assistant edit tool: Replace with flexible indentation (#27039)

Sometimes the model produces SEARCH queries that don't match the
indentation of the source file exactly.

When we can't find an exact match, we'll now attempt to match the lines
while being more flexible about the leading whitespace as long as all
lines are consistently offset from the source, and extend the leading
whitespace in the REPLACE string accordingly.

Release Notes:

- N/A

Cargo.lock

@@ -728,6 +728,7 @@ dependencies = [
  "settings",
  "theme",
  "ui",
+ "unindent",
  "util",
  "workspace",
  "worktree",

crates/assistant_tools/Cargo.toml

@@ -43,3 +43,4 @@ language = { workspace = true, features = ["test-support"] }
 project = { workspace = true, features = ["test-support"] }
 rand.workspace = true
 workspace = { workspace = true, features = ["test-support"] }
+unindent.workspace = true

crates/assistant_tools/src/edit_files_tool.rs

@@ -1,5 +1,6 @@
 mod edit_action;
 pub mod log;
+mod replace;
 
 use anyhow::{anyhow, Context, Result};
 use assistant_tool::{ActionLog, Tool};
@@ -11,12 +12,12 @@ use language_model::{
     LanguageModelRegistry, LanguageModelRequest, LanguageModelRequestMessage, MessageContent, Role,
 };
 use log::{EditToolLog, EditToolRequestId};
-use project::{search::SearchQuery, Project};
+use project::Project;
+use replace::{replace_exact, replace_with_flexible_indent};
 use schemars::JsonSchema;
 use serde::{Deserialize, Serialize};
 use std::fmt::Write;
 use std::sync::Arc;
-use util::paths::PathMatcher;
 use util::ResultExt;
 
 #[derive(Debug, Serialize, Deserialize, JsonSchema)]
@@ -289,41 +290,18 @@ impl EditToolRequest {
         file_path: std::path::PathBuf,
         snapshot: language::BufferSnapshot,
     ) -> Result<DiffResult> {
-        let query = SearchQuery::text(
-            old.clone(),
-            false,
-            true,
-            true,
-            PathMatcher::new(&[])?,
-            PathMatcher::new(&[])?,
-            None,
-        )?;
+        let result =
+            // Try to match exactly
+            replace_exact(&old, &new, &snapshot)
+            .await
+            // If that fails, try being flexible about indentation
+            .or_else(|| replace_with_flexible_indent(&old, &new, &snapshot));
 
-        let matches = query.search(&snapshot, None).await;
-
-        if matches.is_empty() {
-            return Ok(DiffResult::BadSearch(BadSearch {
-                search: old.clone(),
+        let Some(diff) = result else {
+            return anyhow::Ok(DiffResult::BadSearch(BadSearch {
+                search: old,
                 file_path: file_path.display().to_string(),
             }));
-        }
-
-        let edit_range = matches[0].clone();
-        let diff = language::text_diff(&old, &new);
-
-        let edits = diff
-            .into_iter()
-            .map(|(old_range, text)| {
-                let start = edit_range.start + old_range.start;
-                let end = edit_range.start + old_range.end;
-                (start..end, text)
-            })
-            .collect::<Vec<_>>();
-
-        let diff = language::Diff {
-            base_version: snapshot.version().clone(),
-            line_ending: snapshot.line_ending(),
-            edits,
         };
 
         anyhow::Ok(DiffResult::Diff(diff))

