1use agent::{Message, MessageSegment, SerializedThread, ThreadStore};
2use anyhow::{Context as _, 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 LanguageModelToolResultContent, 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::MarkdownCodeBlock;
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 all_messages: 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(
198 &self,
199 model: Arc<dyn LanguageModel>,
200 app_state: Arc<AgentAppState>,
201 cx: &mut App,
202 ) -> Task<Result<RunOutput>> {
203 let project = Project::local(
204 app_state.client.clone(),
205 app_state.node_runtime.clone(),
206 app_state.user_store.clone(),
207 app_state.languages.clone(),
208 app_state.fs.clone(),
209 None,
210 cx,
211 );
212
213 let worktree = project.update(cx, |project, cx| {
214 project.create_worktree(self.worktree_path(), true, cx)
215 });
216
217 let tools = cx.new(|_| ToolWorkingSet::default());
218 let prompt_store = None;
219 let thread_store = ThreadStore::load(
220 project.clone(),
221 tools,
222 prompt_store,
223 app_state.prompt_builder.clone(),
224 cx,
225 );
226 let meta = self.thread.meta();
227 let this = self.clone();
228
229 cx.spawn(async move |cx| {
230 let worktree = worktree.await?;
231
232 // Wait for worktree scan to finish before choosing a file to open.
233 worktree
234 .update(cx, |worktree, _cx| {
235 worktree.as_local().unwrap().scan_complete()
236 })?
237 .await;
238
239 struct LanguageServerState {
240 _lsp_open_handle: OpenLspBufferHandle,
241 language_file_buffer: Entity<Buffer>,
242 }
243
244 let mut diagnostics_before = None;
245 let mut diagnostic_summary_before = DiagnosticSummary::default();
246
247 let lsp = if let Some(language_server) = &meta.language_server {
248 // Open a file that matches the language to cause LSP to start.
249 let language_file = worktree.read_with(cx, |worktree, _cx| {
250 worktree
251 .files(false, 0)
252 .find_map(|e| {
253 if e.path.clone().extension()
254 == Some(&language_server.file_extension)
255 {
256 Some(ProjectPath {
257 worktree_id: worktree.id(),
258 path: e.path.clone(),
259 })
260 } else {
261 None
262 }
263 })
264 .context("Failed to find a file for example language")
265 })??;
266
267 let open_language_file_buffer_task = project.update(cx, |project, cx| {
268 project.open_buffer(language_file.clone(), cx)
269 })?;
270
271 let language_file_buffer = open_language_file_buffer_task.await?;
272
273 let lsp_open_handle = project.update(cx, |project, cx| {
274 project.register_buffer_with_language_servers(&language_file_buffer, cx)
275 })?;
276
277 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
278
279 diagnostic_summary_before = project.read_with(cx, |project, cx| {
280 project.diagnostic_summary(false, cx)
281 })?;
282
283 diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
284 if diagnostics_before.is_some() && language_server.allow_preexisting_diagnostics {
285 anyhow::bail!("Example has pre-existing diagnostics. If you want to run this example regardless, set `allow_preexisting_diagnostics` to `true` in `base.toml`");
286 }
287
288 Some(LanguageServerState {
289 _lsp_open_handle: lsp_open_handle,
290 language_file_buffer,
291 })
292 } else {
293 None
294 };
295
296 anyhow::ensure!(std::env::var("ZED_EVAL_SETUP_ONLY").is_err(), "Setup only mode");
297
298 let last_diff_file_path = this.run_directory.join("last.diff");
299
300 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
301 // history using undo/redo.
