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::udiff;
10use gpui::AsyncApp;
11use similar::DiffableStr;
12use std::ops::Range;
13use std::sync::Arc;
14use zeta_prompt::{
15 ZetaFormat, excerpt_range_for_format, format_zeta_prompt, resolve_cursor_region,
16};
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 match args.provider {
35 PredictionProvider::Teacher(_) | PredictionProvider::TeacherNonBatching(_) => {
36 step_progress.set_substatus("formatting teacher prompt");
37
38 let zeta_format = ZetaFormat::default();
39 let excerpt_ranges = prompt_inputs
40 .excerpt_ranges
41 .as_ref()
42 .context("prompt_inputs must have excerpt_ranges")?;
43 let (editable_range, context_range) =
44 excerpt_range_for_format(zeta_format, excerpt_ranges);
45
46 let prompt = TeacherPrompt::format_prompt(example, editable_range, context_range);
47 example.prompt = Some(ExamplePrompt {
48 input: prompt,
49 expected_output: String::new(),
50 rejected_output: None,
51 prefill: None,
52 provider: args.provider,
53 });
54 }
55 PredictionProvider::Zeta2(zeta_format) => {
56 step_progress.set_substatus("formatting zeta2 prompt");
57
58 let prompt = format_zeta_prompt(prompt_inputs, zeta_format);
59 let prefill = zeta_prompt::get_prefill(prompt_inputs, zeta_format);
60 let (expected_patch, expected_cursor_offset) = example
61 .spec
62 .expected_patches_with_cursor_positions()
63 .into_iter()
64 .next()
65 .context("expected patches is empty")?;
66 let expected_output = zeta2_output_for_patch(
67 prompt_inputs,
68 &expected_patch,
69 expected_cursor_offset,
70 zeta_format,
71 )?;
72 let rejected_output = example.spec.rejected_patch.as_ref().and_then(|patch| {
73 zeta2_output_for_patch(prompt_inputs, patch, None, zeta_format).ok()
74 });
75
76 example.prompt = Some(ExamplePrompt {
77 input: prompt,
78 expected_output,
79 rejected_output,
80 provider: args.provider,
81 prefill: Some(prefill),
82 });
83 }
84 _ => {
85 panic!("Cannot format prompt for {:?}", args.provider);
86 }
87 };
88 Ok(())
89}
90
91pub fn zeta2_output_for_patch(
92 input: &zeta_prompt::ZetaPromptInput,
93 patch: &str,
94 cursor_offset: Option<usize>,
95 version: ZetaFormat,
96) -> Result<String> {
97 let (context, editable_range, _) = resolve_cursor_region(input, version);
98 let mut old_editable_region = context[editable_range].to_string();
99
100 if !old_editable_region.ends_with_newline() {
101 old_editable_region.push('\n');
102 }
103
104 let (mut result, first_hunk_offset) =
105 udiff::apply_diff_to_string_with_hunk_offset(patch, &old_editable_region).with_context(
106 || {
107 format!(
108 "Patch:\n```\n{}```\n\nEditable region:\n```\n{}```",
109 patch, old_editable_region
110 )
111 },
112 )?;
113
114 if let Some(cursor_offset) = cursor_offset {
115 // The cursor_offset is relative to the start of the hunk's new text (context + additions).
116 // We need to add where the hunk context matched in the editable region to compute
117 // the actual cursor position in the result.
118 let hunk_start = first_hunk_offset.unwrap_or(0);
119 let offset = result.floor_char_boundary((hunk_start + cursor_offset).min(result.len()));
120 result.insert_str(offset, zeta_prompt::CURSOR_MARKER);
121 }
122
123 match version {
124 ZetaFormat::V0120GitMergeMarkers
125 | ZetaFormat::V0131GitMergeMarkersPrefix
126 | ZetaFormat::V0211SeedCoder => {
127 if !result.ends_with('\n') {
128 result.push('\n');
129 }
130 result.push_str(zeta_prompt::v0120_git_merge_markers::END_MARKER);
131 }
132 _ => (),
133 }
134
135 Ok(result)
136}
137
138pub struct TeacherPrompt;
139
140impl TeacherPrompt {
141 pub(crate) const EDITABLE_REGION_START: &str = "<|editable_region_start|>\n";
142 pub(crate) const EDITABLE_REGION_END: &str = "\n<|editable_region_end|>";
143 pub(crate) const USER_CURSOR_MARKER: &str = "<|user_cursor|>";
144 pub(crate) const NO_EDITS: &str = "NO_EDITS";
145
146 /// Truncate edit history to this number of last lines
147 const MAX_HISTORY_LINES: usize = 128;
148
149 pub fn format_prompt(
150 example: &Example,
151 editable_range: Range<usize>,
152 context_range: Range<usize>,
153 ) -> String {
154 let edit_history = Self::format_edit_history(&example.spec.edit_history);
155 let context = Self::format_context(example);
156 let cursor_excerpt = Self::format_cursor_excerpt(example, editable_range, context_range);
157
158 let prompt_template = crate::prompt_assets::get_prompt("teacher.md");
159 let prompt = prompt_template
160 .replace("{{context}}", &context)
161 .replace("{{edit_history}}", &edit_history)
162 .replace("{{cursor_excerpt}}", &cursor_excerpt);
163
164 prompt
165 }
166
167 pub fn parse(example: &Example, response: &str) -> Result<(String, Option<ActualCursor>)> {
168 // Check if the model indicated no edits are needed
169 let no_edits = (String::new(), None);
170 if let Some(last_codeblock) = extract_last_codeblock(&response) {
171 if last_codeblock.trim() == Self::NO_EDITS {
172 return Ok(no_edits);
173 }
174 }
175
176 if response.trim().ends_with(Self::NO_EDITS) {
177 return Ok(no_edits);
178 }
179
180 // Extract updated (new) editable region from the model response.
