1//! `ep split-commit` implementation.
2//!
3//! This command generates a single evaluation example JSON object from a
4//! chronologically-ordered unified diff (a "commit").
5//!
6//! TODO: Port Python code to generate chronologically-ordered commits
7use crate::FailedHandling;
8use crate::reorder_patch::{Patch, PatchLine, extract_edits, locate_edited_line};
9
10/// Find the largest valid UTF-8 char boundary at or before `index` in `s`.
11fn floor_char_boundary(s: &str, index: usize) -> usize {
12 if index >= s.len() {
13 s.len()
14 } else if s.is_char_boundary(index) {
15 index
16 } else {
17 // Find the nearest valid character boundary at or before index
18 (0..index)
19 .rev()
20 .find(|&i| s.is_char_boundary(i))
21 .unwrap_or(0)
22 }
23}
24use anyhow::{Context as _, Result};
25use clap::Args;
26use edit_prediction::example_spec::ExampleSpec;
27use rand::Rng;
28use rand::SeedableRng;
29use serde::{Deserialize, Serialize};
30use similar::{DiffTag, TextDiff};
31use std::collections::BTreeSet;
32use std::fs;
33use std::io::{self, Write};
34use std::path::Path;
35use std::path::PathBuf;
36
37/// `ep split-commit` CLI args.
38#[derive(Debug, Args, Clone)]
39pub struct SplitCommitArgs {
40 /// Split point (float 0.0-1.0 for fraction, or integer for index)
41 #[arg(long, short = 's')]
42 pub split_point: Option<String>,
43
44 /// Random seed for reproducibility
45 #[arg(long)]
46 pub seed: Option<u64>,
47
48 /// Pretty-print JSON output
49 #[arg(long, short = 'p')]
50 pub pretty: bool,
51
52 /// Number of samples to generate per commit (samples random split points)
53 #[arg(long, short = 'n')]
54 pub num_samples: Option<usize>,
55}
56
57/// Input format for annotated commits (JSON Lines).
58#[derive(Debug, Clone, Deserialize)]
59#[allow(dead_code)]
60pub struct AnnotatedCommit {
61 /// Repository path (e.g., "repos/zed")
62 pub repo: String,
63 /// Repository URL (e.g., "https://github.com/zed-industries/zed")
64 pub repo_url: String,
65 /// Commit SHA
66 pub commit_sha: String,
67 /// Chronologically reordered commit diff
68 pub reordered_commit: String,
69 /// Original commit diff
70 pub original_commit: String,
71 /// Whether diff stats match between original and reordered
72 pub diff_stats_match: bool,
73}
74
75/// Cursor position in a file.
76#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
77pub struct CursorPosition {
78 pub file: String,
79 pub line: usize,
80 pub column: usize,
81}
82
83impl std::fmt::Display for CursorPosition {
84 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
85 write!(f, "{}:{}:{}", self.file, self.line, self.column)
86 }
87}
88
89/// Represents a split commit with source and target patches.
90#[derive(Debug, Clone)]
91pub struct SplitCommit {
92 pub source_patch: String,
93 pub target_patch: String,
94}
95
96/// Split point specification for evaluation generation.
97#[derive(Debug, Clone)]
98pub enum SplitPoint {
99 /// Fraction of total edits (0.0 to 1.0)
100 Fraction(f64),
101 /// Absolute index
102 Index(usize),
103}
104
105fn parse_split_point(value: &str) -> Option<SplitPoint> {
106 if value.contains('.') {
107 value.parse::<f64>().ok().map(SplitPoint::Fraction)
108 } else {
109 value.parse::<usize>().ok().map(SplitPoint::Index)
110 }
111}
112
113/// Entry point for the `ep split-commit` subcommand.
114///
115/// This runs synchronously and outputs JSON Lines (one output per input line).
116pub fn run_split_commit(
117 args: &SplitCommitArgs,
118 inputs: &[PathBuf],
119 output_path: Option<&PathBuf>,
120 failed: FailedHandling,
121) -> Result<()> {
122 use std::collections::HashSet;
123 use std::io::BufRead;
124
125 let stdin_path = PathBuf::from("-");
126 let inputs = if inputs.is_empty() {
127 std::slice::from_ref(&stdin_path)
128 } else {
129 inputs
130 };
131
132 let split_point = args.split_point.as_deref().and_then(parse_split_point);
133 let mut output_lines = Vec::new();
134
135 for input_path in inputs {
136 let input: Box<dyn BufRead> = if input_path.as_os_str() == "-" {
137 Box::new(io::BufReader::new(io::stdin()))
138 } else {
139 let file = fs::File::open(input_path)
140 .with_context(|| format!("failed to open input file {}", input_path.display()))?;
141 Box::new(io::BufReader::new(file))
142 };
143
144 for (line_num, line_result) in input.lines().enumerate() {
145 let line =
146 line_result.with_context(|| format!("failed to read line {}", line_num + 1))?;
147
148 if line.trim().is_empty() {
149 continue;
150 }
151
152 let annotated: AnnotatedCommit = serde_json::from_str(&line)
153 .with_context(|| format!("failed to parse JSON at line {}", line_num + 1))?;
154
155 // Generate multiple samples if num_samples is set
156 if let Some(num_samples) = args.num_samples {
157 let mut seen_samples: HashSet<String> = HashSet::new();
158 let base_seed = args.seed.unwrap_or_else(|| rand::random());
159
160 for sample_idx in 0..num_samples {
161 let sample_seed = base_seed.wrapping_add(sample_idx as u64);
162
163 let case = match generate_evaluation_example_from_ordered_commit(
164 &annotated.reordered_commit,
165 &annotated.repo_url,
166 &annotated.commit_sha,
167 None, // Use random split point for multi-sample mode
168 Some(sample_seed),
169 Some(sample_idx),
170 ) {
171 Ok(case) => case,
172 Err(e) => {
173 let err_msg = format!(
174 "failed to generate evaluation example for commit {} at line {} (sample {}): {}",
175 annotated.commit_sha,
176 line_num + 1,
177 sample_idx,
178 e
179 );
180 match failed {
181 FailedHandling::Skip | FailedHandling::SkipNoFiles => {
182 eprintln!("{}", err_msg);
183 continue;
184 }
185 FailedHandling::Keep => {
186 anyhow::bail!(err_msg);
187 }
188 }
189 }
190 };
191
192 let json = if args.pretty {
193 serde_json::to_string_pretty(&case)
194 } else {
195 serde_json::to_string(&case)
196 }
197 .context("failed to serialize evaluation case as JSON")?;
198
199 // Only add unique samples (different split points may produce same result)
200 if seen_samples.insert(json.clone()) {
201 output_lines.push(json);
202 }
203 }
204 } else {
205 let case = match generate_evaluation_example_from_ordered_commit(
206 &annotated.reordered_commit,
207 &annotated.repo_url,
208 &annotated.commit_sha,
209 split_point.clone(),
210 args.seed,
211 None,
212 ) {
213 Ok(case) => case,
214 Err(e) => {
215 let err_msg = format!(
216 "failed to generate evaluation example for commit {} at line {}: {}",
217 annotated.commit_sha,
218 line_num + 1,
219 e
220 );
221 match failed {
222 FailedHandling::Skip | FailedHandling::SkipNoFiles => {
223 eprintln!("{}", err_msg);
224 continue;
225 }
226 FailedHandling::Keep => {
227 anyhow::bail!(err_msg);
228 }
229 }
230 }
231 };
232
233 let json = if args.pretty {
234 serde_json::to_string_pretty(&case)
235 } else {
236 serde_json::to_string(&case)
237 }
238 .context("failed to serialize evaluation case as JSON")?;
239
240 output_lines.push(json);
241 }
242 }
243 }
244
245 let output_content = output_lines.join("\n") + if output_lines.is_empty() { "" } else { "\n" };
246
247 if let Some(path) = output_path {
248 fs::write(path, &output_content)
249 .with_context(|| format!("failed to write output to {}", path.display()))?;
250 } else {
251 io::stdout()
252 .write_all(output_content.as_bytes())
253 .context("failed to write to stdout")?;
254 }
255
256 Ok(())
257}
258
259/// Main function to generate an evaluation example from an ordered commit.
