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, LanguageModelRequest, LanguageModelRequestMessage, MessageContent, Role,
14 StopReason, TokenUsage,
15};
16use project::{LspStore, Project, ProjectPath};
17use serde::{Deserialize, Serialize};
18use std::fmt::Write as _;
19use std::fs::File;
20use std::io::Write as _;
21use std::sync::{Arc, Mutex};
22use std::time::Duration;
23use std::{
24 fs,
25 path::{Path, PathBuf},
26};
27use unindent::Unindent as _;
28use util::ResultExt as _;
29use util::command::new_smol_command;
30use util::serde::default_true;
31
32use crate::AgentAppState;
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
40#[derive(Clone, Debug, Deserialize)]
41pub struct ExampleBase {
42 pub url: String,
43 pub revision: String,
44 pub language_extension: Option<String>,
45 pub insert_id: Option<String>,
46 #[serde(default = "default_true")]
47 pub require_lsp: bool,
48}
49
50#[derive(Clone, Debug)]
51pub struct Example {
52 pub name: String,
53 /// Content of `base.toml`
54 pub base: ExampleBase,
55 /// Content of `prompt.md`
56 pub prompt: String,
57 /// Content of `criteria.md`
58 pub criteria: String,
59 /// Markdown output file to append to
60 pub output_file: Option<Arc<Mutex<File>>>,
61 /// Path to the output run directory.
62 pub run_dir: PathBuf,
63 /// Path to markdown output file
64 pub output_file_path: PathBuf,
65 /// Prefix used for logging that identifies this example
66 pub log_prefix: String,
67}
68
69#[derive(Debug, Serialize, Deserialize, Clone)]
70pub struct RunOutput {
71 pub repository_diff: String,
72 pub diagnostics: String,
73 pub response_count: usize,
74 pub token_usage: TokenUsage,
75 pub tool_use_counts: HashMap<Arc<str>, u32>,
76}
77
78#[derive(Debug, Clone, Serialize, Deserialize)]
79pub struct JudgeInput {
80 pub repository_diff: String,
81 pub criteria: String,
82}
83
84#[derive(Debug, Clone, Serialize, Deserialize)]
85pub struct JudgeOutput {
86 pub analysis: String,
87 pub score: u32,
88}
89
90impl Example {
91 /// Load an example from a directory containing base.toml, prompt.md, and criteria.md
92 pub fn load_from_directory(dir_path: &Path, run_dir: &Path) -> Result<Self> {
93 let name = Self::name_from_path(dir_path);
94 let base_path = dir_path.join("base.toml");
95 let prompt_path = dir_path.join("prompt.md");
96 let criteria_path = dir_path.join("criteria.md");
97
98 let output_file_path = run_dir.join(format!(
99 "{}.md",
100 dir_path.file_name().unwrap().to_str().unwrap()
101 ));
102
103 Ok(Example {
104 name: name.clone(),
105 base: toml::from_str(&fs::read_to_string(&base_path)?)?,
106 prompt: fs::read_to_string(prompt_path.clone())?,
107 criteria: fs::read_to_string(criteria_path.clone())?,
108 run_dir: run_dir.to_path_buf(),
109 output_file: None,
110 output_file_path,
111 log_prefix: name,
112 })
113 }
114
115 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
116 self.log_prefix = format!(
117 "{}{:<width$}\x1b[0m | ",
118 color,
119 self.name,
120 width = name_width
121 );
122 }
123
124 pub fn name_from_path(path: &Path) -> String {
125 path.file_name().unwrap().to_string_lossy().to_string()
126 }
127
128 pub fn worktree_path(&self) -> PathBuf {
129 Path::new(WORKTREES_DIR)
130 .canonicalize()
131 .context(format!("No such directory {WORKTREES_DIR}"))
132 .unwrap()
133 .join(&self.name)
134 }
135
136 /// Set up the example by checking out the specified Git revision
137 pub async fn setup(&mut self) -> Result<()> {
138 let repo_path = repo_path_for_url(&self.base.url);
139
140 println!("{}Fetching", self.log_prefix);
141
142 run_git(
143 &repo_path,
144 &["fetch", "--depth", "1", "origin", &self.base.revision],
145 )
146 .await?;
147
148 let worktree_path = self.worktree_path();
149
150 if worktree_path.is_dir() {
151 println!("{}Resetting existing worktree", self.log_prefix);
152
153 // TODO: consider including "-x" to remove ignored files. The downside of this is that
154 // it will also remove build artifacts, and so prevent incremental reuse there.
