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 example_output_dir = this.example_output_directory();
327 let last_diff_file_path = example_output_dir.join("last.diff");
328
329 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
330 // history using undo/redo.
331 std::fs::write(&last_diff_file_path, "")?;
332
333 let thread_store = thread_store.await?;
334 let thread =
335 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
336 let last_request = Rc::new(RefCell::new(None));
337
338 thread.update(cx, |thread, _cx| {
339 let mut request_count = 0;
340 let last_request = Rc::clone(&last_request);
341 let previous_diff = Rc::new(RefCell::new("".to_string()));
342 let example_output_dir = example_output_dir.clone();
343 let last_diff_file_path = last_diff_file_path.clone();
344 let this = this.clone();
345 thread.set_request_callback(move |request, response_events| {
346 *last_request.borrow_mut() = Some(request.clone());
347
348 request_count += 1;
349 let messages_file_path = example_output_dir.join(format!("{request_count}.messages.md"));
350 let diff_file_path = example_output_dir.join(format!("{request_count}.diff"));
351 let last_messages_file_path = example_output_dir.join("last.messages.md");
352 let request_markdown = RequestMarkdown::new(request);
353 let response_events_markdown = response_events_to_markdown(response_events);
354
355 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
356 fs::write(&messages_file_path, messages.clone()).expect("failed to write messages file");
357 fs::write(&last_messages_file_path, messages).expect("failed to write last messages file");
358
359 let diff_result = smol::block_on(this.repository_diff());
360 match diff_result {
361 Ok(diff) => {
362 if diff != previous_diff.borrow().clone() {
363 fs::write(&diff_file_path, &diff).expect("failed to write diff file");
364 fs::write(&last_diff_file_path, &diff).expect("failed to write last diff file");
365 *previous_diff.borrow_mut() = diff;
366 }
367 }
368 Err(err) => {
369 let error_message = format!("{err:?}");
370 fs::write(&diff_file_path, &error_message).expect("failed to write diff error to file");
371 fs::write(&last_diff_file_path, &error_message).expect("failed to write last diff file");
372 }
373 }
374
375 if request_count == 1 {
376 let tools_file_path = example_output_dir.join("tools.md");
377 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
378 }
379 });
380 })?;
381
382 let tool_metrics = Arc::new(Mutex::new(ToolMetrics::default()));
383
384 let (thread_event_tx, mut thread_event_rx) = mpsc::unbounded();
385
386 let subscription = cx.subscribe(&thread, move |_thread, event: &ThreadEvent, _cx| {
387 thread_event_tx.unbounded_send(event.clone()).log_err();
388 });
389
390 let event_handler_task = cx.spawn({
391 let log_prefix = this.log_prefix.clone();
392 let tool_metrics = tool_metrics.clone();
393 let thread = thread.downgrade();
394 async move |cx| {
395 loop {
396 let event = select_biased! {
397 event = thread_event_rx.next() => event,
398 _ = cx.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
399 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
400 }
401 };
402 let Some(event) = event else {
403 return Err(anyhow!("ThreadEvent channel ended early"));
404 };
405
406 match event {
407 ThreadEvent::Stopped(reason) => match reason {
408 Ok(StopReason::EndTurn) => {
409 return Ok(());
410 }
411 Ok(StopReason::MaxTokens) => {
412 return Err(anyhow!("Exceeded maximum tokens"));
413 }
414 Ok(StopReason::ToolUse) => {
415 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
416 println!("{}StopReason: Tool use", log_prefix);
417 }
418 }
419 Err(error) => {
420 return Err(anyhow!(error.clone()));
421 }
422 },
423 ThreadEvent::ShowError(thread_error) => {
424 break Err(anyhow!(thread_error.clone()));
425 }
426 ThreadEvent::StreamedAssistantText(_, _)| ThreadEvent::StreamedAssistantThinking(_, _) | ThreadEvent::UsePendingTools { .. } => {
427 }
428 ThreadEvent::ToolFinished {
429 tool_use_id,
430 pending_tool_use,
431 ..