260///
261/// # Arguments
262/// * `commit` - Chronologically ordered unified diff of the commit
263/// * `repository_url` - URL of the repository
264/// * `commit_hash` - Hash of the commit
265/// * `split_point` - Point at which the commit will be split (None for random)
266/// * `seed` - Optional seed for randomness
267/// * `sample_num` - Optional sample number for generating unique names
268pub fn generate_evaluation_example_from_ordered_commit(
269 commit: &str,
270 repository_url: &str,
271 commit_hash: &str,
272 split_point: Option<SplitPoint>,
273 seed: Option<u64>,
274 sample_num: Option<usize>,
275) -> Result<ExampleSpec> {
276 let mut rng: Box<dyn rand::RngCore> = match seed {
277 Some(seed) => Box::new(rand::rngs::StdRng::seed_from_u64(seed)),
278 None => Box::new(rand::rngs::ThreadRng::default()),
279 };
280
281 // Parse and normalize the commit
282 let mut patch = Patch::parse_unified_diff(commit);
283
284 // Filter header to only keep lines starting with "//"
285 let header_lines: Vec<&str> = patch
286 .header
287 .lines()
288 .filter(|line| line.starts_with("//"))
289 .collect();
290 patch.header = if header_lines.is_empty() {
291 String::new()
292 } else {
293 header_lines.join("\n") + "\n"
294 };
295 let commit_normalized = patch.to_string();
296
297 // Compute the split point
298 let stats = patch.stats();
299 let num_edits = stats.added + stats.removed;
300
301 anyhow::ensure!(num_edits != 0, "no edits found in commit");
302
303 let split = match split_point {
304 None => rng.random_range(1..=num_edits),
305 Some(SplitPoint::Fraction(f)) => {
306 let v = (f * num_edits as f64).floor() as usize;
307 v.min(num_edits)
308 }
309 Some(SplitPoint::Index(i)) => i.min(num_edits),
310 };
311
312 // Split the commit into source and target patches
313 let (prefix, suffix) = split_ordered_commit(&commit_normalized, split);
314
315 let mut split_commit = SplitCommit {
316 source_patch: prefix,
317 target_patch: suffix,
318 };
319
320 // Imitate human edits
321 let human_edit_seed = rng.random_range(1..=10000u64);
322 let (src_patch, tgt_patch, cursor_opt) = imitate_human_edits(
323 &split_commit.source_patch,
324 &split_commit.target_patch,
325 human_edit_seed,
326 );
327 split_commit.source_patch = src_patch;
328 split_commit.target_patch = tgt_patch;
329
330 // Sample cursor position
331 let cursor = match cursor_opt {
332 Some(c) => c,
333 None => sample_cursor_position(&patch, &split_commit)
334 .context("failed to sample cursor position")?,
335 };
336
337 // Get cursor excerpt
338 let cursor_excerpt = get_cursor_excerpt(
339 &cursor,
340 &split_commit.source_patch,
341 &split_commit.target_patch,
342 )
343 .context("failed to generate cursor excerpt")?;
344
345 // Handle edge case where split_point == 0
346 if split == 0 {
347 split_commit.target_patch = String::new();
348 }
349
350 let repo_name = repository_url
351 .trim_end_matches('/')
352 .rsplit('/')
353 .next()
354 .unwrap_or("unknown");
355 let short_sha = &commit_hash[..commit_hash.len().min(8)];
356 let name = match sample_num {
357 Some(n) => format!("{}-{}-{}", repo_name, short_sha, n),
358 None => format!("{}-{}", repo_name, short_sha),
359 };
360
361 Ok(ExampleSpec {
362 name,
363 repository_url: repository_url.to_string(),
364 revision: format!("{}~1", commit_hash),
365 edit_history: split_commit.source_patch.clone(),
366 // cursor_position: cursor.to_string(),
367 cursor_path: Path::new(&cursor.file).into(),
368 cursor_position: cursor_excerpt,
369 expected_patches: vec![split_commit.target_patch],
370 tags: vec![],
371 reasoning: None,
372 uncommitted_diff: String::new(),
373 rejected_patch: None,
374 captured_prompt_input: None,
375 })
376}
377
378/// Split an ordered commit into source and target commits.
379///
380/// # Arguments
381/// * `commit` - Ordered commit string
382/// * `split_pos` - Position to split the commit (number of edited lines)
383///
384/// # Returns
385/// A tuple of (source_diff, target_diff)
386pub fn split_ordered_commit(commit: &str, split_pos: usize) -> (String, String) {
387 let patch = Patch::parse_unified_diff(commit);
388 let source_edits: BTreeSet<usize> = (0..split_pos).collect();
389 let (source, target) = extract_edits(&patch, &source_edits);
390
391 let mut source_str = source.to_string();
392 let target_str = target.to_string();
393
394 // Strip last group header from the source (lines starting with "//" at the end)
395 let source_lines: Vec<&str> = source_str.lines().collect();
396 let mut end_idx = source_lines.len();
397 for i in (0..source_lines.len()).rev() {
398 if source_lines[i].starts_with("//") {
399 end_idx = i;
400 } else {
401 break;
402 }
403 }
404 if end_idx < source_lines.len() {
405 source_str = source_lines[..end_idx].join("\n");
406 if !source_str.is_empty() {
407 source_str.push('\n');
408 }
409 }
410
411 (source_str, target_str)
412}
413
414/// Tokenize text into words and non-word characters.
415fn tokenize(text: &str) -> Vec<String> {
416 let mut tokens = Vec::new();
417 let mut current = String::new();
418
419 for ch in text.chars() {
420 if ch.is_alphanumeric() {
421 current.push(ch);
422 } else if ch == '_' {
423 // Include underscore with the current word, then flush
424 current.push(ch);
425 if !current.is_empty() {
426 tokens.push(std::mem::take(&mut current));
427 }
428 } else {
429 // Punctuation or whitespace - flush current word first
430 if !current.is_empty() {
431 tokens.push(std::mem::take(&mut current));
432 }
433 // Each punctuation/whitespace is its own token
434 tokens.push(ch.to_string());
435 }
436 }
437
438 if !current.is_empty() {
439 tokens.push(current);
440 }
441
442 tokens
443}
444
445/// Calculate the weight for a split position based on the character at that position.
446///
447/// Higher weights indicate more natural pause points (e.g., after punctuation,
448/// at identifier boundaries). Lower weights indicate less natural points
449/// (e.g., mid-identifier).
450fn position_weight(text: &str, pos: usize) -> u32 {
451 if pos == 0 || pos > text.len() {
452 return 1;
453 }
454
455 let chars: Vec<char> = text.chars().collect();
456 if pos > chars.len() {
457 return 1;
458 }
459
460 // Get the character just before this position (what we just "typed")
461 let prev_char = chars[pos - 1];
462
463 // High weight: natural pause points (end of statement/argument, opening brackets)
464 if matches!(prev_char, ',' | ';' | ':' | '(' | '[' | '{') {
465 return 10;
466 }
467
468 // High weight: closing brackets (finished a group)
469 if matches!(prev_char, ')' | ']' | '}') {
470 return 8;
471 }
472
473 // Medium weight: operators and method chains
474 if matches!(
475 prev_char,
476 '.' | '+' | '-' | '*' | '/' | '=' | '<' | '>' | '&' | '|' | '!'
