff_output))
556 }
557
558 async fn judge_thread(
559 &self,
560 model: Arc<dyn LanguageModel>,
561 run_output: &RunOutput,
562 judge_number: u32,
563 cx: &AsyncApp,
564 ) -> Result<(String, Option<JudgeResponse>)> {
565 if let Some(criteria) = self.thread_criteria.clone() {
566 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
567 let judge_thread_prompt_name = "judge_thread_prompt";
568 let mut hbs = Handlebars::new();
569 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)?;
570
571 let request_markdown = RequestMarkdown::new(&run_output.last_request);
572 let thread_prompt = hbs.render(
573 judge_thread_prompt_name,
574 &JudgeThreadInput {
575 messages: request_markdown.messages,
576 criteria,
577 },
578 )?;
579
580 let request = LanguageModelRequest {
581 thread_id: None,
582 prompt_id: None,
583 messages: vec![LanguageModelRequestMessage {
584 role: Role::User,
585 content: vec![MessageContent::Text(thread_prompt)],
586 cache: false,
587 }],
588 temperature: None,
589 tools: Vec::new(),
590 stop: Vec::new(),
591 };
592
593 let thread_response = send_language_model_request(model, request, cx).await?;
594 let thread_output = JudgeResponse::parse(&thread_response)?;
595
596 println!(
597 "{}Judge #{judge_number} - Thread score: {}",
598 self.log_prefix, thread_output.score
599 );
600
601 Ok((thread_response, Some(thread_output)))
602 } else {
603 let msg = "There were no criteria specified for this thread, so this example was not judged on its thread.".to_string();
604 Ok((msg, None))
605 }
606 }
607
608 pub async fn judge(
609 &self,
610 model: Arc<dyn LanguageModel>,
611 run_output: &RunOutput,
612 judge_number: u32,
613 cx: &AsyncApp,
614 ) -> Result<JudgeOutput> {
615 let mut output_file = File::create(
616 self.example_output_directory()
617 .join(format!("judge_{}.md", judge_number)),
618 )
619 .expect("failed to create judge.md");
620
621 println!("{}Running judge #{judge_number}", self.log_prefix);
622
623 let diff_task = self.judge_diff(model.clone(), &run_output, judge_number, cx);
624 let thread_task = self.judge_thread(model.clone(), &run_output, judge_number, cx);
625
626 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
627
628 let (diff_response, diff_output) = diff_result?;
629 let (thread_response, thread_output) = thread_result?;
630
631 writeln!(
632 &mut output_file,
633 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
634 )
635 .log_err();
636
637 Ok(JudgeOutput {
638 thread: thread_output,
639 diff: diff_output,
640 })
641 }
642
643 async fn repository_diff(&self) -> Result<String> {
644 let worktree_path = self.worktree_path();
645 run_git(&worktree_path, &["add", "."]).await?;
646 let mut diff_args = vec!["diff", "--staged"];
647 if self.base.url == ZED_REPO_URL {
648 diff_args.push(":(exclude).rules");
649 }
650 run_git(&worktree_path, &diff_args).await
651 }
652}
653
654fn wait_for_lang_server(
655 project: &Entity<Project>,
656 buffer: &Entity<Buffer>,
657 log_prefix: String,
658 cx: &mut AsyncApp,
659) -> Task<Result<()>> {
660 println!("{}⏵ Waiting for language server", log_prefix);
661
662 let (mut tx, mut rx) = mpsc::channel(1);
663
664 let lsp_store = project
665 .update(cx, |project, _| project.lsp_store())
666 .unwrap();
667
668 let has_lang_server = buffer
669 .update(cx, |buffer, cx| {
670 lsp_store.update(cx, |lsp_store, cx| {
671 lsp_store
672 .language_servers_for_local_buffer(&buffer, cx)
673 .next()
674 .is_some()
675 })
676 })
677 .unwrap_or(false);
678
679 if has_lang_server {
680 project
681 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
682 .unwrap()
683 .detach();
684 }
685
686 let subscriptions =
687 [
688 cx.subscribe(&lsp_store, {
689 let log_prefix = log_prefix.clone();
690 move |_, event, _| match event {
691 project::LspStoreEvent::LanguageServerUpdate {
692 message:
693 client::proto::update_language_server::Variant::WorkProgress(
694 LspWorkProgress {
695 message: Some(message),
696 ..