432 } => {
433 thread.update(cx, |thread, _cx| {
434 if let Some(tool_use) = pending_tool_use {
435 let mut tool_metrics = tool_metrics.lock().unwrap();
436 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
437 let message = if tool_result.is_error {
438 format!("TOOL FAILED: {}", tool_use.name)
439 } else {
440 format!("TOOL FINISHED: {}", tool_use.name)
441 };
442 println!("{log_prefix}{message}");
443 tool_metrics.insert(tool_result.tool_name.clone(), !tool_result.is_error);
444 } else {
445 let message = format!("TOOL FINISHED WITHOUT RESULT: {}", tool_use.name);
446 println!("{log_prefix}{message}");
447 tool_metrics.insert(tool_use.name.clone(), true);
448 }
449 }
450 })?;
451 }
452 ThreadEvent::ToolConfirmationNeeded => {
453 panic!("{}Bug: Tool confirmation should not be required in eval", log_prefix);
454 },
455 ThreadEvent::StreamedToolUse { .. } |
456 ThreadEvent::StreamedCompletion |
457 ThreadEvent::MessageAdded(_) |
458 ThreadEvent::MessageEdited(_) |
459 ThreadEvent::MessageDeleted(_) |
460 ThreadEvent::SummaryChanged |
461 ThreadEvent::SummaryGenerated |
462 ThreadEvent::CheckpointChanged |
463 ThreadEvent::UsageUpdated(_) => {
464 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
465 println!("{}Event: {:#?}", log_prefix, event);
466 }
467 }
468 }
469 }
470 }
471 });
472
473 thread.update(cx, |thread, cx| {
474 let context = vec![];
475 thread.insert_user_message(this.prompt.clone(), context, None, cx);
476 thread.send_to_model(model, cx);
477 })?;
478
479 event_handler_task.await?;
480
481 println!("{}Stopped", this.log_prefix);
482
483 if let Some((_, language_file_buffer)) = lsp.as_ref() {
484 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
485 }
486
487 println!("{}Getting repository diff", this.log_prefix);
488 let repository_diff = this.repository_diff().await?;
489 std::fs::write(last_diff_file_path, &repository_diff)?;
490
491 println!("{}Getting diagnostics", this.log_prefix);
492 let diagnostics_after = cx
493 .update(move |cx| {
494 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
495 })?
496 .await?;
497 println!("{}Got diagnostics", this.log_prefix);
498
499 let Some(last_request) = last_request.borrow_mut().take() else {
500 return Err(anyhow!("No requests ran."));
501 };
502
503 drop(subscription);
504 drop(lsp);
505
506 if let Some(diagnostics_before) = &diagnostics_before {
507 fs::write(example_output_dir.join("diagnostics_before.txt"), diagnostics_before)?;
508 }
509
510 if let Some(diagnostics_after) = &diagnostics_after {
511 fs::write(example_output_dir.join("diagnostics_after.txt"), diagnostics_after)?;
512 }
513
514
515 thread.update(cx, |thread, _cx| {
516 let response_count = thread
517 .messages()
518 .filter(|message| message.role == language_model::Role::Assistant)
519 .count();
520 RunOutput {
521 repository_diff,
522 ran_diagnostics_check: this.base.require_lsp,
523 diagnostics_before,
524 diagnostics_after,
525 response_count,
526 token_usage: thread.cumulative_token_usage(),
527 tool_metrics: tool_metrics.lock().unwrap().clone(),
528 last_request,
529 }
530 })
531 })
532 }
533
534 async fn judge_diff(
535 &self,
536 model: Arc<dyn LanguageModel>,
537 run_output: &RunOutput,
538 judge_number: u32,
539 cx: &AsyncApp,
540 ) -> Result<(String, JudgeResponse)> {
541 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
542 let judge_diff_prompt_name = "judge_diff_prompt";
543 let mut hbs = Handlebars::new();
544 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)?;
545
546 let diff_prompt = hbs.render(
547 judge_diff_prompt_name,
