1use agent::ThreadStore;
2use anyhow::{Context, Result, anyhow, bail};
3use assistant_tool::ToolWorkingSet;
4use client::proto::LspWorkProgress;
5use futures::channel::mpsc;
6use futures::{FutureExt as _, StreamExt as _, future};
7use gpui::{App, AppContext as _, AsyncApp, Entity, Task};
8use handlebars::Handlebars;
9use language::{Buffer, DiagnosticSeverity, OffsetRangeExt as _};
10use language_model::{
11 LanguageModel, LanguageModelCompletionEvent, LanguageModelRequest, LanguageModelRequestMessage,
12 MessageContent, Role, TokenUsage,
13};
14use project::lsp_store::OpenLspBufferHandle;
15use project::{DiagnosticSummary, Project, ProjectPath};
16use serde::{Deserialize, Serialize};
17use std::cell::RefCell;
18use std::fmt::Write as _;
19use std::fs;
20use std::fs::File;
21use std::io::Write as _;
22use std::path::Path;
23use std::path::PathBuf;
24use std::rc::Rc;
25use std::sync::Arc;
26use std::time::Duration;
27use unindent::Unindent as _;
28use util::ResultExt as _;
29use util::command::new_smol_command;
30use util::markdown::MarkdownString;
31
32use crate::assertions::{AssertionsReport, RanAssertion, RanAssertionResult};
33use crate::example::{Example, ExampleContext, FailedAssertion, JudgeAssertion};
34use crate::{AgentAppState, ToolMetrics};
35
36pub const ZED_REPO_URL: &str = "https://github.com/zed-industries/zed.git";
37
38#[derive(Clone)]
39pub struct ExampleInstance {
40 pub thread: Rc<dyn Example>,
41 pub name: String,
42 pub run_directory: PathBuf,
43 pub log_prefix: String,
44 /// The repetition number for this example (0-based)
45 /// When running multiple repetitions of the same example, each instance is assigned a unique repetition number.
46 /// This affects the worktree path and log prefix to avoid clobbering results between runs.
47 pub repetition: usize,
48 pub repo_path: PathBuf,
49 /// Path to the directory containing the requests and responses for the agentic loop
50 worktrees_dir: PathBuf,
51}
52
53#[derive(Debug, Serialize, Clone)]
54pub struct RunOutput {
55 pub repository_diff: String,
56 pub diagnostic_summary_before: DiagnosticSummary,
57 pub diagnostic_summary_after: DiagnosticSummary,
58 pub diagnostics_before: Option<String>,
59 pub diagnostics_after: Option<String>,
60 pub response_count: usize,
61 pub token_usage: TokenUsage,
62 pub tool_metrics: ToolMetrics,
63 pub last_request: LanguageModelRequest,
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.to_string();
94 let run_directory = run_dir
95 .join(&name)
96 .join(repetition.to_string())
97 .to_path_buf();
98
99 let repo_path = repo_path_for_url(repos_dir, &thread.meta().url);
100
101 Self {
102 name,
103 thread,
104 log_prefix: String::new(),
105 run_directory,
106 repetition,
107 repo_path,
108 worktrees_dir: worktrees_dir.to_path_buf(),
109 }
110 }
111
112 pub fn repo_url(&self) -> String {
113 self.thread.meta().url
114 }
115
116 pub fn revision(&self) -> String {
117 self.thread.meta().revision
118 }
119
120 pub fn worktree_name(&self) -> String {
121 format!("{}-{}", self.name, self.repetition)
122 }
123
124 pub fn set_log_prefix_style(&mut self, color: &str, name_width: usize) {
125 self.log_prefix = format!(
126 "{}{:<width$}\x1b[0m | ",
127 color,
128 self.worktree_name(),
129 width = name_width
130 );
131 }
132
133 /// Set up the example by checking out the specified Git revision
134 pub async fn fetch(&mut self) -> Result<()> {
135 let meta = self.thread.meta();
136
137 let revision_exists = run_git(
138 &self.repo_path,
139 &["rev-parse", &format!("{}^{{commit}}", &meta.revision)],
140 )
141 .await
142 .is_ok();
143
144 if !revision_exists {
145 println!("{}Fetching revision {}", self.log_prefix, &meta.revision);
146 run_git(
147 &self.repo_path,
148 &["fetch", "--depth", "1", "origin", &meta.revision],
149 )
150 .await?;
151 }
152 Ok(())
153 }
154
155 /// Set up the example by checking out the specified Git revision
156 pub async fn setup(&mut self) -> Result<()> {
157 let worktree_path = self.worktree_path();
158 let meta = self.thread.meta();
159 if worktree_path.is_dir() {
160 println!("{}Resetting existing worktree", self.log_prefix);
161
162 // TODO: consider including "-x" to remove ignored files. The downside of this is that
163 // it will also remove build artifacts, and so prevent incremental reuse there.