697 },
698 ),
699 ..
700 } => println!("{}⟲ {message}", log_prefix),
701 _ => {}
702 }
703 }),
704 cx.subscribe(&project, {
705 let buffer = buffer.clone();
706 move |project, event, cx| match event {
707 project::Event::LanguageServerAdded(_, _, _) => {
708 let buffer = buffer.clone();
709 project
710 .update(cx, |project, cx| project.save_buffer(buffer, cx))
711 .detach();
712 }
713 project::Event::DiskBasedDiagnosticsFinished { .. } => {
714 tx.try_send(()).ok();
715 }
716 _ => {}
717 }
718 }),
719 ];
720
721 cx.spawn(async move |cx| {
722 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
723 let result = futures::select! {
724 _ = rx.next() => {
725 println!("{}⚑ Language server idle", log_prefix);
726 anyhow::Ok(())
727 },
728 _ = timeout.fuse() => {
729 Err(anyhow!("LSP wait timed out after 5 minutes"))
730 }
731 };
732 drop(subscriptions);
733 result
734 })
735}
736
737async fn query_lsp_diagnostics(
738 project: Entity<Project>,
739 cx: &mut AsyncApp,
740) -> Result<Option<String>> {
741 let paths_with_diagnostics = project.update(cx, |project, cx| {
742 project
743 .diagnostic_summaries(true, cx)
744 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
745 .map(|(project_path, _, _)| project_path)
746 .collect::<Vec<_>>()
747 })?;
748
749 if paths_with_diagnostics.is_empty() {
750 return Ok(None);
751 }
752
753 let mut output = String::new();
754 for project_path in paths_with_diagnostics {
755 let buffer = project
756 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
757 .await?;
758 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
759
760 for (_, group) in snapshot.diagnostic_groups(None) {
761 let entry = &group.entries[group.primary_ix];
762 let range = entry.range.to_point(&snapshot);
763 let severity = match entry.diagnostic.severity {
764 DiagnosticSeverity::ERROR => "error",
765 DiagnosticSeverity::WARNING => "warning",
766 _ => continue,
767 };
768
769 writeln!(
770 output,
771 "{} at line {}: {}",
772 severity,
773 range.start.row + 1,
774 entry.diagnostic.message
775 )?;
776 }
777 }
778 anyhow::Ok(Some(output))
779}
780
781impl JudgeResponse {
782 fn parse(response: &str) -> Result<Self> {
783 let analysis = get_tag("analysis", response)?.to_string();
784 let score = get_tag("score", response)?
785 .parse()
786 .context("error parsing score")?;
787
788 Ok(Self { analysis, score })
789 }
790}
791
792fn get_tag(name: &'static str, response: &str) -> Result<String> {
793 let start_tag = format!("<{}>", name);
794 let end_tag = format!("</{}>", name);
795
796 let start_ix = response
797 .find(&start_tag)
798 .context(format!("{} start tag not found", name))?;
799 let content_start_ix = start_ix + start_tag.len();
800
801 let end_ix = content_start_ix
802 + response[content_start_ix..]
