1use crate::{
2 FormatPromptArgs, PredictionProvider,
3 example::{ActualCursor, Example, ExamplePrompt},
4 headless::EpAppState,
5 progress::{ExampleProgress, Step},
6 retrieve_context::run_context_retrieval,
7};
8use anyhow::{Context as _, Result, anyhow};
9use edit_prediction::{cursor_excerpt::editable_and_context_ranges_for_cursor_position, udiff};
10use gpui::{AppContext, AsyncApp};
11use language::{Buffer, OffsetRangeExt, Point};
12use similar::DiffableStr;
13use std::sync::Arc;
14use std::{fmt::Write as _, ops::Range};
15use zeta_prompt::ZetaFormat;
16use zeta_prompt::format_zeta_prompt;
17
18pub async fn run_format_prompt(
19 example: &mut Example,
20 args: &FormatPromptArgs,
21 app_state: Arc<EpAppState>,
22 example_progress: &ExampleProgress,
23 cx: AsyncApp,
24) -> Result<()> {
25 run_context_retrieval(example, app_state.clone(), example_progress, cx.clone()).await?;
26
27 let step_progress = example_progress.start(Step::FormatPrompt);
28
29 let prompt_inputs = example
30 .prompt_inputs
31 .as_ref()
32 .context("prompt_inputs must be set after context retrieval")?;
33
34 let language = app_state
35 .languages
36 .load_language_for_file_path(&example.spec.cursor_path)
37 .await
38 .ok();
39 let snapshot_fut = cx.update(|cx| {
40 Buffer::build_snapshot(
41 prompt_inputs.content.as_str().into(),
42 language,
43 Some(app_state.languages.clone()),
44 cx,
45 )
46 });
47 let cursor_point = Point::new(prompt_inputs.cursor_row, prompt_inputs.cursor_column);
48 let snapshot = cx.background_spawn(snapshot_fut).await;
49
50 match args.provider {
51 PredictionProvider::Teacher(_) | PredictionProvider::TeacherNonBatching(_) => {
52 step_progress.set_substatus("formatting teacher prompt");
53
54 let (editable_range, context_range) = editable_and_context_ranges_for_cursor_position(
55 cursor_point,
56 &snapshot,
57 edit_prediction::zeta2::max_editable_tokens(ZetaFormat::default()),
58 edit_prediction::zeta2::MAX_CONTEXT_TOKENS,
59 );
60 let editable_range = editable_range.to_offset(&snapshot);
61 let context_range = context_range.to_offset(&snapshot);
62
63 let prompt = TeacherPrompt::format_prompt(example, editable_range, context_range);
64 example.prompt = Some(ExamplePrompt {
65 input: prompt,
66 expected_output: String::new(),
67 rejected_output: None,
68 prefill: None,
69 provider: args.provider,
70 });
71 }
72 PredictionProvider::Zeta2(version) => {
73 step_progress.set_substatus("formatting zeta2 prompt");
74
75 let (editable_range, context_range) = editable_and_context_ranges_for_cursor_position(
76 cursor_point,
77 &snapshot,
78 edit_prediction::zeta2::max_editable_tokens(version),
79 edit_prediction::zeta2::MAX_CONTEXT_TOKENS,
80 );
81 let editable_range = editable_range.to_offset(&snapshot);
82 let context_range = context_range.to_offset(&snapshot);
83
84 let context_start = context_range.start;
85 let cursor_offset_in_excerpt = prompt_inputs.cursor_offset - context_start;
86 let editable_range_in_excerpt =
87 (editable_range.start - context_start)..(editable_range.end - context_start);
88 let input = zeta_prompt::ZetaPromptInput {
89 cursor_path: example.spec.cursor_path.clone(),
90 cursor_excerpt: prompt_inputs.content[context_range].to_string().into(),
91 editable_range_in_excerpt,
92 cursor_offset_in_excerpt,
93 excerpt_start_row: prompt_inputs.excerpt_start_row,
94 events: prompt_inputs.edit_history.clone(),
95 related_files: prompt_inputs.related_files.clone().unwrap_or_default(),
96 excerpt_ranges: None,
97 preferred_model: None,
98 in_open_source_repo: example
99 .spec
100 .captured_prompt_input
101 .as_ref()
102 .map_or(false, |input| input.in_open_source_repo),
103 can_collect_data: false,
104 };
105 let prompt = format_zeta_prompt(&input, version);
106 let prefill = zeta_prompt::get_prefill(&input, version);
107 let (expected_patch, expected_cursor_offset) = example
108 .spec
109 .expected_patches_with_cursor_positions()
110 .into_iter()
111 .next()
