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