803 .find(&end_tag)
804 .context(format!("{} end tag not found", name))?;
805
806 let content = response[content_start_ix..end_ix].trim().unindent();
807
808 anyhow::Ok(content)
809}
810
811pub fn repo_path_for_url(repo_url: &str) -> PathBuf {
812 let repo_name = repo_url
813 .trim_start_matches("https://")
814 .replace(|c: char| !c.is_alphanumeric(), "-");
815 Path::new(REPOS_DIR)
816 .canonicalize()
817 .context(format!("No such directory {REPOS_DIR}"))
818 .unwrap()
819 .join(repo_name)
820}
821
822pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
823 let output = new_smol_command("git")
824 .current_dir(repo_path)
825 .args(args)
826 .output()
827 .await?;
828
829 if output.status.success() {
830 Ok(String::from_utf8(output.stdout)?.trim().to_string())
831 } else {
832 Err(anyhow!(
833 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
834 args.join(" "),
835 repo_path.display(),
836 output.status,
837 String::from_utf8_lossy(&output.stderr),
838 String::from_utf8_lossy(&output.stdout),
839 ))
840 }
841}
842
843pub async fn send_language_model_request(
844 model: Arc<dyn LanguageModel>,
845 request: LanguageModelRequest,
846 cx: &AsyncApp,
847) -> anyhow::Result<String> {
848 match model.stream_completion_text(request, &cx).await {
849 Ok(mut stream) => {
850 let mut full_response = String::new();
851 while let Some(chunk_result) = stream.stream.next().await {
852 match chunk_result {
853 Ok(chunk_str) => {
854 full_response.push_str(&chunk_str);
855 }
856 Err(err) => {
857 return Err(anyhow!(
858 "Error receiving response from language model: {err}"
859 ));
860 }
861 }
862 }
863 Ok(full_response)
864 }
865 Err(err) => Err(anyhow!(
866 "Failed to get response from language model. Error was: {err}"
867 )),
868 }
869}
870
871struct RequestMarkdown {
872 tools: String,
873 messages: String,
874}
875
876impl RequestMarkdown {
877 fn new(request: &LanguageModelRequest) -> Self {
878 let mut tools = String::new();
879 let mut messages = String::new();
880 let mut assistant_message_number: u32 = 1;
881
882 // Print the tools
883 if !request.tools.is_empty() {
884 for tool in &request.tools {
885 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
886 write!(&mut tools, "{}\n\n", tool.description).unwrap();
887 write!(
888 &mut tools,
889 "{}\n",
890 MarkdownString::code_block("json", &format!("{:#}", tool.input_schema))
891 )
892 .unwrap();
893 }
894 }
895
896 // Print the messages
897 for message in &request.messages {
898 match message.role {
899 Role::System => messages.push_str("# ⚙️ SYSTEM\n\n"),
900 Role::User => messages.push_str("# 👤 USER\n\n"),
901 Role::Assistant => {
902 messages.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
903 assistant_message_number += 1;
904 }
905 };
906
907 for content in &message.content {
908 match content {
909 MessageContent::Text(text) => {
910 messages.push_str(text);
911 messages.push_str("\n\n");
912 }
913 MessageContent::Image(_) => {
914 messages.push_str("[IMAGE DATA]\n\n");
915 }
916 MessageContent::Thinking { text, signature } => {
917 messages.push_str("**Thinking**:\n\n");
918 if let Some(sig) = signature {
919 messages.push_str(&format!("Signature: {}\n\n", sig));
920 }
921 messages.push_str(text);
922 messages.push_str("\n");
923 }
924 MessageContent::RedactedThinking(items) => {
925 messages.push_str(&format!(
926 "**Redacted Thinking**: {} item(s)\n\n",
927 items.len()
928 ));
929 }
930 MessageContent::ToolUse(tool_use) => {
931 messages.push_str(&format!(
932 "**Tool Use**: {} (ID: {})\n",
933 tool_use.name, tool_use.id
934 ));
935 messages.push_str(&format!(
936 "{}\n",
937 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
938 ));
939 }
940 MessageContent::ToolResult(tool_result) => {
941 messages.push_str(&format!(
942 "**Tool Result**: {} (ID: {})\n\n",
943 tool_result.tool_name, tool_result.tool_use_id
944 ));
945 if tool_result.is_error {
946 messages.push_str("**ERROR:**\n");
947 }
948 messages.push_str(&format!("{}\n\n", tool_result.content));
949 }
950 }
951 }
952 }
953
954 Self { tools, messages }
955 }
956}
957
958fn response_events_to_markdown(
959 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