302 std::fs::write(&last_diff_file_path, "")?;
303
304 let thread_store = thread_store.await?;
305
306
307 let thread =
308 thread_store.update(cx, |thread_store, cx| {
309 let thread = if let Some(json) = &meta.existing_thread_json {
310 let serialized = SerializedThread::from_json(json.as_bytes()).expect("Can't read serialized thread");
311 thread_store.create_thread_from_serialized(serialized, cx)
312 } else {
313 thread_store.create_thread(cx)
314 };
315 thread.update(cx, |thread, cx| {
316 thread.set_profile(meta.profile_id.clone(), cx);
317 });
318 thread
319 })?;
320
321
322 thread.update(cx, |thread, _cx| {
323 let mut request_count = 0;
324 let previous_diff = Rc::new(RefCell::new("".to_string()));
325 let example_output_dir = this.run_directory.clone();
326 let last_diff_file_path = last_diff_file_path.clone();
327 let messages_json_file_path = example_output_dir.join("last.messages.json");
328 let this = this.clone();
329 thread.set_request_callback(move |request, response_events| {
330 request_count += 1;
331 let messages_file_path = example_output_dir.join(format!("{request_count}.messages.md"));
332 let diff_file_path = example_output_dir.join(format!("{request_count}.diff"));
333 let last_messages_file_path = example_output_dir.join("last.messages.md");
334 let request_markdown = RequestMarkdown::new(request);
335 let response_events_markdown = response_events_to_markdown(response_events);
336 let dialog = ThreadDialog::new(request, response_events);
337 let dialog_json = serde_json::to_string_pretty(&dialog.to_combined_request()).unwrap_or_default();
338
339 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
340 fs::write(&messages_file_path, messages.clone()).expect("failed to write messages file");
341 fs::write(&last_messages_file_path, messages).expect("failed to write last messages file");
342 fs::write(&messages_json_file_path, dialog_json).expect("failed to write last.messages.json");
343
344 let diff_result = smol::block_on(this.repository_diff());
345 match diff_result {
346 Ok(diff) => {
347 if diff != previous_diff.borrow().clone() {
348 fs::write(&diff_file_path, &diff).expect("failed to write diff file");
349 fs::write(&last_diff_file_path, &diff).expect("failed to write last diff file");
350 *previous_diff.borrow_mut() = diff;
351 }
352 }
353 Err(err) => {
354 let error_message = format!("{err:?}");
355 fs::write(&diff_file_path, &error_message).expect("failed to write diff error to file");
356 fs::write(&last_diff_file_path, &error_message).expect("failed to write last diff file");
357 }
358 }
359
360 if request_count == 1 {
361 let tools_file_path = example_output_dir.join("tools.md");
362 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
363 }
364 });
365 })?;
366
367 let mut example_cx = ExampleContext::new(
368 meta.clone(),
369 this.log_prefix.clone(),
370 thread.clone(),
371 model.clone(),
372 cx.clone(),
373 );
374 let result = this.thread.conversation(&mut example_cx).await;
375
376 if let Err(err) = result
377 && !err.is::<FailedAssertion>() {
378 return Err(err);
379 }
380
381 println!("{}Stopped", this.log_prefix);
382
383 println!("{}Getting repository diff", this.log_prefix);
384 let repository_diff = this.repository_diff().await?;
385
386 std::fs::write(last_diff_file_path, &repository_diff)?;
387
388
389 let mut diagnostics_after = None;
390 let mut diagnostic_summary_after = Default::default();
391
392 if let Some(language_server_state) = lsp {
393 wait_for_lang_server(&project, &language_server_state.language_file_buffer, this.log_prefix.clone(), cx).await?;
394
395 println!("{}Getting diagnostics", this.log_prefix);
396 diagnostics_after = cx
397 .update(|cx| {
398 let project = project.clone();
399 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
400 })?