164 run_git(&worktree_path, &["clean", "--force", "-d"]).await?;
165 run_git(&worktree_path, &["reset", "--hard", "HEAD"]).await?;
166 run_git(&worktree_path, &["checkout", &meta.revision]).await?;
167 } else {
168 println!("{}Creating worktree", self.log_prefix);
169
170 let worktree_path_string = worktree_path.to_string_lossy().to_string();
171
172 run_git(
173 &self.repo_path,
174 &[
175 "worktree",
176 "add",
177 "-f",
178 &worktree_path_string,
179 &meta.revision,
180 ],
181 )
182 .await?;
183 }
184
185 if meta.url == ZED_REPO_URL {
186 std::fs::write(worktree_path.join(".rules"), std::fs::read(".rules")?)?;
187 }
188
189 std::fs::create_dir_all(&self.run_directory)?;
190
191 Ok(())
192 }
193
194 pub fn worktree_path(&self) -> PathBuf {
195 self.worktrees_dir
196 .join(self.worktree_name())
197 .join(self.thread.meta().repo_name())
198 }
199
200 pub fn run(
201 &self,
202 model: Arc<dyn LanguageModel>,
203 app_state: Arc<AgentAppState>,
204 cx: &mut App,
205 ) -> Task<Result<RunOutput>> {
206 let project = Project::local(
207 app_state.client.clone(),
208 app_state.node_runtime.clone(),
209 app_state.user_store.clone(),
210 app_state.languages.clone(),
211 app_state.fs.clone(),
212 None,
213 cx,
214 );
215
216 let worktree = project.update(cx, |project, cx| {
217 project.create_worktree(self.worktree_path(), true, cx)
218 });
219
220 let tools = cx.new(|_| ToolWorkingSet::default());
221 let thread_store =
222 ThreadStore::load(project.clone(), tools, app_state.prompt_builder.clone(), cx);
223 let meta = self.thread.meta();
224 let this = self.clone();
225
226 cx.spawn(async move |cx| {
227 let worktree = worktree.await?;
228
229 // Wait for worktree scan to finish before choosing a file to open.
230 worktree
231 .update(cx, |worktree, _cx| {
232 worktree.as_local().unwrap().scan_complete()
233 })?
234 .await;
235
236 struct LanguageServerState {
237 _lsp_open_handle: OpenLspBufferHandle,
238 language_file_buffer: Entity<Buffer>,
239 }
240
241 let mut diagnostics_before = None;
242 let mut diagnostic_summary_before = DiagnosticSummary::default();
243
244 let lsp = if let Some(language_server) = &meta.language_server {
245 // Open a file that matches the language to cause LSP to start.
246 let language_file = worktree.read_with(cx, |worktree, _cx| {
247 worktree
248 .files(false, 0)
249 .find_map(|e| {
250 if e.path.clone().extension().and_then(|ext| ext.to_str())
251 == Some(&language_server.file_extension)
252 {
253 Some(ProjectPath {
254 worktree_id: worktree.id(),
255 path: e.path.clone(),
256 })
257 } else {
258 None
259 }
260 })
261 .context("Failed to find a file for example language")
262 })??;
263
264 let open_language_file_buffer_task = project.update(cx, |project, cx| {
265 project.open_buffer(language_file.clone(), cx)
266 })?;
267
268 let language_file_buffer = open_language_file_buffer_task.await?;
269
270 let lsp_open_handle = project.update(cx, |project, cx| {
271 project.register_buffer_with_language_servers(&language_file_buffer, cx)
272 })?;
273
274 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
275
276 diagnostic_summary_before = project.read_with(cx, |project, cx| {
277 project.diagnostic_summary(false, cx)
278 })?;
279
280 diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
281 if diagnostics_before.is_some() && language_server.allow_preexisting_diagnostics {
282 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`"));
283 }
284
285 Some(LanguageServerState {
286 _lsp_open_handle: lsp_open_handle,
287 language_file_buffer,
288 })
289 } else {
290 None
291 };
292
293 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
294 return Err(anyhow!("Setup only mode"));
295 }
296
297 let last_diff_file_path = this.run_directory.join("last.diff");
298
299 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
300 // history using undo/redo.