477 ) {
478 return 5;
479 }
480
481 // Check if we're at the end of an identifier (word char followed by non-word char)
482 let is_prev_word_char = prev_char.is_alphanumeric() || prev_char == '_';
483 let is_next_word_char =
484 pos < chars.len() && (chars[pos].is_alphanumeric() || chars[pos] == '_');
485
486 if is_prev_word_char && !is_next_word_char {
487 // End of identifier - high weight
488 return 8;
489 }
490
491 // Whitespace is a natural pause
492 if prev_char.is_whitespace() {
493 return 6;
494 }
495
496 // Mid-identifier: low weight (rare autocomplete scenarios)
497 if is_prev_word_char && is_next_word_char {
498 return 1;
499 }
500
501 // Default medium-low weight
502 3
503}
504
505/// Select a weighted random index from a list of weights.
506///
507/// Returns an index based on the weights, using the provided seed for
508/// deterministic selection.
509fn weighted_select(weights: &[u32], seed: u64) -> usize {
510 if weights.is_empty() {
511 return 0;
512 }
513
514 let total_weight: u64 = weights.iter().map(|&w| w as u64).sum();
515 if total_weight == 0 {
516 // Fallback to uniform selection if all weights are zero
517 return seed as usize % weights.len();
518 }
519
520 // Use seed to select a value in [0, total_weight)
521 let target = seed % total_weight;
522 let mut cumulative: u64 = 0;
523
524 for (idx, &weight) in weights.iter().enumerate() {
525 cumulative += weight as u64;
526 if target < cumulative {
527 return idx;
528 }
529 }
530
531 // Fallback to last index
532 weights.len() - 1
533}
534
535/// Calculate similarity ratio between two strings (0-100).
536fn fuzzy_ratio(s1: &str, s2: &str) -> u32 {
537 if s1.is_empty() && s2.is_empty() {
538 return 100;
539 }
540 if s1.is_empty() || s2.is_empty() {
541 return 0;
542 }
543
544 let diff = TextDiff::from_chars(s1, s2);
545 let matching: usize = diff
546 .ops()
547 .iter()
548 .filter_map(|op| {
549 if matches!(op.tag(), DiffTag::Equal) {
550 Some(op.new_range().len())
551 } else {
552 None
553 }
554 })
555 .sum();
556
557 let total = s1.len() + s2.len();
558 ((2 * matching * 100) / total) as u32
559}
560
561/// Imitate human edits by introducing partial line edits.
562///
563/// This function simulates how a human might incrementally type code,
564/// rather than making complete line replacements.
565pub fn imitate_human_edits(
566 source_patch: &str,
567 target_patch: &str,
568 seed: u64,
569) -> (String, String, Option<CursorPosition>) {
570 let no_change = (source_patch.to_string(), target_patch.to_string(), None);
571
572 let src_patch = Patch::parse_unified_diff(source_patch);
573 let tgt_patch = Patch::parse_unified_diff(target_patch);
574
575 if tgt_patch.hunks.is_empty() {
576 return no_change;
577 }
578
579 // Try to locate the first edit in target
580 let tgt_edit_loc = match locate_edited_line(&tgt_patch, 0) {
581 Some(loc) => loc,
582 None => return no_change,
583 };
584
585 let tgt_is_addition = matches!(tgt_edit_loc.patch_line, PatchLine::Addition(_));
586 if !tgt_is_addition {
587 return no_change;
588 }
589
590 let tgt_line = match &tgt_edit_loc.patch_line {
591 PatchLine::Addition(s) => s.clone(),
592 _ => return no_change,
593 };
594
595 // Try to locate the last edit in source
596 let src_edit_loc = locate_edited_line(&src_patch, -1);
597
598 // Check if source has ANY edit at the same line as target's first edit
599 // We need to iterate through all edits to check this
600 let src_has_edit_at_target_line = {
601 let mut found = false;
602 let mut idx = 0isize;
603 while let Some(loc) = locate_edited_line(&src_patch, idx) {
604 if loc.filename == tgt_edit_loc.filename
605 && loc.target_line_number == tgt_edit_loc.target_line_number
606 {
607 found = true;
608 break;
609 }
610 idx += 1;
611 }
612 found
613 };
614
615 // Check if this is a replacement (deletion followed by insertion on the same line)
616 // or a pure insertion (no corresponding deletion in source)
617 let is_replacement = src_edit_loc.as_ref().map_or(false, |loc| {
618 matches!(loc.patch_line, PatchLine::Deletion(_))
619 && loc.filename == tgt_edit_loc.filename
620 && loc.target_line_number == tgt_edit_loc.target_line_number
621 });
622
623 // If source has an edit at the same line but it's not a replacement (i.e., it's an addition),
624 // we shouldn't process this as a pure insertion either
625 if !is_replacement && src_has_edit_at_target_line {
626 return no_change;
627 }
628
629 let src_line = if is_replacement {
630 match &src_edit_loc.as_ref().unwrap().patch_line {
631 PatchLine::Deletion(s) => s.clone(),
632 _ => return no_change,
633 }
634 } else {
635 // Pure insertion: source line is empty
636 String::new()
637 };
638
639 // Don't process if source and target are the same
640 if src_line == tgt_line {
641 return no_change;
642 }
643
644 // Tokenize both lines
645 let src_tokens = tokenize(&src_line);
646 let tgt_tokens = tokenize(&tgt_line);
647
648 // Convert to slices for similar
649 let src_refs: Vec<&str> = src_tokens.iter().map(|s| s.as_str()).collect();
650 let tgt_refs: Vec<&str> = tgt_tokens.iter().map(|s| s.as_str()).collect();
651
652 // Use similar to get diff operations
653 let diff = TextDiff::from_slices(&src_refs, &tgt_refs);
654
655 // Build weights for each possible split position
656 let mut position_weights: Vec<u32> = Vec::new();
657
658 // Simulate the edit process to collect weights for all possible split positions
659 {
660 let mut current_text = String::new();
661
662 for op in diff.ops() {
663 match op.tag() {
664 DiffTag::Equal => {
665 for i in op.old_range() {
666 current_text.push_str(&src_tokens[i]);
667 }
668 }
669 DiffTag::Replace => {
670 let ins: String = op.new_range().map(|i| tgt_tokens[i].as_str()).collect();
671 let del: String = op.old_range().map(|i| src_tokens[i].as_str()).collect();
672
673 // For insertion part
674 for ch in ins.chars() {
675 current_text.push(ch);
676 let weight = position_weight(¤t_text, current_text.len());
677 position_weights.push(weight);
678 }
679
680 // For deletion part (we're "untyping" from source)
681 for _ in del.chars() {
682 // Weight deletions lower as they represent removing text
683 position_weights.push(2);
684 }
685 }
686 DiffTag::Insert => {
687 let ins: String = op.new_range().map(|i| tgt_tokens[i].as_str()).collect();
688 for ch in ins.chars() {
689 current_text.push(ch);
690 let weight = position_weight(¤t_text, current_text.len());
691 position_weights.push(weight);
692 }
693 }
694 DiffTag::Delete => {