181 let new_editable_region = Self::extract_editable_region(&response)?;
182 let cursor_offset = new_editable_region.find(Self::USER_CURSOR_MARKER);
183 let mut new_editable_region = new_editable_region.replace(Self::USER_CURSOR_MARKER, "");
184 let old_editable_region = Self::extract_editable_region(
185 &example
186 .prompt
187 .as_ref()
188 .context("example prompt missing")?
189 .input,
190 )?
191 .replace(Self::USER_CURSOR_MARKER, "");
192
193 let prompt_inputs = example
194 .prompt_inputs
195 .as_ref()
196 .context("example is missing prompt inputs")?;
197
198 // Normalize leading newlines: if old starts with newline but new doesn't,
199 // prepend newline to new to preserve whitespace structure.
200 // This handles the case where the model drops the leading blank line.
201 if old_editable_region.starts_with('\n') && !new_editable_region.starts_with('\n') {
202 new_editable_region.insert(0, '\n');
203 }
204
205 let excerpt = prompt_inputs.cursor_excerpt.as_ref();
206 let (editable_region_offset, _) = excerpt
207 .match_indices(&old_editable_region)
208 .min_by_key(|(index, _)| index.abs_diff(prompt_inputs.cursor_offset_in_excerpt))
209 .context("editable region not found in prompt content")?;
210 let editable_region_start_line = excerpt[..editable_region_offset].matches('\n').count();
211
212 // Use full context so cursor offset (relative to editable region start) aligns with diff content
213 let editable_region_lines = old_editable_region.lines().count() as u32;
214 let diff = language::unified_diff_with_context(
215 &old_editable_region,
216 &new_editable_region,
217 editable_region_start_line as u32,
218 editable_region_start_line as u32,
219 editable_region_lines,
220 );
221
222 let diff = indoc::formatdoc! {"
223 --- a/{path}
224 +++ b/{path}
225 {diff}",
226 path = example.spec.cursor_path.to_string_lossy(),
227 diff = diff,
228 };
229
230 let actual_cursor = cursor_offset.map(|editable_region_cursor_offset| {
231 ActualCursor::from_editable_region(
232 &example.spec.cursor_path,
233 editable_region_cursor_offset,
234 &new_editable_region,
235 excerpt,
236 editable_region_offset,
237 editable_region_start_line,
238 )
239 });
240
241 Ok((diff, actual_cursor))
242 }
243
244 fn format_edit_history(edit_history: &str) -> String {
245 let lines: Vec<&str> = edit_history.lines().collect();
246
247 if lines.is_empty() {
248 return "(No edit history)".to_string();
249 }
250
251 if lines.len() > Self::MAX_HISTORY_LINES {
252 let truncated = lines[lines.len() - Self::MAX_HISTORY_LINES..].join("\n");
253 format!("{truncated}\n[...truncated...]")