401 .await?;
402 println!("{}Got diagnostics", this.log_prefix);
403
404 diagnostic_summary_after = project.read_with(cx, |project, cx| {
405 project.diagnostic_summary(false, cx)
406 })?;
407
408 }
409
410 if let Some(diagnostics_before) = &diagnostics_before {
411 fs::write(this.run_directory.join("diagnostics_before.txt"), diagnostics_before)?;
412 }
413
414 if let Some(diagnostics_after) = &diagnostics_after {
415 fs::write(this.run_directory.join("diagnostics_after.txt"), diagnostics_after)?;
416 }
417
418 thread.update(cx, |thread, _cx| {
419 let response_count = thread
420 .messages()
421 .filter(|message| message.role == language_model::Role::Assistant)
422 .count();
423 RunOutput {
424 repository_diff,
425 diagnostic_summary_before,
426 diagnostic_summary_after,
427 diagnostics_before,
428 diagnostics_after,
429 response_count,
430 token_usage: thread.cumulative_token_usage(),
431 tool_metrics: example_cx.tool_metrics.lock().unwrap().clone(),
432 all_messages: messages_to_markdown(thread.messages()),
433 programmatic_assertions: example_cx.assertions,
434 }
435 })
436 })
437 }
438
439 async fn repository_diff(&self) -> Result<String> {
440 let worktree_path = self.worktree_path();
441 run_git(&worktree_path, &["add", "."]).await?;
442 let mut diff_args = vec!["diff", "--staged"];
443 if self.thread.meta().url == ZED_REPO_URL {
444 diff_args.push(":(exclude).rules");
445 }
446 run_git(&worktree_path, &diff_args).await
447 }
448
449 pub async fn judge(
450 &self,
451 model: Arc<dyn LanguageModel>,
452 run_output: &RunOutput,
453 cx: &AsyncApp,
454 ) -> JudgeOutput {
455 let mut output_file =
456 File::create(self.run_directory.join("judge.md")).expect("failed to create judge.md");
457
458 let diff_task = self.judge_diff(model.clone(), run_output, cx);
459 let thread_task = self.judge_thread(model.clone(), run_output, cx);
460
461 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
462
463 let (diff_response, diff_output) = diff_result;
464 let (thread_response, thread_output) = thread_result;
465
466 writeln!(
467 &mut output_file,
468 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
469 )
470 .log_err();
471
472 JudgeOutput {
473 thread: thread_output,
474 diff: diff_output,
475 }
476 }
477
478 async fn judge_diff(
479 &self,
480 model: Arc<dyn LanguageModel>,
481 run_output: &RunOutput,
482 cx: &AsyncApp,
483 ) -> (String, AssertionsReport) {
484 let diff_assertions = self.thread.diff_assertions();
485
486 if diff_assertions.is_empty() {
487 return (
488 "No diff assertions".to_string(),
489 AssertionsReport::default(),
490 );
491 }
492
493 println!("{}Running diff judge", self.log_prefix);
494
495 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
496 let judge_diff_prompt_name = "judge_diff_prompt";
497 let mut hbs = Handlebars::new();
498 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)
499 .unwrap();
500
501 let to_prompt = |assertion: String| {
502 hbs.render(
503 judge_diff_prompt_name,
504 &JudgeDiffInput {
505 repository_diff: run_output.repository_diff.clone(),
506 assertion,
507 },
508 )
509 .unwrap()
510 };
511
512 let (responses, report) = self
513 .judge_assertions(model, diff_assertions, to_prompt, cx)
514 .await;
515
516 println!(
517 "{}Judge - Diff score: {}%",
518 self.log_prefix,
519 report.passed_percentage()
520 );
521
522 (responses, report)
523 }
524
525 async fn judge_thread(
526 &self,
527 model: Arc<dyn LanguageModel>,
528 run_output: &RunOutput,
529 cx: &AsyncApp,
530 ) -> (String, AssertionsReport) {
531 let thread_assertions = self.thread.thread_assertions();
532
533 if thread_assertions.is_empty() {
534 return (
535 "No thread assertions".to_string(),