695 let del: String = op.old_range().map(|i| src_tokens[i].as_str()).collect();
696 for _ in del.chars() {
697 // Weight deletions lower
698 position_weights.push(2);
699 }
700 }
701 }
702 }
703 }
704
705 // Use weighted selection to choose split index
706 if position_weights.is_empty() {
707 return no_change;
708 }
709 let split_index = weighted_select(&position_weights, seed);
710
711 let mut edit_index = 0usize;
712 let mut new_src = String::new();
713 let mut split_found = false;
714 let mut last_old_end = 0usize;
715
716 for op in diff.ops() {
717 match op.tag() {
718 DiffTag::Equal => {
719 for i in op.old_range() {
720 new_src.push_str(&src_tokens[i]);
721 }
722 last_old_end = op.old_range().end;
723 }
724 DiffTag::Replace => {
725 // Handle replace as delete + insert
726 let del: String = op.old_range().map(|i| src_tokens[i].as_str()).collect();
727 let ins: String = op.new_range().map(|i| tgt_tokens[i].as_str()).collect();
728 let repl_len = del.len() + ins.len();
729 if edit_index + repl_len >= split_index {
730 // Split within this replace operation
731 let offset = split_index - edit_index;
732 if offset < ins.len() {
733 let safe_offset = floor_char_boundary(&ins, offset);
734 new_src.push_str(&ins[..safe_offset]);
735 } else {
736 new_src.push_str(&ins);
737 let del_offset = offset - ins.len();
738 let safe_del_offset = floor_char_boundary(&del, del_offset.min(del.len()));
739 new_src.push_str(&del[..safe_del_offset]);
740 }
741 split_found = true;
742 last_old_end = op.old_range().end;
743 break;
744 } else {
745 edit_index += repl_len;
746 new_src.push_str(&ins);
747 last_old_end = op.old_range().end;
748 }
749 }
750 DiffTag::Insert => {
751 let repl: String = op.new_range().map(|i| tgt_tokens[i].as_str()).collect();
752 if edit_index + repl.len() >= split_index {
753 let offset = split_index - edit_index;
754 let safe_offset = floor_char_boundary(&repl, offset);
755 new_src.push_str(&repl[..safe_offset]);
756 split_found = true;
757 break;
758 } else {
759 edit_index += repl.len();
760 new_src.push_str(&repl);
761 }
762 }
763 DiffTag::Delete => {
764 let repl: String = op.old_range().map(|i| src_tokens[i].as_str()).collect();
765 if edit_index + repl.len() >= split_index {
766 let offset = split_index - edit_index;
767 let safe_offset = floor_char_boundary(&repl, offset);
768 new_src.push_str(&repl[..safe_offset]);
769 split_found = true;
770 last_old_end = op.old_range().start + safe_offset.min(op.old_range().len());
771 break;
772 } else {
773 edit_index += repl.len();
774 new_src.push_str(&repl);
775 last_old_end = op.old_range().end;
776 }
777 }
778 }
779 }
780
781 if !split_found {
782 return no_change;
783 }
784
785 // Calculate cursor position
786 let cursor = CursorPosition {
787 file: tgt_edit_loc.filename.clone(),
788 line: if is_replacement {
789 src_edit_loc.as_ref().unwrap().source_line_number
790 } else {
791 tgt_edit_loc.target_line_number
792 },
793 column: new_src.len() + 1,
794 };
795
796 // Add remainder of source if similar enough to target remainder
797 let remainder_src: String = (last_old_end..src_tokens.len())
798 .map(|i| src_tokens[i].as_str())
799 .collect();
800 let remainder_tgt: String = (last_old_end..tgt_tokens.len())
801 .filter_map(|i| tgt_tokens.get(i).map(|s| s.as_str()))
802 .collect();
803
804 let ratio = fuzzy_ratio(&remainder_src, &remainder_tgt);
805 if ratio > 35 {
806 new_src.push_str(&remainder_src);
807 }
808
809 if new_src.trim().is_empty() {
810 return no_change;
811 }
812
813 if new_src == src_line {
814 return no_change;
815 }
816
817 // Build new source patch with the intermediate line
818 let mut new_src_patch = src_patch;
819 if is_replacement {
820 // For replacements, insert after the deletion line
821 let src_loc = src_edit_loc.as_ref().unwrap();
822 if let Some(hunk) = new_src_patch.hunks.get_mut(src_loc.hunk_index) {
823 hunk.lines.insert(
824 src_loc.line_index_within_hunk + 1,
825 PatchLine::Addition(new_src.clone()),
826 );
827 hunk.new_count += 1;
828 }
829 } else {
830 // For pure insertions, insert after the last edit in source patch
831 // This imitates human typing - the intermediate content is what the user is currently typing
832 let last_src_edit = locate_edited_line(&new_src_patch, -1);
833
834 if let Some(src_loc) = last_src_edit {
835 // Insert after the last edit in source
836 if let Some(hunk) = new_src_patch.hunks.get_mut(src_loc.hunk_index) {
837 hunk.lines.insert(
838 src_loc.line_index_within_hunk + 1,
839 PatchLine::Addition(new_src.clone()),
840 );
841 hunk.new_count += 1;
842 }
843 } else {
844 // Source patch is empty or has incompatible hunk structure, create a new hunk based on target
845 if let Some(tgt_hunk) = tgt_patch.hunks.get(tgt_edit_loc.hunk_index) {
846 let mut new_hunk = tgt_hunk.clone();
847 // Replace the full addition with the partial one
848 new_hunk.lines.clear();
849 for (i, line) in tgt_hunk.lines.iter().enumerate() {
850 if i == tgt_edit_loc.line_index_within_hunk {
851 new_hunk.lines.push(PatchLine::Addition(new_src.clone()));
852 } else {
853 match line {
854 PatchLine::Addition(_) => {
855 // Skip other additions from target
856 }
857 _ => new_hunk.lines.push(line.clone()),
858 }
859 }
860 }
861 new_hunk.new_count = new_hunk.old_count + 1;
862 new_src_patch.hunks.push(new_hunk);
863 // Copy header from target if source doesn't have one
864 if new_src_patch.header.is_empty() {
865 new_src_patch.header = tgt_patch.header.clone();
866 }
867 }
868 }
869 }
870
871 // Build new target patch with the intermediate line as deletion
872 let mut new_tgt_patch = tgt_patch;
873 if let Some(hunk) = new_tgt_patch.hunks.get_mut(tgt_edit_loc.hunk_index) {
874 hunk.lines.insert(
875 tgt_edit_loc.line_index_within_hunk,
876 PatchLine::Deletion(new_src),
877 );
878 hunk.old_count += 1;
879 }
880
881 (
882 new_src_patch.to_string(),
883 new_tgt_patch.to_string(),
884 Some(cursor),
885 )
886}
887
888/// Locate the end of the last edit in a patch.
889fn locate_end_of_last_edit(patch: &Patch) -> Option<CursorPosition> {
890 let loc = locate_edited_line(patch, -1)?;
891
892 let (line, col) = match &loc.patch_line {
893 PatchLine::Addition(content) => (loc.target_line_number, content.len()),
894 PatchLine::Deletion(_) => (loc.target_line_number, 1),
895 _ => return None,
896 };
897
898 Some(CursorPosition {
899 file: loc.filename,
900 line,
901 column: col,
902 })
903}
904
905/// Locate the beginning of the first edit in a patch.
