edit_file: Let agent specify locations of edit chunks (#32628)
These changes help the agent edit files when `<old_text>` matches more than one location. First, the agent can specify an optional `<old_text line=XX>` parameter. When this is provided and multiple matches exist, we use this hint to identify the best match. Second, when there is ambiguity in matches, we now return the agent a more helpful message listing the line numbers of all possible matches. Together, these changes should reduce the number of misplaced edits and agent confusion. I have ensured the LLM Worker works with these prompt changes. Release Notes: - Agent: Improved locating edits
This commit is contained in:
Oleksiy Syvokon
GitHub
parent
e8d495806f
commit
5d293ae8ac
6 changed files with 286 additions and 63 deletions
|
|
@ -1,4 +1,5 @@
|
|||
use derive_more::{Add, AddAssign};
|
||||
use regex::Regex;
|
||||
use schemars::JsonSchema;
|
||||
use serde::{Deserialize, Serialize};
|
||||
use smallvec::SmallVec;
|
||||
|
|
@ -11,8 +12,15 @@ const END_TAGS: [&str; 3] = [OLD_TEXT_END_TAG, NEW_TEXT_END_TAG, EDITS_END_TAG];
|
|||
|
||||
#[derive(Debug)]
|
||||
pub enum EditParserEvent {
|
||||
OldTextChunk { chunk: String, done: bool },
|
||||
NewTextChunk { chunk: String, done: bool },
|
||||
OldTextChunk {
|
||||
chunk: String,
|
||||
done: bool,
|
||||
line_hint: Option<u32>,
|
||||
},
|
||||
NewTextChunk {
|
||||
chunk: String,
|
||||
done: bool,
|
||||
},
|
||||
}
|
||||
|
||||
#[derive(
|
||||
|
|
@ -33,7 +41,7 @@ pub struct EditParser {
|
|||
#[derive(Debug, PartialEq)]
|
||||
enum EditParserState {
|
||||
Pending,
|
||||
WithinOldText { start: bool },
|
||||
WithinOldText { start: bool, line_hint: Option<u32> },
|
||||
AfterOldText,
|
||||
WithinNewText { start: bool },
|
||||
}
|
||||
|
|
@ -54,14 +62,24 @@ impl EditParser {
|
|||
loop {
|
||||
match &mut self.state {
|
||||
EditParserState::Pending => {
|
||||
if let Some(start) = self.buffer.find("<old_text>") {
|
||||
self.buffer.drain(..start + "<old_text>".len());
|
||||
self.state = EditParserState::WithinOldText { start: true };
|
||||
if let Some(start) = self.buffer.find("<old_text") {
|
||||
if let Some(tag_end) = self.buffer[start..].find('>') {
|
||||
let tag_end = start + tag_end + 1;
|
||||
let tag = &self.buffer[start..tag_end];
|
||||
let line_hint = self.parse_line_hint(tag);
|
||||
self.buffer.drain(..tag_end);
|
||||
self.state = EditParserState::WithinOldText {
|
||||
start: true,
|
||||
line_hint,
|
||||
};
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
EditParserState::WithinOldText { start } => {
|
||||
EditParserState::WithinOldText { start, line_hint } => {
|
||||
if !self.buffer.is_empty() {
|
||||
if *start && self.buffer.starts_with('\n') {
|
||||
self.buffer.remove(0);
|
||||
|
|
@ -69,6 +87,7 @@ impl EditParser {
|
|||
*start = false;
|
||||
}
|
||||
|
||||
let line_hint = *line_hint;
|
||||
if let Some(tag_range) = self.find_end_tag() {
|
||||
let mut chunk = self.buffer[..tag_range.start].to_string();
|
||||
if chunk.ends_with('\n') {
|
||||
|
|
@ -82,12 +101,17 @@ impl EditParser {
|
|||
|
