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(
307 rand::rng()
308 .sample_iter(&distr::Alphanumeric)
309 .take(7)
310 .map(char::from)
311 .collect::<String>()
312 .into(),
313 );
314
315 let db_thread = agent::DbThread::from_json(json.as_bytes()).expect("Can't read serialized thread");
316 cx.new(|cx| agent::Thread::from_db(session_id, db_thread, project.clone(), project_context, context_server_registry, agent::Templates::new(), cx))
317 } else {
318 cx.new(|cx| agent::Thread::new(project.clone(), project_context, context_server_registry, agent::Templates::new(), None, cx))
319 };
320
321 thread.update(cx, |thread, cx| {
322 thread.add_default_tools(Rc::new(EvalThreadEnvironment {
323 project: project.clone(),
324 }), cx);
325 thread.set_profile(meta.profile_id.clone(), cx);
326 thread.set_model(
327 LanguageModelInterceptor::new(
328 LanguageModelRegistry::read_global(cx).default_model().expect("Missing model").model.clone(),
329 this.run_directory.clone(),
330 last_diff_file_path.clone(),
331 this.run_directory.join("last.messages.json"),
332 this.worktree_path(),
333 this.repo_url(),
334 ),
335 cx,
336 );
337 });
338
339 thread
340 }).unwrap();
341
342 let mut example_cx = ExampleContext::new(
343 meta.clone(),
344 this.log_prefix.clone(),
345 thread.clone(),
346 cx.clone(),
347 );
348 let result = this.thread.conversation(&mut example_cx).await;
349
350 if let Err(err) = result
351 && !err.is::<FailedAssertion>() {
352 return Err(err);
353 }
354
355 println!("{}Stopped", this.log_prefix);
356
357 println!("{}Getting repository diff", this.log_prefix);
358 let repository_diff = Self::repository_diff(this.worktree_path(), &this.repo_url()).await?;
359
360 std::fs::write(last_diff_file_path, &repository_diff)?;
361
362
363 let mut diagnostics_after = None;
364 let mut diagnostic_summary_after = Default::default();
365
366 if let Some(language_server_state) = lsp {
367 wait_for_lang_server(&project, &language_server_state.language_file_buffer, this.log_prefix.clone(), cx).await?;
368
369 println!("{}Getting diagnostics", this.log_prefix);
370 diagnostics_after = cx
371 .update(|cx| {
372 let project = project.clone();
373 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
374 })?
375 .await?;
376 println!("{}Got diagnostics", this.log_prefix);
377
378 diagnostic_summary_after = project.read_with(cx, |project, cx| {
379 project.diagnostic_summary(false, cx)
380 })?;
381
382 }
383
384 if let Some(diagnostics_before) = &diagnostics_before {
385 fs::write(this.run_directory.join("diagnostics_before.txt"), diagnostics_before)?;
386 }
387
388 if let Some(diagnostics_after) = &diagnostics_after {
389 fs::write(this.run_directory.join("diagnostics_after.txt"), diagnostics_after)?;
390 }
391
392 thread.update(cx, |thread, _cx| {
393 RunOutput {
394 repository_diff,
395 diagnostic_summary_before,
396 diagnostic_summary_after,
397 diagnostics_before,
398 diagnostics_after,
399 token_usage: thread.latest_request_token_usage().unwrap(),
400 tool_metrics: example_cx.tool_metrics.lock().unwrap().clone(),
401 thread_markdown: thread.to_markdown(),
402 programmatic_assertions: example_cx.assertions,
403 }
404 })
405 })
406 }
407
408 async fn repository_diff(repository_path: PathBuf, repository_url: &str) -> Result<String> {
409 run_git(&repository_path, &["add", "."]).await?;
410 let mut diff_args = vec!["diff", "--staged"];
411 if repository_url == ZED_REPO_URL {
412 diff_args.push(":(exclude).rules");
413 }
414 run_git(&repository_path, &diff_args).await
415 }
416
417 pub async fn judge(
418 &self,
419 model: Arc<dyn LanguageModel>,
420 run_output: &RunOutput,
421 cx: &AsyncApp,
422 ) -> JudgeOutput {
423 let mut output_file =
424 File::create(self.run_directory.join("judge.md")).expect("failed to create judge.md");
425
426 let diff_task = self.judge_diff(model.clone(), run_output, cx);
427 let thread_task = self.judge_thread(model.clone(), run_output, cx);
428
429 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
430
431 let (diff_response, diff_output) = diff_result;