548 &JudgeDiffInput {
549 repository_diff: run_output.repository_diff.clone(),
550 ran_diagnostics_check: run_output.ran_diagnostics_check,
551 diagnostics_before: run_output.diagnostics_before.clone(),
552 diagnostics_after: run_output.diagnostics_after.clone(),
553 criteria: self.diff_criteria.clone(),
554 },
555 )?;
556
557 let request = LanguageModelRequest {
558 thread_id: None,
559 prompt_id: None,
560 messages: vec![LanguageModelRequestMessage {
561 role: Role::User,
562 content: vec![MessageContent::Text(diff_prompt)],
563 cache: false,
564 }],
565 temperature: None,
566 tools: Vec::new(),
567 stop: Vec::new(),
568 };
569
570 let diff_response = send_language_model_request(model, request, cx).await?;
571 let diff_output = JudgeResponse::parse(&diff_response)?;
572
573 println!(
574 "{}Judge #{judge_number} - Diff score: {}",
575 self.log_prefix, diff_output.score
576 );
577
578 Ok((diff_response, diff_output))
579 }
580
581 async fn judge_thread(
582 &self,
583 model: Arc<dyn LanguageModel>,
584 run_output: &RunOutput,
585 judge_number: u32,
586 cx: &AsyncApp,
587 ) -> Result<(String, Option<JudgeResponse>)> {
588 if let Some(criteria) = self.thread_criteria.clone() {
589 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
590 let judge_thread_prompt_name = "judge_thread_prompt";
591 let mut hbs = Handlebars::new();
592 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)?;
593
594 let request_markdown = RequestMarkdown::new(&run_output.last_request);
595 let thread_prompt = hbs.render(
596 judge_thread_prompt_name,
597 &JudgeThreadInput {
598 messages: request_markdown.messages,
599 criteria,
600 },
601 )?;
602
603 let request = LanguageModelRequest {
604 thread_id: None,
605 prompt_id: None,
606 messages: vec![LanguageModelRequestMessage {
607 role: Role::User,
608 content: vec![MessageContent::Text(thread_prompt)],
609 cache: false,
610 }],
611 temperature: None,
612 tools: Vec::new(),
613 stop: Vec::new(),
614 };
615
616 let thread_response = send_language_model_request(model, request, cx).await?;
617 let thread_output = JudgeResponse::parse(&thread_response)?;
618
619 println!(
620 "{}Judge #{judge_number} - Thread score: {}",
621 self.log_prefix, thread_output.score
622 );
623
624 Ok((thread_response, Some(thread_output)))
625 } else {
626 let msg = "There were no criteria specified for this thread, so this example was not judged on its thread.".to_string();
627 Ok((msg, None))
628 }
629 }
630
631 pub async fn judge(
632 &self,
633 model: Arc<dyn LanguageModel>,
634 run_output: &RunOutput,
635 judge_number: u32,
636 cx: &AsyncApp,
637 ) -> Result<JudgeOutput> {
638 let mut output_file = File::create(
639 self.example_output_directory()
640 .join(format!("judge_{}.md", judge_number)),
641 )
642 .expect("failed to create judge.md");
643
644 println!("{}Running judge #{judge_number}", self.log_prefix);
645
646 let diff_task = self.judge_diff(model.clone(), &run_output, judge_number, cx);
647 let thread_task = self.judge_thread(model.clone(), &run_output, judge_number, cx);
648
649 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
650
651 let (diff_response, diff_output) = diff_result?;
652 let (thread_response, thread_output) = thread_result?;
653
654 writeln!(
655 &mut output_file,
656 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
657 )
658 .log_err();
659
660 Ok(JudgeOutput {
661 thread: thread_output,
662 diff: diff_output,
663 })
664 }
665
666 async fn repository_diff(&self) -> Result<String> {
667 let worktree_path = self.worktree_path();
668 run_git(&worktree_path, &["add", "."]).await?;
