1use agent::{Message, MessageSegment, SerializedThread, ThreadStore};
2use anyhow::{Context, Result, anyhow, bail};
3use assistant_tool::ToolWorkingSet;
4use client::proto::LspWorkProgress;
5use futures::channel::mpsc;
6use futures::{FutureExt as _, StreamExt as _, future};
7use gpui::{App, AppContext as _, AsyncApp, Entity, Task};
8use handlebars::Handlebars;
9use language::{Buffer, DiagnosticSeverity, OffsetRangeExt as _};
10use language_model::{
11 LanguageModel, LanguageModelCompletionEvent, LanguageModelRequest, LanguageModelRequestMessage,
12 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 return Err(anyhow!("Example has pre-existing diagnostics. If you want to run this example regardless, set `allow_preexisting_diagnostics` to `true` in `base.toml`"));
289 }
290
291 Some(LanguageServerState {
292 _lsp_open_handle: lsp_open_handle,
293 language_file_buffer,
294 })
295 } else {
296 None
297 };
298
299 if std::env::var("ZED_EVAL_SETUP_ONLY").is_ok() {
300 return Err(anyhow!("Setup only mode"));
301 }
302
303 let last_diff_file_path = this.run_directory.join("last.diff");
304
305 // Write an empty "last.diff" so that it can be opened in Zed for convenient view of the
306 // history using undo/redo.
307 std::fs::write(&last_diff_file_path, "")?;
308
309 let thread_store = thread_store.await?;
310
311 let profile_id = meta.profile_id.clone();
312 thread_store.update(cx, |thread_store, cx| thread_store.load_profile_by_id(profile_id, cx)).expect("Failed to load profile");
313
314 let thread =
315 thread_store.update(cx, |thread_store, cx| {
316 if let Some(json) = &meta.existing_thread_json {
317 let serialized = SerializedThread::from_json(json.as_bytes()).expect("Can't read serialized thread");
318 thread_store.create_thread_from_serialized(serialized, cx)
319 } else {
320 thread_store.create_thread(cx)
321 }
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(meta.clone(), this.log_prefix.clone(), thread.clone(), model.clone(), cx.clone());
371 let result = this.thread.conversation(&mut example_cx).await;
372
373 if let Err(err) = result {
374 if !err.is::<FailedAssertion>() {
375 return Err(err);
376 }
377 }
378
379 println!("{}Stopped", this.log_prefix);
380
381 println!("{}Getting repository diff", this.log_prefix);
382 let repository_diff = this.repository_diff().await?;
383
384 std::fs::write(last_diff_file_path, &repository_diff)?;
385
386
387 let mut diagnostics_after = None;
388 let mut diagnostic_summary_after = Default::default();
389
390 if let Some(language_server_state) = lsp {
391 wait_for_lang_server(&project, &language_server_state.language_file_buffer, this.log_prefix.clone(), cx).await?;
392
393 println!("{}Getting diagnostics", this.log_prefix);
394 diagnostics_after = cx
395 .update(|cx| {
396 let project = project.clone();
397 cx.spawn(async move |cx| query_lsp_diagnostics(project, cx).await)
398 })?