432 let (thread_response, thread_output) = thread_result;
433
434 writeln!(
435 &mut output_file,
436 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
437 )
438 .log_err();
439
440 JudgeOutput {
441 thread: thread_output,
442 diff: diff_output,
443 }
444 }
445
446 async fn judge_diff(
447 &self,
448 model: Arc<dyn LanguageModel>,
449 run_output: &RunOutput,
450 cx: &AsyncApp,
451 ) -> (String, AssertionsReport) {
452 let diff_assertions = self.thread.diff_assertions();
453
454 if diff_assertions.is_empty() {
455 return (
456 "No diff assertions".to_string(),
457 AssertionsReport::default(),
458 );
459 }
460
461 println!("{}Running diff judge", self.log_prefix);
462
463 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
464 let judge_diff_prompt_name = "judge_diff_prompt";
465 let mut hbs = Handlebars::new();
466 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)
467 .unwrap();
468
469 let to_prompt = |assertion: String| {
470 hbs.render(
471 judge_diff_prompt_name,
472 &JudgeDiffInput {
473 repository_diff: run_output.repository_diff.clone(),
474 assertion,
475 },
476 )
477 .unwrap()
478 };
479
480 let (responses, report) = self
481 .judge_assertions(model, diff_assertions, to_prompt, cx)
482 .await;
483
484 println!(
485 "{}Judge - Diff score: {}%",
486 self.log_prefix,
487 report.passed_percentage()
488 );
489
490 (responses, report)
491 }
492
493 async fn judge_thread(
494 &self,
495 model: Arc<dyn LanguageModel>,
496 run_output: &RunOutput,
497 cx: &AsyncApp,
498 ) -> (String, AssertionsReport) {
499 let thread_assertions = self.thread.thread_assertions();
500
501 if thread_assertions.is_empty() {
502 return (
503 "No thread assertions".to_string(),
504 AssertionsReport::default(),
505 );
506 }
507
508 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
509 let judge_thread_prompt_name = "judge_thread_prompt";
510 let mut hbs = Handlebars::new();
511 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)
512 .unwrap();
513
514 let complete_messages = &run_output.thread_markdown;
515 let to_prompt = |assertion: String| {
516 hbs.render(
517 judge_thread_prompt_name,
518 &JudgeThreadInput {
519 messages: complete_messages.clone(),
520 assertion,
521 },
522 )
523 .unwrap()
524 };
525
526 let (responses, report) = self
527 .judge_assertions(model, thread_assertions, to_prompt, cx)
528 .await;
529
530 println!(
531 "{}Judge - Thread score: {}%",
532 self.log_prefix,
533 report.passed_percentage()
534 );
535
536 (responses, report)
537 }
538
539 async fn judge_assertions(
540 &self,
541 model: Arc<dyn LanguageModel>,
542 assertions: Vec<JudgeAssertion>,
543 to_prompt: impl Fn(String) -> String,
544 cx: &AsyncApp,
545 ) -> (String, AssertionsReport) {
546 let assertions = assertions.into_iter().map(|assertion| {
547 let request = LanguageModelRequest {
548 thread_id: None,
549 prompt_id: None,
550 mode: None,
551 intent: None,
552 messages: vec![LanguageModelRequestMessage {
553 role: Role::User,
554 content: vec![MessageContent::Text(to_prompt(assertion.description))],
555 cache: false,
556 reasoning_details: None,
557 }],
558 temperature: None,
559 tools: Vec::new(),
560 tool_choice: None,
561 stop: Vec::new(),
562 thinking_allowed: true,
563 };
564
565 let model = model.clone();
566 let log_prefix = self.log_prefix.clone();
567 async move {
568 let response = send_language_model_request(model, request, cx).await;
569
570 let (response, result) = match response {
571 Ok(response) => (
572 response.clone(),
573 parse_assertion_result(&response).map_err(|err| err.to_string()),
574 ),
575 Err(err) => (err.to_string(), Err(err.to_string())),
576 };
577
578 if result.is_ok() {
579 println!("{}✅ {}", log_prefix, assertion.id);
580 } else {
581 println!("{}❌ {}", log_prefix, assertion.id);
582 }
583
584 (
585 response,
586 RanAssertion {
587 id: assertion.id,
588 result,
589 },
590 )
591 }
592 });
593
594 let mut responses = String::new();
595 let mut report = AssertionsReport::default();
596
597 for (response, assertion) in future::join_all(assertions).await {