112 .context("expected patches is empty")?;
113 let expected_output =
114 zeta2_output_for_patch(&input, &expected_patch, expected_cursor_offset, version)?;
115 let rejected_output = example
116 .spec
117 .rejected_patch
118 .as_ref()
119 .and_then(|patch| zeta2_output_for_patch(&input, patch, None, version).ok());
120
121 example.prompt = Some(ExamplePrompt {
122 input: prompt,
123 expected_output,
124 rejected_output,
125 provider: args.provider,
126 prefill: Some(prefill),
127 });
128 }
129 _ => {
130 panic!("Cannot format prompt for {:?}", args.provider);
131 }
132 };
133 Ok(())
134}
135
136pub fn zeta2_output_for_patch(
137 input: &zeta_prompt::ZetaPromptInput,
138 patch: &str,
139 cursor_offset: Option<usize>,
140 version: ZetaFormat,
141) -> Result<String> {
142 let mut old_editable_region =
143 input.cursor_excerpt[input.editable_range_in_excerpt.clone()].to_string();
144
145 if !old_editable_region.ends_with_newline() {
146 old_editable_region.push('\n');
147 }
148
149 let (mut result, first_hunk_offset) =
150 udiff::apply_diff_to_string_with_hunk_offset(patch, &old_editable_region).with_context(
151 || {
152 format!(
153 "Patch:\n```\n{}```\n\nEditable region:\n```\n{}```",
154 patch, old_editable_region
155 )
156 },
157 )?;
158
159 if let Some(cursor_offset) = cursor_offset {
160 // The cursor_offset is relative to the start of the hunk's new text (context + additions).
161 // We need to add where the hunk context matched in the editable region to compute
162 // the actual cursor position in the result.
163 let hunk_start = first_hunk_offset.unwrap_or(0);
164 let offset = result.floor_char_boundary((hunk_start + cursor_offset).min(result.len()));
165 result.insert_str(offset, zeta_prompt::CURSOR_MARKER);
166 }
167
168 match version {
169 ZetaFormat::V0120GitMergeMarkers
170 | ZetaFormat::V0131GitMergeMarkersPrefix
171 | ZetaFormat::V0211SeedCoder => {
172 if !result.ends_with('\n') {
173 result.push('\n');
174 }
175 result.push_str(zeta_prompt::v0120_git_merge_markers::END_MARKER);
176 }
177 _ => (),
178 }
179
180 Ok(result)
181}
182
183pub struct TeacherPrompt;
184
185impl TeacherPrompt {
186 pub(crate) const EDITABLE_REGION_START: &str = "<|editable_region_start|>\n";
187 pub(crate) const EDITABLE_REGION_END: &str = "\n<|editable_region_end|>";
188 pub(crate) const USER_CURSOR_MARKER: &str = "<|user_cursor|>";
189 pub(crate) const NO_EDITS: &str = "NO_EDITS";
190
191 /// Truncate edit history to this number of last lines
192 const MAX_HISTORY_LINES: usize = 128;
193
194 pub fn format_prompt(
195 example: &Example,
196 editable_range: Range<usize>,
197 context_range: Range<usize>,
198 ) -> String {
199 let edit_history = Self::format_edit_history(&example.spec.edit_history);
200 let context = Self::format_context(example);
201 let cursor_excerpt = Self::format_cursor_excerpt(example, editable_range, context_range);
202
203 let prompt_template = crate::prompt_assets::get_prompt("teacher.md");
204 let prompt = prompt_template
205 .replace("{{context}}", &context)
206 .replace("{{edit_history}}", &edit_history)
207 .replace("{{cursor_excerpt}}", &cursor_excerpt);
208
209 prompt
210 }
211
212 pub fn parse(example: &Example, response: &str) -> Result<(String, Option<ActualCursor>)> {
213 // Check if the model indicated no edits are needed
214 if let Some(last_codeblock) = extract_last_codeblock(&response) {
215 if last_codeblock.trim() == Self::NO_EDITS {
216 return Ok((String::new(), None));
217 }
218 }
219
220 // Extract updated (new) editable region from the model response.
221 let new_editable_region = Self::extract_editable_region(&response)?;
222 let cursor_offset = new_editable_region.find(Self::USER_CURSOR_MARKER);
223 let mut new_editable_region = new_editable_region.replace(Self::USER_CURSOR_MARKER, "");
224 let old_editable_region = Self::extract_editable_region(
225 &example
226 .prompt
227 .as_ref()
228 .context("example prompt missing")?