536 AssertionsReport::default(),
537 );
538 }
539
540 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
541 let judge_thread_prompt_name = "judge_thread_prompt";
542 let mut hbs = Handlebars::new();
543 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)
544 .unwrap();
545
546 let complete_messages = &run_output.all_messages;
547 let to_prompt = |assertion: String| {
548 hbs.render(
549 judge_thread_prompt_name,
550 &JudgeThreadInput {
551 messages: complete_messages.clone(),
552 assertion,
553 },
554 )
555 .unwrap()
556 };
557
558 let (responses, report) = self
559 .judge_assertions(model, thread_assertions, to_prompt, cx)
560 .await;
561
562 println!(
563 "{}Judge - Thread score: {}%",
564 self.log_prefix,
565 report.passed_percentage()
566 );
567
568 (responses, report)
569 }
570
571 async fn judge_assertions(
572 &self,
573 model: Arc<dyn LanguageModel>,
574 assertions: Vec<JudgeAssertion>,
575 to_prompt: impl Fn(String) -> String,
576 cx: &AsyncApp,
577 ) -> (String, AssertionsReport) {
578 let assertions = assertions.into_iter().map(|assertion| {
579 let request = LanguageModelRequest {
580 thread_id: None,
581 prompt_id: None,
582 mode: None,
583 intent: None,
584 messages: vec![LanguageModelRequestMessage {
585 role: Role::User,
586 content: vec![MessageContent::Text(to_prompt(assertion.description))],
587 cache: false,
588 }],
589 temperature: None,
590 tools: Vec::new(),
591 tool_choice: None,
592 stop: Vec::new(),
593 thinking_allowed: true,
594 };
595
596 let model = model.clone();
597 let log_prefix = self.log_prefix.clone();
598 async move {
599 let response = send_language_model_request(model, request, cx).await;
600
601 let (response, result) = match response {
602 Ok(response) => (
603 response.clone(),
604 parse_assertion_result(&response).map_err(|err| err.to_string()),
605 ),
606 Err(err) => (err.to_string(), Err(err.to_string())),
607 };
608
609 if result.is_ok() {
610 println!("{}✅ {}", log_prefix, assertion.id);
611 } else {
612 println!("{}❌ {}", log_prefix, assertion.id);
613 }
614
615 (
616 response,
617 RanAssertion {
618 id: assertion.id,
619 result,
620 },
621 )
622 }
623 });
624
625 let mut responses = String::new();
626 let mut report = AssertionsReport::default();
627
628 for (response, assertion) in future::join_all(assertions).await {
629 writeln!(&mut responses, "# {}", assertion.id).unwrap();
630 writeln!(&mut responses, "{}\n\n", response).unwrap();
631 report.ran.push(assertion);
632 }
633
634 (responses, report)
635 }
636}
637
638pub fn wait_for_lang_server(
639 project: &Entity<Project>,
640 buffer: &Entity<Buffer>,
641 log_prefix: String,
642 cx: &mut AsyncApp,
643) -> Task<Result<()>> {
644 if std::env::var("ZED_EVAL_SKIP_LS").is_ok() {
645 return Task::ready(Ok(()));
646 }
647
648 println!("{}⏵ Waiting for language server", log_prefix);
649
650 let (mut tx, mut rx) = mpsc::channel(1);
651
652 let lsp_store = project
653 .read_with(cx, |project, _| project.lsp_store())
654 .unwrap();
655
656 let has_lang_server = buffer
657 .update(cx, |buffer, cx| {
658 lsp_store.update(cx, |lsp_store, cx| {
659 lsp_store
660 .language_servers_for_local_buffer(buffer, cx)
661 .next()
662 .is_some()
663 })
664 })
665 .unwrap_or(false);
666
667 if has_lang_server {
668 project
669 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
670 .unwrap()
671 .detach();
672 }
673
674 let subscriptions =
675 [
676 cx.subscribe(&lsp_store, {
677 let log_prefix = log_prefix.clone();
678 move |_, event, _| {
679 if let project::LspStoreEvent::LanguageServerUpdate {
680 message:
681 client::proto::update_language_server::Variant::WorkProgress(
682 LspWorkProgress {
683 message: Some(message),
684 ..
685 },
686 ),
687 ..