598 writeln!(&mut responses, "# {}", assertion.id).unwrap();
599 writeln!(&mut responses, "{}\n\n", response).unwrap();
600 report.ran.push(assertion);
601 }
602
603 (responses, report)
604 }
605}
606
607struct EvalThreadEnvironment {
608 project: Entity<Project>,
609}
610
611struct EvalTerminalHandle {
612 terminal: Entity<acp_thread::Terminal>,
613}
614
615impl agent::TerminalHandle for EvalTerminalHandle {
616 fn id(&self, cx: &AsyncApp) -> Result<acp::TerminalId> {
617 self.terminal.read_with(cx, |term, _cx| term.id().clone())
618 }
619
620 fn wait_for_exit(&self, cx: &AsyncApp) -> Result<Shared<Task<acp::TerminalExitStatus>>> {
621 self.terminal
622 .read_with(cx, |term, _cx| term.wait_for_exit())
623 }
624
625 fn current_output(&self, cx: &AsyncApp) -> Result<acp::TerminalOutputResponse> {
626 self.terminal
627 .read_with(cx, |term, cx| term.current_output(cx))
628 }
629}
630
631impl agent::ThreadEnvironment for EvalThreadEnvironment {
632 fn create_terminal(
633 &self,
634 command: String,
635 cwd: Option<PathBuf>,
636 output_byte_limit: Option<u64>,
637 cx: &mut AsyncApp,
638 ) -> Task<Result<Rc<dyn agent::TerminalHandle>>> {
639 let project = self.project.clone();
640 cx.spawn(async move |cx| {
641 let language_registry =
642 project.read_with(cx, |project, _cx| project.languages().clone())?;
643 let id = acp::TerminalId(uuid::Uuid::new_v4().to_string().into());
644 let terminal =
645 acp_thread::create_terminal_entity(command, &[], vec![], cwd.clone(), &project, cx)
646 .await?;
647 let terminal = cx.new(|cx| {
648 acp_thread::Terminal::new(
649 id,
650 "",
651 cwd,
652 output_byte_limit.map(|limit| limit as usize),
653 terminal,
654 language_registry,
655 cx,
656 )
657 })?;
658 Ok(Rc::new(EvalTerminalHandle { terminal }) as Rc<dyn agent::TerminalHandle>)
659 })
660 }
661}
662
663struct LanguageModelInterceptor {
664 model: Arc<dyn LanguageModel>,
665 request_count: Arc<Mutex<usize>>,
666 previous_diff: Arc<Mutex<String>>,
667 example_output_dir: PathBuf,
668 last_diff_file_path: PathBuf,
669 messages_json_file_path: PathBuf,
670 repository_path: PathBuf,
671 repository_url: String,
672}
673
674impl LanguageModelInterceptor {
675 fn new(
676 model: Arc<dyn LanguageModel>,
677 example_output_dir: PathBuf,
678 last_diff_file_path: PathBuf,
679 messages_json_file_path: PathBuf,
680 repository_path: PathBuf,
681 repository_url: String,
682 ) -> Arc<Self> {
683 Arc::new(Self {
684 model,
685 request_count: Arc::new(Mutex::new(0)),
686 previous_diff: Arc::new(Mutex::new("".to_string())),
687 example_output_dir,
688 last_diff_file_path,
689 messages_json_file_path,
690 repository_path,
691 repository_url,
692 })
693 }
694}
695
696impl language_model::LanguageModel for LanguageModelInterceptor {
697 fn id(&self) -> language_model::LanguageModelId {
698 self.model.id()
699 }
700
701 fn name(&self) -> language_model::LanguageModelName {
702 self.model.name()
703 }
704
705 fn provider_id(&self) -> language_model::LanguageModelProviderId {
706 self.model.provider_id()
707 }
708
709 fn provider_name(&self) -> language_model::LanguageModelProviderName {
710 self.model.provider_name()
711 }
712
713 fn telemetry_id(&self) -> String {
714 self.model.telemetry_id()
715 }
716
717 fn supports_images(&self) -> bool {
718 self.model.supports_images()
719 }
720
721 fn supports_tools(&self) -> bool {
722 self.model.supports_tools()
723 }
724
725 fn supports_tool_choice(&self, choice: language_model::LanguageModelToolChoice) -> bool {
726 self.model.supports_tool_choice(choice)
727 }
728
729 fn max_token_count(&self) -> u64 {
730 self.model.max_token_count()
731 }
732
733 fn count_tokens(
734 &self,
735 request: LanguageModelRequest,
736 cx: &App,
737 ) -> future::BoxFuture<'static, Result<u64>> {
738 self.model.count_tokens(request, cx)
739 }
740
741 fn stream_completion(
742 &self,
743 request: LanguageModelRequest,
744 cx: &AsyncApp,
745 ) -> future::BoxFuture<
746 'static,
747 Result<
748 futures::stream::BoxStream<
749 'static,
750 Result<LanguageModelCompletionEvent, language_model::LanguageModelCompletionError>,