906fn locate_beginning_of_first_edit(patch: &Patch) -> Option<CursorPosition> {
907 let loc = locate_edited_line(patch, 0)?;
908
909 let hunk = patch.hunks.get(loc.hunk_index)?;
910 let column = if loc.line_index_within_hunk > 0 {
911 if let Some(prev_line) = hunk.lines.get(loc.line_index_within_hunk - 1) {
912 let content = match prev_line {
913 PatchLine::Context(s) | PatchLine::Addition(s) | PatchLine::Deletion(s) => s,
914 _ => return None,
915 };
916 content.len().max(1) - 1
917 } else {
918 0
919 }
920 } else {
921 0
922 };
923
924 let line = loc.target_line_number.saturating_sub(1).max(1);
925
926 Some(CursorPosition {
927 file: loc.filename,
928 line,
929 column,
930 })
931}
932
933/// Sample cursor position according to the following rules:
934/// 1. 50% chance of cursor being at the end of the source patch
935/// 2. 50% chance of cursor being at the beginning of the target patch
936pub fn sample_cursor_position(patch: &Patch, split_commit: &SplitCommit) -> Option<CursorPosition> {
937 // Try end of history first
938 let src_patch = Patch::parse_unified_diff(&split_commit.source_patch);
939 if let Some(cursor) = locate_end_of_last_edit(&src_patch) {
940 return Some(cursor);
941 }
942
943 // Try beginning of target
944 let tgt_patch = Patch::parse_unified_diff(&split_commit.target_patch);
945 if let Some(cursor) = locate_beginning_of_first_edit(&tgt_patch) {
946 return Some(cursor);
947 }
948
949 // Fallback: use the original patch
950 locate_end_of_last_edit(patch)
951}
952
953/// Get cursor excerpt from the patches.
954///
955/// This extracts the lines around the cursor position with a cursor marker.
956pub fn get_cursor_excerpt(
957 cursor: &CursorPosition,
958 source_patch: &str,
959 target_patch: &str,
960) -> Option<String> {
961 let mut excerpt_lines: Vec<String> = Vec::new();
962 let mut excerpt_first_line: usize = 0;
963
964 // Search in the last hunk of source patch
965 let src = Patch::parse_unified_diff(source_patch);
966 if let Some(loc) = locate_edited_line(&src, -1) {
967 if loc.filename == cursor.file && loc.target_line_number == cursor.line {
968 if let Some(hunk) = src.hunks.get(loc.hunk_index) {
969 excerpt_first_line = hunk.new_start as usize;
970 for line in &hunk.lines {
971 match line {
972 PatchLine::Addition(s) | PatchLine::Context(s) => {
973 excerpt_lines.push(s.clone());
974 }
975 _ => {}
976 }
977 }
978 // If hunk only has deletions (file deletion), include deletion lines
979 if excerpt_lines.is_empty() {
980 excerpt_first_line = hunk.old_start as usize;
981 for line in &hunk.lines {
982 match line {
983 PatchLine::Deletion(s) => {
984 excerpt_lines.push(s.clone());
985 }
986 _ => {}
987 }
988 }
989 }
990 }
991 }
992 }
993
994 // Search in target patch if not found
995 if excerpt_lines.is_empty() {
996 let tgt = Patch::parse_unified_diff(target_patch);
997 // Search all hunks for the cursor file, not just the first edit's hunk
998 for hunk in &tgt.hunks {
999 if hunk.filename == cursor.file {
1000 excerpt_first_line = hunk.new_start as usize;
1001 // First try to collect deletions and context (what exists before edits)
1002 for line in &hunk.lines {
1003 match line {
1004 PatchLine::Deletion(s) | PatchLine::Context(s) => {
1005 excerpt_lines.push(s.clone());
1006 }
1007 _ => {}
1008 }
1009 }
1010 // If hunk only has additions (no deletions/context), include all lines
1011 // This handles cases like adding to an empty file or section
1012 if excerpt_lines.is_empty() {
1013 for line in &hunk.lines {
1014 match line {
1015 PatchLine::Addition(s)
1016 | PatchLine::Deletion(s)
1017 | PatchLine::Context(s) => {
1018 excerpt_lines.push(s.clone());
1019 }
1020 _ => {}
1021 }
1022 }
1023 }
1024 if !excerpt_lines.is_empty() {
1025 break;
1026 }
1027 }
1028 }
1029 }
1030
1031 // Also search source patch hunks if still not found (for fallback cursor case)
1032 if excerpt_lines.is_empty() {
1033 for hunk in &src.hunks {
1034 if hunk.filename == cursor.file {
1035 excerpt_first_line = hunk.new_start as usize;
1036 for line in &hunk.lines {
1037 match line {
1038 PatchLine::Addition(s) | PatchLine::Context(s) => {
1039 excerpt_lines.push(s.clone());
1040 }
1041 _ => {}
1042 }
1043 }
1044 // If hunk only has deletions, include deletion lines
1045 if excerpt_lines.is_empty() {
1046 excerpt_first_line = hunk.old_start as usize;
1047 for line in &hunk.lines {
1048 match line {
1049 PatchLine::Deletion(s) => {
1050 excerpt_lines.push(s.clone());
1051 }
1052 _ => {}
1053 }
1054 }
1055 }
1056 if !excerpt_lines.is_empty() {
1057 break;
1058 }
1059 }
1060 }
1061 }
1062
1063 if excerpt_lines.is_empty() {
1064 return None;
1065 }
1066
1067 // Add cursor marker
1068 for (i, line) in excerpt_lines.iter_mut().enumerate() {
1069 let line_num = excerpt_first_line + i;
1070 if line_num == cursor.line {
1071 let col = cursor.column.min(line.len());
1072 // Ensure we split at a valid UTF-8 character boundary
1073 let col = if line.is_char_boundary(col) {
1074 col
1075 } else {
1076 // Find the nearest valid character boundary
1077 (0..=col)
1078 .rev()
1079 .find(|&i| line.is_char_boundary(i))
1080 .unwrap_or(0)
1081 };
1082 let (before, after) = line.split_at(col);
1083 *line = format!("{}<|user_cursor|>{}", before, after);
1084 break;
1085 }
1086 }
1087
1088 Some(excerpt_lines.join("\n"))
1089}
1090
1091#[cfg(test)]
1092mod tests {
1093 use std::path::Path;
1094
1095 use edit_prediction::example_spec::ExampleSpec;
1096
1097 use super::*;
1098
1099 #[test]
1100 fn test_tokenize() {
1101 let tokens = tokenize("hello world");
1102 assert_eq!(tokens, vec!["hello", " ", "world"]);
1103
1104 let tokens = tokenize("foo_bar123 + baz");
1105 assert_eq!(tokens, vec!["foo_", "bar123", " ", "+", " ", "baz"]);
1106
1107 let tokens = tokenize("print(\"hello\")");
1108 assert_eq!(tokens, vec!["print", "(", "\"", "hello", "\"", ")"]);
1109
1110 let tokens = tokenize("hello_world");
1111 assert_eq!(tokens, vec!["hello_", "world"]);
1112
1113 let tokens = tokenize("fn();");
1114 assert_eq!(tokens, vec!["fn", "(", ")", ";"]);
1115 }
1116
1117 #[test]
1118 fn test_fuzzy_ratio() {
1119 assert_eq!(fuzzy_ratio("hello", "hello"), 100);
1120 assert_eq!(fuzzy_ratio("", ""), 100);
1121 assert!(fuzzy_ratio("hello", "world") < 50);
1122 assert!(fuzzy_ratio("hello world", "hello worl") > 80);
1123 }
1124
1125 #[test]
1126 fn test_split_ordered_commit() {
1127 let commit = r#"// First change
1128--- a/test.rs
1129+++ b/test.rs
1130@@ -1,3 +1,4 @@
1131 fn main() {
1132+ println!("hello");
1133+ println!("world");
1134 }
1135"#;
1136 let patch = Patch::parse_unified_diff(commit);
1137 let stats = patch.stats();
1138 assert_eq!(stats.added, 2);
1139
1140 let (source, target) = split_ordered_commit(commit, 1);