669 let mut diff_args = vec!["diff", "--staged"];
670 if self.base.url == ZED_REPO_URL {
671 diff_args.push(":(exclude).rules");
672 }
673 run_git(&worktree_path, &diff_args).await
674 }
675}
676
677fn wait_for_lang_server(
678 project: &Entity<Project>,
679 buffer: &Entity<Buffer>,
680 log_prefix: String,
681 cx: &mut AsyncApp,
682) -> Task<Result<()>> {
683 println!("{}⏵ Waiting for language server", log_prefix);
684
685 let (mut tx, mut rx) = mpsc::channel(1);
686
687 let lsp_store = project
688 .update(cx, |project, _| project.lsp_store())
689 .unwrap();
690
691 let has_lang_server = buffer
692 .update(cx, |buffer, cx| {
693 lsp_store.update(cx, |lsp_store, cx| {
694 lsp_store
695 .language_servers_for_local_buffer(&buffer, cx)
696 .next()
697 .is_some()
698 })
699 })
700 .unwrap_or(false);
701
702 if has_lang_server {
703 project
704 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
705 .unwrap()
706 .detach();
707 }
708
709 let subscriptions =
710 [
711 cx.subscribe(&lsp_store, {
712 let log_prefix = log_prefix.clone();
713 move |_, event, _| match event {
714 project::LspStoreEvent::LanguageServerUpdate {
715 message:
716 client::proto::update_language_server::Variant::WorkProgress(
717 LspWorkProgress {
718 message: Some(message),
719 ..
720 },
721 ),
722 ..
723 } => println!("{}⟲ {message}", log_prefix),
724 _ => {}
725 }
726 }),
727 cx.subscribe(&project, {
728 let buffer = buffer.clone();
729 move |project, event, cx| match event {
730 project::Event::LanguageServerAdded(_, _, _) => {
731 let buffer = buffer.clone();
732 project
733 .update(cx, |project, cx| project.save_buffer(buffer, cx))
734 .detach();
735 }
736 project::Event::DiskBasedDiagnosticsFinished { .. } => {
737 tx.try_send(()).ok();
738 }
739 _ => {}
740 }
741 }),
742 ];
743
744 cx.spawn(async move |cx| {
745 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
746 let result = futures::select! {
747 _ = rx.next() => {
748 println!("{}⚑ Language server idle", log_prefix);
749 anyhow::Ok(())
750 },
751 _ = timeout.fuse() => {
752 Err(anyhow!("LSP wait timed out after 5 minutes"))
753 }
754 };
755 drop(subscriptions);
756 result
757 })
758}
759
760async fn query_lsp_diagnostics(
761 project: Entity<Project>,
762 cx: &mut AsyncApp,
763) -> Result<Option<String>> {
764 let paths_with_diagnostics = project.update(cx, |project, cx| {
765 project
766 .diagnostic_summaries(true, cx)
767 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
768 .map(|(project_path, _, _)| project_path)
769 .collect::<Vec<_>>()
770 })?;
771
772 if paths_with_diagnostics.is_empty() {
773 return Ok(None);
774 }
775
776 let mut output = String::new();
777 for project_path in paths_with_diagnostics {
778 let buffer = project
779 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
780 .await?;
781 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
782
783 for (_, group) in snapshot.diagnostic_groups(None) {
784 let entry = &group.entries[group.primary_ix];
785 let range = entry.range.to_point(&snapshot);
786 let severity = match entry.diagnostic.severity {
787 DiagnosticSeverity::ERROR => "error",
788 DiagnosticSeverity::WARNING => "warning",
789 _ => continue,
790 };
791
792 writeln!(
793 output,
794 "{} at line {}: {}",
795 severity,
796 range.start.row + 1,
797 entry.diagnostic.message
798 )?;
799 }
800 }
801 anyhow::Ok(Some(output))
802}
803
804impl JudgeResponse {
805 fn parse(response: &str) -> Result<Self> {
806 let analysis = get_tag("analysis", response)?.to_string();
807 let score = get_tag("score", response)?
