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::{DiagnosticSummary, 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
34const THREAD_EVENT_TIMEOUT: Duration = Duration::from_secs(60 * 2);
35
36const ZED_REPO_URL: &str = "https://github.com/zed-industries/zed.git";
37
38#[derive(Clone, Debug, Deserialize)]
39pub struct ExampleBase {
40 pub url: String,
41 pub revision: String,
42 pub language_extension: Option<String>,
43 pub insert_id: Option<String>,
44 #[serde(default = "default_true")]
45 pub require_lsp: bool,
46 #[serde(default)]
47 pub allow_preexisting_diagnostics: bool,
48}
49
50impl ExampleBase {
51 pub fn repo_name(&self) -> String {
52 self.url
53 .split('/')
54 .next_back()
55 .unwrap_or(&"")
56 .trim_end_matches(".git")
57 .into()
58 }
59}
60
61#[derive(Clone, Debug)]
62pub struct Example {
63 pub name: String,
64 /// Content of `base.toml`
65 pub base: ExampleBase,
66 /// Content of `prompt.md`
67 pub prompt: String,
68 /// Content of `diff_criteria.md`
69 pub diff_criteria: String,
70 /// Content of `thread_criteria.md`, if that file exists (it's optional)
71 pub thread_criteria: String,
72 /// Prefix used for logging that identifies this example
73 pub log_prefix: String,
74 /// The repetition number for this example (0-based)
75 /// When running multiple repetitions of the same example, each instance is assigned a unique repetition number.
76 /// This affects the worktree path and log prefix to avoid clobbering results between runs.
77 pub repetition: usize,
78 pub repo_path: PathBuf,
79 /// Path to the directory containing the requests and responses for the agentic loop
80 run_dir_path: PathBuf,
81 worktrees_dir: PathBuf,
82}
83
84#[derive(Debug, Serialize, Clone)]
85pub struct RunOutput {
86 pub repository_diff: String,
87 pub ran_diagnostics_check: bool,
88 pub diagnostic_summary_before: DiagnosticSummary,
89 pub diagnostic_summary_after: DiagnosticSummary,
90 pub diagnostics_before: Option<String>,
91 pub diagnostics_after: Option<String>,
92 pub response_count: usize,
93 pub token_usage: TokenUsage,
94 pub tool_metrics: ToolMetrics,
95 pub last_request: LanguageModelRequest,
96}
97
98#[derive(Debug, Clone, Serialize, Deserialize)]
99pub struct JudgeDiffInput {
100 pub repository_diff: String,
101 pub criteria: String,
102}
103
104#[derive(Debug, Clone, Serialize, Deserialize)]
105pub struct JudgeThreadInput {
106 pub messages: String,
107 pub criteria: String,
108}
109
110#[derive(Debug, Clone, Serialize, Deserialize)]
111pub struct JudgeResponse {
112 pub analysis: String,
113 pub passing_criteria: u32,
114 pub total_criteria: u32,
115}
116
117#[derive(Debug, Clone, Serialize, Deserialize)]
118pub struct JudgeOutput {
119 pub thread: 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(
126 dir_path: &Path,
127 run_dir: &Path,
128 worktrees_dir: &Path,
129 repos_dir: &Path,
130 ) -> Result<Self> {
131 let name = dir_path.file_name().unwrap().to_string_lossy().to_string();
132 let base_path = dir_path.join("base.toml");
133 let prompt_path = dir_path.join("prompt.md");
134 let diff_criteria_path = dir_path.join("diff_criteria.md");
135 let thread_criteria_path = dir_path.join("thread_criteria.md");
136 let base: ExampleBase = toml::from_str(&fs::read_to_string(&base_path)?)?;
137
138 let repo_path = repo_path_for_url(repos_dir, &base.url);
139
140 Ok(Example {
141 name: name.clone(),
142 base,
143 prompt: fs::read_to_string(prompt_path.clone())?,
144 thread_criteria: fs::read_to_string(thread_criteria_path.clone())?,
145 diff_criteria: fs::read_to_string(diff_criteria_path.clone())?,
146 run_dir_path: run_dir.to_path_buf(),
147 worktrees_dir: worktrees_dir.to_path_buf(),
148 repo_path,
149 log_prefix: name,
150 repetition: 0,
151 })
152 }
153
154 pub fn repetition_name(&self) -> String {
155 format!("{}-{}", self.name, self.repetition)
156 }
157
158 pub fn worktree_path(&self) -> PathBuf {
159 self.worktrees_dir
160 .canonicalize()
161 .unwrap()
162 .join(self.repetition_name())
163 .join(&self.base.repo_name())
164 }
165
166 pub fn run_directory_path(&self) -> PathBuf {
167 self.run_dir_path.join(self.repetition_name())
168 }
169
170 /// Create an iterator that returns copies of this example with different repetition numbers
171 /// Each copy will have a different repetition number and worktree path based on the repetition
172 pub fn repeat(self, repetitions: usize) -> impl Iterator<Item = Self> {
173 (0..repetitions).map(move |repetition| {
174 let mut example = self.clone();
175 example.repetition = repetition;
176 example
177 })
178 }
179
180 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
181 self.log_prefix = format!(
182 "{}{:<width$}\x1b[0m | ",
183 color,
184 self.repetition_name(),
185 width = name_width
186 );
187 }
188
189 /// Set up the example by checking out the specified Git revision
190 pub async fn fetch(&mut self) -> Result<()> {
191 let revision_exists = run_git(
192 &self.repo_path,
193 &["rev-parse", &format!("{}^{{commit}}", self.base.revision)],
194 )
195 .await
196 .is_ok();
197
198 if !revision_exists {
199 println!(
200 "{}Fetching revision {}",
201 self.log_prefix, &self.base.revision
202 );
203 run_git(
204 &self.repo_path,
205 &["fetch", "--depth", "1", "origin", &self.base.revision],
206 )
207 .await?;
208 }
209 Ok(())
210 }
211
212 /// Set up the example by checking out the specified Git revision
213 pub async fn setup(&mut self) -> Result<()> {
214 let worktree_path = self.worktree_path();
215 if worktree_path.is_dir() {
216 println!("{}Resetting existing worktree", self.log_prefix);
217
218 // TODO: consider including "-x" to remove ignored files. The downside of this is that
219 // it will also remove build artifacts, and so prevent incremental reuse there.
220 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
221 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
222 run_git(&worktree_path, &["checkout", &self.base.revision]).await?;
223 } else {
224 println!("{}Creating worktree", self.log_prefix);
225
226 let worktree_path_string = worktree_path.to_string_lossy().to_string();
227
228 run_git(
229 &self.repo_path,
230 &[
231 "worktree",
232 "add",
233 "-f",
234 &worktree_path_string,
235 &self.base.revision,
236 ],
237 )
238 .await?;
239 }
240
241 if self.base.url == ZED_REPO_URL {
242 std::fs::write(worktree_path.join(".rules"), std::fs::read(".rules")?)?;
243 }
244
245 std::fs::create_dir_all(self.run_directory_path())?;
246
247 Ok(())
248 }
249
250 pub fn run(
251 &self,
252 model: Arc<dyn LanguageModel>,
253 app_state: Arc<AgentAppState>,
254 cx: &mut App,
255 ) -> Task<Result<RunOutput>> {
256 let project = Project::local(
257 app_state.client.clone(),
258 app_state.node_runtime.clone(),
259 app_state.user_store.clone(),
260 app_state.languages.clone(),
261 app_state.fs.clone(),
262 None,
263 cx,
264 );
265
266 let worktree = project.update(cx, |project, cx| {
267 project.create_worktree(self.worktree_path(), true, cx)
268 });
269
270 let tools = cx.new(|_| ToolWorkingSet::default());
271 let thread_store =
272 ThreadStore::load(project.clone(), tools, app_state.prompt_builder.clone(), cx);
273 let this = self.clone();
274
275 cx.spawn(async move |cx| {
276 let worktree = worktree.await?;
277
278 // Wait for worktree scan to finish before choosing a file to open.
279 worktree
280 .update(cx, |worktree, _cx| {
281 worktree.as_local().unwrap().scan_complete()
282 })?
283 .await;
284
285 let lsp = if this.base.require_lsp {
286 let language_extension = this.base.language_extension.as_deref().context(
287 "language_extension field is required in base.toml when `require_lsp == true`",
288 )?;
289
290 // Open a file that matches the language to cause LSP to start.
291 let language_file = worktree.read_with(cx, |worktree, _cx| {
292 worktree
293 .files(false, 0)
294 .find_map(|e| {
295 if e.path.clone().extension().and_then(|ext| ext.to_str())
296 == Some(language_extension)
297 {
298 Some(ProjectPath {
299 worktree_id: worktree.id(),
300 path: e.path.clone(),
301 })
302 } else {
303 None
304 }
305 })
306 .context("Failed to find a file for example language")
307 })??;
308
309 let open_language_file_buffer_task = project.update(cx, |project, cx| {
310 project.open_buffer(language_file.clone(), cx)
311 })?;
312
313 let language_file_buffer = open_language_file_buffer_task.await?;
314
315 let lsp_open_handle = project.update(cx, |project, cx| {
316 project.register_buffer_with_language_servers(&language_file_buffer, cx)
317 })?;
318
319 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
320
321 Some((lsp_open_handle, language_file_buffer))
322 } else {
323 None
324 };
325
326 let diagnostic_summary_before = project.read_with(cx, |project, cx| {
327 project.diagnostic_summary(false, cx)
328 })?;
329 let diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
330 if diagnostics_before.is_some() && !this.base.allow_preexisting_diagnostics {
331 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`"));
332 }
333
334 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
335 return Err(anyhow!("Setup only mode"));
336 }
337
338 let example_output_dir = this.run_directory_path();
339 let last_diff_file_path = example_output_dir.join("last.diff");
340
341 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
342 // history using undo/redo.
343 std::fs::write(&last_diff_file_path, "")?;
344
345 let thread_store = thread_store.await?;
346 let thread =
347 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
348 let last_request = Rc::new(RefCell::new(None));
349
350 thread.update(cx, |thread, _cx| {
351 let mut request_count = 0;
352 let last_request = Rc::clone(&last_request);
353 let previous_diff = Rc::new(RefCell::new("".to_string()));
354 let example_output_dir = example_output_dir.clone();
355 let last_diff_file_path = last_diff_file_path.clone();
356 let this = this.clone();
357 thread.set_request_callback(move |request, response_events| {
358 *last_request.borrow_mut() = Some(request.clone());
359
360 request_count += 1;
361 let messages_file_path = example_output_dir.join(format!("{request_count}.messages.md"));
362 let diff_file_path = example_output_dir.join(format!("{request_count}.diff"));
363 let last_messages_file_path = example_output_dir.join("last.messages.md");
364 let request_markdown = RequestMarkdown::new(request);
365 let response_events_markdown = response_events_to_markdown(response_events);
366
367 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
368 fs::write(&messages_file_path, messages.clone()).expect("failed to write messages file");
369 fs::write(&last_messages_file_path, messages).expect("failed to write last messages file");
370
371 let diff_result = smol::block_on(this.repository_diff());
372 match diff_result {
373 Ok(diff) => {
374 if diff != previous_diff.borrow().clone() {
375 fs::write(&diff_file_path, &diff).expect("failed to write diff file");
376 fs::write(&last_diff_file_path, &diff).expect("failed to write last diff file");
377 *previous_diff.borrow_mut() = diff;
378 }
379 }
380 Err(err) => {
381 let error_message = format!("{err:?}");
382 fs::write(&diff_file_path, &error_message).expect("failed to write diff error to file");
383 fs::write(&last_diff_file_path, &error_message).expect("failed to write last diff file");
384 }
385 }
386
387 if request_count == 1 {
388 let tools_file_path = example_output_dir.join("tools.md");
389 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
390 }
391 });
392 })?;
393
394 let tool_metrics = Arc::new(Mutex::new(ToolMetrics::default()));
395
396 let (thread_event_tx, mut thread_event_rx) = mpsc::unbounded();
397
398 let subscription = cx.subscribe(&thread, move |_thread, event: &ThreadEvent, _cx| {
399 thread_event_tx.unbounded_send(event.clone()).log_err();
400 });
401
402 let event_handler_task = cx.spawn({
403 let log_prefix = this.log_prefix.clone();
404 let tool_metrics = tool_metrics.clone();
405 let thread = thread.downgrade();
406 async move |cx| {
407 loop {
408 let event = select_biased! {
409 event = thread_event_rx.next() => event,
410 _ = cx.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
411 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
412 }
413 };
414 let Some(event) = event else {
415 return Err(anyhow!("ThreadEvent channel ended early"));
416 };
417
418 match event {
419 ThreadEvent::Stopped(reason) => match reason {
420 Ok(StopReason::EndTurn) => {
421 return Ok(());
422 }
423 Ok(StopReason::MaxTokens) => {
424 return Err(anyhow!("Exceeded maximum tokens"));
425 }
426 Ok(StopReason::ToolUse) => {
427 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
428 println!("{}StopReason: Tool use", log_prefix);
429 }
430 }
431 Err(error) => {
432 return Err(anyhow!(error.clone()));
433 }
434 },
435 ThreadEvent::ShowError(thread_error) => {
436 break Err(anyhow!(thread_error.clone()));
437 }
438 ThreadEvent::StreamedAssistantText(_, _)| ThreadEvent::StreamedAssistantThinking(_, _) | ThreadEvent::UsePendingTools { .. } => {
439 }
440 ThreadEvent::ToolFinished {
441 tool_use_id,
442 pending_tool_use,
443 ..
444 } => {
445 thread.update(cx, |thread, _cx| {
446 if let Some(tool_use) = pending_tool_use {
447 let mut tool_metrics = tool_metrics.lock().unwrap();
448 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
449 let message = if tool_result.is_error {
450 format!("TOOL FAILED: {}", tool_use.name)
451 } else {
452 format!("TOOL FINISHED: {}", tool_use.name)
453 };
454 println!("{log_prefix}{message}");
455 tool_metrics.insert(tool_result.tool_name.clone(), !tool_result.is_error);
456 } else {
457 let message = format!("TOOL FINISHED WITHOUT RESULT: {}", tool_use.name);
458 println!("{log_prefix}{message}");
459 tool_metrics.insert(tool_use.name.clone(), true);
460 }
461 }
462 })?;
463 }
464 ThreadEvent::ToolConfirmationNeeded => {
465 panic!("{}Bug: Tool confirmation should not be required in eval", log_prefix);
466 },
467 ThreadEvent::StreamedToolUse { .. } |
468 ThreadEvent::StreamedCompletion |
469 ThreadEvent::MessageAdded(_) |
470 ThreadEvent::MessageEdited(_) |
471 ThreadEvent::MessageDeleted(_) |
472 ThreadEvent::SummaryChanged |
473 ThreadEvent::SummaryGenerated |
474 ThreadEvent::CheckpointChanged |
475 ThreadEvent::ReceivedTextChunk |
476 ThreadEvent::UsageUpdated(_) => {
477 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
478 println!("{}Event: {:#?}", log_prefix, event);
479 }
480 }
481 }
482 }
483 }
484 });
485
486 thread.update(cx, |thread, cx| {
487 let context = vec![];
488 thread.insert_user_message(this.prompt.clone(), context, None, cx);
489 thread.send_to_model(model, cx);
490 })?;
491
492 event_handler_task.await?;
493
494 println!("{}Stopped", this.log_prefix);
495
496 if let Some((_, language_file_buffer)) = lsp.as_ref() {
497 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
498 }
499
500 println!("{}Getting repository diff", this.log_prefix);
501 let repository_diff = this.repository_diff().await?;
502 std::fs::write(last_diff_file_path, &repository_diff)?;
503
504 println!("{}Getting diagnostics", this.log_prefix);
505 let diagnostic_summary_after = project.read_with(cx, |project, cx| {
506 project.diagnostic_summary(false, cx)
507 })?;
508 let diagnostics_after = cx
509 .update(move |cx| {
510 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
511 })?
512 .await?;
513 println!("{}Got diagnostics", this.log_prefix);
514
515 let Some(last_request) = last_request.borrow_mut().take() else {
516 return Err(anyhow!("No requests ran."));
517 };
518
519 drop(subscription);
520 drop(lsp);
521
522 if let Some(diagnostics_before) = &diagnostics_before {
523 fs::write(example_output_dir.join("diagnostics_before.txt"), diagnostics_before)?;
524 }
525
526 if let Some(diagnostics_after) = &diagnostics_after {
527 fs::write(example_output_dir.join("diagnostics_after.txt"), diagnostics_after)?;
528 }
529
530
531 thread.update(cx, |thread, _cx| {
532 let response_count = thread
533 .messages()
534 .filter(|message| message.role == language_model::Role::Assistant)
535 .count();
536 RunOutput {
537 repository_diff,
538 ran_diagnostics_check: this.base.require_lsp,
539 diagnostic_summary_before,
540 diagnostic_summary_after,
541 diagnostics_before,
542 diagnostics_after,
543 response_count,
544 token_usage: thread.cumulative_token_usage(),
545 tool_metrics: tool_metrics.lock().unwrap().clone(),
546 last_request,
547 }
548 })
549 })
550 }
551
552 async fn judge_diff(
553 &self,
554 model: Arc<dyn LanguageModel>,
555 run_output: &RunOutput,
556 cx: &AsyncApp,
557 ) -> Result<(String, JudgeResponse)> {
558 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
559 let judge_diff_prompt_name = "judge_diff_prompt";
560 let mut hbs = Handlebars::new();
561 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)?;
562
563 let diff_prompt = hbs.render(
564 judge_diff_prompt_name,
565 &JudgeDiffInput {
566 repository_diff: run_output.repository_diff.clone(),
567 criteria: self.diff_criteria.clone(),
568 },
569 )?;
570
571 let request = LanguageModelRequest {
572 thread_id: None,
573 prompt_id: None,
574 messages: vec![LanguageModelRequestMessage {
575 role: Role::User,
576 content: vec![MessageContent::Text(diff_prompt)],
577 cache: false,
578 }],
579 temperature: None,
580 tools: Vec::new(),
581 stop: Vec::new(),
582 };
583
584 let diff_response = send_language_model_request(model, request, cx).await?;
585 let diff_output = JudgeResponse::parse(&diff_response)?;
586
587 println!(
588 "{}Judge - Diff score: {}%",
589 self.log_prefix,
590 diff_output.score()
591 );
592
593 Ok((diff_response, diff_output))
594 }
595
596 async fn judge_thread(
597 &self,
598 model: Arc<dyn LanguageModel>,
599 run_output: &RunOutput,
600 cx: &AsyncApp,
601 ) -> Result<(String, JudgeResponse)> {
602 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
603 let judge_thread_prompt_name = "judge_thread_prompt";
604 let mut hbs = Handlebars::new();
605 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)?;
606
607 let request_markdown = RequestMarkdown::new(&run_output.last_request);
608 let thread_prompt = hbs.render(
609 judge_thread_prompt_name,
610 &JudgeThreadInput {
611 messages: request_markdown.messages,
612 criteria: self.thread_criteria.clone(),
613 },
614 )?;
615
616 let request = LanguageModelRequest {
617 thread_id: None,
618 prompt_id: None,
619 messages: vec![LanguageModelRequestMessage {
620 role: Role::User,
621 content: vec![MessageContent::Text(thread_prompt)],
622 cache: false,
623 }],
624 temperature: None,
625 tools: Vec::new(),
626 stop: Vec::new(),
627 };
628
629 let thread_response = send_language_model_request(model, request, cx).await?;
630 let thread_output = JudgeResponse::parse(&thread_response)?;
631
632 println!(
633 "{}Judge - Thread score: {}%",
634 self.log_prefix,
635 thread_output.score()
636 );
637
638 Ok((thread_response, thread_output))
639 }
640
641 pub async fn judge(
642 &self,
643 model: Arc<dyn LanguageModel>,
644 run_output: &RunOutput,
645 cx: &AsyncApp,
646 ) -> Result<JudgeOutput> {
647 let mut output_file = File::create(self.run_directory_path().join("judge.md"))
648 .expect("failed to create judge.md");
649
650 println!("{}Running judge", self.log_prefix);
651
652 let diff_task = self.judge_diff(model.clone(), &run_output, cx);
653 let thread_task = self.judge_thread(model.clone(), &run_output, cx);
654
655 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
656
657 let (diff_response, diff_output) = diff_result?;
658 let (thread_response, thread_output) = thread_result?;
659
660 writeln!(
661 &mut output_file,
662 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
663 )
664 .log_err();
665
666 Ok(JudgeOutput {
667 thread: thread_output,
668 diff: diff_output,
669 })
670 }
671
672 async fn repository_diff(&self) -> Result<String> {
673 let worktree_path = self.worktree_path();
674
675 run_git(&worktree_path, &["add", "."]).await?;
676 let mut diff_args = vec!["diff", "--staged"];
677 if self.base.url == ZED_REPO_URL {
678 diff_args.push(":(exclude).rules");
679 }
680 run_git(&worktree_path, &diff_args).await
681 }
682}
683
684fn wait_for_lang_server(
685 project: &Entity<Project>,
686 buffer: &Entity<Buffer>,
687 log_prefix: String,
688 cx: &mut AsyncApp,
689) -> Task<Result<()>> {
690 println!("{}⏵ Waiting for language server", log_prefix);
691
692 let (mut tx, mut rx) = mpsc::channel(1);
693
694 let lsp_store = project
695 .update(cx, |project, _| project.lsp_store())
696 .unwrap();
697
698 let has_lang_server = buffer
699 .update(cx, |buffer, cx| {
700 lsp_store.update(cx, |lsp_store, cx| {
701 lsp_store
702 .language_servers_for_local_buffer(&buffer, cx)
703 .next()
704 .is_some()
705 })
706 })
707 .unwrap_or(false);
708
709 if has_lang_server {
710 project
711 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
712 .unwrap()
713 .detach();
714 }
715
716 let subscriptions =
717 [
718 cx.subscribe(&lsp_store, {
719 let log_prefix = log_prefix.clone();
720 move |_, event, _| match event {
721 project::LspStoreEvent::LanguageServerUpdate {
722 message:
723 client::proto::update_language_server::Variant::WorkProgress(
724 LspWorkProgress {
725 message: Some(message),
726 ..
727 },
728 ),
729 ..
730 } => println!("{}⟲ {message}", log_prefix),
731 _ => {}
732 }
733 }),
734 cx.subscribe(&project, {
735 let buffer = buffer.clone();
736 move |project, event, cx| match event {
737 project::Event::LanguageServerAdded(_, _, _) => {
738 let buffer = buffer.clone();
739 project
740 .update(cx, |project, cx| project.save_buffer(buffer, cx))
741 .detach();
742 }
743 project::Event::DiskBasedDiagnosticsFinished { .. } => {
744 tx.try_send(()).ok();
745 }
746 _ => {}
747 }
748 }),
749 ];
750
751 cx.spawn(async move |cx| {
752 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
753 let result = futures::select! {
754 _ = rx.next() => {
755 println!("{}⚑ Language server idle", log_prefix);
756 anyhow::Ok(())
757 },
758 _ = timeout.fuse() => {
759 Err(anyhow!("LSP wait timed out after 5 minutes"))
760 }
761 };
762 drop(subscriptions);
763 result
764 })
765}
766
767async fn query_lsp_diagnostics(
768 project: Entity<Project>,
769 cx: &mut AsyncApp,
770) -> Result<Option<String>> {
771 let paths_with_diagnostics = project.update(cx, |project, cx| {
772 project
773 .diagnostic_summaries(true, cx)
774 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
775 .map(|(project_path, _, _)| project_path)
776 .collect::<Vec<_>>()
777 })?;
778
779 if paths_with_diagnostics.is_empty() {
780 return Ok(None);
781 }
782
783 let mut output = String::new();
784 for project_path in paths_with_diagnostics {
785 let buffer = project
786 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
787 .await?;
788 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
789
790 for (_, group) in snapshot.diagnostic_groups(None) {
791 let entry = &group.entries[group.primary_ix];
792 let range = entry.range.to_point(&snapshot);
793 let severity = match entry.diagnostic.severity {
794 DiagnosticSeverity::ERROR => "error",
795 DiagnosticSeverity::WARNING => "warning",
796 _ => continue,
797 };
798
799 writeln!(
800 output,
801 "{} at line {}: {}",
802 severity,
803 range.start.row + 1,
804 entry.diagnostic.message
805 )?;
806 }
807 }
808 anyhow::Ok(Some(output))
809}
810
811impl JudgeResponse {
812 fn parse(response: &str) -> Result<Self> {
813 let analysis = get_tag("analysis", response)?.to_string();
814 let passing_criteria = get_tag("passing_criteria", response)?
815 .parse()
816 .context("error parsing score")?;
817 let total_criteria = get_tag("total_criteria", response)?
818 .parse()
819 .context("error parsing score")?;
820 Ok(Self {
821 analysis,
822 total_criteria,
823 passing_criteria,
824 })
825 }
826
827 pub fn score(&self) -> u32 {
828 (100.0 * self.passing_criteria as f32 / self.total_criteria as f32).round() as u32
829 }
830}
831
832fn get_tag(name: &'static str, response: &str) -> Result<String> {
833 let start_tag = format!("<{}>", name);
834 let end_tag = format!("</{}>", name);
835
836 let start_ix = response
837 .find(&start_tag)
838 .context(format!("{} start tag not found", name))?;
839 let content_start_ix = start_ix + start_tag.len();
840
841 let end_ix = content_start_ix
842 + response[content_start_ix..]
843 .find(&end_tag)
844 .context(format!("{} end tag not found", name))?;
845
846 let content = response[content_start_ix..end_ix].trim().unindent();
847
848 anyhow::Ok(content)
849}
850
851pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
852 let repo_name = repo_url
853 .trim_start_matches("https://")
854 .replace(|c: char| !c.is_alphanumeric(), "-");
855 Path::new(repos_dir)
856 .canonicalize()
857 .context(format!("No such directory {}", repos_dir.display()))
858 .unwrap()
859 .join(repo_name)
860}
861
862pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
863 let output = new_smol_command("git")
864 .current_dir(repo_path)
865 .args(args)
866 .output()
867 .await?;
868
869 if output.status.success() {
870 Ok(String::from_utf8(output.stdout)?.trim().to_string())
871 } else {
872 Err(anyhow!(
873 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
874 args.join(" "),
875 repo_path.display(),
876 output.status,
877 String::from_utf8_lossy(&output.stderr),
878 String::from_utf8_lossy(&output.stdout),
879 ))
880 }
881}
882
883pub async fn send_language_model_request(
884 model: Arc<dyn LanguageModel>,
885 request: LanguageModelRequest,
886 cx: &AsyncApp,
887) -> anyhow::Result<String> {
888 match model.stream_completion_text(request, &cx).await {
889 Ok(mut stream) => {
890 let mut full_response = String::new();
891 while let Some(chunk_result) = stream.stream.next().await {
892 match chunk_result {
893 Ok(chunk_str) => {
894 full_response.push_str(&chunk_str);
895 }
896 Err(err) => {
897 return Err(anyhow!(
898 "Error receiving response from language model: {err}"
899 ));
900 }
901 }
902 }
903 Ok(full_response)
904 }
905 Err(err) => Err(anyhow!(
906 "Failed to get response from language model. Error was: {err}"
907 )),
908 }
909}
910
911struct RequestMarkdown {
912 tools: String,
913 messages: String,
914}
915
916impl RequestMarkdown {
917 fn new(request: &LanguageModelRequest) -> Self {
918 let mut tools = String::new();
919 let mut messages = String::new();
920 let mut assistant_message_number: u32 = 1;
921
922 // Print the tools
923 if !request.tools.is_empty() {
924 for tool in &request.tools {
925 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
926 write!(&mut tools, "{}\n\n", tool.description).unwrap();
927 write!(
928 &mut tools,
929 "{}\n",
930 MarkdownString::code_block("json", &format!("{:#}", tool.input_schema))
931 )
932 .unwrap();
933 }
934 }
935
936 // Print the messages
937 for message in &request.messages {
938 match message.role {
939 Role::System => messages.push_str("# ⚙️ SYSTEM\n\n"),
940 Role::User => messages.push_str("# 👤 USER\n\n"),
941 Role::Assistant => {
942 messages.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
943 assistant_message_number += 1;
944 }
945 };
946
947 for content in &message.content {
948 match content {
949 MessageContent::Text(text) => {
950 messages.push_str(text);
951 messages.push_str("\n\n");
952 }
953 MessageContent::Image(_) => {
954 messages.push_str("[IMAGE DATA]\n\n");
955 }
956 MessageContent::Thinking { text, signature } => {
957 messages.push_str("**Thinking**:\n\n");
958 if let Some(sig) = signature {
959 messages.push_str(&format!("Signature: {}\n\n", sig));
960 }
961 messages.push_str(text);
962 messages.push_str("\n");
963 }
964 MessageContent::RedactedThinking(items) => {
965 messages.push_str(&format!(
966 "**Redacted Thinking**: {} item(s)\n\n",
967 items.len()
968 ));
969 }
970 MessageContent::ToolUse(tool_use) => {
971 messages.push_str(&format!(
972 "**Tool Use**: {} (ID: {})\n",
973 tool_use.name, tool_use.id
974 ));
975 messages.push_str(&format!(
976 "{}\n",
977 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
978 ));
979 }
980 MessageContent::ToolResult(tool_result) => {
981 messages.push_str(&format!(
982 "**Tool Result**: {} (ID: {})\n\n",
983 tool_result.tool_name, tool_result.tool_use_id
984 ));
985 if tool_result.is_error {
986 messages.push_str("**ERROR:**\n");
987 }
988 messages.push_str(&format!("{}\n\n", tool_result.content));
989 }
990 }
991 }
992 }
993
994 Self { tools, messages }
995 }
996}
997
998fn response_events_to_markdown(
999 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1000) -> String {
1001 let mut response = String::new();
1002 // Print the response events if any
1003 response.push_str("# Response\n\n");
1004 let mut text_buffer = String::new();
1005 let mut thinking_buffer = String::new();
1006
1007 let flush_buffers =
1008 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
1009 if !text_buffer.is_empty() {
1010 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
1011 text_buffer.clear();
1012 }
1013 if !thinking_buffer.is_empty() {
1014 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
1015 thinking_buffer.clear();
1016 }
1017 };
1018
1019 for event in response_events {
1020 match event {
1021 Ok(LanguageModelCompletionEvent::Text(text)) => {
1022 text_buffer.push_str(text);
1023 }
1024 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1025 thinking_buffer.push_str(text);
1026 }
1027 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1028 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1029 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1030 }
1031 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1032 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1033 response.push_str(&format!(
1034 "**Tool Use**: {} (ID: {})\n",
1035 tool_use.name, tool_use.id
1036 ));
1037 response.push_str(&format!(
1038 "{}\n",
1039 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
1040 ));
1041 }
1042 Ok(
1043 LanguageModelCompletionEvent::UsageUpdate(_)
1044 | LanguageModelCompletionEvent::StartMessage { .. },
1045 ) => {}
1046 Err(error) => {
1047 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1048 response.push_str(&format!("**Error**: {}\n\n", error));
1049 }
1050 }
1051 }
1052
1053 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1054
1055 response
1056}
1057
1058#[cfg(test)]
1059mod test {
1060 use super::*;
1061
1062 #[test]
1063 fn test_parse_judge_output() {
1064 let response = r#"
1065 <analysis>The model did a good job but there were still compilations errors.</analysis>
1066 <passing_criteria>3</passing_criteria>
1067 <total_criteria>5</total_criteria>
1068 "#
1069 .unindent();
1070
1071 let output = JudgeResponse::parse(&response).unwrap();
1072 assert_eq!(
1073 output.analysis,
1074 "The model did a good job but there were still compilations errors."
1075 );
1076 assert_eq!(output.passing_criteria, 3);
1077 assert_eq!(output.total_criteria, 5);
1078
1079 let response = r#"
1080 Text around ignored
1081
1082 <analysis>
1083 Failed to compile:
1084 - Error 1
1085 - Error 2
1086 </analysis>
1087
1088 <passing_criteria>1</passing_criteria>
1089
1090 <total_criteria>3</total_criteria>
1091 "#
1092 .unindent();
1093
1094 let output = JudgeResponse::parse(&response).unwrap();
1095 assert_eq!(output.analysis, "Failed to compile:\n- Error 1\n- Error 2");
1096 assert_eq!(output.passing_criteria, 1);
1097 assert_eq!(output.total_criteria, 3);
1098 }
1099}