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