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 = (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 // Extract updated (new) editable region from the model response.
214 // The model may include editable region markers in its output, so we need to strip them.
215 let new_editable_region = extract_last_codeblock(response);
216
217 // Check if the model indicated no edits are needed
218 if new_editable_region.trim() == Self::NO_EDITS {
219 return Ok((String::new(), None));
220 }
221
222 let new_editable_region = Self::extract_editable_region(&new_editable_region)?;
223 let cursor_offset = new_editable_region.find(Self::USER_CURSOR_MARKER);
224 let mut new_editable_region = new_editable_region.replace(Self::USER_CURSOR_MARKER, "");
225 let old_editable_region = Self::extract_editable_region(
226 &example
227 .prompt
228 .as_ref()
229 .context("example prompt missing")?
230 .input,
231 )?
232 .replace(Self::USER_CURSOR_MARKER, "");
233
234 let prompt_inputs = example
235 .prompt_inputs
236 .as_ref()
237 .context("example is missing prompt inputs")?;
238
239 // Normalize leading newlines: if old starts with newline but new doesn't,
240 // prepend newline to new to preserve whitespace structure.
241 // This handles the case where the model drops the leading blank line.
242 if old_editable_region.starts_with('\n') && !new_editable_region.starts_with('\n') {
243 new_editable_region.insert(0, '\n');
244 }
245
246 let (editable_region_offset, _) = prompt_inputs
247 .content
248 .match_indices(&old_editable_region)
249 .min_by_key(|(index, _)| index.abs_diff(prompt_inputs.cursor_offset))
250 .context("editable region not found in prompt content")?;
251 let editable_region_start_line = prompt_inputs.content[..editable_region_offset]
252 .matches('\n')
253 .count();
254
255 // Use full context so cursor offset (relative to editable region start) aligns with diff content
256 let editable_region_lines = old_editable_region.lines().count() as u32;
257 let diff = language::unified_diff_with_context(
258 &old_editable_region,
259 &new_editable_region,
260 editable_region_start_line as u32,
261 editable_region_start_line as u32,
262 editable_region_lines,
263 );
264
265 let diff = indoc::formatdoc! {"
266 --- a/{path}
267 +++ b/{path}
268 {diff}",
269 path = example.spec.cursor_path.to_string_lossy(),
270 diff = diff,
271 };
272
273 let actual_cursor = cursor_offset.map(|editable_region_cursor_offset| {
274 ActualCursor::from_editable_region(
275 &example.spec.cursor_path,
276 editable_region_cursor_offset,
277 &new_editable_region,
278 &prompt_inputs.content,
279 editable_region_offset,
280 editable_region_start_line,
281 )
282 });
283
284 Ok((diff, actual_cursor))
285 }
286
287 fn format_edit_history(edit_history: &str) -> String {
288 // Strip comments ("garbage lines") from edit history
289 let lines = edit_history
290 .lines()
291 .filter(|&s| Self::is_udiff_content_line(s))
292 .collect::<Vec<_>>();
293
294 let history_lines = if lines.len() > Self::MAX_HISTORY_LINES {
295 &lines[lines.len() - Self::MAX_HISTORY_LINES..]
