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