751 >,
752 language_model::LanguageModelCompletionError,
753 >,
754 > {
755 let stream = self.model.stream_completion(request.clone(), cx);
756 let request_count = self.request_count.clone();
757 let previous_diff = self.previous_diff.clone();
758 let example_output_dir = self.example_output_dir.clone();
759 let last_diff_file_path = self.last_diff_file_path.clone();
760 let messages_json_file_path = self.messages_json_file_path.clone();
761 let repository_path = self.repository_path.clone();
762 let repository_url = self.repository_url.clone();
763
764 Box::pin(async move {
765 let stream = stream.await?;
766
767 let response_events = Arc::new(Mutex::new(Vec::new()));
768 let request_clone = request.clone();
769
770 let wrapped_stream = stream.then(move |event| {
771 let response_events = response_events.clone();
772 let request = request_clone.clone();
773 let request_count = request_count.clone();
774 let previous_diff = previous_diff.clone();
775 let example_output_dir = example_output_dir.clone();
776 let last_diff_file_path = last_diff_file_path.clone();
777 let messages_json_file_path = messages_json_file_path.clone();
778 let repository_path = repository_path.clone();
779 let repository_url = repository_url.clone();
780
781 async move {
782 let event_result = match &event {
783 Ok(ev) => Ok(ev.clone()),
784 Err(err) => Err(err.to_string()),
785 };
786 response_events.lock().unwrap().push(event_result);
787
788 let should_execute = matches!(
789 &event,
790 Ok(LanguageModelCompletionEvent::Stop { .. }) | Err(_)
791 );
792
793 if should_execute {
794 let current_request_count = {
795 let mut count = request_count.lock().unwrap();
796 *count += 1;
797 *count
798 };
799
800 let messages_file_path =
801 example_output_dir.join(format!("{current_request_count}.messages.md"));
802 let diff_file_path =
803 example_output_dir.join(format!("{current_request_count}.diff"));
804 let last_messages_file_path = example_output_dir.join("last.messages.md");
805
806 let collected_events = response_events.lock().unwrap().clone();
807 let request_markdown = RequestMarkdown::new(&request);
808 let response_events_markdown =
809 response_events_to_markdown(&collected_events);
810 let dialog = ThreadDialog::new(&request, &collected_events);
811 let dialog_json =
812 serde_json::to_string_pretty(&dialog.to_combined_request())
813 .unwrap_or_default();
814
815 let messages = format!(
816 "{}\n\n{}",
817 request_markdown.messages, response_events_markdown
818 );
819 fs::write(&messages_file_path, messages.clone())
820 .expect("failed to write messages file");
821 fs::write(&last_messages_file_path, messages)
822 .expect("failed to write last messages file");
823 fs::write(&messages_json_file_path, dialog_json)
824 .expect("failed to write last.messages.json");
825
826 // Get repository diff
827 let diff_result =
828 ExampleInstance::repository_diff(repository_path, &repository_url)
829 .await;
830
831 match diff_result {
832 Ok(diff) => {
833 let prev_diff = previous_diff.lock().unwrap().clone();
834 if diff != prev_diff {
835 fs::write(&diff_file_path, &diff)
836 .expect("failed to write diff file");
837 fs::write(&last_diff_file_path, &diff)
838 .expect("failed to write last diff file");
839 *previous_diff.lock().unwrap() = diff;
840 }
841 }
842 Err(err) => {
843 let error_message = format!("{err:?}");
844 fs::write(&diff_file_path, &error_message)
845 .expect("failed to write diff error to file");
846 fs::write(&last_diff_file_path, &error_message)
847 .expect("failed to write last diff file");
848 }
849 }
850
851 if current_request_count == 1 {
852 let tools_file_path = example_output_dir.join("tools.md");
853 fs::write(tools_file_path, request_markdown.tools)
854 .expect("failed to write tools file");
855 }
856 }
857
858 event
859 }
860 });
861
862 Ok(Box::pin(wrapped_stream)
863 as futures::stream::BoxStream<
864 'static,
865 Result<
866 LanguageModelCompletionEvent,
867 language_model::LanguageModelCompletionError,
868 >,
869 >)
870 })
871 }
872}
873
874pub fn wait_for_lang_server(