301 std::fs::write(&last_diff_file_path, "")?;
302
303 let thread_store = thread_store.await?;
304 let thread =
305 thread_store.update(cx, |thread_store, cx| thread_store.create_thread(cx))?;
306 let last_request = Rc::new(RefCell::new(None));
307
308 thread.update(cx, |thread, _cx| {
309 let mut request_count = 0;
310 let last_request = Rc::clone(&last_request);
311 let previous_diff = Rc::new(RefCell::new("".to_string()));
312 let example_output_dir = this.run_directory.clone();
313 let last_diff_file_path = last_diff_file_path.clone();
314 let this = this.clone();
315 thread.set_request_callback(move |request, response_events| {
316 *last_request.borrow_mut() = Some(request.clone());
317
318 request_count += 1;
319 let messages_file_path = example_output_dir.join(format!("{request_count}.messages.md"));
320 let diff_file_path = example_output_dir.join(format!("{request_count}.diff"));
321 let last_messages_file_path = example_output_dir.join("last.messages.md");
322 let request_markdown = RequestMarkdown::new(request);
323 let response_events_markdown = response_events_to_markdown(response_events);
324
325 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
326 fs::write(&messages_file_path, messages.clone()).expect("failed to write messages file");
327 fs::write(&last_messages_file_path, messages).expect("failed to write last messages file");
328
329 let diff_result = smol::block_on(this.repository_diff());
330 match diff_result {
331 Ok(diff) => {
332 if diff != previous_diff.borrow().clone() {
333 fs::write(&diff_file_path, &diff).expect("failed to write diff file");
334 fs::write(&last_diff_file_path, &diff).expect("failed to write last diff file");
335 *previous_diff.borrow_mut() = diff;
336 }
337 }
338 Err(err) => {
339 let error_message = format!("{err:?}");
340 fs::write(&diff_file_path, &error_message).expect("failed to write diff error to file");
341 fs::write(&last_diff_file_path, &error_message).expect("failed to write last diff file");
342 }
343 }
344
345 if request_count == 1 {
346 let tools_file_path = example_output_dir.join("tools.md");
347 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
348 }
349 });
350 })?;
351
352 let mut example_cx = ExampleContext::new(meta.clone(), this.log_prefix.clone(), thread.clone(), model.clone(), cx.clone());
353 let result = this.thread.conversation(&mut example_cx).await;
354
355 if let Err(err) = result {
356 if !err.is::<FailedAssertion>() {
357 return Err(err);
358 }
359 }
360
361 println!("{}Stopped", this.log_prefix);
362
363 println!("{}Getting repository diff", this.log_prefix);
364 let repository_diff = this.repository_diff().await?;
365
366 std::fs::write(last_diff_file_path, &repository_diff)?;
367
368
369 let mut diagnostics_after = None;
370 let mut diagnostic_summary_after = Default::default();
371
372 if let Some(language_server_state) = lsp {
373 wait_for_lang_server(&project, &language_server_state.language_file_buffer, this.log_prefix.clone(), cx).await?;
374
375 println!("{}Getting diagnostics", this.log_prefix);
376 diagnostics_after = cx
377 .update(|cx| {
378 let project = project.clone();
379 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
380 })?