||||
self.buffer.drain(..tag_range.end);
|
||||
self.state = EditParserState::AfterOldText;
|
||||
edit_events.push(EditParserEvent::OldTextChunk { chunk, done: true });
|
||||
edit_events.push(EditParserEvent::OldTextChunk {
|
||||
chunk,
|
||||
done: true,
|
||||
line_hint,
|
||||
});
|
||||
} else {
|
||||
if !self.ends_with_tag_prefix() {
|
||||
edit_events.push(EditParserEvent::OldTextChunk {
|
||||
chunk: mem::take(&mut self.buffer),
|
||||
done: false,
|
||||
line_hint,
|
||||
});
|
||||
}
|
||||
break;
|
||||
|
|
@ -154,6 +178,16 @@ impl EditParser {
|
|||
end_prefixes.any(|prefix| self.buffer.ends_with(&prefix))
|
||||
}
|
||||
|
||||
fn parse_line_hint(&self, tag: &str) -> Option<u32> {
|
||||
static LINE_HINT_REGEX: std::sync::LazyLock<Regex> =
|
||||
std::sync::LazyLock::new(|| Regex::new(r#"line=(?:"?)(\d+)"#).unwrap());
|
||||
|
||||
LINE_HINT_REGEX
|
||||
.captures(tag)
|
||||
.and_then(|caps| caps.get(1))
|
||||
.and_then(|m| m.as_str().parse::<u32>().ok())
|
||||
}
|
||||
|
||||
pub fn finish(self) -> EditParserMetrics {
|
||||
self.metrics
|
||||
}
|
||||
|
|
@ -178,6 +212,7 @@ mod tests {
|
|||
vec![Edit {
|
||||
old_text: "original".to_string(),
|
||||
new_text: "updated".to_string(),
|
||||
line_hint: None,
|
||||
}]
|
||||
);
|
||||
assert_eq!(
|
||||
|
|
@ -209,10 +244,12 @@ mod tests {
|
|||
Edit {
|
||||
old_text: "first old".to_string(),
|
||||
new_text: "first new".to_string(),
|
||||
line_hint: None,
|
||||
},
|
||||
Edit {
|
||||
old_text: "second old".to_string(),
|
||||
new_text: "second new".to_string(),
|
||||
line_hint: None,
|
||||
},
|
||||
]
|
||||
);
|
||||
|
|
@ -244,14 +281,17 @@ mod tests {
|
|||
Edit {
|
||||
old_text: "content".to_string(),
|
||||
new_text: "updated content".to_string(),
|
||||
line_hint: None,
|
||||
},
|
||||
Edit {
|
||||
old_text: "second item".to_string(),
|
||||
new_text: "modified second item".to_string(),
|
||||
line_hint: None,
|
||||
},
|
||||
Edit {
|
||||
old_text: "third case".to_string(),
|
||||
new_text: "improved third case".to_string(),
|
||||
line_hint: None,
|
||||
},
|
||||
]
|
||||
);
|
||||
|
|
@ -276,6 +316,7 @@ mod tests {
|
|||
vec![Edit {
|
||||
old_text: "code with <tag>nested</tag> elements".to_string(),
|
||||
new_text: "new <code>content</code>".to_string(),
|
||||
line_hint: None,
|
||||
}]
|
||||
);
|
||||
assert_eq!(
|
||||
|
|
@ -299,6 +340,7 @@ mod tests {
|
|||
vec![Edit {
|
||||
old_text: "".to_string(),
|
||||
new_text: "".to_string(),
|
||||
line_hint: None,
|
||||
}]
|
||||
);
|
||||
assert_eq!(
|
||||
|
|
@ -322,6 +364,7 @@ mod tests {
|
|||
vec![Edit {
|
||||
old_text: "line1\nline2\nline3".to_string(),
|
||||
new_text: "line1\nmodified line2\nline3".to_string(),
|
||||
line_hint: None,
|
||||
}]
|
||||
);
|
||||
assert_eq!(
|
||||
|
|
@ -368,10 +411,12 @@ mod tests {
|
|||
Edit {
|
||||
old_text: "a\nb\nc".to_string(),
|
||||
new_text: "a\nB\nc".to_string(),
|
||||
line_hint: None,
|
||||
},
|
||||
Edit {
|
||||
old_text: "d\ne\nf".to_string(),
|
||||
new_text: "D\ne\nF".to_string(),
|
||||
line_hint: None,
|
||||
}
|
||||
]
|
||||
);
|
||||
|
|