808 .parse()
809 .context("error parsing score")?;
810
811 Ok(Self { analysis, score })
812 }
813}
814
815fn get_tag(name: &'static str, response: &str) -> Result<String> {
816 let start_tag = format!("<{}>", name);
817 let end_tag = format!("</{}>", name);
818
819 let start_ix = response
820 .find(&start_tag)
821 .context(format!("{} start tag not found", name))?;
822 let content_start_ix = start_ix + start_tag.len();
823
824 let end_ix = content_start_ix
825 + response[content_start_ix..]
826 .find(&end_tag)
827 .context(format!("{} end tag not found", name))?;
828
829 let content = response[content_start_ix..end_ix].trim().unindent();
830
831 anyhow::Ok(content)
832}
833
834pub fn repo_path_for_url(repo_url: &str) -> PathBuf {
835 let repo_name = repo_url
836 .trim_start_matches("https://")
837 .replace(|c: char| !c.is_alphanumeric(), "-");
838 Path::new(REPOS_DIR)
839 .canonicalize()
840 .context(format!("No such directory {REPOS_DIR}"))
841 .unwrap()
842 .join(repo_name)
843}
844
845pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
846 let output = new_smol_command("git")
847 .current_dir(repo_path)
848 .args(args)
849 .output()
850 .await?;
851
852 if output.status.success() {
853 Ok(String::from_utf8(output.stdout)?.trim().to_string())
854 } else {
855 Err(anyhow!(
856 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
857 args.join(" "),
858 repo_path.display(),
859 output.status,
860 String::from_utf8_lossy(&output.stderr),
861 String::from_utf8_lossy(&output.stdout),
862 ))
863 }
864}
865
866pub async fn send_language_model_request(
867 model: Arc<dyn LanguageModel>,
868 request: LanguageModelRequest,
869 cx: &AsyncApp,
870) -> anyhow::Result<String> {
871 match model.stream_completion_text(request, &cx).await {
872 Ok(mut stream) => {
873 let mut full_response = String::new();
874 while let Some(chunk_result) = stream.stream.next().await {
875 match chunk_result {
876 Ok(chunk_str) => {
877 full_response.push_str(&chunk_str);
878 }
879 Err(err) => {
880 return Err(anyhow!(
881 "Error receiving response from language model: {err}"
882 ));
883 }
884 }
885 }
886 Ok(full_response)
887 }
888 Err(err) => Err(anyhow!(
889 "Failed to get response from language model. Error was: {err}"
890 )),
891 }
892}
893
894struct RequestMarkdown {
895 tools: String,
896 messages: String,
897}
898
899impl RequestMarkdown {
900 fn new(request: &LanguageModelRequest) -> Self {
901 let mut tools = String::new();
902 let mut messages = String::new();
903 let mut assistant_message_number: u32 = 1;
904
905 // Print the tools
906 if !request.tools.is_empty() {
907 for tool in &request.tools {
908 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
909 write!(&mut tools, "{}\n\n", tool.description).unwrap();
910 write!(
911 &mut tools,
912 "{}\n",
913 MarkdownString::code_block("json", &format!("{:#}", tool.input_schema))
914 )
915 .unwrap();
916 }
917 }
918
919 // Print the messages
920 for message in &request.messages {
921 match message.role {
922 Role::System => messages.push_str("# ⚙️ SYSTEM\n\n"),
923 Role::User => messages.push_str("# 👤 USER\n\n"),
924 Role::Assistant => {
925 messages.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
926 assistant_message_number += 1;
927 }
928 };
929
930 for content in &message.content {
931 match content {
932 MessageContent::Text(text) => {
933 messages.push_str(text);
934 messages.push_str("\n\n");
935 }
936 MessageContent::Image(_) => {
937 messages.push_str("[IMAGE DATA]\n\n");
938 }
939 MessageContent::Thinking { text, signature } => {
940 messages.push_str("**Thinking**:\n\n");
941 if let Some(sig) = signature {
942 messages.push_str(&format!("Signature: {}\n\n", sig));
943 }
944 messages.push_str(text);
945 messages.push_str("\n");
946 }
947 MessageContent::RedactedThinking(items) => {
948 messages.push_str(&format!(
949 "**Redacted Thinking**: {} item(s)\n\n",
950 items.len()
951 ));
952 }
953 MessageContent::ToolUse(tool_use) => {
954 messages.push_str(&format!(
955 "**Tool Use**: {} (ID: {})\n",
956 tool_use.name, tool_use.id
957 ));
958 messages.push_str(&format!(
959 "{}\n",
960 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
961 ));
962 }
963 MessageContent::ToolResult(tool_result) => {
964 messages.push_str(&format!(
965 "**Tool Result**: {} (ID: {})\n\n",
966 tool_result.tool_name, tool_result.tool_use_id