1141
1142 // Source should have 1 addition
1143 let src_patch = Patch::parse_unified_diff(&source);
1144 assert_eq!(src_patch.stats().added, 1);
1145
1146 // Target should have 1 addition
1147 let tgt_patch = Patch::parse_unified_diff(&target);
1148 assert_eq!(tgt_patch.stats().added, 1);
1149 }
1150
1151 #[test]
1152 fn test_split_ordered_commit_with_deletions() {
1153 let commit = r#"// Change
1154--- a/test.rs
1155+++ b/test.rs
1156@@ -1,3 +1,3 @@
1157 fn main() {
1158- println!("old");
1159+ println!("new");
1160 }
1161"#;
1162 let patch = Patch::parse_unified_diff(commit);
1163 let stats = patch.stats();
1164 assert_eq!(stats.added, 1);
1165 assert_eq!(stats.removed, 1);
1166
1167 // Split at position 1 (after the deletion)
1168 let (source, target) = split_ordered_commit(commit, 1);
1169
1170 let src_patch = Patch::parse_unified_diff(&source);
1171 let tgt_patch = Patch::parse_unified_diff(&target);
1172
1173 // Source should have the deletion
1174 assert_eq!(src_patch.stats().removed, 1);
1175 // Target should have the addition
1176 assert_eq!(tgt_patch.stats().added, 1);
1177 }
1178
1179 #[test]
1180 fn test_generate_evaluation_example() {
1181 let commit = r#"commit abc123
1182Author: Test <test@example.com>
1183Date: Mon Jan 1 00:00:00 2024
1184
1185 Test commit
1186
1187////////////////////////////////////////////////////////////////////////////////
1188// Add greeting
1189////////////////////////////////////////////////////////////////////////////////
1190--- a/test.rs
1191+++ b/test.rs
1192@@ -1,3 +1,5 @@
1193 fn main() {
1194+ println!("hello");
1195+ println!("world");
1196 }
1197"#;
1198
1199 let result = generate_evaluation_example_from_ordered_commit(
1200 commit,
1201 "https://github.com/test/repo",
1202 "abc123",
1203 Some(SplitPoint::Fraction(0.5)),
1204 Some(42),
1205 None,
1206 );
1207
1208 assert!(result.is_ok());
1209 let case = result.unwrap();
1210 assert_eq!(case.repository_url, "https://github.com/test/repo");
1211 assert_eq!(case.revision, "abc123~1");
1212 assert!(!case.edit_history.is_empty());
1213 }
1214
1215 #[test]
1216 fn test_generate_evaluation_example_reproducible() {
1217 let commit = r#"////////////////////////////////////////////////////////////////////////////////
1218// Add greeting
1219////////////////////////////////////////////////////////////////////////////////
1220--- a/test.rs
1221+++ b/test.rs
1222@@ -1,3 +1,5 @@
1223 fn main() {
1224+ println!("hello");
1225+ println!("world");
1226 }
1227"#;
1228
1229 // Run twice with the same seed
1230 let result1 = generate_evaluation_example_from_ordered_commit(
1231 commit,
1232 "https://github.com/test/repo",
1233 "abc123",
1234 Some(SplitPoint::Fraction(0.5)),
1235 Some(12345),
1236 None,
1237 )
1238 .unwrap();
1239
1240 let result2 = generate_evaluation_example_from_ordered_commit(
1241 commit,
1242 "https://github.com/test/repo",
1243 "abc123",
1244 Some(SplitPoint::Fraction(0.5)),
1245 Some(12345),
1246 None,
1247 )
1248 .unwrap();
1249
1250 // Results should be identical
1251 assert_eq!(result1.edit_history, result2.edit_history);
1252 assert_eq!(result1.expected_patches, result2.expected_patches);
1253 assert_eq!(result1.cursor_position, result2.cursor_position);
1254 }
1255
1256 #[test]
1257 fn test_cursor_position_display() {
1258 let cursor = CursorPosition {
1259 file: "src/main.rs".to_string(),
1260 line: 42,
1261 column: 10,
1262 };
1263 assert_eq!(cursor.to_string(), "src/main.rs:42:10");
1264 }
1265
1266 #[test]
1267 fn test_imitate_human_edits_no_change_when_no_replacement() {
1268 // Source and target patches that don't form a replacement pattern
1269 let source = r#"--- a/test.rs
1270+++ b/test.rs
1271@@ -1,3 +1,4 @@
1272 fn main() {
1273+ println!("hello");
1274 }
1275"#;
1276 let target = r#"--- a/test.rs
1277+++ b/test.rs
1278@@ -1,3 +1,4 @@
1279 fn main() {
1280+ println!("world");
1281 }
1282"#;
1283
1284 let (new_src, new_tgt, cursor) = imitate_human_edits(source, target, 42);
1285
1286 // Should return unchanged when not a replacement pattern
1287 assert_eq!(new_src, source);
1288 assert_eq!(new_tgt, target);
1289 assert!(cursor.is_none());
1290 }
1291
1292 #[test]
1293 fn test_split_point_fraction() {
1294 let commit = r#"// Change
1295--- a/test.rs
1296+++ b/test.rs
1297@@ -1,5 +1,10 @@
1298 fn main() {
1299+ line1();
1300+ line2();
1301+ line3();
1302+ line4();
1303+ line5();
1304 }
1305"#;
1306
1307 // Split at 20% should give first edit in source
1308 let result = generate_evaluation_example_from_ordered_commit(
1309 commit,
1310 "",
1311 "hash",
1312 Some(SplitPoint::Fraction(0.2)),
1313 Some(1),
1314 None,
1315 );
1316
1317 assert!(result.is_ok());
1318 let case = result.unwrap();
1319
1320 // Source should have some edits
1321 let src_patch = Patch::parse_unified_diff(&case.edit_history);
1322 assert!(src_patch.stats().added > 0);
1323 }
1324
1325 #[test]
1326 fn test_split_point_index() {
1327 let commit = r#"// Change
1328--- a/test.rs
1329+++ b/test.rs
1330@@ -1,5 +1,10 @@
1331 fn main() {
1332+ line1();
1333+ line2();
1334+ line3();
1335+ line4();
1336+ line5();
1337 }
1338"#;
1339
1340 // Split at index 2 should give first 2 edits in source
1341 // With pure insertion handling, source gets 2 original + 1 partial = 3 additions
1342 let result = generate_evaluation_example_from_ordered_commit(
1343 commit,
1344 "",
1345 "hash",
1346 Some(SplitPoint::Index(2)),
1347 Some(1),
1348 None,
1349 );
1350
1351 assert!(result.is_ok());
1352 let case = result.unwrap();
1353
1354 let src_patch = Patch::parse_unified_diff(&case.edit_history);
1355 // Pure insertion adds a partial line, so we expect 3 (2 original + 1 partial)
1356 assert_eq!(src_patch.stats().added, 3);
1357 }
1358
1359 #[test]
1360 fn test_cursor_excerpt_contains_marker() {
1361 let commit = r#"////////////////////////////////////////////////////////////////////////////////
1362// Add code
1363////////////////////////////////////////////////////////////////////////////////
1364--- a/test.rs
1365+++ b/test.rs
1366@@ -1,3 +1,5 @@
1367 fn main() {
1368+ println!("hello");
1369+ println!("world");
1370 }
1371"#;
1372
1373 let result = generate_evaluation_example_from_ordered_commit(
1374 commit,
1375 "",
1376 "hash",
1377 Some(SplitPoint::Fraction(0.5)),
1378 Some(42),
1379 None,
1380 )
1381 .unwrap();
1382
1383 // Cursor excerpt should contain the cursor marker
1384 assert!(
1385 result.cursor_position.contains("<|user_cursor|>"),
1386 "Cursor excerpt should contain marker: {}",
1387 result.cursor_position
1388 );
1389 }
1390
1391 #[test]
1392 fn test_evaluation_case_json_serialization() {
1393 let case = ExampleSpec {
1394 name: "test-abc123".to_string(),
1395 repository_url: "https://github.com/test/repo".to_string(),
