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.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 prompt_store = None;
222 let thread_store = ThreadStore::load(
223 project.clone(),
224 tools,
225 prompt_store,
226 app_state.prompt_builder.clone(),
227 cx,
228 );
229 let meta = self.thread.meta();
230 let this = self.clone();
231
232 cx.spawn(async move |cx| {
233 let worktree = worktree.await?;
234
235 // Wait for worktree scan to finish before choosing a file to open.
236 worktree
237 .update(cx, |worktree, _cx| {
238 worktree.as_local().unwrap().scan_complete()
239 })?
240 .await;
241
242 struct LanguageServerState {
243 _lsp_open_handle: OpenLspBufferHandle,
244 language_file_buffer: Entity<Buffer>,
245 }
246
247 let mut diagnostics_before = None;
248 let mut diagnostic_summary_before = DiagnosticSummary::default();
249
250 let lsp = if let Some(language_server) = &meta.language_server {
251 // Open a file that matches the language to cause LSP to start.
252 let language_file = worktree.read_with(cx, |worktree, _cx| {
253 worktree
254 .files(false, 0)
255 .find_map(|e| {
256 if e.path.clone().extension().and_then(|ext| ext.to_str())
257 == Some(&language_server.file_extension)
258 {
259 Some(ProjectPath {
260 worktree_id: worktree.id(),
261 path: e.path.clone(),
262 })
263 } else {
264 None
265 }
266 })
267 .context("Failed to find a file for example language")
268 })??;
269
270 let open_language_file_buffer_task = project.update(cx, |project, cx| {
271 project.open_buffer(language_file.clone(), cx)
272 })?;
273
274 let language_file_buffer = open_language_file_buffer_task.await?;
275
276 let lsp_open_handle = project.update(cx, |project, cx| {
277 project.register_buffer_with_language_servers(&language_file_buffer, cx)
278 })?;
279
280 wait_for_lang_server(&project, &language_file_buffer, this.log_prefix.clone(), cx).await?;
281
282 diagnostic_summary_before = project.read_with(cx, |project, cx| {
283 project.diagnostic_summary(false, cx)
284 })?;
285
286 diagnostics_before = query_lsp_diagnostics(project.clone(), cx).await?;
287 if diagnostics_before.is_some() && language_server.allow_preexisting_diagnostics {
288 anyhow::bail!("Example has pre-existing diagnostics. If you want to run this example regardless, set `allow_preexisting_diagnostics` to `true` in `base.toml`");
289 }
290
291 Some(LanguageServerState {
292 _lsp_open_handle: lsp_open_handle,
293 language_file_buffer,
294 })
295 } else {
296 None
297 };
298
299 anyhow::ensure!(std::env::var("ZED_EVAL_SETUP_ONLY").is_err(), "Setup only mode");
300
301 let last_diff_file_path = this.run_directory.join("last.diff");
302
303 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
304 // history using undo/redo.