688 } = event
689 {
690 println!("{}⟲ {message}", log_prefix)
691 }
692 }
693 }),
694 cx.subscribe(project, {
695 let buffer = buffer.clone();
696 move |project, event, cx| match event {
697 project::Event::LanguageServerAdded(_, _, _) => {
698 let buffer = buffer.clone();
699 project
700 .update(cx, |project, cx| project.save_buffer(buffer, cx))
701 .detach();
702 }
703 project::Event::DiskBasedDiagnosticsFinished { .. } => {
704 tx.try_send(()).ok();
705 }
706 _ => {}
707 }
708 }),
709 ];
710
711 cx.spawn(async move |cx| {
712 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
713 let result = futures::select! {
714 _ = rx.next() => {
715 println!("{}⚑ Language server idle", log_prefix);
716 anyhow::Ok(())
717 },
718 _ = timeout.fuse() => {
719 anyhow::bail!("LSP wait timed out after 5 minutes");
720 }
721 };
722 drop(subscriptions);
723 result
724 })
725}
726
727pub async fn query_lsp_diagnostics(
728 project: Entity<Project>,
729 cx: &mut AsyncApp,
730) -> Result<Option<String>> {
731 let paths_with_diagnostics = project.update(cx, |project, cx| {
732 project
733 .diagnostic_summaries(true, cx)
734 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
735 .map(|(project_path, _, _)| project_path)
736 .collect::<Vec<_>>()
737 })?;
738
739 if paths_with_diagnostics.is_empty() {
740 return Ok(None);
741 }
742
743 let mut output = String::new();
744 for project_path in paths_with_diagnostics {
745 let buffer = project
746 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
747 .await?;
748 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
749
750 for (_, group) in snapshot.diagnostic_groups(None) {
751 let entry = &group.entries[group.primary_ix];
752 let range = entry.range.to_point(&snapshot);
753 let severity = match entry.diagnostic.severity {
754 DiagnosticSeverity::ERROR => "error",
755 DiagnosticSeverity::WARNING => "warning",
756 _ => continue,
757 };
758
759 writeln!(
760 output,
761 "{} at line {}: {}",
762 severity,
763 range.start.row + 1,
764 entry.diagnostic.message
765 )?;
766 }
767 }
768 anyhow::Ok(Some(output))
769}
770
771fn parse_assertion_result(response: &str) -> Result<RanAssertionResult> {
772 let analysis = get_tag("analysis", response)?;
773 let passed = match get_tag("passed", response)?.to_lowercase().as_str() {
774 "true" => true,
775 "false" => false,
776 value @ _ => bail!("invalid judge `passed` tag: {value}"),
777 };
778 Ok(RanAssertionResult {
779 analysis: Some(analysis),
780 passed,
781 })
782}
783
784fn get_tag(name: &'static str, response: &str) -> Result<String> {
785 let start_tag = format!("<{}>", name);
786 let end_tag = format!("</{}>", name);
787
788 let start_ix = response
789 .find(&start_tag)
790 .context(format!("{} start tag not found", name))?;
791 let content_start_ix = start_ix + start_tag.len();
792
793 let end_ix = content_start_ix
794 + response[content_start_ix..]