875 project: &Entity<Project>,
876 buffer: &Entity<Buffer>,
877 log_prefix: String,
878 cx: &mut AsyncApp,
879) -> Task<Result<()>> {
880 if std::env::var("ZED_EVAL_SKIP_LS").is_ok() {
881 return Task::ready(Ok(()));
882 }
883
884 println!("{}⏵ Waiting for language server", log_prefix);
885
886 let (mut tx, mut rx) = mpsc::channel(1);
887
888 let lsp_store = project
889 .read_with(cx, |project, _| project.lsp_store())
890 .unwrap();
891
892 let has_lang_server = buffer
893 .update(cx, |buffer, cx| {
894 lsp_store.update(cx, |lsp_store, cx| {
895 lsp_store
896 .language_servers_for_local_buffer(buffer, cx)
897 .next()
898 .is_some()
899 })
900 })
901 .unwrap_or(false);
902
903 if has_lang_server {
904 project
905 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
906 .unwrap()
907 .detach();
908 }
909
910 let subscriptions =
911 [
912 cx.subscribe(&lsp_store, {
913 let log_prefix = log_prefix.clone();
914 move |_, event, _| {
915 if let project::LspStoreEvent::LanguageServerUpdate {
916 message:
917 client::proto::update_language_server::Variant::WorkProgress(
918 LspWorkProgress {
919 message: Some(message),
920 ..
921 },
922 ),
923 ..
924 } = event
925 {
926 println!("{}⟲ {message}", log_prefix)
927 }
928 }
929 }),
930 cx.subscribe(project, {
931 let buffer = buffer.clone();
932 move |project, event, cx| match event {
933 project::Event::LanguageServerAdded(_, _, _) => {
934 let buffer = buffer.clone();
935 project
936 .update(cx, |project, cx| project.save_buffer(buffer, cx))
937 .detach();
938 }
939 project::Event::DiskBasedDiagnosticsFinished { .. } => {
940 tx.try_send(()).ok();
941 }
942 _ => {}
943 }
944 }),
945 ];
946
947 cx.spawn(async move |cx| {
948 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
949 let result = futures::select! {
950 _ = rx.next() => {
951 println!("{}⚑ Language server idle", log_prefix);
952 anyhow::Ok(())
953 },
954 _ = timeout.fuse() => {
955 anyhow::bail!("LSP wait timed out after 5 minutes");
956 }
957 };
958 drop(subscriptions);
959 result
960 })
961}
962
963pub async fn query_lsp_diagnostics(
964 project: Entity<Project>,
965 cx: &mut AsyncApp,
966) -> Result<Option<String>> {
967 let paths_with_diagnostics = project.update(cx, |project, cx| {
968 project
969 .diagnostic_summaries(true, cx)
970 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
971 .map(|(project_path, _, _)| project_path)
972 .collect::<Vec<_>>()
973 })?;
974
975 if paths_with_diagnostics.is_empty() {
976 return Ok(None);
977 }
978
979 let mut output = String::new();
980 for project_path in paths_with_diagnostics {
981 let buffer = project
982 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
983 .await?;
984 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
985
986 for (_, group) in snapshot.diagnostic_groups(None) {
987 let entry = &group.entries[group.primary_ix];
988 let range = entry.range.to_point(&snapshot);
989 let severity = match entry.diagnostic.severity {
990 DiagnosticSeverity::ERROR => "error",
991 DiagnosticSeverity::WARNING => "warning",
992 _ => continue,
993 };
994
995 writeln!(
996 output,
997 "{} at line {}: {}",
998 severity,
999 range.start.row + 1,
1000 entry.diagnostic.message
1001 )?;
1002 }
1003 }
1004 anyhow::Ok(Some(output))
1005}
1006
1007fn parse_assertion_result(response: &str) -> Result<RanAssertionResult> {
1008 let analysis = get_tag("analysis", response)?;
1009 let passed = match get_tag("passed", response)?.to_lowercase().as_str() {
1010 "true" => true,
1011 "false" => false,
1012 value @ _ => bail!("invalid judge `passed` tag: {value}"),
1013 };
1014 Ok(RanAssertionResult {
1015 analysis: Some(analysis),
1016 passed,
1017 })
1018}
1019
1020fn get_tag(name: &'static str, response: &str) -> Result<String> {
1021 let start_tag = format!("<{}>", name);
1022 let end_tag = format!("</{}>", name);
1023
1024 let start_ix = response
1025 .find(&start_tag)
1026 .context(format!("{} start tag not found", name))?;
1027 let content_start_ix = start_ix + start_tag.len();
1028
1029 let end_ix = content_start_ix
1030 + response[content_start_ix..]