155 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
156 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
157 run_git(&worktree_path, &["checkout", &self.base.revision]).await?;
158 } else {
159 println!("{}Creating worktree", self.log_prefix);
160
161 let worktree_path_string = worktree_path.to_string_lossy().to_string();
162
163 run_git(
164 &repo_path,
165 &[
166 "worktree",
167 "add",
168 "-f",
169 &worktree_path_string,
170 &self.base.revision,
171 ],
172 )
173 .await?;
174 }
175
176 // Create the output file
177 let output_file = Arc::new(Mutex::new(File::create(&self.output_file_path)?));
178 self.output_file = Some(output_file);
179
180 Ok(())
181 }
182
183 /// Returns the output file, panicking if it's not set
184 fn output_file(&self) -> Arc<Mutex<File>> {
185 self.output_file
186 .clone()
187 .expect("Output file not created. Call setup() first.")
188 }
189
190 pub fn run(
191 &self,
192 model: Arc<dyn LanguageModel>,
193 app_state: Arc<AgentAppState>,
194 cx: &mut App,
195 ) -> Task<Result<RunOutput>> {
196 let project = Project::local(
197 app_state.client.clone(),
198 app_state.node_runtime.clone(),
199 app_state.user_store.clone(),
200 app_state.languages.clone(),
201 Arc::new(DapRegistry::default()),
202 app_state.fs.clone(),
203 None,
204 cx,
205 );
206
207 let worktree_path = self.worktree_path();
208 let worktree = project.update(cx, |project, cx| {
209 project.create_worktree(&worktree_path, true, cx)
210 });
211
212 let tools = cx.new(|_| ToolWorkingSet::default());
213 let thread_store =
214 ThreadStore::load(project.clone(), tools, app_state.prompt_builder.clone(), cx);
215 let this = self.clone();
216
217 cx.spawn(async move |cx| {
218 let worktree = worktree.await?;
219
220 // Wait for worktree scan to finish before choosing a file to open.
221 worktree
222 .update(cx, |worktree, _cx| {
223 worktree.as_local().unwrap().scan_complete()
224 })?
225 .await;
226
227 let lsp_open_handle_and_store = if this.base.require_lsp {
228 let language_extension = this.base.language_extension.as_deref().context(
229 "language_extension field is required in base.toml when `require_lsp == true`",
230 )?;
231
232 // Open a file that matches the language to cause LSP to start.