305 std::fs::write(&last_diff_file_path, "")?;
306
307 let thread_store = thread_store.await?;
308
309
310 let thread =
311 thread_store.update(cx, |thread_store, cx| {
312 let thread = if let Some(json) = &meta.existing_thread_json {
313 let serialized = SerializedThread::from_json(json.as_bytes()).expect("Can't read serialized thread");
314 thread_store.create_thread_from_serialized(serialized, cx)
315 } else {
316 thread_store.create_thread(cx)
317 };
318 thread.update(cx, |thread, cx| {
319 thread.set_profile(meta.profile_id.clone(), cx);
320 });
321 thread
322 })?;
323
324
325 thread.update(cx, |thread, _cx| {
326 let mut request_count = 0;
327 let previous_diff = Rc::new(RefCell::new("".to_string()));
328 let example_output_dir = this.run_directory.clone();
329 let last_diff_file_path = last_diff_file_path.clone();
330 let messages_json_file_path = example_output_dir.join("last.messages.json");
331 let this = this.clone();
332 thread.set_request_callback(move |request, response_events| {
333 request_count += 1;
334 let messages_file_path = example_output_dir.join(format!("{request_count}.messages.md"));
335 let diff_file_path = example_output_dir.join(format!("{request_count}.diff"));
336 let last_messages_file_path = example_output_dir.join("last.messages.md");
337 let request_markdown = RequestMarkdown::new(request);
338 let response_events_markdown = response_events_to_markdown(response_events);
339 let dialog = ThreadDialog::new(request, response_events);
340 let dialog_json = serde_json::to_string_pretty(&dialog.to_combined_request()).unwrap_or_default();
341
342 let messages = format!("{}\n\n{}", request_markdown.messages, response_events_markdown);
343 fs::write(&messages_file_path, messages.clone()).expect("failed to write messages file");
344 fs::write(&last_messages_file_path, messages).expect("failed to write last messages file");
345 fs::write(&messages_json_file_path, dialog_json).expect("failed to write last.messages.json");
346
347 let diff_result = smol::block_on(this.repository_diff());
348 match diff_result {
349 Ok(diff) => {
350 if diff != previous_diff.borrow().clone() {
351 fs::write(&diff_file_path, &diff).expect("failed to write diff file");
352 fs::write(&last_diff_file_path, &diff).expect("failed to write last diff file");
353 *previous_diff.borrow_mut() = diff;
354 }
355 }
356 Err(err) => {
357 let error_message = format!("{err:?}");
358 fs::write(&diff_file_path, &error_message).expect("failed to write diff error to file");
359 fs::write(&last_diff_file_path, &error_message).expect("failed to write last diff file");
360 }
361 }
362
363 if request_count == 1 {
364 let tools_file_path = example_output_dir.join("tools.md");
365 fs::write(tools_file_path, request_markdown.tools).expect("failed to write tools file");
366 }
367 });
368 })?;
369
370 let mut example_cx = ExampleContext::new(
371 meta.clone(),
372 this.log_prefix.clone(),
373 thread.clone(),
374 model.clone(),
375 cx.clone(),
376 );
377 let result = this.thread.conversation(&mut example_cx).await;
378
379 if let Err(err) = result {
380 if !err.is::<FailedAssertion>() {
381 return Err(err);
382 }
383 }
384
385 println!("{}Stopped", this.log_prefix);
386
387 println!("{}Getting repository diff", this.log_prefix);
388 let repository_diff = this.repository_diff().await?;
389
390 std::fs::write(last_diff_file_path, &repository_diff)?;
391
392
393 let mut diagnostics_after = None;
394 let mut diagnostic_summary_after = Default::default();
395
396 if let Some(language_server_state) = lsp {
397 wait_for_lang_server(&project, &language_server_state.language_file_buffer, this.log_prefix.clone(), cx).await?;
398
399 println!("{}Getting diagnostics", this.log_prefix);
400 diagnostics_after = cx
401 .update(|cx| {
402 let project = project.clone();
403 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
404 })?