960) -> String {
961 let mut response = String::new();
962 // Print the response events if any
963 response.push_str("# Response\n\n");
964 let mut text_buffer = String::new();
965 let mut thinking_buffer = String::new();
966
967 let flush_buffers =
968 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
969 if !text_buffer.is_empty() {
970 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
971 text_buffer.clear();
972 }
973 if !thinking_buffer.is_empty() {
974 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
975 thinking_buffer.clear();
976 }
977 };
978
979 for event in response_events {
980 match event {
981 Ok(LanguageModelCompletionEvent::Text(text)) => {
982 text_buffer.push_str(text);
983 }
984 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
985 thinking_buffer.push_str(text);
986 }
987 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
988 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
989 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
990 }
991 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
992 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
993 response.push_str(&format!(
994 "**Tool Use**: {} (ID: {})\n",
995 tool_use.name, tool_use.id
996 ));
997 response.push_str(&format!(
998 "{}\n",
999 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
1000 ));
1001 }
1002 Ok(
1003 LanguageModelCompletionEvent::UsageUpdate(_)
1004 | LanguageModelCompletionEvent::StartMessage { .. },
1005 ) => {}
1006 Err(error) => {
1007 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1008 response.push_str(&format!("**Error**: {}\n\n", error));
1009 }
1010 }
1011 }
1012
1013 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1014
1015 response
1016}
1017
1018#[cfg(test)]
1019mod test {
1020 use super::*;
1021 use handlebars::Handlebars;
1022
1023 #[test]
1024 fn test_parse_judge_output() {
1025 let response = r#"
1026 <analysis>The model did a good job but there were still compilations errors.</analysis>
1027 <score>3</score>
1028 "#
1029 .unindent();
1030
1031 let output = JudgeResponse::parse(&response).unwrap();
1032 assert_eq!(
1033 output.analysis,
1034 "The model did a good job but there were still compilations errors."
1035 );
1036 assert_eq!(output.score, 3);
1037
1038 let response = r#"
1039 Text around ignored
1040
1041 <analysis>
1042 Failed to compile:
1043 - Error 1
1044 - Error 2
1045 </analysis>
1046
1047 <score>1</score>
1048 "#
1049 .unindent();
1050
1051 let output = JudgeResponse::parse(&response).unwrap();
1052 assert_eq!(output.analysis, "Failed to compile:\n- Error 1\n- Error 2");
1053 assert_eq!(output.score, 1);
1054 }
1055
1056 #[test]
1057 fn test_judge_prompt_with_diagnostics() {
1058 // Case 1: Both diagnostics before and after are present
1059 let input = JudgeDiffInput {
1060 repository_diff: "diff content goes here".to_string(),
1061 ran_diagnostics_check: true,
1062 diagnostics_before: Some("Error at line 10: variable not found".to_string()),
1063 diagnostics_after: Some("Error at line 15: missing semicolon".to_string()),
1064 criteria: "Fix all bugs".to_string(),
1065 };
1066
1067 let rendered = templates().render(JUDGE_PROMPT_NAME, &input).unwrap();
1068
1069 let expected_diagnostics_section = r#"
1070 Take into account the diagnostics before and after applying the change:
1071
1072 <diagnostics_before>
1073 Error at line 10: variable not found
1074 </diagnostics_before>
1075
1076 <diagnostics_after>
1077 Error at line 15: missing semicolon
1078 </diagnostics_after>
1079 "#
1080 .unindent();
1081
1082 assert!(rendered.contains(&expected_diagnostics_section));
1083 }
1084
1085 #[test]
1086 fn test_judge_prompt_with_empty_diagnostics() {
1087 // Case 2: Diagnostics check run but no diagnostics found
1088 let input = JudgeDiffInput {
1089 repository_diff: "diff content goes here".to_string(),
1090 ran_diagnostics_check: true,
1091 diagnostics_before: None,
1092 diagnostics_after: None,
1093 criteria: "Fix all bugs".to_string(),
1094 };
1095
1096 let rendered = templates().render(JUDGE_PROMPT_NAME, &input).unwrap();
1097
1098 let expected_diagnostics_section = r#"