1396 revision: "abc123~1".to_string(),
1397 edit_history: "patch1".to_string(),
1398 // cursor_position: "file.rs:10:5".to_string(),
1399 cursor_path: Path::new("file.rs").into(),
1400 cursor_position: "some code<|user_cursor|>".to_string(),
1401 expected_patches: vec!["patch".to_string()],
1402 tags: vec![],
1403 reasoning: None,
1404 uncommitted_diff: String::new(),
1405 rejected_patch: None,
1406 captured_prompt_input: None,
1407 };
1408
1409 let json = serde_json::to_string(&case).unwrap();
1410 let deserialized: ExampleSpec = serde_json::from_str(&json).unwrap();
1411
1412 assert_eq!(case.repository_url, deserialized.repository_url);
1413 assert_eq!(case.revision, deserialized.revision);
1414 assert_eq!(case.cursor_position, deserialized.cursor_position);
1415 }
1416
1417 #[test]
1418 fn test_empty_commit_returns_error() {
1419 let commit = "";
1420
1421 let result = generate_evaluation_example_from_ordered_commit(
1422 commit,
1423 "",
1424 "hash",
1425 Some(SplitPoint::Fraction(0.5)),
1426 Some(1),
1427 None,
1428 );
1429
1430 assert!(result.is_err());
1431 }
1432
1433 #[test]
1434 fn test_header_filtering() {
1435 let commit = r#"commit abc123
1436Author: Test
1437Date: Today
1438
1439 Message
1440
1441diff --git a/test.rs b/test.rs
1442index 123..456 789
1443////////////////////////////////////////////////////////////////////////////////
1444// First group
1445////////////////////////////////////////////////////////////////////////////////
1446--- a/test.rs
1447+++ b/test.rs
1448@@ -1,3 +1,4 @@
1449 fn main() {
1450+ code();
1451 }
1452"#;
1453
1454 let result = generate_evaluation_example_from_ordered_commit(
1455 commit,
1456 "",
1457 "hash",
1458 Some(SplitPoint::Index(1)),
1459 Some(1),
1460 None,
1461 );
1462
1463 assert!(result.is_ok());
1464 let case = result.unwrap();
1465
1466 // The edit history should contain the group header (// lines)
1467 // but not the commit metadata
1468 assert!(!case.edit_history.contains("Author:"));
1469 assert!(!case.edit_history.contains("Date:"));
1470 }
1471
1472 #[test]
1473 fn test_position_weight() {
1474 // High weight positions (natural pause points)
1475 assert_eq!(position_weight("foo(", 4), 10); // After '('
1476 assert_eq!(position_weight("a, b", 2), 10); // After ','
1477 assert_eq!(position_weight("x;", 2), 10); // After ';'
1478 assert_eq!(position_weight("a: b", 2), 10); // After ':'
1479 assert_eq!(position_weight("[", 1), 10); // After '['
1480 assert_eq!(position_weight("{", 1), 10); // After '{'
1481
1482 // High weight for closing brackets
1483 assert_eq!(position_weight("foo)", 4), 8); // After ')'
1484 assert_eq!(position_weight("]", 1), 8); // After ']'
1485 assert_eq!(position_weight("}", 1), 8); // After '}'
1486
1487 // High weight at end of identifier
1488 assert_eq!(position_weight("foo ", 3), 8); // End of 'foo' before space
1489 assert_eq!(position_weight("bar(", 3), 8); // End of 'bar' before '('
1490
1491 // Medium weight for operators
1492 assert_eq!(position_weight("a + b", 3), 5); // After '+'
1493 assert_eq!(position_weight("x.", 2), 5); // After '.'
1494 assert_eq!(position_weight("a=b", 2), 5); // After '='
1495
1496 // Medium weight for whitespace
1497 assert_eq!(position_weight("a ", 2), 6); // After space
1498
1499 // Low weight mid-identifier
1500 assert_eq!(position_weight("foobar", 3), 1); // Mid-identifier 'foo|bar'
1501
1502 // Edge cases
1503 assert_eq!(position_weight("", 0), 1); // Empty string
1504 assert_eq!(position_weight("a", 0), 1); // Position 0
1505 }
1506
1507 #[test]
1508 fn test_weighted_select() {
1509 // Test that weighted selection returns correct indices
1510 let weights = vec![1, 10, 1];
1511
1512 // With total weight 12, seed 0 should select index 0
1513 // seed 0 % 12 = 0, cumulative: 1 at idx 0, so returns 0
1514 assert_eq!(weighted_select(&weights, 0), 0);
1515
1516 // seed 1 % 12 = 1, cumulative: 1 at idx 0 (1 < 1 is false), 11 at idx 1 (1 < 11 is true)
1517 assert_eq!(weighted_select(&weights, 1), 1);
1518
1519 // seed 10 % 12 = 10, cumulative: 1, 11 at idx 1 (10 < 11 is true)
1520 assert_eq!(weighted_select(&weights, 10), 1);
1521
1522 // seed 11 % 12 = 11, cumulative: 1, 11 at idx 1 (11 < 11 is false), 12 at idx 2 (11 < 12 is true)
1523 assert_eq!(weighted_select(&weights, 11), 2);
1524
1525 // Empty weights should return 0
1526 let empty: Vec<u32> = vec![];
1527 assert_eq!(weighted_select(&empty, 42), 0);
1528
1529 // Single weight should always return index 0
1530 let single = vec![10];
1531 assert_eq!(weighted_select(&single, 0), 0);
1532 assert_eq!(weighted_select(&single, 100), 0);
1533 }
1534
1535 #[test]
1536 fn test_weighted_split_prefers_natural_boundaries() {
1537 // Test that with different seeds, weighted selection tends to prefer
1538 // positions after punctuation over mid-identifier positions
1539 let text_with_punctuation = "foo(bar, baz)";
1540 let text_mid_identifier = "foobar";
1541
1542 // Position after '(' should have high weight
1543 let weight_after_paren = position_weight(text_with_punctuation, 4);
1544 // Position after ',' should have high weight
1545 let weight_after_comma = position_weight(text_with_punctuation, 8);
1546 // Position mid-identifier should have low weight
1547 let weight_mid_ident = position_weight(text_mid_identifier, 3);
1548
1549 assert!(
1550 weight_after_paren > weight_mid_ident,
1551 "After '(' ({}) should be weighted higher than mid-identifier ({})",
1552 weight_after_paren,
1553 weight_mid_ident
1554 );
1555 assert!(
1556 weight_after_comma > weight_mid_ident,
1557 "After ',' ({}) should be weighted higher than mid-identifier ({})",
1558 weight_after_comma,
1559 weight_mid_ident
1560 );
1561 }
1562
1563 #[test]
1564 fn test_imitate_human_edits_pure_insertion() {
1565 // Source patch is empty (no edits yet)
1566 // Target patch has a pure insertion (adding a new line)
1567 let source = r#"--- a/test.rs
1568+++ b/test.rs
1569@@ -1,2 +1,2 @@
1570 fn main() {
1571 }
1572"#;
1573 let target = r#"--- a/test.rs
1574+++ b/test.rs
1575@@ -1,2 +1,3 @@
1576 fn main() {
1577+ println!("debug");
1578 }
1579"#;
1580
1581 let (new_src, new_tgt, cursor) = imitate_human_edits(source, target, 42);
1582
1583 // Should have transformed the patches
1584 assert_ne!(
1585 new_src, source,
1586 "Source should be modified for pure insertion"
1587 );
1588 assert_ne!(
1589 new_tgt, target,
1590 "Target should be modified for pure insertion"
1591 );
1592 assert!(cursor.is_some(), "Cursor should be set");
1593
1594 // Source should now have a partial addition
1595 let src_patch = Patch::parse_unified_diff(&new_src);
1596 assert!(
1597 src_patch.stats().added > 0,
1598 "Source should have added lines"
1599 );
1600