405 .await?;
406 println!("{}Got diagnostics", this.log_prefix);
407
408 diagnostic_summary_after = project.read_with(cx, |project, cx| {
409 project.diagnostic_summary(false, cx)
410 })?;
411
412 }
413
414 if let Some(diagnostics_before) = &diagnostics_before {
415 fs::write(this.run_directory.join("diagnostics_before.txt"), diagnostics_before)?;
416 }
417
418 if let Some(diagnostics_after) = &diagnostics_after {
419 fs::write(this.run_directory.join("diagnostics_after.txt"), diagnostics_after)?;
420 }
421
422 thread.update(cx, |thread, _cx| {
423 let response_count = thread
424 .messages()
425 .filter(|message| message.role == language_model::Role::Assistant)
426 .count();
427 RunOutput {
428 repository_diff,
429 diagnostic_summary_before,
430 diagnostic_summary_after,
431 diagnostics_before,
432 diagnostics_after,
433 response_count,
434 token_usage: thread.cumulative_token_usage(),
435 tool_metrics: example_cx.tool_metrics.lock().unwrap().clone(),
436 all_messages: messages_to_markdown(thread.messages()),
437 programmatic_assertions: example_cx.assertions,
438 }
439 })
440 })
441 }
442
443 async fn repository_diff(&self) -> Result<String> {
444 let worktree_path = self.worktree_path();
445 run_git(&worktree_path, &["add", "."]).await?;
446 let mut diff_args = vec!["diff", "--staged"];
447 if self.thread.meta().url == ZED_REPO_URL {
448 diff_args.push(":(exclude).rules");
449 }
450 run_git(&worktree_path, &diff_args).await
451 }
452
453 pub async fn judge(
454 &self,
455 model: Arc<dyn LanguageModel>,
456 run_output: &RunOutput,
457 cx: &AsyncApp,
458 ) -> JudgeOutput {
459 let mut output_file =
460 File::create(self.run_directory.join("judge.md")).expect("failed to create judge.md");
461
462 let diff_task = self.judge_diff(model.clone(), &run_output, cx);
463 let thread_task = self.judge_thread(model.clone(), &run_output, cx);
464
465 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
466
467 let (diff_response, diff_output) = diff_result;
468 let (thread_response, thread_output) = thread_result;
469
470 writeln!(
471 &mut output_file,
472 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
473 )
474 .log_err();
475
476 JudgeOutput {
477 thread: thread_output,
478 diff: diff_output,
479 }
480 }
481
482 async fn judge_diff(
483 &self,
484 model: Arc<dyn LanguageModel>,
485 run_output: &RunOutput,
486 cx: &AsyncApp,
487 ) -> (String, AssertionsReport) {
488 let diff_assertions = self.thread.diff_assertions();
489
490 if diff_assertions.is_empty() {
491 return (
492 "No diff assertions".to_string(),
493 AssertionsReport::default(),
494 );
495 }
496
497 println!("{}Running diff judge", self.log_prefix);
498
499 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
500 let judge_diff_prompt_name = "judge_diff_prompt";
501 let mut hbs = Handlebars::new();
502 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)
503 .unwrap();
504
505 let to_prompt = |assertion: String| {
506 hbs.render(
507 judge_diff_prompt_name,
508 &JudgeDiffInput {
509 repository_diff: run_output.repository_diff.clone(),
510 assertion,
511 },
512 )
513 .unwrap()
514 };
515
516 let (responses, report) = self
517 .judge_assertions(model, diff_assertions, to_prompt, cx)
518 .await;
519
520 println!(
521 "{}Judge - Diff score: {}%",
522 self.log_prefix,
523 report.passed_percentage()
524 );
525
526 (responses, report)
527 }
528
529 async fn judge_thread(
530 &self,
531 model: Arc<dyn LanguageModel>,
532 run_output: &RunOutput,
533 cx: &AsyncApp,
534 ) -> (String, AssertionsReport) {
535 let thread_assertions = self.thread.thread_assertions();
536
537 if thread_assertions.is_empty() {
538 return (
539 "No thread assertions".to_string(),
540 AssertionsReport::default(),