254 } else {
255 lines.join("\n")
256 }
257 }
258
259 pub fn format_context(example: &Example) -> String {
260 let related_files = example.prompt_inputs.as_ref().map(|pi| &pi.related_files);
261 let Some(related_files) = related_files else {
262 return "(No context)".to_string();
263 };
264
265 if related_files.is_empty() {
266 return "(No context)".to_string();
267 }
268
269 let prefix = "`````";
270 let suffix = "`````\n\n";
271 let max_tokens = 1024;
272 zeta_prompt::format_related_files_within_budget(related_files, &prefix, &suffix, max_tokens)
273 }
274
275 fn format_cursor_excerpt(
276 example: &Example,
277 editable_range: Range<usize>,
278 context_range: Range<usize>,
279 ) -> String {
280 let mut result = String::new();
281
282 let prompt_inputs = example.prompt_inputs.as_ref().unwrap();
283 let excerpt = prompt_inputs.cursor_excerpt.as_ref();
284 let cursor_offset = prompt_inputs.cursor_offset_in_excerpt;
285
286 let path_str = example.spec.cursor_path.to_string_lossy();
287 result.push_str(&format!("`````{path_str}\n"));
288 result.push_str(&excerpt[context_range.start..editable_range.start]);
289 result.push_str(Self::EDITABLE_REGION_START);
290 result.push_str(&excerpt[editable_range.start..cursor_offset]);
291 result.push_str(Self::USER_CURSOR_MARKER);
292 result.push_str(&excerpt[cursor_offset..editable_range.end]);
293 result.push_str(Self::EDITABLE_REGION_END);
294 result.push_str(&excerpt[editable_range.end..context_range.end]);
295 result.push_str("\n`````");
296
297 result
298 }
299
300 pub fn extract_editable_region(text: &str) -> Result<String> {
301 let start = text
302 .rfind(Self::EDITABLE_REGION_START)
303 .map_or(0, |pos| pos + Self::EDITABLE_REGION_START.len());
304 let end = text.rfind(Self::EDITABLE_REGION_END).unwrap_or(text.len());
305
306 if start >= end {
307 return Err(anyhow!("Invalid editable region markers"));
308 }
309
310 let region = &text[start..end];
311 Ok(region.strip_suffix('\n').unwrap_or(region).to_string())
312 }
313}
314
315/// Extract the cursor excerpt from an example.
316/// First tries to extract from an existing prompt, then falls back to constructing from prompt_inputs.
317pub fn extract_cursor_excerpt_from_example(example: &Example) -> Option<String> {
318 // If we have the original prompt, extract the cursor excerpt from it
319 if let Some(prompt) = &example.prompt {
320 // Find "# 3. Current File" section and extract the content
321 if let Some(start) = prompt.input.find("# 3. Current File") {
322 let content_start = prompt.input[start..].find('`').map(|i| start + i)?;
323 let backtick_count = prompt.input[content_start..]
324 .chars()
325 .take_while(|&c| c == '`')
326 .count();
327 let content_start = content_start + backtick_count;
328
329 // Find the path line and skip it
330 let newline_pos = prompt.input[content_start..].find('\n')?;
331 let text_start = content_start + newline_pos + 1;
332
333 // Find the closing backticks
334 let closing_pattern = "`".repeat(backtick_count);
335 let text_end = prompt.input[text_start..].find(&closing_pattern)?;
336 let cursor_excerpt = &prompt.input[text_start..text_start + text_end];
337
338 let path_str = example.spec.cursor_path.to_string_lossy();
339 return Some(format!("`````{path_str}\n{cursor_excerpt}`````"));
340 }
341 }
342
343 // Fallback: construct from prompt_inputs if available
344 let prompt_inputs = example.prompt_inputs.as_ref()?;
345 let excerpt = prompt_inputs.cursor_excerpt.as_ref();
346 let cursor_offset = prompt_inputs.cursor_offset_in_excerpt;
347
348 // Simple fallback: just show content around cursor with markers
349 let path_str = example.spec.cursor_path.to_string_lossy();
350 let mut result = format!("`````{path_str}\n");
351 result.push_str(TeacherPrompt::EDITABLE_REGION_START);
352 result.push_str(&excerpt[..cursor_offset]);
353 result.push_str(TeacherPrompt::USER_CURSOR_MARKER);
354 result.push_str(&excerpt[cursor_offset..]);
355 result.push_str(TeacherPrompt::EDITABLE_REGION_END);
356 result.push_str("\n`````");
357
358 Some(result)
359}
360
361pub(crate) fn extract_last_codeblock(text: &str) -> Option<String> {
362 let lines: Vec<&str> = text.lines().collect();
363
364 // Search from the end for a closing fence (line containing only backticks, 3+)
365 let mut closing_line_idx = None;
366 let mut backtick_count = 0;
367
368 for i in (0..lines.len()).rev() {
369 let line = lines[i].trim();
370 if line.len() >= 3 && line.chars().all(|c| c == '`') {
371 closing_line_idx = Some(i);
372 backtick_count = line.len();
373 break;
374 }
375 }
376
377 let closing_idx = closing_line_idx?;
378
379 // Search backwards for matching opening fence
380 // Opening fence starts with same backtick count, possibly followed by language/metadata