crates/assistant_tools/src/edit_files_tool/replace.rs

@@ -0,0 +1,525 @@
+use language::{BufferSnapshot, Diff, Point, ToOffset};
+use project::search::SearchQuery;
+use util::{paths::PathMatcher, ResultExt as _};
+
+/// Performs an exact string replacement in a buffer, requiring precise character-for-character matching.
+/// Uses the search functionality to locate the first occurrence of the exact string.
+/// Returns None if no exact match is found in the buffer.
+pub async fn replace_exact(old: &str, new: &str, snapshot: &BufferSnapshot) -> Option<Diff> {
+    let query = SearchQuery::text(
+        old,
+        false,
+        true,
+        true,
+        PathMatcher::new(&[]).ok()?,
+        PathMatcher::new(&[]).ok()?,
+        None,
+    )
+    .log_err()?;
+
+    let matches = query.search(&snapshot, None).await;
+
+    if matches.is_empty() {
+        return None;
+    }
+
+    let edit_range = matches[0].clone();
+    let diff = language::text_diff(&old, &new);
+
+    let edits = diff
+        .into_iter()
+        .map(|(old_range, text)| {
+            let start = edit_range.start + old_range.start;
+            let end = edit_range.start + old_range.end;
+            (start..end, text)
+        })
+        .collect::<Vec<_>>();
+
+    let diff = language::Diff {
+        base_version: snapshot.version().clone(),
+        line_ending: snapshot.line_ending(),
+        edits,
+    };
+
+    Some(diff)
+}
+
+/// Performs a replacement that's indentation-aware - matches text content ignoring leading whitespace differences.
+/// When replacing, preserves the indentation level found in the buffer at each matching line.
+/// Returns None if no match found or if indentation is offset inconsistently across matched lines.
+pub fn replace_with_flexible_indent(old: &str, new: &str, buffer: &BufferSnapshot) -> Option<Diff> {
+    let (old_lines, old_min_indent) = lines_with_min_indent(old);
+    let (new_lines, new_min_indent) = lines_with_min_indent(new);
+    let min_indent = old_min_indent.min(new_min_indent);
+
+    let old_lines = drop_lines_prefix(&old_lines, min_indent);
+    let new_lines = drop_lines_prefix(&new_lines, min_indent);
+
+    let max_row = buffer.max_point().row;
+
+    'windows: for start_row in 0..max_row.saturating_sub(old_lines.len() as u32 - 1) {
+        let mut common_leading = None;
+
+        let end_row = start_row + old_lines.len() as u32 - 1;
+
+        if end_row > max_row {
+            // The buffer ends before fully matching the pattern
+            return None;
+        }
+
+        let start_point = Point::new(start_row, 0);
+        let end_point = Point::new(end_row, buffer.line_len(end_row));
+        let range = start_point.to_offset(buffer)..end_point.to_offset(buffer);
+
+        let window_text = buffer.text_for_range(range.clone());
+        let mut window_lines = window_text.lines();
+        let mut old_lines_iter = old_lines.iter();
+
+        while let (Some(window_line), Some(old_line)) = (window_lines.next(), old_lines_iter.next())
+        {
+            let line_trimmed = window_line.trim_start();
+
+            if line_trimmed != old_line.trim_start() {
+                continue 'windows;
+            }
+
+            if line_trimmed.is_empty() {
+                continue;
+            }
+
+            let line_leading = &window_line[..window_line.len() - old_line.len()];
+
+            match &common_leading {
+                Some(common_leading) if common_leading != line_leading => {
+                    continue 'windows;
+                }
+                Some(_) => (),
+                None => common_leading = Some(line_leading.to_string()),
+            }
+        }
+
+        if let Some(common_leading) = common_leading {
+            let line_ending = buffer.line_ending();
+            let replacement = new_lines
+                .iter()
+                .map(|new_line| {
+                    if new_line.trim().is_empty() {
+                        new_line.to_string()
+                    } else {
+                        common_leading.to_string() + new_line
+                    }
+                })
+                .collect::<Vec<_>>()
+                .join(line_ending.as_str());
+
+            let diff = Diff {
+                base_version: buffer.version().clone(),
+                line_ending,
+                edits: vec![(range, replacement.into())],
+            };
+
+            return Some(diff);
+        }
+    }
+
+    None
+}
+
+fn drop_lines_prefix<'a>(lines: &'a [&str], prefix_len: usize) -> Vec<&'a str> {
+    lines
+        .iter()
+        .map(|line| line.get(prefix_len..).unwrap_or(""))
+        .collect()
+}
+
+fn lines_with_min_indent(input: &str) -> (Vec<&str>, usize) {
+    let mut lines = Vec::new();
+    let mut min_indent: Option<usize> = None;
+
+    for line in input.lines() {
+        lines.push(line);
+        if !line.trim().is_empty() {
+            let indent = line.len() - line.trim_start().len();
+            min_indent = Some(min_indent.map_or(indent, |m| m.min(indent)));
+        }
+    }
+