1031 .find(&end_tag)
1032 .context(format!("{} end tag not found", name))?;
1033
1034 let content = response[content_start_ix..end_ix].trim().unindent();
1035
1036 anyhow::Ok(content)
1037}
1038
1039pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
1040 let repo_name = repo_url
1041 .trim_start_matches("https://")
1042 .replace(|c: char| !c.is_alphanumeric(), "-");
1043 Path::new(repos_dir).join(repo_name)
1044}
1045
1046pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
1047 let output = new_smol_command("git")
1048 .current_dir(repo_path)
1049 .args(args)
1050 .output()
1051 .await?;
1052
1053 anyhow::ensure!(
1054 output.status.success(),
1055 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
1056 args.join(" "),
1057 repo_path.display(),
1058 output.status,
1059 String::from_utf8_lossy(&output.stderr),
1060 String::from_utf8_lossy(&output.stdout),
1061 );
1062 Ok(String::from_utf8(output.stdout)?.trim().to_string())
1063}
1064
1065fn push_role(role: &Role, buf: &mut String, assistant_message_number: &mut u32) {
1066 match role {
1067 Role::System => buf.push_str("# ⚙️ SYSTEM\n\n"),
1068 Role::User => buf.push_str("# 👤 USER\n\n"),
1069 Role::Assistant => {
1070 buf.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
1071 *assistant_message_number = *assistant_message_number + 1;
1072 }
1073 }
1074}
1075
1076pub async fn send_language_model_request(
1077 model: Arc<dyn LanguageModel>,
1078 request: LanguageModelRequest,
1079 cx: &AsyncApp,
1080) -> anyhow::Result<String> {
1081 match model.stream_completion_text(request, cx).await {
1082 Ok(mut stream) => {
1083 let mut full_response = String::new();
1084 while let Some(chunk_result) = stream.stream.next().await {
1085 match chunk_result {
1086 Ok(chunk_str) => {
1087 full_response.push_str(&chunk_str);
1088 }
1089 Err(err) => {
1090 anyhow::bail!("Error receiving response from language model: {err}");
1091 }
1092 }
1093 }
1094 Ok(full_response)
1095 }
1096 Err(err) => Err(anyhow!(
1097 "Failed to get response from language model. Error was: {err}"
1098 )),
1099 }
1100}
1101
1102pub struct RequestMarkdown {
1103 pub tools: String,
1104 pub messages: String,
1105}
1106
1107impl RequestMarkdown {
1108 pub fn new(request: &LanguageModelRequest) -> Self {
1109 let mut tools = String::new();
1110 let mut messages = String::new();
1111 let mut assistant_message_number: u32 = 1;
1112
1113 // Print the tools
1114 if !request.tools.is_empty() {
1115 for tool in &request.tools {
1116 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
1117 write!(&mut tools, "{}\n\n", tool.description).unwrap();
1118 writeln!(
1119 &mut tools,
1120 "{}",
1121 MarkdownCodeBlock {
1122 tag: "json",
1123 text: &format!("{:#}", tool.input_schema)
1124 }
1125 )
1126 .unwrap();
1127 }
1128 }
1129
1130 // Print the messages
1131 for message in &request.messages {
1132 push_role(&message.role, &mut messages, &mut assistant_message_number);
1133
1134 for content in &message.content {
1135 match content {
1136 MessageContent::Text(text) => {
1137 messages.push_str(text);
1138 messages.push_str("\n\n");
1139 }
1140 MessageContent::Image(_) => {
1141 messages.push_str("[IMAGE DATA]\n\n");
1142 }
1143 MessageContent::Thinking { text, signature } => {
1144 messages.push_str("**Thinking**:\n\n");
1145 if let Some(sig) = signature {
1146 messages.push_str(&format!("Signature: {}\n\n", sig));
1147 }
1148 messages.push_str(text);
1149 messages.push_str("\n");
1150 }
1151 MessageContent::RedactedThinking(items) => {
1152 messages.push_str(&format!(
1153 "**Redacted Thinking**: {} item(s)\n\n",
1154 items.len()
1155 ));
1156 }
1157 MessageContent::ToolUse(tool_use) => {
1158 messages.push_str(&format!(
1159 "**Tool Use**: {} (ID: {})\n",
1160 tool_use.name, tool_use.id
1161 ));
1162 messages.push_str(&format!(
1163 "{}\n",
1164 MarkdownCodeBlock {
1165 tag: "json",
1166 text: &format!("{:#}", tool_use.input)
1167 }