381 let opening_pattern = "`".repeat(backtick_count);
382
383 for i in (0..closing_idx).rev() {
384 let line = lines[i];
385 if line.starts_with(&opening_pattern) {
386 // Ensure it's exactly the right number of backticks (not more)
387 let rest = &line[backtick_count..];
388 if rest.is_empty() || !rest.starts_with('`') {
389 // Found matching opening fence
390 // Extract content between opening and closing (exclusive)
391 if closing_idx > i + 1 {
392 let content = lines[i + 1..closing_idx].join("\n");
393 // Preserve trailing newline to match previous behavior
394 return Some(format!("{}\n", content));
395 } else {
396 // Empty block
397 return Some(String::new());
398 }
399 }
400 }
401 }
402
403 None
404}
405
406#[cfg(test)]
407mod tests {
408 use super::*;
409
410 #[test]
411 fn test_extract_last_code_block() {
412 let text = indoc::indoc! {"
413 Some thinking
414
415 ```
416 first block
417 ```
418
419 `````path='something' lines=1:2
420 last block
421 `````
422 "};
423 let last_block = extract_last_codeblock(text).unwrap();
424 assert_eq!(last_block, "last block\n");
425 }
426
427 #[test]
428 fn test_extract_codeblock_with_nested_fences() {
429 let text = indoc::indoc! {"
430 `````
431 content with ``` inline
432 and ```python nested
433 more content
434 `````
435 "};
436 let last_block = extract_last_codeblock(text).unwrap();
437 assert_eq!(
438 last_block,
439 "content with ``` inline\nand ```python nested\nmore content\n"
440 );
441 }
442
443 #[test]
444 fn test_extract_codeblock_ignores_inline_backticks() {
445 let text = indoc::indoc! {"
446 `````
447 here is some `code` with inline backticks
448 and here```more```stuff
449 `````
450 "};
451 let last_block = extract_last_codeblock(text).unwrap();
452 assert_eq!(
453 last_block,
454 "here is some `code` with inline backticks\nand here```more```stuff\n"
455 );
456 }
457
458 #[test]
459 fn test_extract_editable_region() {
460 let text = indoc::indoc! {"
461 some lines
462 are
463 here
464 <|editable_region_start|>
465 one
466 two three
467
468 <|editable_region_end|>
469 more
470 lines here
471 "};
472 let parsed = TeacherPrompt::extract_editable_region(text).unwrap();
473 assert_eq!(
474 parsed,
475 indoc::indoc! {"
476 one
477 two three"}
478 );
479 }
480
481 #[test]
482 fn test_extract_last_codeblock_nested_bibtex() {
483 let text = indoc::indoc! {r#"
484 Looking at the edit history, I can see that a Citation section was just added.
485
486 `````
487 ## Collaborations
488 Our mission is to create a 4D generative model.
489
490 ## Citation
491
492 If you found Unique3D helpful, please cite our report:
493 ```bibtex
494 @misc{wu2024unique3d,
495 title={Unique3D},
496 }
497 ```
498 `````
499 "#};
500 let last_block = extract_last_codeblock(text).unwrap();
501 assert_eq!(
502 last_block,
503 indoc::indoc! {r#"
504 ## Collaborations
505 Our mission is to create a 4D generative model.
506
507 ## Citation
508
509 If you found Unique3D helpful, please cite our report:
510 ```bibtex
511 @misc{wu2024unique3d,
512 title={Unique3D},
513 }
514 ```
515 "#}
516 );
517 }
518
519 #[test]
520 fn test_extract_editable_region_no_markers() {
521 let text = indoc::indoc! {"
522 one
523 two three"};
524 let parsed = TeacherPrompt::extract_editable_region(text).unwrap();
525 assert_eq!(
526 parsed,
527 indoc::indoc! {"
528 one
529 two three"}
530 );
531 }
532
533 #[test]
534 fn test_parse_no_edits_response() {
535 let response = indoc::indoc! {"
536 The code is already complete. There is no clear next edit to make.
537
538 `````
539 NO_EDITS
540 `````
541 "};
542 let codeblock = extract_last_codeblock(response).unwrap();
543 assert_eq!(codeblock.trim(), TeacherPrompt::NO_EDITS);
544 }
545
546 #[test]
547 fn test_extract_codeblock_no_valid_block() {
548 // Text with no code blocks should return None
549 let text = "Just some plain text without any code blocks";
550 assert!(extract_last_codeblock(text).is_none());
551
552 // Unclosed code block should return None
553 let text = indoc::indoc! {"
554 ```
555 unclosed block
556 "};
557 assert!(extract_last_codeblock(text).is_none());
558
559 // Analysis text with nested markdown but no proper outer block
560 let text = indoc::indoc! {"
561 # Analysis
562 Looking at this:
563 ```
564 some code
565 ```
566 But then more analysis without wrapping block
567 "};
568 // This should find the inner block
569 let result = extract_last_codeblock(text).unwrap();
570 assert_eq!(result, "some code\n");
571 }
572
573 #[test]
574 fn test_extract_codeblock_no_trailing_newline() {
575 // Text ending without trailing newline after closing fence
576 let text = "`````\ncontent here\n`````";
577 let result = extract_last_codeblock(text).unwrap();
578 assert_eq!(result, "content here\n");
579 }
580}