541 );
542 }
543
544 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
545 let judge_thread_prompt_name = "judge_thread_prompt";
546 let mut hbs = Handlebars::new();
547 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)
548 .unwrap();
549
550 let complete_messages = &run_output.all_messages;
551 let to_prompt = |assertion: String| {
552 hbs.render(
553 judge_thread_prompt_name,
554 &JudgeThreadInput {
555 messages: complete_messages.clone(),
556 assertion,
557 },
558 )
559 .unwrap()
560 };
561
562 let (responses, report) = self
563 .judge_assertions(model, thread_assertions, to_prompt, cx)
564 .await;
565
566 println!(
567 "{}Judge - Thread score: {}%",
568 self.log_prefix,
569 report.passed_percentage()
570 );
571
572 (responses, report)
573 }
574
575 async fn judge_assertions(
576 &self,
577 model: Arc<dyn LanguageModel>,
578 assertions: Vec<JudgeAssertion>,
579 to_prompt: impl Fn(String) -> String,
580 cx: &AsyncApp,
581 ) -> (String, AssertionsReport) {
582 let assertions = assertions.into_iter().map(|assertion| {
583 let request = LanguageModelRequest {
584 thread_id: None,
585 prompt_id: None,
586 mode: None,
587 intent: None,
588 messages: vec![LanguageModelRequestMessage {
589 role: Role::User,
590 content: vec![MessageContent::Text(to_prompt(assertion.description))],
591 cache: false,
592 }],
593 temperature: None,
594 tools: Vec::new(),
595 tool_choice: None,
596 stop: Vec::new(),
597 thinking_allowed: true,
598 };
599
600 let model = model.clone();
601 let log_prefix = self.log_prefix.clone();
602 async move {
603 let response = send_language_model_request(model, request, cx).await;
604
605 let (response, result) = match response {
606 Ok(response) => (
607 response.clone(),
608 parse_assertion_result(&response).map_err(|err| err.to_string()),
609 ),
610 Err(err) => (err.to_string(), Err(err.to_string())),
611 };
612
613 if result.is_ok() {
614 println!("{}✅ {}", log_prefix, assertion.id);
615 } else {
616 println!("{}❌ {}", log_prefix, assertion.id);
617 }
618
619 (
620 response,
621 RanAssertion {
622 id: assertion.id,
623 result,
624 },
625 )
626 }
627 });
628
629 let mut responses = String::new();
630 let mut report = AssertionsReport::default();
631
632 for (response, assertion) in future::join_all(assertions).await {
633 writeln!(&mut responses, "# {}", assertion.id).unwrap();
634 writeln!(&mut responses, "{}\n\n", response).unwrap();
635 report.ran.push(assertion);
636 }
637
638 (responses, report)
639 }
640}
641
642pub fn wait_for_lang_server(
643 project: &Entity<Project>,
644 buffer: &Entity<Buffer>,
645 log_prefix: String,
646 cx: &mut AsyncApp,
647) -> Task<Result<()>> {
648 if std::env::var("ZED_EVAL_SKIP_LS").is_ok() {
649 return Task::ready(Ok(()));
650 }
651
652 println!("{}⏵ Waiting for language server", log_prefix);
653
654 let (mut tx, mut rx) = mpsc::channel(1);
655
656 let lsp_store = project
657 .read_with(cx, |project, _| project.lsp_store())
658 .unwrap();
659
660 let has_lang_server = buffer
661 .update(cx, |buffer, cx| {
662 lsp_store.update(cx, |lsp_store, cx| {
663 lsp_store
664 .language_servers_for_local_buffer(&buffer, cx)
665 .next()
666 .is_some()
667 })
668 })
669 .unwrap_or(false);
670
671 if has_lang_server {
672 project
673 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
674 .unwrap()
675 .detach();
676 }
677
678 let subscriptions =
679 [
680 cx.subscribe(&lsp_store, {
681 let log_prefix = log_prefix.clone();
682 move |_, event, _| match event {
683 project::LspStoreEvent::LanguageServerUpdate {
684 message:
685 client::proto::update_language_server::Variant::WorkProgress(
686 LspWorkProgress {
687 message: Some(message),
688 ..
689 },
690 ),
691 ..