399 .await?;
400 println!("{}Got diagnostics", this.log_prefix);
401
402 diagnostic_summary_after = project.read_with(cx, |project, cx| {
403 project.diagnostic_summary(false, cx)
404 })?;
405
406 }
407
408 if let Some(diagnostics_before) = &diagnostics_before {
409 fs::write(this.run_directory.join("diagnostics_before.txt"), diagnostics_before)?;
410 }
411
412 if let Some(diagnostics_after) = &diagnostics_after {
413 fs::write(this.run_directory.join("diagnostics_after.txt"), diagnostics_after)?;
414 }
415
416 thread.update(cx, |thread, _cx| {
417 let response_count = thread
418 .messages()
419 .filter(|message| message.role == language_model::Role::Assistant)
420 .count();
421 RunOutput {
422 repository_diff,
423 diagnostic_summary_before,
424 diagnostic_summary_after,
425 diagnostics_before,
426 diagnostics_after,
427 response_count,
428 token_usage: thread.cumulative_token_usage(),
429 tool_metrics: example_cx.tool_metrics.lock().unwrap().clone(),
430 all_messages: messages_to_markdown(thread.messages()),
431 programmatic_assertions: example_cx.assertions,
432 }
433 })
434 })
435 }
436
437 async fn repository_diff(&self) -> Result<String> {
438 let worktree_path = self.worktree_path();
439 run_git(&worktree_path, &["add", "."]).await?;
440 let mut diff_args = vec!["diff", "--staged"];
441 if self.thread.meta().url == ZED_REPO_URL {
442 diff_args.push(":(exclude).rules");
443 }
444 run_git(&worktree_path, &diff_args).await
445 }
446
447 pub async fn judge(
448 &self,
449 model: Arc<dyn LanguageModel>,
450 run_output: &RunOutput,
451 cx: &AsyncApp,
452 ) -> JudgeOutput {
453 let mut output_file =
454 File::create(self.run_directory.join("judge.md")).expect("failed to create judge.md");
455
456 let diff_task = self.judge_diff(model.clone(), &run_output, cx);
457 let thread_task = self.judge_thread(model.clone(), &run_output, cx);
458
459 let (diff_result, thread_result) = futures::join!(diff_task, thread_task);
460
461 let (diff_response, diff_output) = diff_result;
462 let (thread_response, thread_output) = thread_result;
463
464 writeln!(
465 &mut output_file,
466 "# Judgment\n\n## Thread\n\n{thread_response}\n\n## Diff\n\n{diff_response}",
467 )
468 .log_err();
469
470 JudgeOutput {
471 thread: thread_output,
472 diff: diff_output,
473 }
474 }
475
476 async fn judge_diff(
477 &self,
478 model: Arc<dyn LanguageModel>,
479 run_output: &RunOutput,
480 cx: &AsyncApp,
481 ) -> (String, AssertionsReport) {
482 let diff_assertions = self.thread.diff_assertions();
483
484 if diff_assertions.is_empty() {
485 return (
486 "No diff assertions".to_string(),
487 AssertionsReport::default(),
488 );
489 }
490
491 println!("{}Running diff judge", self.log_prefix);
492
493 let judge_diff_prompt = include_str!("judge_diff_prompt.hbs");
494 let judge_diff_prompt_name = "judge_diff_prompt";
495 let mut hbs = Handlebars::new();
496 hbs.register_template_string(judge_diff_prompt_name, judge_diff_prompt)
497 .unwrap();
498
499 let to_prompt = |assertion: String| {
500 hbs.render(
501 judge_diff_prompt_name,
502 &JudgeDiffInput {
503 repository_diff: run_output.repository_diff.clone(),
504 assertion,
505 },
506 )
507 .unwrap()
508 };
509
510 let (responses, report) = self
511 .judge_assertions(model, diff_assertions, to_prompt, cx)
512 .await;
513
514 println!(
515 "{}Judge - Diff score: {}%",
516 self.log_prefix,
517 report.passed_percentage()
518 );
519
520 (responses, report)
521 }
522
523 async fn judge_thread(
524 &self,
525 model: Arc<dyn LanguageModel>,
526 run_output: &RunOutput,
527 cx: &AsyncApp,
528 ) -> (String, AssertionsReport) {
529 let thread_assertions = self.thread.thread_assertions();
530
531 if thread_assertions.is_empty() {
532 return (