795 .find(&end_tag)
796 .context(format!("{} end tag not found", name))?;
797
798 let content = response[content_start_ix..end_ix].trim().unindent();
799
800 anyhow::Ok(content)
801}
802
803pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
804 let repo_name = repo_url
805 .trim_start_matches("https://")
806 .replace(|c: char| !c.is_alphanumeric(), "-");
807 Path::new(repos_dir).join(repo_name)
808}
809
810pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
811 let output = new_smol_command("git")
812 .current_dir(repo_path)
813 .args(args)
814 .output()
815 .await?;
816
817 anyhow::ensure!(
818 output.status.success(),
819 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
820 args.join(" "),
821 repo_path.display(),
822 output.status,
823 String::from_utf8_lossy(&output.stderr),
824 String::from_utf8_lossy(&output.stdout),
825 );
826 Ok(String::from_utf8(output.stdout)?.trim().to_string())
827}
828
829fn messages_to_markdown<'a>(message_iter: impl IntoIterator<Item = &'a Message>) -> String {
830 let mut messages = String::new();
831 let mut assistant_message_number: u32 = 1;
832
833 for message in message_iter {
834 push_role(&message.role, &mut messages, &mut assistant_message_number);
835
836 for segment in &message.segments {
837 match segment {
838 MessageSegment::Text(text) => {
839 messages.push_str(text);
840 messages.push_str("\n\n");
841 }
842 MessageSegment::Thinking { text, signature } => {
843 messages.push_str("**Thinking**:\n\n");
844 if let Some(sig) = signature {
845 messages.push_str(&format!("Signature: {}\n\n", sig));
846 }
847 messages.push_str(text);
848 messages.push_str("\n");
849 }
850 MessageSegment::RedactedThinking(items) => {
851 messages.push_str(&format!(
852 "**Redacted Thinking**: {} item(s)\n\n",
853 items.len()
854 ));
855 }
856 }
857 }
858 }
859
860 messages
861}
862
863fn push_role(role: &Role, buf: &mut String, assistant_message_number: &mut u32) {
864 match role {
865 Role::System => buf.push_str("# ⚙️ SYSTEM\n\n"),
866 Role::User => buf.push_str("# 👤 USER\n\n"),
867 Role::Assistant => {
868 buf.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
869 *assistant_message_number = *assistant_message_number + 1;
870 }
871 }
872}
873
874pub async fn send_language_model_request(
875 model: Arc<dyn LanguageModel>,
876 request: LanguageModelRequest,
877 cx: &AsyncApp,
878) -> anyhow::Result<String> {
879 match model.stream_completion_text(request, cx).await {
880 Ok(mut stream) => {
881 let mut full_response = String::new();
882 while let Some(chunk_result) = stream.stream.next().await {
883 match chunk_result {
884 Ok(chunk_str) => {
885 full_response.push_str(&chunk_str);
886 }
887 Err(err) => {
888 anyhow::bail!("Error receiving response from language model: {err}");
889 }
890 }
891 }
892 Ok(full_response)
893 }
894 Err(err) => Err(anyhow!(
895 "Failed to get response from language model. Error was: {err}"
896 )),
897 }
898}
899
900pub struct RequestMarkdown {
901 pub tools: String,
902 pub messages: String,
903}
904
905impl RequestMarkdown {
906 pub fn new(request: &LanguageModelRequest) -> Self {
907 let mut tools = String::new();
908 let mut messages = String::new();
909 let mut assistant_message_number: u32 = 1;
910
911 // Print the tools
912 if !request.tools.is_empty() {
913 for tool in &request.tools {
914 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
915 write!(&mut tools, "{}\n\n", tool.description).unwrap();
916 writeln!(
917 &mut tools,
918 "{}",
919 MarkdownCodeBlock {
920 tag: "json",
921 text: &format!("{:#}", tool.input_schema)
922 }
923 )
924 .unwrap();
925 }
926 }
927
928 // Print the messages
929 for message in &request.messages {
930 push_role(&message.role, &mut messages, &mut assistant_message_number);
931
932 for content in &message.content {
933 match content {
934 MessageContent::Text(text) => {
935 messages.push_str(text);
936 messages.push_str("\n\n");
937 }
938 MessageContent::Image(_) => {
939 messages.push_str("[IMAGE DATA]\n\n");
940 }
941 MessageContent::Thinking { text, signature } => {
942 messages.push_str("**Thinking**:\n\n");
943 if let Some(sig) = signature {
944 messages.push_str(&format!("Signature: {}\n\n", sig));
