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