533 "No thread assertions".to_string(),
534 AssertionsReport::default(),
535 );
536 }
537
538 let judge_thread_prompt = include_str!("judge_thread_prompt.hbs");
539 let judge_thread_prompt_name = "judge_thread_prompt";
540 let mut hbs = Handlebars::new();
541 hbs.register_template_string(judge_thread_prompt_name, judge_thread_prompt)
542 .unwrap();
543
544 let complete_messages = &run_output.all_messages;
545 let to_prompt = |assertion: String| {
546 hbs.render(
547 judge_thread_prompt_name,
548 &JudgeThreadInput {
549 messages: complete_messages.clone(),
550 assertion,
551 },
552 )
553 .unwrap()
554 };
555
556 let (responses, report) = self
557 .judge_assertions(model, thread_assertions, to_prompt, cx)
558 .await;
559
560 println!(
561 "{}Judge - Thread score: {}%",
562 self.log_prefix,
563 report.passed_percentage()
564 );
565
566 (responses, report)
567 }
568
569 async fn judge_assertions(
570 &self,
571 model: Arc<dyn LanguageModel>,
572 assertions: Vec<JudgeAssertion>,
573 to_prompt: impl Fn(String) -> String,
574 cx: &AsyncApp,
575 ) -> (String, AssertionsReport) {
576 let assertions = assertions.into_iter().map(|assertion| {
577 let request = LanguageModelRequest {
578 thread_id: None,
579 prompt_id: None,
580 mode: None,
581 messages: vec![LanguageModelRequestMessage {
582 role: Role::User,
583 content: vec![MessageContent::Text(to_prompt(assertion.description))],
584 cache: false,
585 }],
586 temperature: None,
587 tools: Vec::new(),
588 tool_choice: None,
589 stop: Vec::new(),
590 };
591
592 let model = model.clone();
593 let log_prefix = self.log_prefix.clone();
594 async move {
595 let response = send_language_model_request(model, request, cx).await;
596
597 let (response, result) = match response {
598 Ok(response) => (
599 response.clone(),
600 parse_assertion_result(&response).map_err(|err| err.to_string()),
601 ),
602 Err(err) => (err.to_string(), Err(err.to_string())),
603 };
604
605 if result.is_ok() {
606 println!("{}✅ {}", log_prefix, assertion.id);
607 } else {
608 println!("{}❌ {}", log_prefix, assertion.id);
609 }
610
611 (
612 response,
613 RanAssertion {
614 id: assertion.id,
615 result,
616 },
617 )
618 }
619 });
620
621 let mut responses = String::new();
622 let mut report = AssertionsReport::default();
623
624 for (response, assertion) in future::join_all(assertions).await {
625 writeln!(&mut responses, "# {}", assertion.id).unwrap();
626 writeln!(&mut responses, "{}\n\n", response).unwrap();
627 report.ran.push(assertion);
628 }
629
630 (responses, report)
631 }
632}
633
634pub fn wait_for_lang_server(
635 project: &Entity<Project>,
636 buffer: &Entity<Buffer>,
637 log_prefix: String,
638 cx: &mut AsyncApp,
639) -> Task<Result<()>> {
640 if std::env::var("ZED_EVAL_SKIP_LS").is_ok() {
641 return Task::ready(Ok(()));
642 }
643
644 println!("{}⏵ Waiting for language server", log_prefix);
645
646 let (mut tx, mut rx) = mpsc::channel(1);
647
648 let lsp_store = project
649 .update(cx, |project, _| project.lsp_store())
650 .unwrap();
651
652 let has_lang_server = buffer
653 .update(cx, |buffer, cx| {
654 lsp_store.update(cx, |lsp_store, cx| {
655 lsp_store
656 .language_servers_for_local_buffer(&buffer, cx)
657 .next()
658 .is_some()
659 })
660 })
661 .unwrap_or(false);
662
663 if has_lang_server {
664 project
665 .update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))
666 .unwrap()
667 .detach();
668 }
669
670 let subscriptions =
671 [
672 cx.subscribe(&lsp_store, {
673 let log_prefix = log_prefix.clone();
674 move |_, event, _| match event {
675 project::LspStoreEvent::LanguageServerUpdate {
676 message:
677 client::proto::update_language_server::Variant::WorkProgress(
678 LspWorkProgress {
679 message: Some(message),
680 ..
681 },
682 ),
683 ..