381 .await?;
382 println!("{}Got diagnostics", this.log_prefix);
383
384 diagnostic_summary_after = project.read_with(cx, |project, cx| {
385 project.diagnostic_summary(false, cx)
386 })?;
387
388 }
389
390 let Some(last_request) = last_request.borrow_mut().take() else {
391 return Err(anyhow!("No requests ran."));
392 };
393
394 if let Some(diagnostics_before) = &diagnostics_before {
395 fs::write(this.run_directory.join("diagnostics_before.txt"), diagnostics_before)?;
396 }
397
398 if let Some(diagnostics_after) = &diagnostics_after {
399 fs::write(this.run_directory.join("diagnostics_after.txt"), diagnostics_after)?;
400 }
401
402 thread.update(cx, |thread, _cx| {
403 let response_count = thread
404 .messages()
405 .filter(|message| message.role == language_model::Role::Assistant)
406 .count();
407 RunOutput {
408 repository_diff,
409 diagnostic_summary_before,
410 diagnostic_summary_after,
411 diagnostics_before,
412 diagnostics_after,
413 response_count,
414 token_usage: thread.cumulative_token_usage(),
415 tool_metrics: example_cx.tool_metrics.lock().unwrap().clone(),
416 last_request,
417 programmatic_assertions: example_cx.assertions,
418 }
419 })
420 })
421 }
422
423 async fn repository_diff(&self) -> Result<String> {
424 let worktree_path = self.worktree_path();
425 run_git(&worktree_path, &["add", "."]).await?;
426 let mut diff_args = vec!["diff", "--staged"];
427 if self.thread.meta().url == ZED_REPO_URL {
428 diff_args.push(":(exclude).rules");
429 }
430 run_git(&worktree_path, &diff_args).await
431 }
432
433 pub async fn judge(
434 &self,
435 model: Arc<dyn LanguageModel>,
436 run_output: &RunOutput,
437 cx: &AsyncApp,
438 ) -> JudgeOutput {
439 let mut output_file =
440 File::create(self.run_directory.join("judge.md")).expect("failed to create judge.md");
441
442 let diff_task = self.judge_diff(model.clone(), &run_output, cx);
443 let thread_task = self.judge_thread(model.clone(), &run_output, cx);
444
445 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
446
447 let (diff_response, diff_output) = diff_result;
448 let (thread_response, thread_output) = thread_result;
449
450 writeln!(
451 &mut output_file,
452 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
453 )
454 .log_err();
455
456 JudgeOutput {
457 thread: thread_output,
458 diff: diff_output,
459 }
460 }
461
462 async fn judge_diff(
463 &self,
464 model: Arc<dyn LanguageModel>,
465 run_output: &RunOutput,
466 cx: &AsyncApp,
467 ) -> (String, AssertionsReport) {
468 let diff_assertions = self.thread.diff_assertions();
469
470 if diff_assertions.is_empty() {
471 return (
472 "No diff assertions".to_string(),
473 AssertionsReport::default(),
474 );
475 }
476
477 println!("{}Running diff judge", self.log_prefix);
478
479 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
480 let judge_diff_prompt_name = "judge_diff_prompt";
481 let mut hbs = Handlebars::new();
482 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)
483 .unwrap();
484
485 let to_prompt = |assertion: String| {
486 hbs.render(
487 judge_diff_prompt_name,
488 &JudgeDiffInput {
489 repository_diff: run_output.repository_diff.clone(),
490 assertion,
491 },
492 )
493 .unwrap()
494 };
495
496 let (responses, report) = self
497 .judge_assertions(model, diff_assertions, to_prompt, cx)
498 .await;
499
500 println!(
501 "{}Judge - Diff score: {}%",
502 self.log_prefix,
503 report.passed_percentage()
504 );
505
506 (responses, report)
507 }
508
509 async fn judge_thread(
510 &self,
511 model: Arc<dyn LanguageModel>,
512 run_output: &RunOutput,
513 cx: &AsyncApp,
514 ) -> (String, AssertionsReport) {
515 let thread_assertions = self.thread.thread_assertions();
516
517 if thread_assertions.is_empty() {