296 } else {
297 &lines
298 };
299
300 if history_lines.is_empty() {
301 return "(No edit history)".to_string();
302 }
303
304 history_lines.join("\n")
305 }
306
307 pub fn format_context(example: &Example) -> String {
308 let related_files = example
309 .prompt_inputs
310 .as_ref()
311 .and_then(|pi| pi.related_files.as_ref());
312
313 let Some(related_files) = related_files else {
314 return "(No context)".to_string();
315 };
316
317 if related_files.is_empty() {
318 return "(No context)".to_string();
319 }
320
321 let mut prompt = String::new();
322 for file in related_files {
323 let path_str = file.path.to_string_lossy();
324 writeln!(&mut prompt, "`````{path_str}").ok();
325
326 let mut prev_row = 0;
327 for excerpt in &file.excerpts {
328 if excerpt.row_range.start > prev_row {
329 prompt.push_str("…\n");
330 }
331 prompt.push_str(&excerpt.text);
332 prompt.push('\n');
333 prev_row = excerpt.row_range.end;
334 }
335 if prev_row < file.max_row {
336 prompt.push_str("…\n");
337 }
338 prompt.push_str("\n`````\n");
339 }
340
341 prompt
342 }
343
344 fn format_cursor_excerpt(
345 example: &Example,
346 editable_range: Range<usize>,
347 context_range: Range<usize>,
348 ) -> String {
349 let mut result = String::new();
350
351 let prompt_inputs = example.prompt_inputs.as_ref().unwrap();
352
353 let path_str = example.spec.cursor_path.to_string_lossy();
354 result.push_str(&format!("`````{path_str}\n"));
355 result.push_str(&prompt_inputs.content[context_range.start..editable_range.start]);
356 result.push_str(Self::EDITABLE_REGION_START);
357 result.push_str(&prompt_inputs.content[editable_range.start..prompt_inputs.cursor_offset]);
358 result.push_str(Self::USER_CURSOR_MARKER);
359 result.push_str(&prompt_inputs.content[prompt_inputs.cursor_offset..editable_range.end]);
360 result.push_str(Self::EDITABLE_REGION_END);
361 result.push_str(&prompt_inputs.content[editable_range.end..context_range.end]);
362 result.push_str("\n`````");
363
364 result
365 }
366
367 pub fn extract_editable_region(text: &str) -> Result<String> {
368 let start = text
369 .rfind(Self::EDITABLE_REGION_START)
370 .map_or(0, |pos| pos + Self::EDITABLE_REGION_START.len());
371 let end = text.rfind(Self::EDITABLE_REGION_END).unwrap_or(text.len());
372
373 if start >= end {
374 return Err(anyhow!("Invalid editable region markers"));
375 }
376
377 let region = &text[start..end];
378 Ok(region.strip_suffix('\n').unwrap_or(region).to_string())
379 }
380
381 fn is_udiff_content_line(s: &str) -> bool {
382 s.starts_with("-")
383 || s.starts_with("+")
384 || s.starts_with(" ")
385 || s.starts_with("---")
386 || s.starts_with("+++")
387 || s.starts_with("@@")
388 }
389}
390
391/// Extract the cursor excerpt from an example.
392/// First tries to extract from an existing prompt, then falls back to constructing from prompt_inputs.
393pub fn extract_cursor_excerpt_from_example(example: &Example) -> Option<String> {
394 // If we have the original prompt, extract the cursor excerpt from it
395 if let Some(prompt) = &example.prompt {
396 // Find "# 3. Current File" section and extract the content
397 if let Some(start) = prompt.input.find("# 3. Current File") {
398 let content_start = prompt.input[start..].find('`').map(|i| start + i)?;
399 let backtick_count = prompt.input[content_start..]
400 .chars()
401 .take_while(|&c| c == '`')
402 .count();
403 let content_start = content_start + backtick_count;
404
405 // Find the path line and skip it
406 let newline_pos = prompt.input[content_start..].find('\n')?;
407 let text_start = content_start + newline_pos + 1;
408
409 // Find the closing backticks
410 let closing_pattern = "`".repeat(backtick_count);
411 let text_end = prompt.input[text_start..].find(&closing_pattern)?;
412 let cursor_excerpt = &prompt.input[text_start..text_start + text_end];
413
414 let path_str = example.spec.cursor_path.to_string_lossy();
415 return Some(format!("`````{path_str}\n{cursor_excerpt}`````"));
416 }
417 }
418
419 // Fallback: construct from prompt_inputs if available
420 let prompt_inputs = example.prompt_inputs.as_ref()?;
421 let content = &prompt_inputs.content;
422 let cursor_offset = prompt_inputs.cursor_offset;
423
424 // Simple fallback: just show content around cursor with markers
425 let path_str = example.spec.cursor_path.to_string_lossy();