233 let language_file = worktree.read_with(cx, |worktree, _cx| {
234 worktree
235 .files(false, 0)
236 .find_map(|e| {
237 if e.path.clone().extension().and_then(|ext| ext.to_str())
238 == Some(language_extension)
239 {
240 Some(ProjectPath {
241 worktree_id: worktree.id(),
242 path: e.path.clone(),
243 })
244 } else {
245 None
246 }
247 })
248 .context("Failed to find a file for example language")
249 })??;
250
251 let open_language_file_buffer_task = project.update(cx, |project, cx| {
252 project.open_buffer(language_file.clone(), cx)
253 })?;
254
255 let language_file_buffer = open_language_file_buffer_task.await?;
256
257 let (lsp_open_handle, lsp_store) = project.update(cx, |project, cx| {
258 (
259 project.register_buffer_with_language_servers(&language_file_buffer, cx),
260 project.lsp_store().clone(),
261 )
262 })?;
263
264 // TODO: remove this once the diagnostics tool waits for new diagnostics
265 cx.background_executor().timer(Duration::new(5, 0)).await;
266 wait_for_lang_server(&lsp_store, this.log_prefix.clone(), cx).await?;
267
268 lsp_store.update(cx, |lsp_store, cx| {
269 lsp_open_handle.update(cx, |buffer, cx| {
270 buffer.update(cx, |buffer, cx| {
271 let has_language_server = lsp_store
272 .language_servers_for_local_buffer(buffer, cx)
273 .next()
274 .is_some();
275 if has_language_server {
276 Ok(())
277 } else {
278 Err(anyhow!(
279 "`{:?}` was opened to cause the language server to start, \
280 but no language servers are registered for its buffer. \
281 Set `require_lsp = false` in `base.toml` to skip this.",
282 language_file
283 ))
284 }
285 })
286 })
287 })??;
288
289 Some((lsp_open_handle, lsp_store))
290 } else {
291 None
292 };
293
294 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
295 return Err(anyhow!("Setup only mode"));
296 }
297
298 let thread_store = thread_store.await;
299 let thread =
300 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
301
302 {
303 let output_file_ref = this.output_file();
304 let mut output_file = output_file_ref.lock().unwrap();
305 writeln!(&mut output_file, "👤 USER:").log_err();
306 writeln!(&mut output_file, "{}", this.prompt).log_err();
307 writeln!(&mut output_file, "🤖 ASSISTANT:").log_err();
308 output_file.flush().log_err();
309 }
310
311 let tool_use_counts: Arc<Mutex<HashMap<Arc<str>, u32>>> =
312 Mutex::new(HashMap::default()).into();
313
314 let (thread_event_tx, mut thread_event_rx) = mpsc::unbounded();
315
316 let subscription = cx.subscribe(&thread, move |_thread, event: &ThreadEvent, _cx| {
317 thread_event_tx.unbounded_send(event.clone()).log_err();
318 });
319
320 let event_handler_task = cx.spawn({
321 // Need to clone the Arc here because the reference from output_file() won't live long enough
322 let output_file = this.output_file.clone().unwrap();
323 let log_prefix = this.log_prefix.clone();
324 let tool_use_counts = tool_use_counts.clone();
325 let thread = thread.downgrade();
326 async move |cx| {
327 loop {
328 let event = select_biased! {
329 event = thread_event_rx.next() => event,
330 _ = cx.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
331 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
332 }
333 };
334 let Some(event) = event else {
335 return Err(anyhow!("ThreadEvent channel ended early"));
336 };
337
338 let mut output_file = output_file.lock().unwrap();
339
340 match event {
341 ThreadEvent::Stopped(reason) => match reason {
342 Ok(StopReason::EndTurn) => {
343 return Ok(());
344 }
345 Ok(StopReason::MaxTokens) => {
346 return Err(anyhow!("Exceeded maximum tokens"));
347 }
348 Ok(StopReason::ToolUse) => {
349 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
350 println!("{}StopReason: Tool use", log_prefix);
351 }
352 }
353 Err(error) => {
354 return Err(anyhow!(error.clone()));
355 }
356 },
357 ThreadEvent::ShowError(thread_error) => {
358 break Err(anyhow!(thread_error.clone()));
359 }
360 ThreadEvent::StreamedAssistantText(_, chunk) => {
361 write!(&mut output_file, "{}", chunk).log_err();
362 }
363 ThreadEvent::StreamedAssistantThinking(_, chunk) => {
364 write!(&mut output_file, "{}", chunk).log_err();
365 }
366 ThreadEvent::UsePendingTools { tool_uses } => {
367 writeln!(&mut output_file, "\n\nUSING TOOLS:").log_err();
368 for tool_use in tool_uses {
369 writeln!(&mut output_file, "{}: {}", tool_use.name, tool_use.input)
370 .log_err();
371 }
372 }
373 ThreadEvent::ToolFinished {
374 tool_use_id,
375 pending_tool_use,
376 ..