518 return (
519 "No diff assertions".to_string(),
520 AssertionsReport::default(),
521 );
522 }
523
524 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
525 let judge_diff_prompt_name = "judge_thread_prompt";
526 let mut hbs = Handlebars::new();
527 hbs.register_template_string(judge_diff_prompt_name, judge_thread_prompt)
528 .unwrap();
529
530 let request_markdown = RequestMarkdown::new(&run_output.last_request);
531 let to_prompt = |assertion: String| {
532 hbs.render(
533 judge_diff_prompt_name,
534 &JudgeThreadInput {
535 messages: request_markdown.messages.clone(),
536 assertion,
537 },
538 )
539 .unwrap()
540 };
541
542 let (responses, report) = self
543 .judge_assertions(model, thread_assertions, to_prompt, cx)
544 .await;
545
546 println!(
547 "{}Judge - Thread score: {}%",
548 self.log_prefix,
549 report.passed_percentage()
550 );
551
552 (responses, report)
553 }
554
555 async fn judge_assertions(
556 &self,
557 model: Arc<dyn LanguageModel>,
558 assertions: Vec<JudgeAssertion>,
559 to_prompt: impl Fn(String) -> String,
560 cx: &AsyncApp,
561 ) -> (String, AssertionsReport) {
562 let assertions = assertions.into_iter().map(|assertion| {
563 let request = LanguageModelRequest {
564 thread_id: None,
565 prompt_id: None,
566 messages: vec![LanguageModelRequestMessage {
567 role: Role::User,
568 content: vec![MessageContent::Text(to_prompt(assertion.description))],
569 cache: false,
570 }],
571 temperature: None,
572 tools: Vec::new(),
573 stop: Vec::new(),
574 };
575
576 let model = model.clone();
577 let log_prefix = self.log_prefix.clone();
578 async move {
579 let response = send_language_model_request(model, request, cx).await;
580
581 let (response, result) = match response {
582 Ok(response) => (
583 response.clone(),
584 parse_assertion_result(&response).map_err(|err| err.to_string()),
585 ),
586 Err(err) => (err.to_string(), Err(err.to_string())),
587 };
588
589 if result.is_ok() {
590 println!("{}✅ {}", log_prefix, assertion.id);
591 } else {
592 println!("{}❌ {}", log_prefix, assertion.id);
593 }
594
595 (
596 response,
597 RanAssertion {
598 id: assertion.id,
599 result,
600 },
601 )
602 }
603 });
604
605 let mut responses = String::new();
606 let mut report = AssertionsReport::default();
607
608 for (response, assertion) in future::join_all(assertions).await {
609 writeln!(&mut responses, "# {}", assertion.id).unwrap();
610 writeln!(&mut responses, "{}\n\n", response).unwrap();
611 report.ran.push(assertion);
612 }
613
614 (responses, report)
615 }
616}
617
618pub fn wait_for_lang_server(
619 project: &Entity<Project>,
620 buffer: &Entity<Buffer>,
621 log_prefix: String,
622 cx: &mut AsyncApp,
623) -> Task<Result<()>> {
624 if std::env::var("ZED_EVAL_SKIP_LS").is_ok() {
625 return Task::ready(Ok(()));
626 }
627
628 println!("{}⏵ Waiting for language server", log_prefix);
629
630 let (mut tx, mut rx) = mpsc::channel(1);
631
632 let lsp_store = project
633 .update(cx, |project, _| project.lsp_store())
634 .unwrap();
635
636 let has_lang_server = buffer
637 .update(cx, |buffer, cx| {
638 lsp_store.update(cx, |lsp_store, cx| {
639 lsp_store
640 .language_servers_for_local_buffer(&buffer, cx)
641 .next()
642 .is_some()
643 })
644 })
645 .unwrap_or(false);
646
647 if has_lang_server {
648 project
649 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
650 .unwrap()
651 .detach();
652 }
653
654 let subscriptions =
655 [
656 cx.subscribe(&lsp_store, {
657 let log_prefix = log_prefix.clone();
658 move |_, event, _| match event {
659 project::LspStoreEvent::LanguageServerUpdate {
660 message:
661 client::proto::update_language_server::Variant::WorkProgress(
662 LspWorkProgress {
663 message: Some(message),
664 ..
665 },
666 ),
667 ..