692 } => println!("{}⟲ {message}", log_prefix),
693 _ => {}
694 }
695 }),
696 cx.subscribe(&project, {
697 let buffer = buffer.clone();
698 move |project, event, cx| match event {
699 project::Event::LanguageServerAdded(_, _, _) => {
700 let buffer = buffer.clone();
701 project
702 .update(cx, |project, cx| project.save_buffer(buffer, cx))
703 .detach();
704 }
705 project::Event::DiskBasedDiagnosticsFinished { .. } => {
706 tx.try_send(()).ok();
707 }
708 _ => {}
709 }
710 }),
711 ];
712
713 cx.spawn(async move |cx| {
714 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
715 let result = futures::select! {
716 _ = rx.next() => {
717 println!("{}⚑ Language server idle", log_prefix);
718 anyhow::Ok(())
719 },
720 _ = timeout.fuse() => {
721 anyhow::bail!("LSP wait timed out after 5 minutes");
722 }
723 };
724 drop(subscriptions);
725 result
726 })
727}
728
729pub async fn query_lsp_diagnostics(
730 project: Entity<Project>,
731 cx: &mut AsyncApp,
732) -> Result<Option<String>> {
733 let paths_with_diagnostics = project.update(cx, |project, cx| {
734 project
735 .diagnostic_summaries(true, cx)
736 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
737 .map(|(project_path, _, _)| project_path)
738 .collect::<Vec<_>>()
739 })?;
740
741 if paths_with_diagnostics.is_empty() {
742 return Ok(None);
743 }
744
745 let mut output = String::new();
746 for project_path in paths_with_diagnostics {
747 let buffer = project
748 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
749 .await?;
750 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
751
752 for (_, group) in snapshot.diagnostic_groups(None) {
753 let entry = &group.entries[group.primary_ix];
754 let range = entry.range.to_point(&snapshot);
755 let severity = match entry.diagnostic.severity {
756 DiagnosticSeverity::ERROR => "error",
757 DiagnosticSeverity::WARNING => "warning",
758 _ => continue,
759 };
760
761 writeln!(
762 output,
763 "{} at line {}: {}",
764 severity,
765 range.start.row + 1,
766 entry.diagnostic.message
767 )?;
768 }
769 }
770 anyhow::Ok(Some(output))
771}
772
773fn parse_assertion_result(response: &str) -> Result<RanAssertionResult> {
774 let analysis = get_tag("analysis", response)?.to_string();
775 let passed = match get_tag("passed", response)?.to_lowercase().as_str() {
776 "true" => true,
777 "false" => false,
778 value @ _ => bail!("invalid judge `passed` tag: {value}"),
779 };
780 Ok(RanAssertionResult {
781 analysis: Some(analysis),
782 passed,
783 })
784}
785
786fn get_tag(name: &'static str, response: &str) -> Result<String> {
787 let start_tag = format!("<{}>", name);
788 let end_tag = format!("</{}>", name);
789
790 let start_ix = response
791 .find(&start_tag)
792 .context(format!("{} start tag not found", name))?;
793 let content_start_ix = start_ix + start_tag.len();
794
795 let end_ix = content_start_ix
796 + response[content_start_ix..]