1601 // Target should have both a deletion (of partial) and addition (of full)
1602 let tgt_patch = Patch::parse_unified_diff(&new_tgt);
1603 assert!(
1604 tgt_patch.stats().removed > 0,
1605 "Target should have removed lines (partial)"
1606 );
1607 assert!(
1608 tgt_patch.stats().added > 0,
1609 "Target should have added lines (full)"
1610 );
1611
1612 // The cursor should be in test.rs
1613 let cursor = cursor.unwrap();
1614 assert_eq!(cursor.file, "test.rs");
1615 }
1616
1617 #[test]
1618 fn test_imitate_human_edits_pure_insertion_empty_source() {
1619 // Source patch has no hunks at all
1620 let source = "";
1621 let target = r#"--- a/test.rs
1622+++ b/test.rs
1623@@ -1,2 +1,3 @@
1624 fn main() {
1625+ println!("hello");
1626 }
1627"#;
1628
1629 let (new_src, _new_tgt, cursor) = imitate_human_edits(source, target, 123);
1630
1631 // Should have created a source patch with partial insertion
1632 assert!(!new_src.is_empty(), "Source should not be empty");
1633 assert!(cursor.is_some(), "Cursor should be set");
1634
1635 let src_patch = Patch::parse_unified_diff(&new_src);
1636 assert!(
1637 src_patch.stats().added > 0,
1638 "Source should have added lines"
1639 );
1640 }
1641
1642 #[test]
1643 fn test_imitate_human_edits_pure_insertion_intermediate_content() {
1644 // Verify the actual intermediate content is a realistic partial typing state
1645 let source = "";
1646 let target = r#"--- a/test.rs
1647+++ b/test.rs
1648@@ -1,2 +1,3 @@
1649 fn main() {
1650+ println!("hello world");
1651 }
1652"#;
1653
1654 // Test with multiple seeds to see different split points
1655 let mut found_partial = false;
1656 for seed in 1..=50 {
1657 let (new_src, new_tgt, cursor) = imitate_human_edits(source, target, seed);
1658
1659 if cursor.is_some() {
1660 let src_patch = Patch::parse_unified_diff(&new_src);
1661 let tgt_patch = Patch::parse_unified_diff(&new_tgt);
1662
1663 // Find the added line in source
1664 for hunk in &src_patch.hunks {
1665 for line in &hunk.lines {
1666 if let PatchLine::Addition(content) = line {
1667 // The partial line should be a prefix of the full line
1668 let full_line = " println!(\"hello world\");";
1669 if content != full_line && full_line.starts_with(content) {
1670 found_partial = true;
1671
1672 // Verify target has the partial as deletion
1673 let mut has_deletion = false;
1674 for tgt_hunk in &tgt_patch.hunks {
1675 for tgt_line in &tgt_hunk.lines {
1676 if let PatchLine::Deletion(del_content) = tgt_line {
1677 if del_content == content {
1678 has_deletion = true;
1679 }
1680 }
1681 }
1682 }
1683 assert!(
1684 has_deletion,
1685 "Target should have deletion of partial line"
1686 );
1687 }
1688 }
1689 }
1690 }
1691 }
1692 }
1693
1694 assert!(
1695 found_partial,
1696 "At least one seed should produce a partial intermediate state"
1697 );
1698 }
1699
1700 #[test]
1701 fn test_imitate_human_edits_inserts_after_last_source_edit() {
1702 // Regression test: intermediate content should appear after the last edit
1703 // in the source patch, not at the position of the first target edit.
1704 // This ensures the diff output correctly imitates human typing order.
1705 //
1706 // The bug was: when source has edits and target has a pure insertion,
1707 // the intermediate content was inserted at tgt_edit_loc.line_index_within_hunk
1708 // (position of first target edit) instead of after the last source edit.
1709 //
1710 // Source patch has edits at lines 1-4, target has a new edit at line 10
1711 // (different location to avoid the "same line" early return)
1712 let source = r#"--- a/test.py
1713+++ b/test.py
1714@@ -1,4 +1,5 @@
1715+import foo
1716 import bar
1717-import old
1718 import baz
1719+import qux
1720"#;
1721 // Target has a pure insertion at a different line (line 10, not overlapping with source)
1722 let target = r#"--- a/test.py
1723+++ b/test.py
1724@@ -10,3 +10,4 @@
1725 def main():
1726+ print("hello world")
1727 pass
1728"#;
1729
1730 // Use a seed that produces a partial result
1731 let (new_src, _new_tgt, cursor) = imitate_human_edits(source, target, 42);
1732
1733 // The function should produce a modified patch
1734 assert!(cursor.is_some(), "Should produce intermediate state");
1735
1736 let src_patch = Patch::parse_unified_diff(&new_src);
1737 let all_additions: Vec<_> = src_patch
1738 .hunks
1739 .iter()
1740 .flat_map(|h| h.lines.iter())
1741 .filter_map(|l| match l {
1742 PatchLine::Addition(s) => Some(s.as_str()),
1743 _ => None,
1744 })
1745 .collect();
1746
1747 // The intermediate content (partial 'print("hello world")') should be
1748 // the LAST addition, appearing after "+import qux" (the last source edit)
1749 let last_addition = all_additions.last().expect("Should have additions");
1750 assert!(
1751 last_addition.trim_start().starts_with("pr"),
1752 "Intermediate content should be the last addition (partial 'print'), but last was: {:?}",
1753 last_addition
1754 );
1755
1756 // Verify the original source edits are still in order before the intermediate
1757 let foo_pos = all_additions.iter().position(|s| *s == "import foo");
1758 let qux_pos = all_additions.iter().position(|s| *s == "import qux");
1759 let intermediate_pos = all_additions
1760 .iter()
1761 .position(|s| s.trim_start().starts_with("pr"));
1762
1763 assert!(foo_pos.is_some(), "Should have 'import foo'");
1764 assert!(qux_pos.is_some(), "Should have 'import qux'");
1765 assert!(
1766 intermediate_pos.is_some(),
1767 "Should have intermediate content"
1768 );
1769
1770 assert!(
1771 foo_pos < qux_pos && qux_pos < intermediate_pos,
1772 "Order should be: foo < qux < intermediate. Got foo={:?}, qux={:?}, intermediate={:?}",
1773 foo_pos,
1774 qux_pos,
1775 intermediate_pos
1776 );
1777 }
1778
1779 #[test]
1780 fn test_cursor_excerpt_with_multibyte_utf8() {
1781 // Test that cursor excerpt handles multi-byte UTF-8 characters correctly
1782 // The Chinese character '第' is 3 bytes (0..3)
1783 let cursor = CursorPosition {
1784 file: "test.md".to_string(),
1785 line: 1,
1786 column: 1, // Byte index 1 is inside '第' (bytes 0..3)
1787 };
1788
1789 let source_patch = r#"--- a/test.md
1790+++ b/test.md
1791@@ -1,1 +1,1 @@
1792+第 14 章 Flask 工作原理与机制解析**
1793"#;
1794
1795 let target_patch = "";
1796
1797 // This should not panic even though column=1 is not a char boundary
1798 let result = get_cursor_excerpt(&cursor, source_patch, target_patch);
1799
1800 // The function should handle the invalid byte index gracefully
1801 if let Some(excerpt) = result {
1802 assert!(
1803 excerpt.contains("<|user_cursor|>"),
1804 "Cursor excerpt should contain marker"
1805 );
1806 // The marker should be placed at a valid character boundary
1807 // (either at the start or after '第')
1808 }
1809 }
1810}