945 }
946 messages.push_str(text);
947 messages.push_str("\n");
948 }
949 MessageContent::RedactedThinking(items) => {
950 messages.push_str(&format!(
951 "**Redacted Thinking**: {} item(s)\n\n",
952 items.len()
953 ));
954 }
955 MessageContent::ToolUse(tool_use) => {
956 messages.push_str(&format!(
957 "**Tool Use**: {} (ID: {})\n",
958 tool_use.name, tool_use.id
959 ));
960 messages.push_str(&format!(
961 "{}\n",
962 MarkdownCodeBlock {
963 tag: "json",
964 text: &format!("{:#}", tool_use.input)
965 }
966 ));
967 }
968 MessageContent::ToolResult(tool_result) => {
969 messages.push_str(&format!(
970 "**Tool Result**: {} (ID: {})\n\n",
971 tool_result.tool_name, tool_result.tool_use_id
972 ));
973 if tool_result.is_error {
974 messages.push_str("**ERROR:**\n");
975 }
976
977 match &tool_result.content {
978 LanguageModelToolResultContent::Text(text) => {
979 writeln!(messages, "{text}\n").ok();
980 }
981 LanguageModelToolResultContent::Image(image) => {
982 writeln!(messages, "\n", image.source).ok();
983 }
984 }
985
986 if let Some(output) = tool_result.output.as_ref() {
987 writeln!(
988 messages,
989 "**Debug Output**:\n\n```json\n{}\n```\n",
990 serde_json::to_string_pretty(output).unwrap()
991 )
992 .unwrap();
993 }
994 }
995 }
996 }
997 }
998
999 Self { tools, messages }
1000 }
1001}
1002
1003pub fn response_events_to_markdown(
1004 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1005) -> String {
1006 let mut response = String::new();
1007 // Print the response events if any
1008 response.push_str("# Response\n\n");
1009 let mut text_buffer = String::new();
1010 let mut thinking_buffer = String::new();
1011
1012 let flush_buffers =
1013 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
1014 if !text_buffer.is_empty() {
1015 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
1016 text_buffer.clear();
1017 }
1018 if !thinking_buffer.is_empty() {
1019 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
1020 thinking_buffer.clear();
1021 }
1022 };
1023
1024 for event in response_events {
1025 match event {
1026 Ok(LanguageModelCompletionEvent::Text(text)) => {
1027 text_buffer.push_str(text);
1028 }
1029 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1030 thinking_buffer.push_str(text);
1031 }
1032 Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => {}
1033 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1034 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1035 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1036 }
1037 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1038 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1039 response.push_str(&format!(
1040 "**Tool Use**: {} (ID: {})\n",
1041 tool_use.name, tool_use.id
1042 ));
1043 response.push_str(&format!(
1044 "{}\n",
1045 MarkdownCodeBlock {
1046 tag: "json",
1047 text: &format!("{:#}", tool_use.input)
1048 }
1049 ));
1050 }
1051 Ok(
1052 LanguageModelCompletionEvent::UsageUpdate(_)
1053 | LanguageModelCompletionEvent::StartMessage { .. }
1054 | LanguageModelCompletionEvent::StatusUpdate { .. },
1055 ) => {}
1056 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1057 json_parse_error, ..
1058 }) => {
1059 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1060 response.push_str(&format!(
1061 "**Error**: parse error in tool use JSON: {}\n\n",
1062 json_parse_error
1063 ));
1064 }
1065 Err(error) => {
1066 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1067 response.push_str(&format!("**Error**: {}\n\n", error));
1068 }
1069 }
1070 }
1071
1072 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1073
1074 response
1075}
1076
1077#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
1078pub struct ThreadDialog {
1079 pub request: LanguageModelRequest,
1080 pub response_events: Vec<std::result::Result<LanguageModelCompletionEvent, String>>,
1081}
1082
1083impl ThreadDialog {
1084 pub fn new(
1085 request: &LanguageModelRequest,
1086 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1087 ) -> Self {
1088 Self {
1089 request: request.clone(),
1090 response_events: response_events.to_vec(),
1091 }
1092 }
1093
1094 /// Represents all request and response messages in a unified format.