+    (lines, min_indent.unwrap_or(0))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use gpui::prelude::*;
+    use gpui::TestAppContext;
+    use unindent::Unindent;
+
+    #[gpui::test]
+    fn test_replace_consistent_indentation(cx: &mut TestAppContext) {
+        let whole = r#"
+            fn test() {
+                let x = 5;
+                println!("x = {}", x);
+                let y = 10;
+            }
+        "#
+        .unindent();
+
+        let old = r#"
+            let x = 5;
+            println!("x = {}", x);
+        "#
+        .unindent();
+
+        let new = r#"
+            let x = 42;
+            println!("New value: {}", x);
+        "#
+        .unindent();
+
+        let expected = r#"
+            fn test() {
+                let x = 42;
+                println!("New value: {}", x);
+                let y = 10;
+            }
+        "#
+        .unindent();
+
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            Some(expected.to_string())
+        );
+    }
+
+    #[gpui::test]
+    fn test_replace_inconsistent_indentation(cx: &mut TestAppContext) {
+        let whole = r#"
+            fn test() {
+                if condition {
+                    println!("{}", 43);
+                }
+            }
+        "#
+        .unindent();
+
+        let old = r#"
+            if condition {
+            println!("{}", 43);
+        "#
+        .unindent();
+
+        let new = r#"
+            if condition {
+            println!("{}", 42);
+        "#
+        .unindent();
+
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            None
+        );
+    }
+
+    #[gpui::test]
+    fn test_replace_with_empty_lines(cx: &mut TestAppContext) {
+        // Test with empty lines
+        let whole = r#"
+            fn test() {
+                let x = 5;
+
+                println!("x = {}", x);
+            }
+        "#
+        .unindent();
+
+        let old = r#"
+            let x = 5;
+
+            println!("x = {}", x);
+        "#
+        .unindent();
+
+        let new = r#"
+            let x = 10;
+
+            println!("New x: {}", x);
+        "#
+        .unindent();
+
+        let expected = r#"
+            fn test() {
+                let x = 10;
+
+                println!("New x: {}", x);
+            }
+        "#
+        .unindent();
+
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            Some(expected.to_string())
+        );
+    }
+
+    #[gpui::test]
+    fn test_replace_no_match(cx: &mut TestAppContext) {
+        // Test with no match
+        let whole = r#"
+            fn test() {
+                let x = 5;
+            }
+        "#
+        .unindent();
+
+        let old = r#"
+            let y = 10;
+        "#
+        .unindent();
+
+        let new = r#"
+            let y = 20;
+        "#
+        .unindent();
+
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            None
+        );
+    }
+
+    #[gpui::test]
+    fn test_replace_whole_ends_before_matching_old(cx: &mut TestAppContext) {
+        let whole = r#"
+            fn test() {
+                let x = 5;
+        "#
+        .unindent();
+
+        let old = r#"
+            let x = 5;
+            println!("x = {}", x);
+        "#
+        .unindent();
+
+        let new = r#"
+            let x = 10;
+            println!("x = {}", x);
+        "#
+        .unindent();
+
+        // Should return None because whole doesn't fully contain the old text
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            None
+        );
+    }
+
+    #[test]
+    fn test_lines_with_min_indent() {
+        // Empty string
+        assert_eq!(lines_with_min_indent(""), (vec![], 0));
+
+        // Single line without indentation
+        assert_eq!(lines_with_min_indent("hello"), (vec!["hello"], 0));
+
+        // Multiple lines with no indentation
+        assert_eq!(
+            lines_with_min_indent("line1\nline2\nline3"),
+            (vec!["line1", "line2", "line3"], 0)
+        );
+
+        // Multiple lines with consistent indentation
+        assert_eq!(
+            lines_with_min_indent("  line1\n  line2\n  line3"),
+            (vec!["  line1", "  line2", "  line3"], 2)
+        );
+
+        // Multiple lines with varying indentation
+        assert_eq!(
+            lines_with_min_indent("    line1\n  line2\n      line3"),
+            (vec!["    line1", "  line2", "      line3"], 2)
+        );
+
+        // Lines with mixed indentation and empty lines
+        assert_eq!(
+            lines_with_min_indent("    line1\n\n  line2"),
+            (vec!["    line1", "", "  line2"], 2)
+        );
+    }
+
+    #[gpui::test]
+    fn test_replace_with_missing_indent_uneven_match(cx: &mut TestAppContext) {
+        let whole = r#"
+            fn test() {
+                if true {
+                        let x = 5;
+                        println!("x = {}", x);
+                }
+            }
+        "#
+        .unindent();
+
+        let old = r#"
+            let x = 5;
+            println!("x = {}", x);
+        "#