967 ));
968 if tool_result.is_error {
969 messages.push_str("**ERROR:**\n");
970 }
971 messages.push_str(&format!("{}\n\n", tool_result.content));
972 }
973 }
974 }
975 }
976
977 Self { tools, messages }
978 }
979}
980
981fn response_events_to_markdown(
982 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
983) -> String {
984 let mut response = String::new();
985 // Print the response events if any
986 response.push_str("# Response\n\n");
987 let mut text_buffer = String::new();
988 let mut thinking_buffer = String::new();
989
990 let flush_buffers =
991 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
992 if !text_buffer.is_empty() {
993 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
994 text_buffer.clear();
995 }
996 if !thinking_buffer.is_empty() {
997 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
998 thinking_buffer.clear();
999 }
1000 };
1001
1002 for event in response_events {
1003 match event {
1004 Ok(LanguageModelCompletionEvent::Text(text)) => {
1005 text_buffer.push_str(text);
1006 }
1007 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1008 thinking_buffer.push_str(text);
1009 }
1010 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1011 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1012 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1013 }
1014 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1015 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1016 response.push_str(&format!(
1017 "**Tool Use**: {} (ID: {})\n",
1018 tool_use.name, tool_use.id
1019 ));
1020 response.push_str(&format!(
1021 "{}\n",
1022 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
1023 ));
1024 }
1025 Ok(
1026 LanguageModelCompletionEvent::UsageUpdate(_)
1027 | LanguageModelCompletionEvent::StartMessage { .. },
1028 ) => {}
1029 Err(error) => {
1030 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1031 response.push_str(&format!("**Error**: {}\n\n", error));
1032 }
1033 }
1034 }
1035
1036 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1037
1038 response
1039}
1040
1041#[cfg(test)]
1042mod test {
1043 use super::*;
1044 use handlebars::Handlebars;
1045
1046 #[test]
1047 fn test_parse_judge_output() {
1048 let response = r#"
1049 <analysis>The model did a good job but there were still compilations errors.</analysis>
1050 <score>3</score>
1051 "#
1052 .unindent();
1053
1054 let output = JudgeResponse::parse(&response).unwrap();
1055 assert_eq!(
1056 output.analysis,
1057 "The model did a good job but there were still compilations errors."
1058 );
1059 assert_eq!(output.score, 3);
1060
1061 let response = r#"
1062 Text around ignored
1063
1064 <analysis>
1065 Failed to compile:
1066 - Error 1
1067 - Error 2
1068 </analysis>
1069
1070 <score>1</score>
1071 "#
1072 .unindent();
1073
1074 let output = JudgeResponse::parse(&response).unwrap();
1075 assert_eq!(output.analysis, "Failed to compile:\n- Error 1\n- Error 2");
1076 assert_eq!(output.score, 1);
1077 }
1078
1079 #[test]
1080 fn test_judge_prompt_with_diagnostics() {
1081 // Case 1: Both diagnostics before and after are present
1082 let input = JudgeDiffInput {
1083 repository_diff: "diff content goes here".to_string(),
1084 ran_diagnostics_check: true,
1085 diagnostics_before: Some("Error at line 10: variable not found".to_string()),
1086 diagnostics_after: Some("Error at line 15: missing semicolon".to_string()),
1087 criteria: "Fix all bugs".to_string(),
1088 };
1089
1090 let rendered = templates().render(JUDGE_PROMPT_NAME, &input).unwrap();
1091
1092 let expected_diagnostics_section = r#"
1093 Take into account the diagnostics before and after applying the change:
1094
1095 <diagnostics_before>
1096 Error at line 10: variable not found
1097 </diagnostics_before>
1098
1099 <diagnostics_after>
1100 Error at line 15: missing semicolon
1101 </diagnostics_after>
1102 "#
1103 .unindent();
1104
1105 assert!(rendered.contains(&expected_diagnostics_section));
1106 }
1107
1108 #[test]
1109 fn test_judge_prompt_with_empty_diagnostics() {
1110 // Case 2: Diagnostics check run but no diagnostics found
1111 let input = JudgeDiffInput {
1112 repository_diff: "diff content goes here".to_string(),
1113 ran_diagnostics_check: true,
1114 diagnostics_before: None,
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 No diagnostics before applying the edits.