668 } => println!("{}⟲ {message}", log_prefix),
669 _ => {}
670 }
671 }),
672 cx.subscribe(&project, {
673 let buffer = buffer.clone();
674 move |project, event, cx| match event {
675 project::Event::LanguageServerAdded(_, _, _) => {
676 let buffer = buffer.clone();
677 project
678 .update(cx, |project, cx| project.save_buffer(buffer, cx))
679 .detach();
680 }
681 project::Event::DiskBasedDiagnosticsFinished { .. } => {
682 tx.try_send(()).ok();
683 }
684 _ => {}
685 }
686 }),
687 ];
688
689 cx.spawn(async move |cx| {
690 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
691 let result = futures::select! {
692 _ = rx.next() => {
693 println!("{}⚑ Language server idle", log_prefix);
694 anyhow::Ok(())
695 },
696 _ = timeout.fuse() => {
697 Err(anyhow!("LSP wait timed out after 5 minutes"))
698 }
699 };
700 drop(subscriptions);
701 result
702 })
703}
704
705pub async fn query_lsp_diagnostics(
706 project: Entity<Project>,
707 cx: &mut AsyncApp,
708) -> Result<Option<String>> {
709 let paths_with_diagnostics = project.update(cx, |project, cx| {
710 project
711 .diagnostic_summaries(true, cx)
712 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
713 .map(|(project_path, _, _)| project_path)
714 .collect::<Vec<_>>()
715 })?;
716
717 if paths_with_diagnostics.is_empty() {
718 return Ok(None);
719 }
720
721 let mut output = String::new();
722 for project_path in paths_with_diagnostics {
723 let buffer = project
724 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
725 .await?;
726 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
727
728 for (_, group) in snapshot.diagnostic_groups(None) {
729 let entry = &group.entries[group.primary_ix];
730 let range = entry.range.to_point(&snapshot);
731 let severity = match entry.diagnostic.severity {
732 DiagnosticSeverity::ERROR => "error",
733 DiagnosticSeverity::WARNING => "warning",
734 _ => continue,
735 };
736
737 writeln!(
738 output,
739 "{} at line {}: {}",
740 severity,
741 range.start.row + 1,
742 entry.diagnostic.message
743 )?;
744 }
745 }
746 anyhow::Ok(Some(output))
747}
748
749fn parse_assertion_result(response: &str) -> Result<RanAssertionResult> {
750 let analysis = get_tag("analysis", response)?.to_string();
751 let passed = match get_tag("passed", response)?.to_lowercase().as_str() {
752 "true" => true,
753 "false" => false,
754 value @ _ => bail!("invalid judge `passed` tag: {value}"),
755 };
756 Ok(RanAssertionResult {
757 analysis: Some(analysis),
758 passed,
759 })
760}
761
762fn get_tag(name: &'static str, response: &str) -> Result<String> {
763 let start_tag = format!("<{}>", name);
764 let end_tag = format!("</{}>", name);
765
766 let start_ix = response
767 .find(&start_tag)
768 .context(format!("{} start tag not found", name))?;
769 let content_start_ix = start_ix + start_tag.len();
770
771 let end_ix = content_start_ix
772 + response[content_start_ix..]
773 .find(&end_tag)
774 .context(format!("{} end tag not found", name))?;
775
776 let content = response[content_start_ix..end_ix].trim().unindent();
777
778 anyhow::Ok(content)
779}
780
781pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
782 let repo_name = repo_url
783 .trim_start_matches("https://")
784 .replace(|c: char| !c.is_alphanumeric(), "-");
785 Path::new(repos_dir).join(repo_name)
786}
787
788pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
789 let output = new_smol_command("git")
790 .current_dir(repo_path)
791 .args(args)
792 .output()
793 .await?;
794
795 if output.status.success() {
796 Ok(String::from_utf8(output.stdout)?.trim().to_string())
797 } else {
798 Err(anyhow!(
799 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
800 args.join(" "),
801 repo_path.display(),
802 output.status,
803 String::from_utf8_lossy(&output.stderr),
804 String::from_utf8_lossy(&output.stdout),
805 ))
806 }
807}
808
809pub async fn send_language_model_request(
810 model: Arc<dyn LanguageModel>,
811 request: LanguageModelRequest,
812 cx: &AsyncApp,
813) -> anyhow::Result<String> {
814 match model.stream_completion_text(request, &cx).await {
815 Ok(mut stream) => {
816 let mut full_response = String::new();
817 while let Some(chunk_result) = stream.stream.next().await {
818 match chunk_result {
819 Ok(chunk_str) => {
820 full_response.push_str(&chunk_str);
821 }
822 Err(err) => {
823 return Err(anyhow!(
824 "Error receiving response from language model: {err}"
825 ));
826 }
827 }
828 }
829 Ok(full_response)
830 }
831 Err(err) => Err(anyhow!(
832 "Failed to get response from language model. Error was: {err}"
833 )),
834 }
835}
836
837pub struct RequestMarkdown {
838 pub tools: String,
839 pub messages: String,
840}
841
842impl RequestMarkdown {
843 pub fn new(request: &LanguageModelRequest) -> Self {
844 let mut tools = String::new();
845 let mut messages = String::new();