229 .input,
230 )?
231 .replace(Self::USER_CURSOR_MARKER, "");
232
233 let prompt_inputs = example
234 .prompt_inputs
235 .as_ref()
236 .context("example is missing prompt inputs")?;
237
238 // Normalize leading newlines: if old starts with newline but new doesn't,
239 // prepend newline to new to preserve whitespace structure.
240 // This handles the case where the model drops the leading blank line.
241 if old_editable_region.starts_with('\n') && !new_editable_region.starts_with('\n') {
242 new_editable_region.insert(0, '\n');
243 }
244
245 let (editable_region_offset, _) = prompt_inputs
246 .content
247 .match_indices(&old_editable_region)
248 .min_by_key(|(index, _)| index.abs_diff(prompt_inputs.cursor_offset))
249 .context("editable region not found in prompt content")?;
250 let editable_region_start_line = prompt_inputs.content[..editable_region_offset]
251 .matches('\n')
252 .count();
253
254 // Use full context so cursor offset (relative to editable region start) aligns with diff content
255 let editable_region_lines = old_editable_region.lines().count() as u32;
256 let diff = language::unified_diff_with_context(
257 &old_editable_region,
258 &new_editable_region,
259 editable_region_start_line as u32,
260 editable_region_start_line as u32,
261 editable_region_lines,
262 );
263
264 let diff = indoc::formatdoc! {"
265 --- a/{path}
266 +++ b/{path}
267 {diff}",
268 path = example.spec.cursor_path.to_string_lossy(),
269 diff = diff,
270 };
271
272 let actual_cursor = cursor_offset.map(|editable_region_cursor_offset| {
273 ActualCursor::from_editable_region(
274 &example.spec.cursor_path,
275 editable_region_cursor_offset,
276 &new_editable_region,
277 &prompt_inputs.content,
278 editable_region_offset,
279 editable_region_start_line,
280 )
281 });
282
283 Ok((diff, actual_cursor))
284 }
285
286 fn format_edit_history(edit_history: &str) -> String {
287 let lines: Vec<&str> = edit_history.lines().collect();
288
289 if lines.is_empty() {
290 return "(No edit history)".to_string();
291 }
292
293 if lines.len() > Self::MAX_HISTORY_LINES {
294 let truncated = lines[lines.len() - Self::MAX_HISTORY_LINES..].join("\n");
295 format!("{truncated}\n[...truncated...]")
296 } else {
297 lines.join("\n")
298 }
299 }
300
301 pub fn format_context(example: &Example) -> String {
302 let related_files = example
303 .prompt_inputs
304 .as_ref()
305 .and_then(|pi| pi.related_files.as_ref());
306
307 let Some(related_files) = related_files else {
308 return "(No context)".to_string();
309 };
310
311 if related_files.is_empty() {
312 return "(No context)".to_string();
313 }
314
315 let mut prompt = String::new();
316 for file in related_files {
317 let path_str = file.path.to_string_lossy();
318 writeln!(&mut prompt, "`````{path_str}").ok();
319
320 let mut prev_row = 0;
321 for excerpt in &file.excerpts {
322 if excerpt.row_range.start > prev_row {
323 prompt.push_str("…\n");
324 }
325 prompt.push_str(&excerpt.text);
326 prompt.push('\n');
327 prev_row = excerpt.row_range.end;
328 }
329 if prev_row < file.max_row {
330 prompt.push_str("…\n");
331 }
332 prompt.push_str("\n`````\n");
333 }
334
335 prompt
336 }
337
338 fn format_cursor_excerpt(
339 example: &Example,
340 editable_range: Range<usize>,
341 context_range: Range<usize>,
342 ) -> String {
343 let mut result = String::new();
344
345 let prompt_inputs = example.prompt_inputs.as_ref().unwrap();
346
347 let path_str = example.spec.cursor_path.to_string_lossy();
348 result.push_str(&format!("`````{path_str}\n"));
349 result.push_str(&prompt_inputs.content[context_range.start..editable_range.start]);
350 result.push_str(Self::EDITABLE_REGION_START);
351 result.push_str(&prompt_inputs.content[editable_range.start..prompt_inputs.cursor_offset]);
352 result.push_str(Self::USER_CURSOR_MARKER);
353 result.push_str(&prompt_inputs.content[prompt_inputs.cursor_offset..editable_range.end]);
354 result.push_str(Self::EDITABLE_REGION_END);
355 result.push_str(&prompt_inputs.content[editable_range.end..context_range.end]);
356 result.push_str("\n`````");
357
358 result
359 }
360
361 pub fn extract_editable_region(text: &str) -> Result<String> {
362 let start = text
363 .rfind(Self::EDITABLE_REGION_START)
364 .map_or(0, |pos| pos + Self::EDITABLE_REGION_START.len());
365 let end = text.rfind(Self::EDITABLE_REGION_END).unwrap_or(text.len());
366
367 if start >= end {
368 return Err(anyhow!("Invalid editable region markers"));
369 }
370
371 let region = &text[start..end];
372 Ok(region.strip_suffix('\n').unwrap_or(region).to_string())
373 }
374}
375
376/// Extract the cursor excerpt from an example.