1168 ));
1169 }
1170 MessageContent::ToolResult(tool_result) => {
1171 messages.push_str(&format!(
1172 "**Tool Result**: {} (ID: {})\n\n",
1173 tool_result.tool_name, tool_result.tool_use_id
1174 ));
1175 if tool_result.is_error {
1176 messages.push_str("**ERROR:**\n");
1177 }
1178
1179 match &tool_result.content {
1180 LanguageModelToolResultContent::Text(text) => {
1181 writeln!(messages, "{text}\n").ok();
1182 }
1183 LanguageModelToolResultContent::Image(image) => {
1184 writeln!(messages, "\n", image.source).ok();
1185 }
1186 }
1187
1188 if let Some(output) = tool_result.output.as_ref() {
1189 writeln!(
1190 messages,
1191 "**Debug Output**:\n\n```json\n{}\n```\n",
1192 serde_json::to_string_pretty(output).unwrap()
1193 )
1194 .unwrap();
1195 }
1196 }
1197 }
1198 }
1199 }
1200
1201 Self { tools, messages }
1202 }
1203}
1204
1205pub fn response_events_to_markdown(
1206 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1207) -> String {
1208 let mut response = String::new();
1209 // Print the response events if any
1210 response.push_str("# Response\n\n");
1211 let mut text_buffer = String::new();
1212 let mut thinking_buffer = String::new();
1213
1214 let flush_buffers =
1215 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
1216 if !text_buffer.is_empty() {
1217 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
1218 text_buffer.clear();
1219 }
1220 if !thinking_buffer.is_empty() {
1221 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
1222 thinking_buffer.clear();
1223 }
1224 };
1225
1226 for event in response_events {
1227 match event {
1228 Ok(LanguageModelCompletionEvent::Text(text)) => {
1229 text_buffer.push_str(text);
1230 }
1231 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1232 thinking_buffer.push_str(text);
1233 }
1234 Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => {}
1235 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1236 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1237 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1238 }
1239 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1240 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1241 response.push_str(&format!(
1242 "**Tool Use**: {} (ID: {})\n",
1243 tool_use.name, tool_use.id
1244 ));
1245 response.push_str(&format!(
1246 "{}\n",
1247 MarkdownCodeBlock {
1248 tag: "json",
1249 text: &format!("{:#}", tool_use.input)
1250 }
1251 ));
1252 }
1253 Ok(
1254 LanguageModelCompletionEvent::UsageUpdate(_)
1255 | LanguageModelCompletionEvent::ToolUseLimitReached
1256 | LanguageModelCompletionEvent::StartMessage { .. }
1257 | LanguageModelCompletionEvent::UsageUpdated { .. }
1258 | LanguageModelCompletionEvent::Queued { .. }
1259 | LanguageModelCompletionEvent::Started
1260 | LanguageModelCompletionEvent::ReasoningDetails(_),
1261 ) => {}
1262 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1263 json_parse_error, ..
1264 }) => {
1265 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1266 response.push_str(&format!(
1267 "**Error**: parse error in tool use JSON: {}\n\n",
1268 json_parse_error
1269 ));
1270 }
1271 Err(error) => {
1272 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1273 response.push_str(&format!("**Error**: {}\n\n", error));
1274 }
1275 }
1276 }
1277
1278 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1279
1280 response
1281}
1282
1283#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
1284pub struct ThreadDialog {
1285 pub request: LanguageModelRequest,
1286 pub response_events: Vec<std::result::Result<LanguageModelCompletionEvent, String>>,
1287}
1288
1289impl ThreadDialog {
1290 pub fn new(
1291 request: &LanguageModelRequest,
1292 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1293 ) -> Self {
1294 Self {
1295 request: request.clone(),
1296 response_events: response_events.to_vec(),
1297 }
1298 }
1299
1300 /// Represents all request and response messages in a unified format.