+        .unindent();
+
+        let new = r#"
+            let x = 42;
+            println!("x = {}", x);
+        "#
+        .unindent();
+
+        let expected = r#"
+            fn test() {
+                if true {
+                        let x = 42;
+                        println!("x = {}", x);
+                }
+            }
+        "#
+        .unindent();
+
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            Some(expected.to_string())
+        );
+    }
+
+    #[gpui::test]
+    fn test_replace_big_example(cx: &mut TestAppContext) {
+        let whole = r#"
+            #[cfg(test)]
+            mod tests {
+                use super::*;
+
+                #[test]
+                fn test_is_valid_age() {
+                    assert!(is_valid_age(0));
+                    assert!(!is_valid_age(151));
+                }
+            }
+        "#
+        .unindent();
+
+        let old = r#"
+            #[test]
+            fn test_is_valid_age() {
+                assert!(is_valid_age(0));
+                assert!(!is_valid_age(151));
+            }
+        "#
+        .unindent();
+
+        let new = r#"
+            #[test]
+            fn test_is_valid_age() {
+                assert!(is_valid_age(0));
+                assert!(!is_valid_age(151));
+            }
+
+            #[test]
+            fn test_group_people_by_age() {
+                let people = vec![
+                    Person::new("Young One", 5, "young@example.com").unwrap(),
+                    Person::new("Teen One", 15, "teen@example.com").unwrap(),
+                    Person::new("Teen Two", 18, "teen2@example.com").unwrap(),
+                    Person::new("Adult One", 25, "adult@example.com").unwrap(),
+                ];
+
+                let groups = group_people_by_age(&people);
+
+                assert_eq!(groups.get(&0).unwrap().len(), 1);  // One person in 0-9
+                assert_eq!(groups.get(&10).unwrap().len(), 2); // Two people in 10-19
+                assert_eq!(groups.get(&20).unwrap().len(), 1); // One person in 20-29
+            }
+        "#
+        .unindent();
+        let expected = r#"
+            #[cfg(test)]
+            mod tests {
+                use super::*;
+
+                #[test]
+                fn test_is_valid_age() {
+                    assert!(is_valid_age(0));
+                    assert!(!is_valid_age(151));
+                }
+
+                #[test]
+                fn test_group_people_by_age() {
+                    let people = vec![
+                        Person::new("Young One", 5, "young@example.com").unwrap(),
+                        Person::new("Teen One", 15, "teen@example.com").unwrap(),
+                        Person::new("Teen Two", 18, "teen2@example.com").unwrap(),
+                        Person::new("Adult One", 25, "adult@example.com").unwrap(),
+                    ];
+
+                    let groups = group_people_by_age(&people);
+
+                    assert_eq!(groups.get(&0).unwrap().len(), 1);  // One person in 0-9
+                    assert_eq!(groups.get(&10).unwrap().len(), 2); // Two people in 10-19
+                    assert_eq!(groups.get(&20).unwrap().len(), 1); // One person in 20-29
+                }
+            }
+        "#
+        .unindent();
+        assert_eq!(
+            test_replace_with_flexible_indent(cx, &whole, &old, &new),
+            Some(expected.to_string())
+        );
+    }
+
+    #[test]
+    fn test_drop_lines_prefix() {
+        // Empty array
+        assert_eq!(drop_lines_prefix(&[], 2), Vec::<&str>::new());
+
+        // Zero prefix length
+        assert_eq!(
+            drop_lines_prefix(&["line1", "line2"], 0),
+            vec!["line1", "line2"]
+        );
+
+        // Normal prefix drop
+        assert_eq!(
+            drop_lines_prefix(&["  line1", "  line2"], 2),
+            vec!["line1", "line2"]
+        );
+
+        // Prefix longer than some lines
+        assert_eq!(drop_lines_prefix(&["  line1", "a"], 2), vec!["line1", ""]);
+
+        // Prefix longer than all lines
+        assert_eq!(drop_lines_prefix(&["a", "b"], 5), vec!["", ""]);
+
+        // Mixed length lines
+        assert_eq!(
+            drop_lines_prefix(&["    line1", "  line2", "      line3"], 2),
+            vec!["  line1", "line2", "    line3"]
+        );
+    }
+
+    fn test_replace_with_flexible_indent(
+        cx: &mut TestAppContext,
+        whole: &str,
+        old: &str,
+        new: &str,
+    ) -> Option<String> {
+        // Create a local buffer with the test content
+        let buffer = cx.new(|cx| language::Buffer::local(whole, cx));
+
+        // Get the buffer snapshot
+        let buffer_snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot());
+
+        // Call replace_flexible and transform the result
+        replace_with_flexible_indent(old, new, &buffer_snapshot).map(|diff| {
+            buffer.update(cx, |buffer, cx| {
+                let _ = buffer.apply_diff(diff, cx);
+                buffer.text()
+            })
+        })
+    }
+}