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