1301 ///
1302 /// Specifically, it appends the assistant's response (derived from response events)
1303 /// as a new message to existing messages in the request.
1304 pub fn to_combined_request(&self) -> LanguageModelRequest {
1305 let mut request = self.request.clone();
1306 if let Some(assistant_message) = self.response_events_to_message() {
1307 request.messages.push(assistant_message);
1308 }
1309 request
1310 }
1311 fn response_events_to_message(&self) -> Option<LanguageModelRequestMessage> {
1312 let response_events = &self.response_events;
1313 let mut content: Vec<MessageContent> = Vec::new();
1314 let mut current_text = String::new();
1315
1316 let flush_text = |text: &mut String, content: &mut Vec<MessageContent>| {
1317 if !text.is_empty() {
1318 content.push(MessageContent::Text(std::mem::take(text)));
1319 }
1320 };
1321
1322 for event in response_events {
1323 match event {
1324 Ok(LanguageModelCompletionEvent::Text(text)) => {
1325 current_text.push_str(text);
1326 }
1327
1328 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1329 flush_text(&mut current_text, &mut content);
1330 if tool_use.is_input_complete {
1331 content.push(MessageContent::ToolUse(tool_use.clone()));
1332 }
1333 }
1334 Ok(LanguageModelCompletionEvent::Thinking { text, signature }) => {
1335 flush_text(&mut current_text, &mut content);
1336 content.push(MessageContent::Thinking {
1337 text: text.clone(),
1338 signature: signature.clone(),
1339 });
1340 }
1341
1342 // Skip these
1343 Ok(LanguageModelCompletionEvent::UsageUpdate(_))
1344 | Ok(LanguageModelCompletionEvent::RedactedThinking { .. })
1345 | Ok(LanguageModelCompletionEvent::StartMessage { .. })
1346 | Ok(LanguageModelCompletionEvent::ReasoningDetails(_))
1347 | Ok(LanguageModelCompletionEvent::Stop(_))
1348 | Ok(LanguageModelCompletionEvent::Queued { .. })
1349 | Ok(LanguageModelCompletionEvent::Started)
1350 | Ok(LanguageModelCompletionEvent::UsageUpdated { .. })
1351 | Ok(LanguageModelCompletionEvent::ToolUseLimitReached) => {}
1352
1353 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1354 json_parse_error,
1355 ..
1356 }) => {
1357 flush_text(&mut current_text, &mut content);
1358 content.push(MessageContent::Text(format!(
1359 "ERROR: parse error in tool use JSON: {}",
1360 json_parse_error
1361 )));
1362 }
1363
1364 Err(error) => {
1365 flush_text(&mut current_text, &mut content);
1366 content.push(MessageContent::Text(format!("ERROR: {}", error)));
1367 }
1368 }
1369 }
1370
1371 flush_text(&mut current_text, &mut content);
1372
1373 if !content.is_empty() {
1374 Some(LanguageModelRequestMessage {
1375 role: Role::Assistant,
1376 content,
1377 cache: false,
1378 reasoning_details: None,
1379 })
1380 } else {
1381 None
1382 }
1383 }
1384}
1385
1386#[cfg(test)]
1387mod test {
1388 use super::*;
1389
1390 #[test]
1391 fn test_parse_judge_output() {
1392 let response = r#"
1393 <analysis>The model did a good job but there were still compilations errors.</analysis>
1394 <passed>true</passed>
1395 "#
1396 .unindent();
1397
1398 let output = parse_assertion_result(&response).unwrap();
1399 assert_eq!(
1400 output.analysis,
1401 Some("The model did a good job but there were still compilations errors.".into())
1402 );
1403 assert!(output.passed);
1404
1405 let response = r#"
1406 Text around ignored
1407
1408 <analysis>
1409 Failed to compile:
1410 - Error 1
1411 - Error 2
1412 </analysis>
1413
1414 <passed>false</passed>
1415 "#
1416 .unindent();
1417
1418 let output = parse_assertion_result(&response).unwrap();
1419 assert_eq!(
1420 output.analysis,
1421 Some("Failed to compile:\n- Error 1\n- Error 2".into())
1422 );
1423 assert!(!output.passed);
1424 }
1425}