684 } => println!("{}⟲ {message}", log_prefix),
685 _ => {}
686 }
687 }),
688 cx.subscribe(&project, {
689 let buffer = buffer.clone();
690 move |project, event, cx| match event {
691 project::Event::LanguageServerAdded(_, _, _) => {
692 let buffer = buffer.clone();
693 project
694 .update(cx, |project, cx| project.save_buffer(buffer, cx))
695 .detach();
696 }
697 project::Event::DiskBasedDiagnosticsFinished { .. } => {
698 tx.try_send(()).ok();
699 }
700 _ => {}
701 }
702 }),
703 ];
704
705 cx.spawn(async move |cx| {
706 let timeout = cx.background_executor().timer(Duration::new(60 * 5, 0));
707 let result = futures::select! {
708 _ = rx.next() => {
709 println!("{}⚑ Language server idle", log_prefix);
710 anyhow::Ok(())
711 },
712 _ = timeout.fuse() => {
713 Err(anyhow!("LSP wait timed out after 5 minutes"))
714 }
715 };
716 drop(subscriptions);
717 result
718 })
719}
720
721pub async fn query_lsp_diagnostics(
722 project: Entity<Project>,
723 cx: &mut AsyncApp,
724) -> Result<Option<String>> {
725 let paths_with_diagnostics = project.update(cx, |project, cx| {
726 project
727 .diagnostic_summaries(true, cx)
728 .filter(|(_, _, summary)| summary.error_count > 0 || summary.warning_count > 0)
729 .map(|(project_path, _, _)| project_path)
730 .collect::<Vec<_>>()
731 })?;
732
733 if paths_with_diagnostics.is_empty() {
734 return Ok(None);
735 }
736
737 let mut output = String::new();
738 for project_path in paths_with_diagnostics {
739 let buffer = project
740 .update(cx, |project, cx| project.open_buffer(project_path, cx))?
741 .await?;
742 let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
743
744 for (_, group) in snapshot.diagnostic_groups(None) {
745 let entry = &group.entries[group.primary_ix];
746 let range = entry.range.to_point(&snapshot);
747 let severity = match entry.diagnostic.severity {
748 DiagnosticSeverity::ERROR => "error",
749 DiagnosticSeverity::WARNING => "warning",
750 _ => continue,
751 };
752
753 writeln!(
754 output,
755 "{} at line {}: {}",
756 severity,
757 range.start.row + 1,
758 entry.diagnostic.message
759 )?;
760 }
761 }
762 anyhow::Ok(Some(output))
763}
764
765fn parse_assertion_result(response: &str) -> Result<RanAssertionResult> {
766 let analysis = get_tag("analysis", response)?.to_string();
767 let passed = match get_tag("passed", response)?.to_lowercase().as_str() {
768 "true" => true,
769 "false" => false,
770 value @ _ => bail!("invalid judge `passed` tag: {value}"),
771 };
772 Ok(RanAssertionResult {
773 analysis: Some(analysis),
774 passed,
775 })
776}
777
778fn get_tag(name: &'static str, response: &str) -> Result<String> {
779 let start_tag = format!("<{}>", name);
780 let end_tag = format!("</{}>", name);
781
782 let start_ix = response
783 .find(&start_tag)
784 .context(format!("{} start tag not found", name))?;
785 let content_start_ix = start_ix + start_tag.len();
786
787 let end_ix = content_start_ix
788 + response[content_start_ix..]
789 .find(&end_tag)
790 .context(format!("{} end tag not found", name))?;
791
792 let content = response[content_start_ix..end_ix].trim().unindent();
793
794 anyhow::Ok(content)
795}
796
797pub fn repo_path_for_url(repos_dir: &Path, repo_url: &str) -> PathBuf {
798 let repo_name = repo_url
799 .trim_start_matches("https://")
800 .replace(|c: char| !c.is_alphanumeric(), "-");
801 Path::new(repos_dir).join(repo_name)
802}
803
804pub async fn run_git(repo_path: &Path, args: &[&str]) -> Result<String> {
805 let output = new_smol_command("git")
806 .current_dir(repo_path)
807 .args(args)
808 .output()
809 .await?;
810
811 if output.status.success() {
812 Ok(String::from_utf8(output.stdout)?.trim().to_string())
813 } else {
814 Err(anyhow!(
815 "`git {}` within `{}` failed with status: {}\nstderr:\n{}\nstdout:\n{}",
816 args.join(" "),
817 repo_path.display(),
818 output.status,
819 String::from_utf8_lossy(&output.stderr),
820 String::from_utf8_lossy(&output.stdout),
821 ))
822 }
823}
824
825fn messages_to_markdown<'a>(message_iter: impl IntoIterator<Item = &'a Message>) -> String {
826 let mut messages = String::new();
827 let mut assistant_message_number: u32 = 1;
828
829 for message in message_iter {
830 push_role(&message.role, &mut messages, &mut assistant_message_number);
831
832 for segment in &message.segments {
833 match segment {
834 MessageSegment::Text(text) => {
835 messages.push_str(&text);
836 messages.push_str("\n\n");
837 }
838 MessageSegment::Thinking { text, signature } => {
839 messages.push_str("**Thinking**:\n\n");
840 if let Some(sig) = signature {
841 messages.push_str(&format!("Signature: {}\n\n", sig));
842 }
843 messages.push_str(&text);
844 messages.push_str("\n");
845 }
846 MessageSegment::RedactedThinking(items) => {
847 messages.push_str(&format!(
848 "**Redacted Thinking**: {} item(s)\n\n",
849 items.len()
850 ));
851 }
852 }
853 }
854 }
855
856 messages
857}
858
859fn push_role(role: &Role, buf: &mut String, assistant_message_number: &mut u32) {
860 match role {
861 Role::System => buf.push_str("# ⚙️ SYSTEM\n\n"),
862 Role::User => buf.push_str("# 👤 USER\n\n"),
863 Role::Assistant => {
864 buf.push_str(&format!("# 🤖 ASSISTANT {assistant_message_number}\n\n"));
865 *assistant_message_number = *assistant_message_number + 1;
866 }
867 }
868}
869
870pub async fn send_language_model_request(
871 model: Arc<dyn LanguageModel>,
872 request: LanguageModelRequest,
873 cx: &AsyncApp,
874) -> anyhow::Result<String> {
875 match model.stream_completion_text(request, &cx).await {
876 Ok(mut stream) => {
877 let mut full_response = String::new();
878 while let Some(chunk_result) = stream.stream.next().await {
879 match chunk_result {
880 Ok(chunk_str) => {
881 full_response.push_str(&chunk_str);
882 }
883 Err(err) => {
884 return Err(anyhow!(
885 "Error receiving response from language model: {err}"
886 ));
887 }
888 }
889 }
890 Ok(full_response)
891 }
892 Err(err) => Err(anyhow!(
893 "Failed to get response from language model. Error was: {err}"
894 )),
895 }
896}
897
898pub struct RequestMarkdown {
899 pub tools: String,
900 pub messages: String,
901}
902
903impl RequestMarkdown {
904 pub fn new(request: &LanguageModelRequest) -> Self {
905 let mut tools = String::new();
906 let mut messages = String::new();
907 let mut assistant_message_number: u32 = 1;
908
909 // Print the tools
910 if !request.tools.is_empty() {
911 for tool in &request.tools {
912 write!(&mut tools, "# {}\n\n", tool.name).unwrap();
913 write!(&mut tools, "{}\n\n", tool.description).unwrap();
914 write!(
915 &mut tools,
916 "{}\n",
917 MarkdownCodeBlock {
918 tag: "json",
919 text: &format!("{:#}", tool.input_schema)
920 }
921 )
922 .unwrap();
923 }
924 }
925
926 // Print the messages
927 for message in &request.messages {
928 push_role(&message.role, &mut messages, &mut assistant_message_number);
929
930 for content in &message.content {
931 match content {
932 MessageContent::Text(text) => {
933 messages.push_str(text);
934 messages.push_str("\n\n");
935 }
936 MessageContent::Image(_) => {
937 messages.push_str("[IMAGE DATA]\n\n");
938 }
939 MessageContent::Thinking { text, signature } => {
940 messages.push_str("**Thinking**:\n\n");
941 if let Some(sig) = signature {
942 messages.push_str(&format!("Signature: {}\n\n", sig));
943 }
944 messages.push_str(text);
945 messages.push_str("\n");
946 }
947 MessageContent::RedactedThinking(items) => {
948 messages.push_str(&format!(
949 "**Redacted Thinking**: {} item(s)\n\n",
950 items.len()
951 ));
952 }
953 MessageContent::ToolUse(tool_use) => {
954 messages.push_str(&format!(
955 "**Tool Use**: {} (ID: {})\n",
956 tool_use.name, tool_use.id
957 ));
958 messages.push_str(&format!(
959 "{}\n",
960 MarkdownCodeBlock {
961 tag: "json",
962 text: &format!("{:#}", tool_use.input)
963 }
964 ));
965 }
966 MessageContent::ToolResult(tool_result) => {
967 messages.push_str(&format!(
968 "**Tool Result**: {} (ID: {})\n\n",
969 tool_result.tool_name, tool_result.tool_use_id
970 ));
971 if tool_result.is_error {
972 messages.push_str("**ERROR:**\n");
973 }
974
975 match &tool_result.content {
976 LanguageModelToolResultContent::Text(str) => {
977 writeln!(messages, "{}\n", str).ok();
978 }
979 LanguageModelToolResultContent::Image(image) => {
980 writeln!(messages, "\n", image.source).ok();
981 }
982 }
983
984 if let Some(output) = tool_result.output.as_ref() {
985 writeln!(
986 messages,
987 "**Debug Output**:\n\n```json\n{}\n```\n",
988 serde_json::to_string_pretty(output).unwrap()
989 )
990 .unwrap();
991 }
992 }
993 }
994 }
995 }
996
997 Self { tools, messages }
998 }
999}
1000
1001pub fn response_events_to_markdown(
1002 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1003) -> String {
1004 let mut response = String::new();
1005 // Print the response events if any
1006 response.push_str("# Response\n\n");
1007 let mut text_buffer = String::new();
1008 let mut thinking_buffer = String::new();
1009
1010 let flush_buffers =
1011 |output: &mut String, text_buffer: &mut String, thinking_buffer: &mut String| {
1012 if !text_buffer.is_empty() {
1013 output.push_str(&format!("**Text**:\n{}\n\n", text_buffer));
1014 text_buffer.clear();
1015 }
1016 if !thinking_buffer.is_empty() {
1017 output.push_str(&format!("**Thinking**:\n{}\n\n", thinking_buffer));
1018 thinking_buffer.clear();
1019 }
1020 };
1021
1022 for event in response_events {
1023 match event {
1024 Ok(LanguageModelCompletionEvent::Text(text)) => {
1025 text_buffer.push_str(text);
1026 }
1027 Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
1028 thinking_buffer.push_str(text);
1029 }
1030 Ok(LanguageModelCompletionEvent::Stop(reason)) => {
1031 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1032 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
1033 }
1034 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1035 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1036 response.push_str(&format!(
1037 "**Tool Use**: {} (ID: {})\n",
1038 tool_use.name, tool_use.id
1039 ));
1040 response.push_str(&format!(
1041 "{}\n",
1042 MarkdownCodeBlock {
1043 tag: "json",
1044 text: &format!("{:#}", tool_use.input)
1045 }
1046 ));
1047 }
1048 Ok(
1049 LanguageModelCompletionEvent::UsageUpdate(_)
1050 | LanguageModelCompletionEvent::StartMessage { .. }
1051 | LanguageModelCompletionEvent::StatusUpdate { .. },
1052 ) => {}
1053 Err(error) => {
1054 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1055 response.push_str(&format!("**Error**: {}\n\n", error));
1056 }
1057 }
1058 }
1059
1060 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
1061
1062 response
1063}
1064
1065#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
1066pub struct ThreadDialog {
1067 pub request: LanguageModelRequest,
1068 pub response_events: Vec<std::result::Result<LanguageModelCompletionEvent, String>>,
1069}
1070
1071impl ThreadDialog {
1072 pub fn new(
1073 request: &LanguageModelRequest,
1074 response_events: &[std::result::Result<LanguageModelCompletionEvent, String>],
1075 ) -> Self {
1076 Self {
1077 request: request.clone(),
1078 response_events: response_events.to_vec(),
1079 }
1080 }
1081
1082 /// Represents all request and response messages in a unified format.
1083 ///
1084 /// Specifically, it appends the assistant's response (derived from response events)
1085 /// as a new message to existing messages in the request.
1086 pub fn to_combined_request(&self) -> LanguageModelRequest {
1087 let mut request = self.request.clone();
1088 if let Some(assistant_message) = self.response_events_to_message() {
1089 request.messages.push(assistant_message);
1090 }
1091 request
1092 }
1093 fn response_events_to_message(&self) -> Option<LanguageModelRequestMessage> {
1094 let response_events = &self.response_events;
1095 let mut content: Vec<MessageContent> = Vec::new();
1096 let mut current_text = String::new();
1097
1098 let flush_text = |text: &mut String, content: &mut Vec<MessageContent>| {
1099 if !text.is_empty() {
1100 content.push(MessageContent::Text(std::mem::take(text)));
1101 }
1102 };
1103
1104 for event in response_events {
1105 match event {
1106 Ok(LanguageModelCompletionEvent::Text(text)) => {
1107 current_text.push_str(text);
1108 }
1109
1110 Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) => {
1111 flush_text(&mut current_text, &mut content);
1112 if tool_use.is_input_complete {
1113 content.push(MessageContent::ToolUse(tool_use.clone()));
1114 }
1115 }
1116 Ok(LanguageModelCompletionEvent::Thinking { text, signature }) => {
1117 flush_text(&mut current_text, &mut content);
1118 content.push(MessageContent::Thinking {
1119 text: text.clone(),
1120 signature: signature.clone(),
1121 });
1122 }
1123
1124 // Skip these
1125 Ok(LanguageModelCompletionEvent::UsageUpdate(_))
1126 | Ok(LanguageModelCompletionEvent::StatusUpdate { .. })
1127 | Ok(LanguageModelCompletionEvent::StartMessage { .. })
1128 | Ok(LanguageModelCompletionEvent::Stop(_)) => {}
1129
1130 Err(error) => {
1131 flush_text(&mut current_text, &mut content);
1132 content.push(MessageContent::Text(format!("ERROR: {}", error)));
1133 }
1134 }
1135 }
1136
1137 flush_text(&mut current_text, &mut content);
1138
1139 if !content.is_empty() {
1140 Some(LanguageModelRequestMessage {
1141 role: Role::Assistant,
1142 content,
1143 cache: false,
1144 })
1145 } else {
1146 None
1147 }
1148 }
1149}
1150
1151#[cfg(test)]
1152mod test {
1153 use super::*;
1154
1155 #[test]
1156 fn test_parse_judge_output() {
1157 let response = r#"
1158 <analysis>The model did a good job but there were still compilations errors.</analysis>
1159 <passed>true</passed>
1160 "#
1161 .unindent();
1162
1163 let output = parse_assertion_result(&response).unwrap();
1164 assert_eq!(
1165 output.analysis,
1166 Some("The model did a good job but there were still compilations errors.".into())
1167 );
1168 assert_eq!(output.passed, true);
1169
1170 let response = r#"
1171 Text around ignored
1172
1173 <analysis>
1174 Failed to compile:
1175 - Error 1
1176 - Error 2
1177 </analysis>
1178
1179 <passed>false</passed>
1180 "#
1181 .unindent();
1182
1183 let output = parse_assertion_result(&response).unwrap();
1184 assert_eq!(
1185 output.analysis,
1186 Some("Failed to compile:\n- Error 1\n- Error 2".into())
1187 );
1188 assert_eq!(output.passed, false);
1189 }
1190}