797 .find(&end_tag)
798 .context(format!("{} end tag not found", name))?;
799
800 let content = response[content_start_ix..end_ix].trim().unindent();
801
802 anyhow::Ok(content)
803}
804
805pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
806 let repo_name = repo_url
807 .trim_start_matches("https://")
808 .replace(|c: char| !c.is_alphanumeric(), "-");
809 Path::new(repos_dir).join(repo_name)
810}
811
812pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
813 let output = new_smol_command("git")
814 .current_dir(repo_path)
815 .args(args)
816 .output()
817 .await?;
818
819 anyhow::ensure!(
820 output.status.success(),
821 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
822 args.join(" "),
823 repo_path.display(),
824 output.status,
825 String::from_utf8_lossy(&output.stderr),
826 String::from_utf8_lossy(&output.stdout),
827 );
828 Ok(String::from_utf8(output.stdout)?.trim().to_string())
829}
830
831fn messages_to_markdown<'a>(message_iter: impl IntoIterator<Item = &'a Message>) -> String {
832 let mut messages = String::new();
833 let mut assistant_message_number: u32 = 1;
834
835 for message in message_iter {
836 push_role(&message.role, &mut messages, &mut assistant_message_number);
837
838 for segment in &message.segments {
839 match segment {
840 MessageSegment::Text(text) => {
841 messages.push_str(&text);
842 messages.push_str("\n\n");
843 }
844 MessageSegment::Thinking { text, signature } => {
845 messages.push_str("**Thinking**:\n\n");
846 if let Some(sig) = signature {
847 messages.push_str(&format!("Signature: {}\n\n", sig));
848 }
849 messages.push_str(&text);
850 messages.push_str("\n");
851 }
852 MessageSegment::RedactedThinking(items) => {
853 messages.push_str(&format!(
854 "**Redacted Thinking**: {} item(s)\n\n",
855 items.len()
856 ));
857 }
858 }
859 }
860 }
861
862 messages
863}
864
865fn push_role(role: &Role, buf: &mut String, assistant_message_number: &mut u32) {
866 match role {
867 Role::System => buf.push_str("# ⚙️ SYSTEM\n\n"),
868 Role::User => buf.push_str("# 👤 USER\n\n"),
869 Role::Assistant => {
870 buf.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
871 *assistant_message_number = *assistant_message_number + 1;
872 }
873 }
874}
875
876pub async fn send_language_model_request(
877 model: Arc<dyn LanguageModel>,
878 request: LanguageModelRequest,
879 cx: &AsyncApp,
880) -> anyhow::Result<String> {
881 match model.stream_completion_text(request, &cx).await {
882 Ok(mut stream) => {
883 let mut full_response = String::new();
884 while let Some(chunk_result) = stream.stream.next().await {
885 match chunk_result {
886 Ok(chunk_str) => {
887 full_response.push_str(&chunk_str);
888 }
889 Err(err) => {
890 anyhow::bail!("Error receiving response from language model: {err}");
891 }
892 }
893 }
894 Ok(full_response)
895 }
896 Err(err) => Err(anyhow!(
897 "Failed to get response from language model. Error was: {err}"
898 )),
899 }
900}
901
902pub struct RequestMarkdown {
903 pub tools: String,
904 pub messages: String,
905}
906
907impl RequestMarkdown {
908 pub fn new(request: &LanguageModelRequest) -> Self {
909 let mut tools = String::new();
910 let mut messages = String::new();
911 let mut assistant_message_number: u32 = 1;
912
913 // Print the tools
914 if !request.tools.is_empty() {
915 for tool in &request.tools {
916 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
917 write!(&mut tools, "{}\n\n", tool.description).unwrap();
918 write!(
919 &mut tools,
920 "{}\n",
921 MarkdownCodeBlock {
922 tag: "json",
923 text: &format!("{:#}", tool.input_schema)
924 }
925 )
926 .unwrap();
927 }
928 }
929
930 // Print the messages
931 for message in &request.messages {
932 push_role(&message.role, &mut messages, &mut assistant_message_number);
933
934 for content in &message.content {
935 match content {
936 MessageContent::Text(text) => {
937 messages.push_str(text);
938 messages.push_str("\n\n");
939 }
940 MessageContent::Image(_) => {
941 messages.push_str("[IMAGE DATA]\n\n");
942 }
943 MessageContent::Thinking { text, signature } => {
944 messages.push_str("**Thinking**:\n\n");
945 if let Some(sig) = signature {
946 messages.push_str(&format!("Signature: {}\n\n", sig));