1099 Take into account the diagnostics before and after applying the change:
1100
1101 <diagnostics_before>
1102 No diagnostics before applying the edits.
1103 </diagnostics_before>
1104
1105 <diagnostics_after>
1106 No diagnostics after applying the edits.
1107 </diagnostics_after>
1108 "#
1109 .unindent();
1110
1111 assert!(rendered.contains(&expected_diagnostics_section));
1112 }
1113
1114 #[test]
1115 fn test_judge_prompt_with_mixed_diagnostics() {
1116 let templates = templates();
1117
1118 // Case 3: Before diagnostics present, after diagnostics absent
1119 let input = JudgeDiffInput {
1120 repository_diff: "diff content goes here".to_string(),
1121 ran_diagnostics_check: true,
1122 diagnostics_before: Some("Error at line 10: variable not found".to_string()),
1123 diagnostics_after: None,
1124 criteria: "Fix all bugs".to_string(),
1125 };
1126
1127 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1128
1129 let expected_diagnostics_section = r#"
1130 Take into account the diagnostics before and after applying the change:
1131
1132 <diagnostics_before>
1133 Error at line 10: variable not found
1134 </diagnostics_before>
1135
1136 <diagnostics_after>
1137 No diagnostics after applying the edits.
1138 </diagnostics_after>
1139 "#
1140 .unindent();
1141
1142 assert!(rendered.contains(&expected_diagnostics_section));
1143
1144 // Case 4: Before diagnostics absent, after diagnostics present
1145 let input = JudgeDiffInput {
1146 repository_diff: "diff content goes here".to_string(),
1147 ran_diagnostics_check: true,
1148 diagnostics_before: None,
1149 diagnostics_after: Some("Error at line 15: missing semicolon".to_string()),
1150 criteria: "Fix all bugs".to_string(),
1151 };
1152
1153 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1154
1155 let expected_diagnostics_section = r#"
1156 Take into account the diagnostics before and after applying the change:
1157
1158 <diagnostics_before>
1159 No diagnostics before applying the edits.
1160 </diagnostics_before>
1161
1162 <diagnostics_after>
1163 Error at line 15: missing semicolon
1164 </diagnostics_after>
1165 "#
1166 .unindent();
1167
1168 assert!(rendered.contains(&expected_diagnostics_section));
1169 }
1170
1171 #[test]
1172 fn test_judge_prompt_without_diagnostics() {
1173 let templates = templates();
1174
1175 // Case 5: No diagnostics check run
1176 let input = JudgeDiffInput {
1177 repository_diff: "diff content goes here".to_string(),
1178 ran_diagnostics_check: false,
1179 diagnostics_before: None,
1180 diagnostics_after: None,
1181 criteria: "Fix all bugs".to_string(),
1182 };
1183
1184 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1185
1186 // Check for the message when no diagnostics were performed
1187 let diagnostics_message = "No diagnostic checks were performed.";
1188
1189 assert!(rendered.contains(diagnostics_message));
1190 assert!(!rendered.contains("<diagnostics_before>"));
1191 assert!(!rendered.contains("<diagnostics_after>"));
1192 }
1193
1194 const JUDGE_PROMPT_NAME: &str = "judge_prompt";
1195
1196 fn templates() -> Handlebars<'static> {
1197 let mut judge_prompt = include_str!("judge_diff_prompt.hbs").to_string();
1198 language::LineEnding::normalize(&mut judge_prompt);
1199 let mut handlebars = Handlebars::new();
1200 handlebars
1201 .register_template_string(JUDGE_PROMPT_NAME, judge_prompt)
1202 .unwrap();
1203 handlebars
1204 }
1205}