846 let mut assistant_message_number: u32 = 1;
847
848 // Print the tools
849 if !request.tools.is_empty() {
850 for tool in &request.tools {
851 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
852 write!(&mut tools, "{}\n\n", tool.description).unwrap();
853 write!(
854 &mut tools,
855 "{}\n",
856 MarkdownString::code_block("json", &format!("{:#}", tool.input_schema))
857 )
858 .unwrap();
859 }
860 }
861
862 // Print the messages
863 for message in &request.messages {
864 match message.role {
865 Role::System => messages.push_str("# ⚙️ SYSTEM\n\n"),
866 Role::User => messages.push_str("# 👤 USER\n\n"),
867 Role::Assistant => {
868 messages.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
869 assistant_message_number += 1;
870 }
871 };
872
873 for content in &message.content {
874 match content {
875 MessageContent::Text(text) => {
876 messages.push_str(text);
877 messages.push_str("\n\n");
878 }
879 MessageContent::Image(_) => {
880 messages.push_str("[IMAGE DATA]\n\n");
881 }
882 MessageContent::Thinking { text, signature } => {
883 messages.push_str("**Thinking**:\n\n");
884 if let Some(sig) = signature {
885 messages.push_str(&format!("Signature: {}\n\n", sig));
886 }
887 messages.push_str(text);
888 messages.push_str("\n");
889 }
890 MessageContent::RedactedThinking(items) => {
891 messages.push_str(&format!(
892 "**Redacted Thinking**: {} item(s)\n\n",
893 items.len()
894 ));
895 }
896 MessageContent::ToolUse(tool_use) => {
897 messages.push_str(&format!(
898 "**Tool Use**: {} (ID: {})\n",
899 tool_use.name, tool_use.id
900 ));
901 messages.push_str(&format!(
902 "{}\n",
903 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
904 ));
905 }
906 MessageContent::ToolResult(tool_result) => {
907 messages.push_str(&format!(
908 "**Tool Result**: {} (ID: {})\n\n",
909 tool_result.tool_name, tool_result.tool_use_id
910 ));
911 if tool_result.is_error {
912 messages.push_str("**ERROR:**\n");
913 }
914 messages.push_str(&format!("{}\n\n", tool_result.content));
915 }
916 }
917 }
918 }
919
920 Self { tools, messages }
921 }
922}
923
924pub fn response_events_to_markdown(
925 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
926) -> String {
927 let mut response = String::new();
928 // Print the response events if any
929 response.push_str("# Response\n\n");
930 let mut text_buffer = String::new();
931 let mut thinking_buffer = String::new();
932
933 let flush_buffers =
934 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
935 if !text_buffer.is_empty() {
936 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
937 text_buffer.clear();
938 }
939 if !thinking_buffer.is_empty() {
940 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
941 thinking_buffer.clear();
942 }
943 };
944
945 for event in response_events {
946 match event {
947 Ok(LanguageModelCompletionEvent::Text(text)) => {
948 text_buffer.push_str(text);
949 }
950 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
951 thinking_buffer.push_str(text);
952 }
953 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
954 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
955 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
956 }
957 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
958 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
959 response.push_str(&format!(
960 "**Tool Use**: {} (ID: {})\n",
961 tool_use.name, tool_use.id
962 ));
963 response.push_str(&format!(
964 "{}\n",
965 MarkdownString::code_block("json", &format!("{:#}", tool_use.input))
966 ));
967 }
968 Ok(
969 LanguageModelCompletionEvent::UsageUpdate(_)
970 | LanguageModelCompletionEvent::StartMessage { .. },
971 ) => {}
972 Err(error) => {
973 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
974 response.push_str(&format!("**Error**: {}\n\n", error));
975 }
976 }
977 }
978
979 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
980
981 response
982}
983
984#[cfg(test)]
985mod test {
986 use super::*;
987
988 #[test]
989 fn test_parse_judge_output() {
990 let response = r#"
991 <analysis>The model did a good job but there were still compilations errors.</analysis>
992 <passed>true</passed>
993 "#
994 .unindent();
995
996 let output = parse_assertion_result(&response).unwrap();
997 assert_eq!(
998 output.analysis,
999 Some("The model did a good job but there were still compilations errors.".into())
1000 );
1001 assert_eq!(output.passed, true);
1002
1003 let response = r#"
1004 Text around ignored
1005
1006 <analysis>
1007 Failed to compile:
1008 - Error 1
1009 - Error 2
1010 </analysis>
1011
1012 <passed>false</passed>
1013 "#
1014 .unindent();
1015
1016 let output = parse_assertion_result(&response).unwrap();
1017 assert_eq!(
1018 output.analysis,
1019 Some("Failed to compile:\n- Error 1\n- Error 2".into())
1020 );
1021 assert_eq!(output.passed, false);
1022 }
1023}