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