377 } => {
378 if let Some(tool_use) = pending_tool_use {
379 let message = format!("TOOL FINISHED: {}", tool_use.name);
380 println!("{}{message}", log_prefix);
381 writeln!(&mut output_file, "\n{}", message).log_err();
382 }
383 thread.update(cx, |thread, _cx| {
384 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
385 writeln!(&mut output_file, "\n{}\n", tool_result.content).log_err();
386 let mut tool_use_counts = tool_use_counts.lock().unwrap();
387 *tool_use_counts
388 .entry(tool_result.tool_name.clone())
389 .or_insert(0) += 1;
390 }
391 })?;
392 }
393 ThreadEvent::ToolConfirmationNeeded => {
394 panic!("{}Bug: Tool confirmation should not be required in eval", log_prefix);
395 },
396 ThreadEvent::StreamedCompletion |
397 ThreadEvent::MessageAdded(_) |
398 ThreadEvent::MessageEdited(_) |
399 ThreadEvent::MessageDeleted(_) |
400 ThreadEvent::SummaryChanged |
401 ThreadEvent::SummaryGenerated |
402 ThreadEvent::CheckpointChanged => {
403 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
404 println!("{}Event: {:#?}", log_prefix, event);
405 }
406 }
407 }
408
409 output_file.flush().log_err();
410 }
411 }
412 });
413
414 thread.update(cx, |thread, cx| {
415 let context = vec![];
416 thread.insert_user_message(this.prompt.clone(), context, None, cx);
417 thread.send_to_model(model, RequestKind::Chat, cx);
418 })?;
419
420 event_handler_task.await?;
421
422 println!("{}Stopped", this.log_prefix);
423
424 if let Some((_, lsp_store)) = lsp_open_handle_and_store.as_ref() {
425 wait_for_lang_server(lsp_store, this.log_prefix.clone(), cx).await?;
426 }
427
428 println!("{}Getting repository diff", this.log_prefix);
429 let repository_diff = this.repository_diff().await?;
430
431 let repository_diff_path = this.run_dir.join(format!("{}.diff", this.name));
432 let mut repository_diff_output_file = File::create(&repository_diff_path)?;
433 writeln!(&mut repository_diff_output_file, "{}", &repository_diff).log_err();
434
435 println!("{}Getting diagnostics", this.log_prefix);
436 let diagnostics = cx
437 .update(move |cx| {
438 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
439 })?
440 .await?;
441 println!("{}Got diagnostics", this.log_prefix);
442
443 drop(subscription);
444 drop(lsp_open_handle_and_store);
445
446 thread.update(cx, |thread, _cx| {
447 let response_count = thread
448 .messages()
449 .filter(|message| message.role == language_model::Role::Assistant)
450 .count();
451 RunOutput {
452 repository_diff,
453 diagnostics,
454 response_count,
455 token_usage: thread.cumulative_token_usage(),
456 tool_use_counts: tool_use_counts.lock().unwrap().clone(),
457 }
458 })
459 })
460 }
461
462 pub async fn judge(
463 &self,
464 model: Arc<dyn LanguageModel>,
465 repository_diff: String,
466 judge_repetitions: u32,
467 cx: &AsyncApp,
468 ) -> Result<JudgeOutput> {
469 let judge_prompt = include_str!("judge_prompt.hbs");
470 let judge_prompt_name = "judge_prompt";
471 let mut handlebars = Handlebars::new();
472 handlebars.register_template_string(judge_prompt_name, judge_prompt)?;
473 let prompt = handlebars.render(
474 judge_prompt_name,
475 &JudgeInput {
476 repository_diff,
477 criteria: self.criteria.clone(),
478 },
479 )?;
480
481 let request = LanguageModelRequest {
482 messages: vec![LanguageModelRequestMessage {
483 role: Role::User,
484 content: vec![MessageContent::Text(prompt)],
485 cache: false,
486 }],
487 temperature: None,
488 tools: Vec::new(),
489 stop: Vec::new(),
490 };