1095 ///
1096 /// Specifically, it appends the assistant's response (derived from response events)
1097 /// as a new message to existing messages in the request.
1098 pub fn to_combined_request(&self) -> LanguageModelRequest {
1099 let mut request = self.request.clone();
1100 if let Some(assistant_message) = self.response_events_to_message() {
1101 request.messages.push(assistant_message);
1102 }
1103 request
1104 }
1105 fn response_events_to_message(&self) -> Option<LanguageModelRequestMessage> {
1106 let response_events = &self.response_events;
1107 let mut content: Vec<MessageContent> = Vec::new();
1108 let mut current_text = String::new();
1109
1110 let flush_text = |text: &mut String, content: &mut Vec<MessageContent>| {
1111 if !text.is_empty() {
1112 content.push(MessageContent::Text(std::mem::take(text)));
1113 }
1114 };
1115
1116 for event in response_events {
1117 match event {
1118 Ok(LanguageModelCompletionEvent::Text(text)) => {
1119 current_text.push_str(text);
1120 }
1121
1122 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1123 flush_text(&mut current_text, &mut content);
1124 if tool_use.is_input_complete {
1125 content.push(MessageContent::ToolUse(tool_use.clone()));
1126 }
1127 }
1128 Ok(LanguageModelCompletionEvent::Thinking { text, signature }) => {
1129 flush_text(&mut current_text, &mut content);
1130 content.push(MessageContent::Thinking {
1131 text: text.clone(),
1132 signature: signature.clone(),
1133 });
1134 }
1135
1136 // Skip these
1137 Ok(LanguageModelCompletionEvent::UsageUpdate(_))
1138 | Ok(LanguageModelCompletionEvent::RedactedThinking { .. })
1139 | Ok(LanguageModelCompletionEvent::StatusUpdate { .. })
1140 | Ok(LanguageModelCompletionEvent::StartMessage { .. })
1141 | Ok(LanguageModelCompletionEvent::Stop(_)) => {}
1142
1143 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1144 json_parse_error,
1145 ..
1146 }) => {
1147 flush_text(&mut current_text, &mut content);
1148 content.push(MessageContent::Text(format!(
1149 "ERROR: parse error in tool use JSON: {}",
1150 json_parse_error
1151 )));
1152 }
1153
1154 Err(error) => {
1155 flush_text(&mut current_text, &mut content);
1156 content.push(MessageContent::Text(format!("ERROR: {}", error)));
1157 }
1158 }
1159 }
1160
1161 flush_text(&mut current_text, &mut content);
1162
1163 if !content.is_empty() {
1164 Some(LanguageModelRequestMessage {
1165 role: Role::Assistant,
1166 content,
1167 cache: false,
1168 })
1169 } else {
1170 None
1171 }
1172 }
1173}
1174
1175#[cfg(test)]
1176mod test {
1177 use super::*;
1178
1179 #[test]
1180 fn test_parse_judge_output() {
1181 let response = r#"
1182 <analysis>The model did a good job but there were still compilations errors.</analysis>
1183 <passed>true</passed>
1184 "#
1185 .unindent();
1186
1187 let output = parse_assertion_result(&response).unwrap();
1188 assert_eq!(
1189 output.analysis,
1190 Some("The model did a good job but there were still compilations errors.".into())
1191 );
1192 assert!(output.passed);
1193
1194 let response = r#"
1195 Text around ignored
1196
1197 <analysis>
1198 Failed to compile:
1199 - Error 1
1200 - Error 2
1201 </analysis>
1202
1203 <passed>false</passed>
1204 "#
1205 .unindent();
1206
1207 let output = parse_assertion_result(&response).unwrap();
1208 assert_eq!(
1209 output.analysis,
1210 Some("Failed to compile:\n- Error 1\n- Error 2".into())
1211 );
1212 assert!(!output.passed);
1213 }
1214}