426 let mut result = format!("`````{path_str}\n");
427 result.push_str(TeacherPrompt::EDITABLE_REGION_START);
428 result.push_str(&content[..cursor_offset]);
429 result.push_str(TeacherPrompt::USER_CURSOR_MARKER);
430 result.push_str(&content[cursor_offset..]);
431 result.push_str(TeacherPrompt::EDITABLE_REGION_END);
432 result.push_str("\n`````");
433
434 Some(result)
435}
436
437pub(crate) fn extract_last_codeblock(text: &str) -> String {
438 let mut last_block = None;
439 let mut search_start = 0;
440
441 while let Some(start) = text[search_start..].find("```") {
442 let start = start + search_start;
443 let bytes = text.as_bytes();
444 let mut backtick_end = start;
445
446 while backtick_end < bytes.len() && bytes[backtick_end] == b'`' {
447 backtick_end += 1;
448 }
449
450 let backtick_count = backtick_end - start;
451 let closing_pattern = format!("\n{}", "`".repeat(backtick_count));
452
453 while backtick_end < bytes.len() && bytes[backtick_end] != b'\n' {
454 backtick_end += 1;
455 }
456
457 if let Some(end_pos) = text[backtick_end..].find(&closing_pattern) {
458 let code_block = &text[backtick_end + 1..backtick_end + end_pos + 1];
459 last_block = Some(code_block.to_string());
460 search_start = backtick_end + end_pos + closing_pattern.len();
461 } else {
462 break;
463 }
464 }
465
466 last_block.unwrap_or_else(|| text.to_string())
467}
468
469#[cfg(test)]
470mod tests {
471 use super::*;
472
473 #[test]
474 fn test_extract_last_code_block() {
475 let text = indoc::indoc! {"
476 Some thinking
477
478 ```
479 first block
480 ```
481
482 `````path='something' lines=1:2
483 last block
484 `````
485 "};
486 let last_block = extract_last_codeblock(text);
487 assert_eq!(last_block, "last block\n");
488 }
489
490 #[test]
491 fn test_extract_codeblock_with_nested_fences() {
492 let text = indoc::indoc! {"
493 `````
494 content with ``` inline
495 and ```python nested
496 more content
497 `````
498 "};
499 let last_block = extract_last_codeblock(text);
500 assert_eq!(
501 last_block,
502 "content with ``` inline\nand ```python nested\nmore content\n"
503 );
504 }
505
506 #[test]
507 fn test_extract_codeblock_ignores_inline_backticks() {
508 let text = indoc::indoc! {"
509 `````
510 here is some `code` with inline backticks
511 and here```more```stuff
512 `````
513 "};
514 let last_block = extract_last_codeblock(text);
515 assert_eq!(
516 last_block,
517 "here is some `code` with inline backticks\nand here```more```stuff\n"
518 );
519 }
520
521 #[test]
522 fn test_extract_editable_region() {
523 let text = indoc::indoc! {"
524 some lines
525 are
526 here
527 <|editable_region_start|>
528 one
529 two three
530
531 <|editable_region_end|>
532 more
533 lines here
534 "};
535 let parsed = TeacherPrompt::extract_editable_region(text).unwrap();
536 assert_eq!(
537 parsed,
538 indoc::indoc! {"
539 one
540 two three"}
541 );
542 }
543
544 #[test]
545 fn test_extract_last_codeblock_nested_bibtex() {
546 let text = indoc::indoc! {r#"
547 Looking at the edit history, I can see that a Citation section was just added.
548
549 `````
550 ## Collaborations
551 Our mission is to create a 4D generative model.
552
553 ## Citation
554
555 If you found Unique3D helpful, please cite our report:
556 ```bibtex
557 @misc{wu2024unique3d,
558 title={Unique3D},
559 }
560 ```
561 `````
562 "#};
563 let last_block = extract_last_codeblock(text);
564 assert_eq!(
565 last_block,
566 indoc::indoc! {r#"
567 ## Collaborations
568 Our mission is to create a 4D generative model.
569
570 ## Citation
571
572 If you found Unique3D helpful, please cite our report:
573 ```bibtex
574 @misc{wu2024unique3d,
575 title={Unique3D},
576 }
577 ```
578 "#}
579 );
580 }
581
582 #[test]
583 fn test_extract_editable_region_no_markers() {
584 let text = indoc::indoc! {"
585 one
586 two three"};
587 let parsed = TeacherPrompt::extract_editable_region(text).unwrap();
588 assert_eq!(
589 parsed,
590 indoc::indoc! {"
591 one
592 two three"}
593 );
594 }
595
596 #[test]
597 fn test_parse_no_edits_response() {
598 let response = indoc::indoc! {"
599 The code is already complete. There is no clear next edit to make.
600
601 `````
602 NO_EDITS
603 `````
604 "};
605 let codeblock = extract_last_codeblock(response);
606 assert_eq!(codeblock.trim(), TeacherPrompt::NO_EDITS);
607 }
608}