377/// First tries to extract from an existing prompt, then falls back to constructing from prompt_inputs.
378pub fn extract_cursor_excerpt_from_example(example: &Example) -> Option<String> {
379 // If we have the original prompt, extract the cursor excerpt from it
380 if let Some(prompt) = &example.prompt {
381 // Find "# 3. Current File" section and extract the content
382 if let Some(start) = prompt.input.find("# 3. Current File") {
383 let content_start = prompt.input[start..].find('`').map(|i| start + i)?;
384 let backtick_count = prompt.input[content_start..]
385 .chars()
386 .take_while(|&c| c == '`')
387 .count();
388 let content_start = content_start + backtick_count;
389
390 // Find the path line and skip it
391 let newline_pos = prompt.input[content_start..].find('\n')?;
392 let text_start = content_start + newline_pos + 1;
393
394 // Find the closing backticks
395 let closing_pattern = "`".repeat(backtick_count);
396 let text_end = prompt.input[text_start..].find(&closing_pattern)?;
397 let cursor_excerpt = &prompt.input[text_start..text_start + text_end];
398
399 let path_str = example.spec.cursor_path.to_string_lossy();
400 return Some(format!("`````{path_str}\n{cursor_excerpt}`````"));
401 }
402 }
403
404 // Fallback: construct from prompt_inputs if available
405 let prompt_inputs = example.prompt_inputs.as_ref()?;
406 let content = &prompt_inputs.content;
407 let cursor_offset = prompt_inputs.cursor_offset;
408
409 // Simple fallback: just show content around cursor with markers
410 let path_str = example.spec.cursor_path.to_string_lossy();
411 let mut result = format!("`````{path_str}\n");
412 result.push_str(TeacherPrompt::EDITABLE_REGION_START);
413 result.push_str(&content[..cursor_offset]);
414 result.push_str(TeacherPrompt::USER_CURSOR_MARKER);
415 result.push_str(&content[cursor_offset..]);
416 result.push_str(TeacherPrompt::EDITABLE_REGION_END);
417 result.push_str("\n`````");
418
419 Some(result)
420}
421
422pub(crate) fn extract_last_codeblock(text: &str) -> Option<String> {
423 let lines: Vec<&str> = text.lines().collect();
424
425 // Search from the end for a closing fence (line containing only backticks, 3+)
426 let mut closing_line_idx = None;
427 let mut backtick_count = 0;
428
429 for i in (0..lines.len()).rev() {
430 let line = lines[i].trim();
431 if line.len() >= 3 && line.chars().all(|c| c == '`') {
432 closing_line_idx = Some(i);
433 backtick_count = line.len();
434 break;
435 }
436 }
437
438 let closing_idx = closing_line_idx?;
439
440 // Search backwards for matching opening fence
441 // Opening fence starts with same backtick count, possibly followed by language/metadata
442 let opening_pattern = "`".repeat(backtick_count);
443
444 for i in (0..closing_idx).rev() {
445 let line = lines[i];
446 if line.starts_with(&opening_pattern) {
447 // Ensure it's exactly the right number of backticks (not more)
448 let rest = &line[backtick_count..];
449 if rest.is_empty() || !rest.starts_with('`') {
450 // Found matching opening fence
451 // Extract content between opening and closing (exclusive)
452 if closing_idx > i + 1 {
453 let content = lines[i + 1..closing_idx].join("\n");
454 // Preserve trailing newline to match previous behavior
455 return Some(format!("{}\n", content));
456 } else {
457 // Empty block
458 return Some(String::new());
459 }
460 }
461 }
462 }
463
464 None
465}
466
467#[cfg(test)]
468mod tests {
469 use super::*;
470
471 #[test]
472 fn test_extract_last_code_block() {
473 let text = indoc::indoc! {"
474 Some thinking
475
476 ```
477 first block
478 ```
479
480 `````path='something' lines=1:2
481 last block
482 `````
483 "};
484 let last_block = extract_last_codeblock(text).unwrap();
485 assert_eq!(last_block, "last block\n");
486 }
487
488 #[test]
489 fn test_extract_codeblock_with_nested_fences() {
490 let text = indoc::indoc! {"
491 `````
492 content with ``` inline
493 and ```python nested
494 more content
495 `````
496 "};
497 let last_block = extract_last_codeblock(text).unwrap();
498 assert_eq!(
499 last_block,
500 "content with ``` inline\nand ```python nested\nmore content\n"
501 );
502 }
503
504 #[test]
505 fn test_extract_codeblock_ignores_inline_backticks() {
506 let text = indoc::indoc! {"
507 `````
508 here is some `code` with inline backticks
509 and here```more```stuff
510 `````
511 "};
512 let last_block = extract_last_codeblock(text).unwrap();
513 assert_eq!(
514 last_block,
515 "here is some `code` with inline backticks\nand here```more```stuff\n"
516 );
517 }
518
519 #[test]
520 fn test_extract_editable_region() {
521 let text = indoc::indoc! {"
522 some lines
523 are
524 here
525 <|editable_region_start|>
526 one
527 two three
528
529 <|editable_region_end|>
530 more
531 lines here
532 "};
533 let parsed = TeacherPrompt::extract_editable_region(text).unwrap();
534 assert_eq!(
535 parsed,
536 indoc::indoc! {"
537 one
538 two three"}
539 );
540 }
541
542 #[test]
543 fn test_extract_last_codeblock_nested_bibtex() {
544 let text = indoc::indoc! {r#"
545 Looking at the edit history, I can see that a Citation section was just added.
546
547 `````
548 ## Collaborations
549 Our mission is to create a 4D generative model.
550
551 ## Citation
552
553 If you found Unique3D helpful, please cite our report:
554 ```bibtex
555 @misc{wu2024unique3d,
556 title={Unique3D},
557 }
558 ```
559 `````
560 "#};
561 let last_block = extract_last_codeblock(text).unwrap();
562 assert_eq!(
563 last_block,
564 indoc::indoc! {r#"
565 ## Collaborations
566 Our mission is to create a 4D generative model.
567
568 ## Citation
569
570 If you found Unique3D helpful, please cite our report:
571 ```bibtex
572 @misc{wu2024unique3d,
573 title={Unique3D},
574 }
575 ```
576 "#}
577 );
578 }
579
580 #[test]
581 fn test_extract_editable_region_no_markers() {
582 let text = indoc::indoc! {"
583 one
584 two three"};
585 let parsed = TeacherPrompt::extract_editable_region(text).unwrap();
586 assert_eq!(
587 parsed,
588 indoc::indoc! {"
589 one
590 two three"}
591 );
592 }
593
594 #[test]
595 fn test_parse_no_edits_response() {
596 let response = indoc::indoc! {"
597 The code is already complete. There is no clear next edit to make.
598
599 `````
600 NO_EDITS
601 `````
602 "};
603 let codeblock = extract_last_codeblock(response).unwrap();
604 assert_eq!(codeblock.trim(), TeacherPrompt::NO_EDITS);
605 }
606
607 #[test]
608 fn test_extract_codeblock_no_valid_block() {
609 // Text with no code blocks should return None
610 let text = "Just some plain text without any code blocks";
611 assert!(extract_last_codeblock(text).is_none());
612
613 // Unclosed code block should return None
614 let text = indoc::indoc! {"
615 ```
616 unclosed block
617 "};
618 assert!(extract_last_codeblock(text).is_none());
619
620 // Analysis text with nested markdown but no proper outer block
621 let text = indoc::indoc! {"
622 # Analysis
623 Looking at this:
624 ```
625 some code
626 ```
627 But then more analysis without wrapping block
628 "};
629 // This should find the inner block
630 let result = extract_last_codeblock(text).unwrap();
631 assert_eq!(result, "some code\n");
632 }
633
634 #[test]
635 fn test_extract_codeblock_no_trailing_newline() {
636 // Text ending without trailing newline after closing fence
637 let text = "`````\ncontent here\n`````";
638 let result = extract_last_codeblock(text).unwrap();
639 assert_eq!(result, "content here\n");
640 }
641}