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
1137 #[test]
1138 fn test_judge_prompt_with_mixed_diagnostics() {
1139 let templates = templates();
1140
1141 // Case 3: Before diagnostics present, after diagnostics absent
1142 let input = JudgeDiffInput {
1143 repository_diff: "diff content goes here".to_string(),
1144 ran_diagnostics_check: true,
1145 diagnostics_before: Some("Error at line 10: variable not found".to_string()),
1146 diagnostics_after: None,
1147 criteria: "Fix all bugs".to_string(),
1148 };
1149
1150 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1151
1152 let expected_diagnostics_section = r#"
1153 Take into account the diagnostics before and after applying the change:
1154
1155 <diagnostics_before>
1156 Error at line 10: variable not found
1157 </diagnostics_before>
1158
1159 <diagnostics_after>
1160 No diagnostics after applying the edits.
1161 </diagnostics_after>
1162 "#
1163 .unindent();
1164
1165 assert!(rendered.contains(&expected_diagnostics_section));
1166
1167 // Case 4: Before diagnostics absent, after diagnostics present
1168 let input = JudgeDiffInput {
1169 repository_diff: "diff content goes here".to_string(),
1170 ran_diagnostics_check: true,
1171 diagnostics_before: None,
1172 diagnostics_after: Some("Error at line 15: missing semicolon".to_string()),
1173 criteria: "Fix all bugs".to_string(),
1174 };
1175
1176 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1177
1178 let expected_diagnostics_section = r#"
1179 Take into account the diagnostics before and after applying the change:
1180
1181 <diagnostics_before>
1182 No diagnostics before applying the edits.
1183 </diagnostics_before>
1184
1185 <diagnostics_after>
1186 Error at line 15: missing semicolon
1187 </diagnostics_after>
1188 "#
1189 .unindent();
1190
1191 assert!(rendered.contains(&expected_diagnostics_section));
1192 }
1193
1194 #[test]
1195 fn test_judge_prompt_without_diagnostics() {
1196 let templates = templates();
1197
1198 // Case 5: No diagnostics check run
1199 let input = JudgeDiffInput {
1200 repository_diff: "diff content goes here".to_string(),
1201 ran_diagnostics_check: false,
1202 diagnostics_before: None,
1203 diagnostics_after: None,
1204 criteria: "Fix all bugs".to_string(),
1205 };
1206
1207 let rendered = templates.render(JUDGE_PROMPT_NAME, &input).unwrap();
1208
1209 // Check for the message when no diagnostics were performed
1210 let diagnostics_message = "No diagnostic checks were performed.";
1211
1212 assert!(rendered.contains(diagnostics_message));
1213 assert!(!rendered.contains("<diagnostics_before>"));
1214 assert!(!rendered.contains("<diagnostics_after>"));
1215 }
1216
1217 const JUDGE_PROMPT_NAME: &str = "judge_prompt";
1218
1219 fn templates() -> Handlebars<'static> {
1220 let mut judge_prompt = include_str!("judge_diff_prompt.hbs").to_string();
1221 language::LineEnding::normalize(&mut judge_prompt);
1222 let mut handlebars = Handlebars::new();
1223 handlebars
1224 .register_template_string(JUDGE_PROMPT_NAME, judge_prompt)
1225 .unwrap();
1226 handlebars
1227 }
1228}