@ -402,6 +447,7 @@ mod tests {
|
|||
vec![Edit {
|
||||
old_text: "Lorem".to_string(),
|
||||
new_text: "LOREM".to_string(),
|
||||
line_hint: None,
|
||||
},]
|
||||
);
|
||||
assert_eq!(
|
||||
|
|
@ -413,10 +459,77 @@ mod tests {
|
|||
);
|
||||
}
|
||||
|
||||
#[gpui::test(iterations = 100)]
|
||||
fn test_line_hints(mut rng: StdRng) {
|
||||
// Line hint is a single quoted line number
|
||||
let mut parser = EditParser::new();
|
||||
|
||||
let edits = parse_random_chunks(
|
||||
r#"
|
||||
<old_text line="23">original code</old_text>
|
||||
<new_text>updated code</new_text>"#,
|
||||
&mut parser,
|
||||
&mut rng,
|
||||
);
|
||||
|
||||
assert_eq!(edits.len(), 1);
|
||||
assert_eq!(edits[0].old_text, "original code");
|
||||
assert_eq!(edits[0].line_hint, Some(23));
|
||||
assert_eq!(edits[0].new_text, "updated code");
|
||||
|
||||
// Line hint is a single unquoted line number
|
||||
let mut parser = EditParser::new();
|
||||
|
||||
let edits = parse_random_chunks(
|
||||
r#"
|
||||
<old_text line=45>original code</old_text>
|
||||
<new_text>updated code</new_text>"#,
|
||||
&mut parser,
|
||||
&mut rng,
|
||||
);
|
||||
|
||||
assert_eq!(edits.len(), 1);
|
||||
assert_eq!(edits[0].old_text, "original code");
|
||||
assert_eq!(edits[0].line_hint, Some(45));
|
||||
assert_eq!(edits[0].new_text, "updated code");
|
||||
|
||||
// Line hint is a range
|
||||
let mut parser = EditParser::new();
|
||||
|
||||
let edits = parse_random_chunks(
|
||||
r#"
|
||||
<old_text line="23:50">original code</old_text>
|
||||
<new_text>updated code</new_text>"#,
|
||||
&mut parser,
|
||||
&mut rng,
|
||||
);
|
||||
|
||||
assert_eq!(edits.len(), 1);
|
||||
assert_eq!(edits[0].old_text, "original code");
|
||||
assert_eq!(edits[0].line_hint, Some(23));
|
||||
assert_eq!(edits[0].new_text, "updated code");
|
||||
|
||||
// No line hint
|
||||
let mut parser = EditParser::new();
|
||||
let edits = parse_random_chunks(
|
||||
r#"
|
||||
<old_text>old</old_text>
|
||||
<new_text>new</new_text>"#,
|
||||
&mut parser,
|
||||
&mut rng,
|
||||
);
|
||||
|
||||
assert_eq!(edits.len(), 1);
|
||||
assert_eq!(edits[0].old_text, "old");
|
||||
assert_eq!(edits[0].line_hint, None);
|
||||
assert_eq!(edits[0].new_text, "new");
|
||||
}
|
||||
|
||||
#[derive(Default, Debug, PartialEq, Eq)]
|
||||
struct Edit {
|
||||
old_text: String,
|
||||
new_text: String,
|
||||
line_hint: Option<u32>,
|
||||
}
|
||||
|
||||
fn parse_random_chunks(input: &str, parser: &mut EditParser, rng: &mut StdRng) -> Vec<Edit> {
|
||||
|
|
@ -433,10 +546,15 @@ mod tests {
|
|||
for chunk_ix in chunk_indices {
|
||||
for event in parser.push(&input[last_ix..chunk_ix]) {
|
||||
match event {
|
||||
EditParserEvent::OldTextChunk { chunk, done } => {
|
||||
EditParserEvent::OldTextChunk {
|
||||
chunk,
|
||||
done,
|
||||
line_hint,
|
||||
} => {
|
||||
old_text.as_mut().unwrap().push_str(&chunk);
|
||||
if done {
|
||||
pending_edit.old_text = old_text.take().unwrap();
|
||||
pending_edit.line_hint = line_hint;
|
||||
new_text = Some(String::new());
|
||||
}
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue