1use agent::ContextServerRegistry;
2use agent_client_protocol as acp;
3use anyhow::{Context as _, Result, anyhow, bail};
4use client::proto::LspWorkProgress;
5use futures::channel::mpsc;
6use futures::future::Shared;
7use futures::{FutureExt as _, StreamExt as _, future};
8use gpui::{App, AppContext as _, AsyncApp, Entity, Task};
9use handlebars::Handlebars;
10use language::{Buffer, DiagnosticSeverity, OffsetRangeExt as _};
11use language_model::{
12 LanguageModel, LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
13 LanguageModelRequestMessage, LanguageModelToolResultContent, MessageContent, Role, TokenUsage,
14};
15use project::{DiagnosticSummary, Project, ProjectPath, lsp_store::OpenLspBufferHandle};
16use prompt_store::{ProjectContext, WorktreeContext};
17use rand::{distr, prelude::*};
18use serde::{Deserialize, Serialize};
19use std::{
20 fmt::Write as _,
21 fs::{self, File},
22 io::Write as _,
23 path::{Path, PathBuf},
24 rc::Rc,
25 sync::{Arc, Mutex},
26 time::Duration,
27};
28use unindent::Unindent as _;
29use util::{ResultExt as _, command::new_smol_command, markdown::MarkdownCodeBlock};
30
31use crate::{
32 AgentAppState, ToolMetrics,
33 assertions::{AssertionsReport, RanAssertion, RanAssertionResult},
34 example::{Example, ExampleContext, FailedAssertion, JudgeAssertion},
35};
36
37pub const ZED_REPO_URL: &str = "https://github.com/zed-industries/zed.git";
38
39#[derive(Clone)]
40pub struct ExampleInstance {
41 pub thread: Rc<dyn Example>,
42 pub name: String,
43 pub run_directory: PathBuf,
44 pub log_prefix: String,
45 /// The repetition number for this example (0-based)
46 /// When running multiple repetitions of the same example, each instance is assigned a unique repetition number.
47 /// This affects the worktree path and log prefix to avoid clobbering results between runs.
48 pub repetition: usize,
49 pub repo_path: PathBuf,
50 /// Path to the directory containing the requests and responses for the agentic loop
51 worktrees_dir: PathBuf,
52}
53
54#[derive(Debug, Serialize, Clone)]
55pub struct RunOutput {
56 pub repository_diff: String,
57 pub diagnostic_summary_before: DiagnosticSummary,
58 pub diagnostic_summary_after: DiagnosticSummary,
59 pub diagnostics_before: Option<String>,
60 pub diagnostics_after: Option<String>,
61 pub token_usage: TokenUsage,
62 pub tool_metrics: ToolMetrics,
63 pub thread_markdown: String,
64 pub programmatic_assertions: AssertionsReport,
65}
66
67#[derive(Debug, Clone, Serialize, Deserialize)]
68pub struct JudgeDiffInput {
69 pub repository_diff: String,
70 pub assertion: String,
71}
72
73#[derive(Debug, Clone, Serialize, Deserialize)]
74pub struct JudgeThreadInput {
75 pub messages: String,
76 pub assertion: String,
77}
78
79#[derive(Debug, Clone, Serialize, Deserialize)]
80pub struct JudgeOutput {
81 pub thread: AssertionsReport,
82 pub diff: AssertionsReport,
83}
84
85impl ExampleInstance {
86 pub fn new(
87 thread: Rc<dyn Example>,
88 repos_dir: &Path,
89 run_dir: &Path,
90 worktrees_dir: &Path,
91 repetition: usize,
92 ) -> Self {
93 let name = thread.meta().name;
94 let run_directory = run_dir.join(&name).join(repetition.to_string());
95
96 let repo_path = repo_path_for_url(repos_dir, &thread.meta().url);
97
98 Self {
99 name,
100 thread,
101 log_prefix: String::new(),
102 run_directory,
103 repetition,
104 repo_path,
105 worktrees_dir: worktrees_dir.to_path_buf(),
106 }
107 }
108
109 pub fn repo_url(&self) -> String {
110 self.thread.meta().url
111 }
112
113 pub fn revision(&self) -> String {
114 self.thread.meta().revision
115 }
116
117 pub fn worktree_name(&self) -> String {
118 format!("{}-{}", self.name, self.repetition)
119 }
120
121 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
122 self.log_prefix = format!(
123 "{}{:<width$}\x1b[0m | ",
124 color,
125 self.worktree_name(),
126 width = name_width
127 );
128 }
129
130 /// Set up the example by checking out the specified Git revision
131 pub async fn fetch(&mut self) -> Result<()> {
132 let meta = self.thread.meta();
133
134 let revision_exists = run_git(
135 &self.repo_path,
136 &["rev-parse", &format!("{}^{{commit}}", &meta.revision)],
137 )
138 .await
139 .is_ok();
140
141 if !revision_exists {
142 println!("{}Fetching revision {}", self.log_prefix, &meta.revision);
143 run_git(
144 &self.repo_path,
145 &["fetch", "--depth", "1", "origin", &meta.revision],
146 )
147 .await?;
148 }
149 Ok(())
150 }
151
152 /// Set up the example by checking out the specified Git revision
153 pub async fn setup(&mut self) -> Result<()> {
154 let worktree_path = self.worktree_path();
155 let meta = self.thread.meta();
156 if worktree_path.is_dir() {
157 println!("{}Resetting existing worktree", self.log_prefix);
158
159 // TODO: consider including "-x" to remove ignored files. The downside of this is that
160 // it will also remove build artifacts, and so prevent incremental reuse there.
161 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
162 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
163 run_git(&worktree_path, &["checkout", &meta.revision]).await?;
164 } else {
165 println!("{}Creating worktree", self.log_prefix);
166
167 let worktree_path_string = worktree_path.to_string_lossy().into_owned();
168
169 run_git(
170 &self.repo_path,
171 &[
172 "worktree",
173 "add",
174 "-f",
175 &worktree_path_string,
176 &meta.revision,
177 ],
178 )
179 .await?;
180 }
181
182 if meta.url == ZED_REPO_URL {
183 std::fs::write(worktree_path.join(".rules"), std::fs::read(".rules")?)?;
184 }
185
186 std::fs::create_dir_all(&self.run_directory)?;
187
188 Ok(())
189 }
190
191 pub fn worktree_path(&self) -> PathBuf {
192 self.worktrees_dir
193 .join(self.worktree_name())
194 .join(self.thread.meta().repo_name())
195 }
196
197 pub fn run(&self, app_state: Arc<AgentAppState>, cx: &mut App) -> Task<Result<RunOutput>> {
198 let project = Project::local(
199 app_state.client.clone(),
200 app_state.node_runtime.clone(),
201 app_state.user_store.clone(),
202 app_state.languages.clone(),
203 app_state.fs.clone(),
204 None,
205 cx,
206 );
207
208 let worktree = project.update(cx, |project, cx| {
209 project.create_worktree(self.worktree_path(), true, cx)
210 });
211
212 let meta = self.thread.meta();
213 let this = self.clone();
214
215 cx.spawn(async move |cx| {
216 let worktree = worktree.await?;
217
218 // Wait for worktree scan to finish before choosing a file to open.
219 worktree
220 .update(cx, |worktree, _cx| {
221 worktree.as_local().unwrap().scan_complete()
222 })?
223 .await;
224
225 struct LanguageServerState {
226 _lsp_open_handle: OpenLspBufferHandle,
227 language_file_buffer: Entity<Buffer>,
228 }
229
230 let mut diagnostics_before = None;
231 let mut diagnostic_summary_before = DiagnosticSummary::default();
232
233 let lsp = if let Some(language_server) = &meta.language_server {
234 // Open a file that matches the language to cause LSP to start.
235 let language_file = worktree.read_with(cx, |worktree, _cx| {
236 worktree
237 .files(false, 0)
238 .find_map(|e| {
239 if e.path.clone().extension()
240 == Some(&language_server.file_extension)
241 {
242 Some(ProjectPath {
243 worktree_id: worktree.id(),
244 path: e.path.clone(),
245 })
246 } else {
247 None
248 }
249 })
250 .context("Failed to find a file for example language")
251 })??;
252
253 let open_language_file_buffer_task = project.update(cx, |project, cx| {
254 project.open_buffer(language_file.clone(), cx)
255 })?;
256
257 let language_file_buffer = open_language_file_buffer_task.await?;
258
259 let lsp_open_handle = project.update(cx, |project, cx| {
260 project.register_buffer_with_language_servers(&language_file_buffer, cx)
261 })?;
262
263 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
264
265 diagnostic_summary_before = project.read_with(cx, |project, cx| {
266 project.diagnostic_summary(false, cx)
267 })?;
268
269 diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
270 if diagnostics_before.is_some() && language_server.allow_preexisting_diagnostics {
271 anyhow::bail!("Example has pre-existing diagnostics. If you want to run this example regardless, set `allow_preexisting_diagnostics` to `true` in `base.toml`");
272 }
273
274 Some(LanguageServerState {
275 _lsp_open_handle: lsp_open_handle,
276 language_file_buffer,
277 })
278 } else {
279 None
280 };
281
282 anyhow::ensure!(std::env::var("ZED_EVAL_SETUP_ONLY").is_err(), "Setup only mode");
283
284 let last_diff_file_path = this.run_directory.join("last.diff");
285
286 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
287 // history using undo/redo.
288 std::fs::write(&last_diff_file_path, "")?;
289
290 let thread = cx.update(|cx| {
291 //todo: Do we want to load rules files here?
292 let worktrees = project.read(cx).visible_worktrees(cx).map(|worktree| {
293 let root_name = worktree.read(cx).root_name_str().into();
294 let abs_path = worktree.read(cx).abs_path();
295
296 WorktreeContext {
297 root_name,
298 abs_path,
299 rules_file: None,
300 }
301 }).collect::<Vec<_>>();
302 let project_context = cx.new(|_cx| ProjectContext::new(worktrees, vec![]));
303 let context_server_registry = cx.new(|cx| ContextServerRegistry::new(project.read(cx).context_server_store(), cx));
304
305 let thread = if let Some(json) = &meta.existing_thread_json {
306 let session_id = acp::SessionId::new(
307 rand::rng()
308 .sample_iter(&distr::Alphanumeric)
309 .take(7)
310 .map(char::from)
311 .collect::<String>(),
312 );
313
314 let db_thread = agent::DbThread::from_json(json.as_bytes()).expect("Can't read serialized thread");
315 cx.new(|cx| agent::Thread::from_db(session_id, db_thread, project.clone(), project_context, context_server_registry, agent::Templates::new(), cx))
316 } else {
317 cx.new(|cx| agent::Thread::new(project.clone(), project_context, context_server_registry, agent::Templates::new(), None, cx))
318 };
319
320 thread.update(cx, |thread, cx| {
321 thread.add_default_tools(Rc::new(EvalThreadEnvironment {
322 project: project.clone(),
323 }), cx);
324 thread.set_profile(meta.profile_id.clone(), cx);
325 thread.set_model(
326 LanguageModelInterceptor::new(
327 LanguageModelRegistry::read_global(cx).default_model().expect("Missing model").model.clone(),
328 this.run_directory.clone(),
329 last_diff_file_path.clone(),
330 this.run_directory.join("last.messages.json"),
331 this.worktree_path(),
332 this.repo_url(),
333 ),
334 cx,
335 );
336 });
337
338 thread
339 }).unwrap();
340
341 let mut example_cx = ExampleContext::new(
342 meta.clone(),
343 this.log_prefix.clone(),
344 thread.clone(),
345 cx.clone(),
346 );
347 let result = this.thread.conversation(&mut example_cx).await;
348
349 if let Err(err) = result
350 && !err.is::<FailedAssertion>() {
351 return Err(err);
352 }
353
354 println!("{}Stopped", this.log_prefix);
355
356 println!("{}Getting repository diff", this.log_prefix);
357 let repository_diff = Self::repository_diff(this.worktree_path(), &this.repo_url()).await?;
358
359 std::fs::write(last_diff_file_path, &repository_diff)?;
360
361
362 let mut diagnostics_after = None;
363 let mut diagnostic_summary_after = Default::default();
364
365 if let Some(language_server_state) = lsp {
366 wait_for_lang_server(&project, &language_server_state.language_file_buffer, this.log_prefix.clone(), cx).await?;
367
368 println!("{}Getting diagnostics", this.log_prefix);
369 diagnostics_after = cx
370 .update(|cx| {
371 let project = project.clone();
372 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
373 })?
374 .await?;
375 println!("{}Got diagnostics", this.log_prefix);
376
377 diagnostic_summary_after = project.read_with(cx, |project, cx| {
378 project.diagnostic_summary(false, cx)
379 })?;
380
381 }
382
383 if let Some(diagnostics_before) = &diagnostics_before {
384 fs::write(this.run_directory.join("diagnostics_before.txt"), diagnostics_before)?;
385 }
386
387 if let Some(diagnostics_after) = &diagnostics_after {
388 fs::write(this.run_directory.join("diagnostics_after.txt"), diagnostics_after)?;
389 }
390
391 thread.update(cx, |thread, _cx| {
392 RunOutput {
393 repository_diff,
394 diagnostic_summary_before,
395 diagnostic_summary_after,
396 diagnostics_before,
397 diagnostics_after,
398 token_usage: thread.latest_request_token_usage().unwrap(),
399 tool_metrics: example_cx.tool_metrics.lock().unwrap().clone(),
400 thread_markdown: thread.to_markdown(),
401 programmatic_assertions: example_cx.assertions,
402 }
403 })
404 })
405 }
406
407 async fn repository_diff(repository_path: PathBuf, repository_url: &str) -> Result<String> {
408 run_git(&repository_path, &["add", "."]).await?;
409 let mut diff_args = vec!["diff", "--staged"];
410 if repository_url == ZED_REPO_URL {
411 diff_args.push(":(exclude).rules");
412 }
413 run_git(&repository_path, &diff_args).await
414 }
415
416 pub async fn judge(
417 &self,
418 model: Arc<dyn LanguageModel>,
419 run_output: &RunOutput,
420 cx: &AsyncApp,
421 ) -> JudgeOutput {
422 let mut output_file =
423 File::create(self.run_directory.join("judge.md")).expect("failed to create judge.md");
424
425 let diff_task = self.judge_diff(model.clone(), run_output, cx);
426 let thread_task = self.judge_thread(model.clone(), run_output, cx);
427
428 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
429
430 let (diff_response, diff_output) = diff_result;
431 let (thread_response, thread_output) = thread_result;
432
433 writeln!(
434 &mut output_file,
435 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
436 )
437 .log_err();
438
439 JudgeOutput {
440 thread: thread_output,
441 diff: diff_output,
442 }
443 }
444
445 async fn judge_diff(
446 &self,
447 model: Arc<dyn LanguageModel>,
448 run_output: &RunOutput,
449 cx: &AsyncApp,
450 ) -> (String, AssertionsReport) {
451 let diff_assertions = self.thread.diff_assertions();
452
453 if diff_assertions.is_empty() {
454 return (
455 "No diff assertions".to_string(),
456 AssertionsReport::default(),
457 );
458 }
459
460 println!("{}Running diff judge", self.log_prefix);
461
462 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
463 let judge_diff_prompt_name = "judge_diff_prompt";
464 let mut hbs = Handlebars::new();
465 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)
466 .unwrap();
467
468 let to_prompt = |assertion: String| {
469 hbs.render(
470 judge_diff_prompt_name,
471 &JudgeDiffInput {
472 repository_diff: run_output.repository_diff.clone(),
473 assertion,
474 },
475 )
476 .unwrap()
477 };
478
479 let (responses, report) = self
480 .judge_assertions(model, diff_assertions, to_prompt, cx)
481 .await;
482
483 println!(
484 "{}Judge - Diff score: {}%",
485 self.log_prefix,
486 report.passed_percentage()
487 );
488
489 (responses, report)
490 }
491
492 async fn judge_thread(
493 &self,
494 model: Arc<dyn LanguageModel>,
495 run_output: &RunOutput,
496 cx: &AsyncApp,
497 ) -> (String, AssertionsReport) {
498 let thread_assertions = self.thread.thread_assertions();
499
500 if thread_assertions.is_empty() {
501 return (
502 "No thread assertions".to_string(),
503 AssertionsReport::default(),
504 );
505 }
506
507 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
508 let judge_thread_prompt_name = "judge_thread_prompt";
509 let mut hbs = Handlebars::new();
510 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)
511 .unwrap();
512
513 let complete_messages = &run_output.thread_markdown;
514 let to_prompt = |assertion: String| {
515 hbs.render(
516 judge_thread_prompt_name,
517 &JudgeThreadInput {
518 messages: complete_messages.clone(),
519 assertion,
520 },
521 )
522 .unwrap()
523 };
524
525 let (responses, report) = self
526 .judge_assertions(model, thread_assertions, to_prompt, cx)
527 .await;
528
529 println!(
530 "{}Judge - Thread score: {}%",
531 self.log_prefix,
532 report.passed_percentage()
533 );
534
535 (responses, report)
536 }
537
538 async fn judge_assertions(
539 &self,
540 model: Arc<dyn LanguageModel>,
541 assertions: Vec<JudgeAssertion>,
542 to_prompt: impl Fn(String) -> String,
543 cx: &AsyncApp,
544 ) -> (String, AssertionsReport) {
545 let assertions = assertions.into_iter().map(|assertion| {
546 let request = LanguageModelRequest {
547 thread_id: None,
548 prompt_id: None,
549 mode: None,
550 intent: None,
551 messages: vec![LanguageModelRequestMessage {
552 role: Role::User,
553 content: vec![MessageContent::Text(to_prompt(assertion.description))],
554 cache: false,
555 reasoning_details: None,
556 }],
557 temperature: None,
558 tools: Vec::new(),
559 tool_choice: None,
560 stop: Vec::new(),
561 thinking_allowed: true,
562 };
563
564 let model = model.clone();
565 let log_prefix = self.log_prefix.clone();
566 async move {
567 let response = send_language_model_request(model, request, cx).await;
568
569 let (response, result) = match response {
570 Ok(response) => (
571 response.clone(),
572 parse_assertion_result(&response).map_err(|err| err.to_string()),
573 ),
574 Err(err) => (err.to_string(), Err(err.to_string())),
575 };
576
577 if result.is_ok() {
578 println!("{}✅ {}", log_prefix, assertion.id);
579 } else {
580 println!("{}❌ {}", log_prefix, assertion.id);
581 }
582
583 (
584 response,
585 RanAssertion {
586 id: assertion.id,
587 result,
588 },
589 )
590 }
591 });
592
593 let mut responses = String::new();
594 let mut report = AssertionsReport::default();
595
596 for (response, assertion) in future::join_all(assertions).await {
597 writeln!(&mut responses, "# {}", assertion.id).unwrap();
598 writeln!(&mut responses, "{}\n\n", response).unwrap();
599 report.ran.push(assertion);
600 }
601
602 (responses, report)
603 }
604}
605
606struct EvalThreadEnvironment {
607 project: Entity<Project>,
608}
609
610struct EvalTerminalHandle {
611 terminal: Entity<acp_thread::Terminal>,
612}
613
614impl agent::TerminalHandle for EvalTerminalHandle {
615 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId> {
616 self.terminal.read_with(cx, |term, _cx| term.id().clone())
617 }
618
619 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>> {
620 self.terminal
621 .read_with(cx, |term, _cx| term.wait_for_exit())
622 }
623
624 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse> {
625 self.terminal
626 .read_with(cx, |term, cx| term.current_output(cx))
627 }
628
629 fn kill(&self, cx: &AsyncApp) -> Result<()> {
630 cx.update(|cx| {
631 self.terminal.update(cx, |terminal, cx| {
632 terminal.kill(cx);
633 });
634 })?;
635 Ok(())
636 }
637}
638
639impl agent::ThreadEnvironment for EvalThreadEnvironment {
640 fn create_terminal(
641 &self,
642 command: String,
643 cwd: Option<PathBuf>,
644 output_byte_limit: Option<u64>,
645 cx: &mut AsyncApp,
646 ) -> Task<Result<Rc<dyn agent::TerminalHandle>>> {
647 let project = self.project.clone();
648 cx.spawn(async move |cx| {
649 let language_registry =
650 project.read_with(cx, |project, _cx| project.languages().clone())?;
651 let id = acp::TerminalId::new(uuid::Uuid::new_v4().to_string());
652 let terminal =
653 acp_thread::create_terminal_entity(command, &[], vec![], cwd.clone(), &project, cx)
654 .await?;
655 let terminal = cx.new(|cx| {
656 acp_thread::Terminal::new(
657 id,
658 "",
659 cwd,
660 output_byte_limit.map(|limit| limit as usize),
661 terminal,
662 language_registry,
663 cx,
664 )
665 })?;
666 Ok(Rc::new(EvalTerminalHandle { terminal }) as Rc<dyn agent::TerminalHandle>)
667 })
668 }
669}
670
671struct LanguageModelInterceptor {
672 model: Arc<dyn LanguageModel>,
673 request_count: Arc<Mutex<usize>>,
674 previous_diff: Arc<Mutex<String>>,
675 example_output_dir: PathBuf,
676 last_diff_file_path: PathBuf,
677 messages_json_file_path: PathBuf,
678 repository_path: PathBuf,
679 repository_url: String,
680}
681
682impl LanguageModelInterceptor {
683 fn new(
684 model: Arc<dyn LanguageModel>,
685 example_output_dir: PathBuf,
686 last_diff_file_path: PathBuf,
687 messages_json_file_path: PathBuf,
688 repository_path: PathBuf,
689 repository_url: String,
690 ) -> Arc<Self> {
691 Arc::new(Self {
692 model,
693 request_count: Arc::new(Mutex::new(0)),
694 previous_diff: Arc::new(Mutex::new("".to_string())),
695 example_output_dir,
696 last_diff_file_path,
697 messages_json_file_path,
698 repository_path,
699 repository_url,
700 })
701 }
702}
703
704impl language_model::LanguageModel for LanguageModelInterceptor {
705 fn id(&self) -> language_model::LanguageModelId {
706 self.model.id()
707 }
708
709 fn name(&self) -> language_model::LanguageModelName {
710 self.model.name()
711 }
712
713 fn provider_id(&self) -> language_model::LanguageModelProviderId {
714 self.model.provider_id()
715 }
716
717 fn provider_name(&self) -> language_model::LanguageModelProviderName {
718 self.model.provider_name()
719 }
720
721 fn telemetry_id(&self) -> String {
722 self.model.telemetry_id()
723 }
724
725 fn supports_images(&self) -> bool {
726 self.model.supports_images()
727 }
728
729 fn supports_tools(&self) -> bool {
730 self.model.supports_tools()
731 }
732
733 fn supports_tool_choice(&self, choice: language_model::LanguageModelToolChoice) -> bool {
734 self.model.supports_tool_choice(choice)
735 }
736
737 fn max_token_count(&self) -> u64 {
738 self.model.max_token_count()
739 }
740
741 fn count_tokens(
742 &self,
743 request: LanguageModelRequest,
744 cx: &App,
745 ) -> future::BoxFuture<'static, Result<u64>> {
746 self.model.count_tokens(request, cx)
747 }
748
749 fn stream_completion(
750 &self,
751 request: LanguageModelRequest,
752 cx: &AsyncApp,
753 ) -> future::BoxFuture<
754 'static,
755 Result<
756 futures::stream::BoxStream<
757 'static,
758 Result<LanguageModelCompletionEvent, language_model::LanguageModelCompletionError>,
759 >,
760 language_model::LanguageModelCompletionError,
761 >,
762 > {
763 let stream = self.model.stream_completion(request.clone(), cx);
764 let request_count = self.request_count.clone();
765 let previous_diff = self.previous_diff.clone();
766 let example_output_dir = self.example_output_dir.clone();
767 let last_diff_file_path = self.last_diff_file_path.clone();
768 let messages_json_file_path = self.messages_json_file_path.clone();
769 let repository_path = self.repository_path.clone();
770 let repository_url = self.repository_url.clone();
771
772 Box::pin(async move {
773 let stream = stream.await?;
774
775 let response_events = Arc::new(Mutex::new(Vec::new()));
776 let request_clone = request.clone();
777
778 let wrapped_stream = stream.then(move |event| {
779 let response_events = response_events.clone();
780 let request = request_clone.clone();
781 let request_count = request_count.clone();
782 let previous_diff = previous_diff.clone();
783 let example_output_dir = example_output_dir.clone();
784 let last_diff_file_path = last_diff_file_path.clone();
785 let messages_json_file_path = messages_json_file_path.clone();
786 let repository_path = repository_path.clone();
787 let repository_url = repository_url.clone();
788
789 async move {
790 let event_result = match &event {
791 Ok(ev) => Ok(ev.clone()),
792 Err(err) => Err(err.to_string()),
793 };
794 response_events.lock().unwrap().push(event_result);
795
796 let should_execute = matches!(
797 &event,
798 Ok(LanguageModelCompletionEvent::Stop { .. }) | Err(_)
799 );
800
801 if should_execute {
802 let current_request_count = {
803 let mut count = request_count.lock().unwrap();
804 *count += 1;
805 *count
806 };
807
808 let messages_file_path =
809 example_output_dir.join(format!("{current_request_count}.messages.md"));
810 let diff_file_path =
811 example_output_dir.join(format!("{current_request_count}.diff"));
812 let last_messages_file_path = example_output_dir.join("last.messages.md");
813
814 let collected_events = response_events.lock().unwrap().clone();
815 let request_markdown = RequestMarkdown::new(&request);
816 let response_events_markdown =
817 response_events_to_markdown(&collected_events);
818 let dialog = ThreadDialog::new(&request, &collected_events);
819 let dialog_json =
820 serde_json::to_string_pretty(&dialog.to_combined_request())
821 .unwrap_or_default();
822
823 let messages = format!(
824 "{}\n\n{}",
825 request_markdown.messages, response_events_markdown
826 );
827 fs::write(&messages_file_path, messages.clone())
828 .expect("failed to write messages file");
829 fs::write(&last_messages_file_path, messages)
830 .expect("failed to write last messages file");
831 fs::write(&messages_json_file_path, dialog_json)
832 .expect("failed to write last.messages.json");
833
834 // Get repository diff
835 let diff_result =
836 ExampleInstance::repository_diff(repository_path, &repository_url)
837 .await;
838
839 match diff_result {
840 Ok(diff) => {
841 let prev_diff = previous_diff.lock().unwrap().clone();
842 if diff != prev_diff {
843 fs::write(&diff_file_path, &diff)
844 .expect("failed to write diff file");
845 fs::write(&last_diff_file_path, &diff)
846 .expect("failed to write last diff file");
847 *previous_diff.lock().unwrap() = diff;
848 }
849 }
850 Err(err) => {
851 let error_message = format!("{err:?}");
852 fs::write(&diff_file_path, &error_message)
853 .expect("failed to write diff error to file");
854 fs::write(&last_diff_file_path, &error_message)
855 .expect("failed to write last diff file");
856 }
857 }
858
859 if current_request_count == 1 {
860 let tools_file_path = example_output_dir.join("tools.md");
861 fs::write(tools_file_path, request_markdown.tools)
862 .expect("failed to write tools file");
863 }
864 }
865
866 event
867 }
868 });
869
870 Ok(Box::pin(wrapped_stream)
871 as futures::stream::BoxStream<
872 'static,
873 Result<
874 LanguageModelCompletionEvent,
875 language_model::LanguageModelCompletionError,
876 >,
877 >)
878 })
879 }
880}
881
882pub fn wait_for_lang_server(
883 project: &Entity<Project>,
884 buffer: &Entity<Buffer>,
885 log_prefix: String,
886 cx: &mut AsyncApp,
887) -> Task<Result<()>> {
888 if std::env::var("ZED_EVAL_SKIP_LS").is_ok() {
889 return Task::ready(Ok(()));
890 }
891
892 println!("{}⏵ Waiting for language server", log_prefix);
893
894 let (mut tx, mut rx) = mpsc::channel(1);
895
896 let lsp_store = project
897 .read_with(cx, |project, _| project.lsp_store())
898 .unwrap();
899
900 let has_lang_server = buffer
901 .update(cx, |buffer, cx| {
902 lsp_store.update(cx, |lsp_store, cx| {
903 lsp_store
904 .running_language_servers_for_local_buffer(buffer, cx)
905 .next()
906 .is_some()
907 })
908 })
909 .unwrap_or(false);
910
911 if has_lang_server {
912 project
913 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
914 .unwrap()
915 .detach();
916 }
917
918 let subscriptions =
919 [
920 cx.subscribe(&lsp_store, {
921 let log_prefix = log_prefix.clone();
922 move |_, event, _| {
923 if let project::LspStoreEvent::LanguageServerUpdate {
924 message:
925 client::proto::update_language_server::Variant::WorkProgress(
926 LspWorkProgress {
927 message: Some(message),
928 ..
929 },
930 ),
931 ..
932 } = event
933 {
934 println!("{}⟲ {message}", log_prefix)
935 }
936 }
937 }),
938 cx.subscribe(project, {
939 let buffer = buffer.clone();
940 move |project, event, cx| match event {
941 project::Event::LanguageServerAdded(_, _, _) => {
942 let buffer = buffer.clone();
943 project
944 .update(cx, |project, cx| project.save_buffer(buffer, cx))
945 .detach();
946 }
947 project::Event::DiskBasedDiagnosticsFinished { .. } => {
948 tx.try_send(()).ok();
949 }
950 _ => {}
951 }
952 }),
953 ];
954
955 cx.spawn(async move |cx| {
956 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
957 let result = futures::select! {
958 _ = rx.next() => {
959 println!("{}⚑ Language server idle", log_prefix);
960 anyhow::Ok(())
961 },
962 _ = timeout.fuse() => {
963 anyhow::bail!("LSP wait timed out after 5 minutes");
964 }
965 };
966 drop(subscriptions);
967 result
968 })
969}
970
971pub async fn query_lsp_diagnostics(
972 project: Entity<Project>,
973 cx: &mut AsyncApp,
974) -> Result<Option<String>> {
975 let paths_with_diagnostics = project.update(cx, |project, cx| {
976 project
977 .diagnostic_summaries(true, cx)
978 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
979 .map(|(project_path, _, _)| project_path)
980 .collect::<Vec<_>>()
981 })?;
982
983 if paths_with_diagnostics.is_empty() {
984 return Ok(None);
985 }
986
987 let mut output = String::new();
988 for project_path in paths_with_diagnostics {
989 let buffer = project
990 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
991 .await?;
992 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
993
994 for (_, group) in snapshot.diagnostic_groups(None) {
995 let entry = &group.entries[group.primary_ix];
996 let range = entry.range.to_point(&snapshot);
997 let severity = match entry.diagnostic.severity {
998 DiagnosticSeverity::ERROR => "error",
999 DiagnosticSeverity::WARNING => "warning",
1000 _ => continue,
1001 };
1002
1003 writeln!(
1004 output,
1005 "{} at line {}: {}",
1006 severity,
1007 range.start.row + 1,
1008 entry.diagnostic.message
1009 )?;
1010 }
1011 }
1012 anyhow::Ok(Some(output))
1013}
1014
1015fn parse_assertion_result(response: &str) -> Result<RanAssertionResult> {
1016 let analysis = get_tag("analysis", response)?;
1017 let passed = match get_tag("passed", response)?.to_lowercase().as_str() {
1018 "true" => true,
1019 "false" => false,
1020 value @ _ => bail!("invalid judge `passed` tag: {value}"),
1021 };
1022 Ok(RanAssertionResult {
1023 analysis: Some(analysis),
1024 passed,
1025 })
1026}
1027
1028fn get_tag(name: &'static str, response: &str) -> Result<String> {
1029 let start_tag = format!("<{}>", name);
1030 let end_tag = format!("</{}>", name);
1031
1032 let start_ix = response
1033 .find(&start_tag)
1034 .context(format!("{} start tag not found", name))?;
1035 let content_start_ix = start_ix + start_tag.len();
1036
1037 let end_ix = content_start_ix
1038 + response[content_start_ix..]
1039 .find(&end_tag)
1040 .context(format!("{} end tag not found", name))?;
1041
1042 let content = response[content_start_ix..end_ix].trim().unindent();
1043
1044 anyhow::Ok(content)
1045}
1046
1047pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
1048 let repo_name = repo_url
1049 .trim_start_matches("https://")
1050 .replace(|c: char| !c.is_alphanumeric(), "-");
1051 Path::new(repos_dir).join(repo_name)
1052}
1053
1054pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
1055 let output = new_smol_command("git")
1056 .current_dir(repo_path)
1057 .args(args)
1058 .output()
1059 .await?;
1060
1061 anyhow::ensure!(
1062 output.status.success(),
1063 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
1064 args.join(" "),
1065 repo_path.display(),
1066 output.status,
1067 String::from_utf8_lossy(&output.stderr),
1068 String::from_utf8_lossy(&output.stdout),
1069 );
1070 Ok(String::from_utf8(output.stdout)?.trim().to_string())
1071}
1072
1073fn push_role(role: &Role, buf: &mut String, assistant_message_number: &mut u32) {
1074 match role {
1075 Role::System => buf.push_str("# ⚙️ SYSTEM\n\n"),
1076 Role::User => buf.push_str("# 👤 USER\n\n"),
1077 Role::Assistant => {
1078 buf.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
1079 *assistant_message_number = *assistant_message_number + 1;
1080 }
1081 }
1082}
1083
1084pub async fn send_language_model_request(
1085 model: Arc<dyn LanguageModel>,
1086 request: LanguageModelRequest,
1087 cx: &AsyncApp,
1088) -> anyhow::Result<String> {
1089 match model.stream_completion_text(request, cx).await {
1090 Ok(mut stream) => {
1091 let mut full_response = String::new();
1092 while let Some(chunk_result) = stream.stream.next().await {
1093 match chunk_result {
1094 Ok(chunk_str) => {
1095 full_response.push_str(&chunk_str);
1096 }
1097 Err(err) => {
1098 anyhow::bail!("Error receiving response from language model: {err}");
1099 }
1100 }
1101 }
1102 Ok(full_response)
1103 }
1104 Err(err) => Err(anyhow!(
1105 "Failed to get response from language model. Error was: {err}"
1106 )),
1107 }
1108}
1109
1110pub struct RequestMarkdown {
1111 pub tools: String,
1112 pub messages: String,
1113}
1114
1115impl RequestMarkdown {
1116 pub fn new(request: &LanguageModelRequest) -> Self {
1117 let mut tools = String::new();
1118 let mut messages = String::new();
1119 let mut assistant_message_number: u32 = 1;
1120
1121 // Print the tools
1122 if !request.tools.is_empty() {
1123 for tool in &request.tools {
1124 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
1125 write!(&mut tools, "{}\n\n", tool.description).unwrap();
1126 writeln!(
1127 &mut tools,
1128 "{}",
1129 MarkdownCodeBlock {
1130 tag: "json",
1131 text: &format!("{:#}", tool.input_schema)
1132 }
1133 )
1134 .unwrap();
1135 }
1136 }
1137
1138 // Print the messages
1139 for message in &request.messages {
1140 push_role(&message.role, &mut messages, &mut assistant_message_number);
1141
1142 for content in &message.content {
1143 match content {
1144 MessageContent::Text(text) => {
1145 messages.push_str(text);
1146 messages.push_str("\n\n");
1147 }
1148 MessageContent::Image(_) => {
1149 messages.push_str("[IMAGE DATA]\n\n");
1150 }
1151 MessageContent::Thinking { text, signature } => {
1152 messages.push_str("**Thinking**:\n\n");
1153 if let Some(sig) = signature {
1154 messages.push_str(&format!("Signature: {}\n\n", sig));
1155 }
1156 messages.push_str(text);
1157 messages.push_str("\n");
1158 }
1159 MessageContent::RedactedThinking(items) => {
1160 messages.push_str(&format!(
1161 "**Redacted Thinking**: {} item(s)\n\n",
1162 items.len()
1163 ));
1164 }
1165 MessageContent::ToolUse(tool_use) => {
1166 messages.push_str(&format!(
1167 "**Tool Use**: {} (ID: {})\n",
1168 tool_use.name, tool_use.id
1169 ));
1170 messages.push_str(&format!(
1171 "{}\n",
1172 MarkdownCodeBlock {
1173 tag: "json",
1174 text: &format!("{:#}", tool_use.input)
1175 }
1176 ));
1177 }
1178 MessageContent::ToolResult(tool_result) => {
1179 messages.push_str(&format!(
1180 "**Tool Result**: {} (ID: {})\n\n",
1181 tool_result.tool_name, tool_result.tool_use_id
1182 ));
1183 if tool_result.is_error {
1184 messages.push_str("**ERROR:**\n");
1185 }
1186
1187 match &tool_result.content {
1188 LanguageModelToolResultContent::Text(text) => {
1189 writeln!(messages, "{text}\n").ok();
1190 }
1191 LanguageModelToolResultContent::Image(image) => {
1192 writeln!(messages, "\n", image.source).ok();
1193 }
1194 }
1195
1196 if let Some(output) = tool_result.output.as_ref() {
1197 writeln!(
1198 messages,
1199 "**Debug Output**:\n\n```json\n{}\n```\n",
1200 serde_json::to_string_pretty(output).unwrap()
1201 )
1202 .unwrap();
1203 }
1204 }
1205 }
1206 }
1207 }
1208
1209 Self { tools, messages }
1210 }
1211}
1212
1213pub fn response_events_to_markdown(
1214 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1215) -> String {
1216 let mut response = String::new();
1217 // Print the response events if any
1218 response.push_str("# Response\n\n");
1219 let mut text_buffer = String::new();
1220 let mut thinking_buffer = String::new();
1221
1222 let flush_buffers =
1223 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
1224 if !text_buffer.is_empty() {
1225 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
1226 text_buffer.clear();
1227 }
1228 if !thinking_buffer.is_empty() {
1229 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
1230 thinking_buffer.clear();
1231 }
1232 };
1233
1234 for event in response_events {
1235 match event {
1236 Ok(LanguageModelCompletionEvent::Text(text)) => {
1237 text_buffer.push_str(text);
1238 }
1239 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1240 thinking_buffer.push_str(text);
1241 }
1242 Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => {}
1243 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1244 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1245 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1246 }
1247 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1248 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1249 response.push_str(&format!(
1250 "**Tool Use**: {} (ID: {})\n",
1251 tool_use.name, tool_use.id
1252 ));
1253 response.push_str(&format!(
1254 "{}\n",
1255 MarkdownCodeBlock {
1256 tag: "json",
1257 text: &format!("{:#}", tool_use.input)
1258 }
1259 ));
1260 }
1261 Ok(
1262 LanguageModelCompletionEvent::UsageUpdate(_)
1263 | LanguageModelCompletionEvent::ToolUseLimitReached
1264 | LanguageModelCompletionEvent::StartMessage { .. }
1265 | LanguageModelCompletionEvent::UsageUpdated { .. }
1266 | LanguageModelCompletionEvent::Queued { .. }
1267 | LanguageModelCompletionEvent::Started
1268 | LanguageModelCompletionEvent::ReasoningDetails(_),
1269 ) => {}
1270 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1271 json_parse_error, ..
1272 }) => {
1273 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1274 response.push_str(&format!(
1275 "**Error**: parse error in tool use JSON: {}\n\n",
1276 json_parse_error
1277 ));
1278 }
1279 Err(error) => {
1280 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1281 response.push_str(&format!("**Error**: {}\n\n", error));
1282 }
1283 }
1284 }
1285
1286 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1287
1288 response
1289}
1290
1291#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
1292pub struct ThreadDialog {
1293 pub request: LanguageModelRequest,
1294 pub response_events: Vec<std::result::Result<LanguageModelCompletionEvent, String>>,
1295}
1296
1297impl ThreadDialog {
1298 pub fn new(
1299 request: &LanguageModelRequest,
1300 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1301 ) -> Self {
1302 Self {
1303 request: request.clone(),
1304 response_events: response_events.to_vec(),
1305 }
1306 }
1307
1308 /// Represents all request and response messages in a unified format.
1309 ///
1310 /// Specifically, it appends the assistant's response (derived from response events)
1311 /// as a new message to existing messages in the request.
1312 pub fn to_combined_request(&self) -> LanguageModelRequest {
1313 let mut request = self.request.clone();
1314 if let Some(assistant_message) = self.response_events_to_message() {
1315 request.messages.push(assistant_message);
1316 }
1317 request
1318 }
1319 fn response_events_to_message(&self) -> Option<LanguageModelRequestMessage> {
1320 let response_events = &self.response_events;
1321 let mut content: Vec<MessageContent> = Vec::new();
1322 let mut current_text = String::new();
1323
1324 let flush_text = |text: &mut String, content: &mut Vec<MessageContent>| {
1325 if !text.is_empty() {
1326 content.push(MessageContent::Text(std::mem::take(text)));
1327 }
1328 };
1329
1330 for event in response_events {
1331 match event {
1332 Ok(LanguageModelCompletionEvent::Text(text)) => {
1333 current_text.push_str(text);
1334 }
1335
1336 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1337 flush_text(&mut current_text, &mut content);
1338 if tool_use.is_input_complete {
1339 content.push(MessageContent::ToolUse(tool_use.clone()));
1340 }
1341 }
1342 Ok(LanguageModelCompletionEvent::Thinking { text, signature }) => {
1343 flush_text(&mut current_text, &mut content);
1344 content.push(MessageContent::Thinking {
1345 text: text.clone(),
1346 signature: signature.clone(),
1347 });
1348 }
1349
1350 // Skip these
1351 Ok(LanguageModelCompletionEvent::UsageUpdate(_))
1352 | Ok(LanguageModelCompletionEvent::RedactedThinking { .. })
1353 | Ok(LanguageModelCompletionEvent::StartMessage { .. })
1354 | Ok(LanguageModelCompletionEvent::ReasoningDetails(_))
1355 | Ok(LanguageModelCompletionEvent::Stop(_))
1356 | Ok(LanguageModelCompletionEvent::Queued { .. })
1357 | Ok(LanguageModelCompletionEvent::Started)
1358 | Ok(LanguageModelCompletionEvent::UsageUpdated { .. })
1359 | Ok(LanguageModelCompletionEvent::ToolUseLimitReached) => {}
1360
1361 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1362 json_parse_error,
1363 ..
1364 }) => {
1365 flush_text(&mut current_text, &mut content);
1366 content.push(MessageContent::Text(format!(
1367 "ERROR: parse error in tool use JSON: {}",
1368 json_parse_error
1369 )));
1370 }
1371
1372 Err(error) => {
1373 flush_text(&mut current_text, &mut content);
1374 content.push(MessageContent::Text(format!("ERROR: {}", error)));
1375 }
1376 }
1377 }
1378
1379 flush_text(&mut current_text, &mut content);
1380
1381 if !content.is_empty() {
1382 Some(LanguageModelRequestMessage {
1383 role: Role::Assistant,
1384 content,
1385 cache: false,
1386 reasoning_details: None,
1387 })
1388 } else {
1389 None
1390 }
1391 }
1392}
1393
1394#[cfg(test)]
1395mod test {
1396 use super::*;
1397
1398 #[test]
1399 fn test_parse_judge_output() {
1400 let response = r#"
1401 <analysis>The model did a good job but there were still compilations errors.</analysis>
1402 <passed>true</passed>
1403 "#
1404 .unindent();
1405
1406 let output = parse_assertion_result(&response).unwrap();
1407 assert_eq!(
1408 output.analysis,
1409 Some("The model did a good job but there were still compilations errors.".into())
1410 );
1411 assert!(output.passed);
1412
1413 let response = r#"
1414 Text around ignored
1415
1416 <analysis>
1417 Failed to compile:
1418 - Error 1
1419 - Error 2
1420 </analysis>
1421
1422 <passed>false</passed>
1423 "#
1424 .unindent();
1425
1426 let output = parse_assertion_result(&response).unwrap();
1427 assert_eq!(
1428 output.analysis,
1429 Some("Failed to compile:\n- Error 1\n- Error 2".into())
1430 );
1431 assert!(!output.passed);
1432 }
1433}