947 }
948 messages.push_str(text);
949 messages.push_str("\n");
950 }
951 MessageContent::RedactedThinking(items) => {
952 messages.push_str(&format!(
953 "**Redacted Thinking**: {} item(s)\n\n",
954 items.len()
955 ));
956 }
957 MessageContent::ToolUse(tool_use) => {
958 messages.push_str(&format!(
959 "**Tool Use**: {} (ID: {})\n",
960 tool_use.name, tool_use.id
961 ));
962 messages.push_str(&format!(
963 "{}\n",
964 MarkdownCodeBlock {
965 tag: "json",
966 text: &format!("{:#}", tool_use.input)
967 }
968 ));
969 }
970 MessageContent::ToolResult(tool_result) => {
971 messages.push_str(&format!(
972 "**Tool Result**: {} (ID: {})\n\n",
973 tool_result.tool_name, tool_result.tool_use_id
974 ));
975 if tool_result.is_error {
976 messages.push_str("**ERROR:**\n");
977 }
978
979 match &tool_result.content {
980 LanguageModelToolResultContent::Text(text) => {
981 writeln!(messages, "{text}\n").ok();
982 }
983 LanguageModelToolResultContent::Image(image) => {
984 writeln!(messages, "\n", image.source).ok();
985 }
986 }
987
988 if let Some(output) = tool_result.output.as_ref() {
989 writeln!(
990 messages,
991 "**Debug Output**:\n\n```json\n{}\n```\n",
992 serde_json::to_string_pretty(output).unwrap()
993 )
994 .unwrap();
995 }
996 }
997 }
998 }
999 }
1000
1001 Self { tools, messages }
1002 }
1003}
1004
1005pub fn response_events_to_markdown(
1006 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1007) -> String {
1008 let mut response = String::new();
1009 // Print the response events if any
1010 response.push_str("# Response\n\n");
1011 let mut text_buffer = String::new();
1012 let mut thinking_buffer = String::new();
1013
1014 let flush_buffers =
1015 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
1016 if !text_buffer.is_empty() {
1017 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
1018 text_buffer.clear();
1019 }
1020 if !thinking_buffer.is_empty() {
1021 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
1022 thinking_buffer.clear();
1023 }
1024 };
1025
1026 for event in response_events {
1027 match event {
1028 Ok(LanguageModelCompletionEvent::Text(text)) => {
1029 text_buffer.push_str(text);
1030 }
1031 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1032 thinking_buffer.push_str(text);
1033 }
1034 Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => {}
1035 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1036 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1037 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1038 }
1039 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1040 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1041 response.push_str(&format!(
1042 "**Tool Use**: {} (ID: {})\n",
1043 tool_use.name, tool_use.id
1044 ));
1045 response.push_str(&format!(
1046 "{}\n",
1047 MarkdownCodeBlock {
1048 tag: "json",
1049 text: &format!("{:#}", tool_use.input)
1050 }
1051 ));
1052 }
1053 Ok(
1054 LanguageModelCompletionEvent::UsageUpdate(_)
1055 | LanguageModelCompletionEvent::StartMessage { .. }
1056 | LanguageModelCompletionEvent::StatusUpdate { .. },
1057 ) => {}
1058 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1059 json_parse_error, ..
1060 }) => {
1061 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1062 response.push_str(&format!(
1063 "**Error**: parse error in tool use JSON: {}\n\n",
1064 json_parse_error
1065 ));
1066 }
1067 Err(error) => {
1068 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1069 response.push_str(&format!("**Error**: {}\n\n", error));
1070 }
1071 }
1072 }
1073
1074 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1075
1076 response
1077}
1078
1079#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
1080pub struct ThreadDialog {
1081 pub request: LanguageModelRequest,
1082 pub response_events: Vec<std::result::Result<LanguageModelCompletionEvent, String>>,
1083}
1084
1085impl ThreadDialog {
1086 pub fn new(
1087 request: &LanguageModelRequest,
1088 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1089 ) -> Self {
1090 Self {
1091 request: request.clone(),
1092 response_events: response_events.to_vec(),
1093 }
1094 }
1095
1096 /// Represents all request and response messages in a unified format.
1097 ///
1098 /// Specifically, it appends the assistant's response (derived from response events)
1099 /// as a new message to existing messages in the request.
1100 pub fn to_combined_request(&self) -> LanguageModelRequest {
1101 let mut request = self.request.clone();
1102 if let Some(assistant_message) = self.response_events_to_message() {
1103 request.messages.push(assistant_message);
1104 }
1105 request
1106 }
1107 fn response_events_to_message(&self) -> Option<LanguageModelRequestMessage> {
1108 let response_events = &self.response_events;
1109 let mut content: Vec<MessageContent> = Vec::new();
1110 let mut current_text = String::new();
1111
1112 let flush_text = |text: &mut String, content: &mut Vec<MessageContent>| {
1113 if !text.is_empty() {
1114 content.push(MessageContent::Text(std::mem::take(text)));
1115 }
1116 };
1117
1118 for event in response_events {
1119 match event {
1120 Ok(LanguageModelCompletionEvent::Text(text)) => {
1121 current_text.push_str(text);
1122 }
1123
1124 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1125 flush_text(&mut current_text, &mut content);
1126 if tool_use.is_input_complete {
1127 content.push(MessageContent::ToolUse(tool_use.clone()));
1128 }
1129 }
1130 Ok(LanguageModelCompletionEvent::Thinking { text, signature }) => {
1131 flush_text(&mut current_text, &mut content);
1132 content.push(MessageContent::Thinking {
1133 text: text.clone(),
1134 signature: signature.clone(),
1135 });
1136 }
1137
1138 // Skip these
1139 Ok(LanguageModelCompletionEvent::UsageUpdate(_))
1140 | Ok(LanguageModelCompletionEvent::RedactedThinking { .. })
1141 | Ok(LanguageModelCompletionEvent::StatusUpdate { .. })
1142 | Ok(LanguageModelCompletionEvent::StartMessage { .. })
1143 | Ok(LanguageModelCompletionEvent::Stop(_)) => {}
1144
1145 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1146 json_parse_error,
1147 ..
1148 }) => {
1149 flush_text(&mut current_text, &mut content);
1150 content.push(MessageContent::Text(format!(
1151 "ERROR: parse error in tool use JSON: {}",
1152 json_parse_error
1153 )));
1154 }
1155
1156 Err(error) => {
1157 flush_text(&mut current_text, &mut content);
1158 content.push(MessageContent::Text(format!("ERROR: {}", error)));
1159 }
1160 }
1161 }
1162
1163 flush_text(&mut current_text, &mut content);
1164
1165 if !content.is_empty() {
1166 Some(LanguageModelRequestMessage {
1167 role: Role::Assistant,
1168 content,
1169 cache: false,
1170 })
1171 } else {
1172 None
1173 }
1174 }
1175}
1176
1177#[cfg(test)]
1178mod test {
1179 use super::*;
1180
1181 #[test]
1182 fn test_parse_judge_output() {
1183 let response = r#"
1184 <analysis>The model did a good job but there were still compilations errors.</analysis>
1185 <passed>true</passed>
1186 "#
1187 .unindent();
1188
1189 let output = parse_assertion_result(&response).unwrap();
1190 assert_eq!(
1191 output.analysis,
1192 Some("The model did a good job but there were still compilations errors.".into())
1193 );
1194 assert_eq!(output.passed, true);
1195
1196 let response = r#"
1197 Text around ignored
1198
1199 <analysis>
1200 Failed to compile:
1201 - Error 1
1202 - Error 2
1203 </analysis>
1204
1205 <passed>false</passed>
1206 "#
1207 .unindent();
1208
1209 let output = parse_assertion_result(&response).unwrap();
1210 assert_eq!(
1211 output.analysis,
1212 Some("Failed to compile:\n- Error 1\n- Error 2".into())
1213 );
1214 assert_eq!(output.passed, false);
1215 }
1216}