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