491
492 let response = send_language_model_request(model, request, cx).await?;
493
494 let judge_file_path = self.run_dir.join(format!(
495 "{}_judge_{}.md",
496 self.name, // This is the eval_name
497 judge_repetitions
498 ));
499
500 let mut judge_output_file = File::create(&judge_file_path)?;
501 writeln!(&mut judge_output_file, "{}", &response).log_err();
502
503 parse_judge_output(&response)
504 }
505
506 pub async fn repository_diff(&self) -> Result<String> {
507 let worktree_path = self.worktree_path();
508 run_git(&worktree_path, &["add", "-N"]).await?;
509 run_git(&worktree_path, &["diff"]).await
510 }
511}
512
513fn wait_for_lang_server(
514 lsp_store: &Entity<LspStore>,
515 log_prefix: String,
516 cx: &mut AsyncApp,
517) -> Task<Result<()>> {
518 if cx
519 .update(|cx| !has_pending_lang_server_work(lsp_store, cx))
520 .unwrap()
521 || std::env::var("ZED_EVAL_SKIP_LS_WAIT").is_ok()
522 {
523 return Task::ready(anyhow::Ok(()));
524 }
525
526 println!("{}⏵ Waiting for language server", log_prefix);
527
528 let (mut tx, mut rx) = mpsc::channel(1);
529
530 let subscription =
531 cx.subscribe(&lsp_store, {
532 let log_prefix = log_prefix.clone();
533 move |lsp_store, event, cx| {
534 match event {
535 project::LspStoreEvent::LanguageServerUpdate {
536 message:
537 client::proto::update_language_server::Variant::WorkProgress(
538 LspWorkProgress {
539 message: Some(message),
540 ..
541 },
542 ),
543 ..
544 } => println!("{}⟲ {message}", log_prefix),
545 _ => {}
546 }
547
548 if !has_pending_lang_server_work(&lsp_store, cx) {
549 tx.try_send(()).ok();
550 }
551 }
552 });
553
554 cx.spawn(async move |cx| {
555 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
556 let result = futures::select! {
557 _ = rx.next() => {
558 println!("{}⚑ Language server idle", log_prefix);
559 anyhow::Ok(())
560 },
561 _ = timeout.fuse() => {
562 Err(anyhow!("LSP wait timed out after 5 minutes"))
563 }
564 };
565 drop(subscription);
566 result
567 })
568}
569
570fn has_pending_lang_server_work(lsp_store: &Entity<LspStore>, cx: &App) -> bool {
571 lsp_store
572 .read(cx)
573 .language_server_statuses()
574 .any(|(_, status)| !status.pending_work.is_empty())
575}
576
577async fn query_lsp_diagnostics(project: Entity<Project>, cx: &mut AsyncApp) -> Result<String> {
578 let paths_with_diagnostics = project.update(cx, |project, cx| {
579 project
580 .diagnostic_summaries(true, cx)
581 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
582 .map(|(project_path, _, _)| project_path)
583 .collect::<Vec<_>>()
584 })?;
585
586 let mut output = String::new();
587 for project_path in paths_with_diagnostics {
588 let buffer = project
589 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
590 .await?;
591 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
592
593 for (_, group) in snapshot.diagnostic_groups(None) {
594 let entry = &group.entries[group.primary_ix];
595 let range = entry.range.to_point(&snapshot);
596 let severity = match entry.diagnostic.severity {
597 DiagnosticSeverity::ERROR => "error",
598 DiagnosticSeverity::WARNING => "warning",
599 _ => continue,
600 };
601
602 writeln!(
603 output,
604 "{} at line {}: {}",
605 severity,
606 range.start.row + 1,
607 entry.diagnostic.message
608 )?;
609 }
610 }
611 anyhow::Ok(output)
612}
613
614fn parse_judge_output(response: &str) -> Result<JudgeOutput> {
615 let analysis = get_tag("analysis", response)?.to_string();
616 let score = get_tag("score", response)?
617 .parse()
618 .context("error parsing score")?;
619
620 Ok(JudgeOutput { analysis, score })
621}
622
623fn get_tag(name: &'static str, response: &str) -> Result<String> {
624 let start_tag = format!("<{}>", name);
625 let end_tag = format!("</{}>", name);
626
627 let start_ix = response
628 .find(&start_tag)
629 .context(format!("{} start tag not found", name))?;
630 let content_start_ix = start_ix + start_tag.len();
631
632 let end_ix = content_start_ix
633 + response[content_start_ix..]
634 .find(&end_tag)
635 .context(format!("{} end tag not found", name))?;
636
637 let content = response[content_start_ix..end_ix].trim().unindent();
638
639 anyhow::Ok(content)
640}
641
642pub fn repo_path_for_url(repo_url: &str) -> PathBuf {
643 let repo_name = repo_url
644 .trim_start_matches("https://")
645 .replace(|c: char| !c.is_alphanumeric(), "-");
646 Path::new(REPOS_DIR)
647 .canonicalize()
648 .context(format!("No such directory {REPOS_DIR}"))
649 .unwrap()
650 .join(repo_name)
651}
652
653pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
654 let output = new_smol_command("git")
655 .current_dir(repo_path)
656 .args(args)
657 .output()
658 .await?;
659
660 if output.status.success() {
661 Ok(String::from_utf8(output.stdout)?.trim().to_string())
662 } else {
663 Err(anyhow!(
664 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
665 args.join(" "),
666 repo_path.display(),
667 output.status,
668 String::from_utf8_lossy(&output.stderr),
669 String::from_utf8_lossy(&output.stdout),
670 ))
671 }
672}
673
674pub async fn send_language_model_request(
675 model: Arc<dyn LanguageModel>,
676 request: LanguageModelRequest,
677 cx: &AsyncApp,
678) -> anyhow::Result<String> {
679 match model.stream_completion_text(request, &cx).await {
680 Ok(mut stream) => {
681 let mut full_response = String::new();
682 while let Some(chunk_result) = stream.stream.next().await {
683 match chunk_result {
684 Ok(chunk_str) => {
685 full_response.push_str(&chunk_str);
686 }
687 Err(err) => {
688 return Err(anyhow!(
689 "Error receiving response from language model: {err}"
690 ));
691 }
692 }
693 }
694 Ok(full_response)
695 }
696 Err(err) => Err(anyhow!(
697 "Failed to get response from language model. Error was: {err}"
698 )),
699 }
700}
701
702#[cfg(test)]
703mod test {
704 use super::*;
705
706 #[test]
707 fn test_parse_judge_output() {
708 let response = r#"
709 <analysis>The model did a good job but there were still compilations errors.</analysis>
710 <score>3</score>
711 "#
712 .unindent();
713
714 let output = parse_judge_output(&response).unwrap();
715 assert_eq!(
716 output.analysis,
717 "The model did a good job but there were still compilations errors."
718 );
719 assert_eq!(output.score, 3);
720
721 let response = r#"
722 Text around ignored
723
724 <analysis>
725 Failed to compile:
726 - Error 1
727 - Error 2
728 </analysis>
729
730 <score>1</score>
731 "#
732 .unindent();
733
734 let output = parse_judge_output(&response).unwrap();
735 assert_eq!(output.analysis, "Failed to compile:\n- Error 1\n- Error 2");
736 assert_eq!(output.score, 1);
737 }
738}