ZIm/crates/assistant_tools/src/project_notifications_tool.rs
Oleksiy Syvokon e339566dab
agent: Limit the size of patches generated from user edits (#34548)
Gradually remove details from a patch to keep it within the size limit.
This helps avoid using too much context when the user pastes large
files, generates files, or just makes many changes between agent
notifications.

Release Notes:

- N/A
2025-07-16 15:46:13 +00:00

359 lines
11 KiB
Rust

use crate::schema::json_schema_for;
use anyhow::Result;
use assistant_tool::{ActionLog, Tool, ToolResult};
use gpui::{AnyWindowHandle, App, Entity, Task};
use language_model::{LanguageModel, LanguageModelRequest, LanguageModelToolSchemaFormat};
use project::Project;
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};
use std::{fmt::Write, sync::Arc};
use ui::IconName;
#[derive(Debug, Serialize, Deserialize, JsonSchema)]
pub struct ProjectUpdatesToolInput {}
pub struct ProjectNotificationsTool;
impl Tool for ProjectNotificationsTool {
fn name(&self) -> String {
"project_notifications".to_string()
}
fn needs_confirmation(&self, _: &serde_json::Value, _: &App) -> bool {
false
}
fn may_perform_edits(&self) -> bool {
false
}
fn description(&self) -> String {
include_str!("./project_notifications_tool/description.md").to_string()
}
fn icon(&self) -> IconName {
IconName::ToolNotification
}
fn input_schema(&self, format: LanguageModelToolSchemaFormat) -> Result<serde_json::Value> {
json_schema_for::<ProjectUpdatesToolInput>(format)
}
fn ui_text(&self, _input: &serde_json::Value) -> String {
"Check project notifications".into()
}
fn run(
self: Arc<Self>,
_input: serde_json::Value,
_request: Arc<LanguageModelRequest>,
_project: Entity<Project>,
action_log: Entity<ActionLog>,
_model: Arc<dyn LanguageModel>,
_window: Option<AnyWindowHandle>,
cx: &mut App,
) -> ToolResult {
let Some(user_edits_diff) =
action_log.update(cx, |log, cx| log.flush_unnotified_user_edits(cx))
else {
return result("No new notifications");
};
// NOTE: Changes to this prompt require a symmetric update in the LLM Worker
const HEADER: &str = include_str!("./project_notifications_tool/prompt_header.txt");
const MAX_BYTES: usize = 8000;
let diff = fit_patch_to_size(&user_edits_diff, MAX_BYTES);
result(&format!("{HEADER}\n\n```diff\n{diff}\n```\n").replace("\r\n", "\n"))
}
}
fn result(response: &str) -> ToolResult {
Task::ready(Ok(response.to_string().into())).into()
}
/// Make sure that the patch fits into the size limit (in bytes).
/// Compress the patch by omitting some parts if needed.
/// Unified diff format is assumed.
fn fit_patch_to_size(patch: &str, max_size: usize) -> String {
if patch.len() <= max_size {
return patch.to_string();
}
// Compression level 1: remove context lines in diff bodies, but
// leave the counts and positions of inserted/deleted lines
let mut current_size = patch.len();
let mut file_patches = split_patch(&patch);
file_patches.sort_by_key(|patch| patch.len());
let compressed_patches = file_patches
.iter()
.rev()
.map(|patch| {
if current_size > max_size {
let compressed = compress_patch(patch).unwrap_or_else(|_| patch.to_string());
current_size -= patch.len() - compressed.len();
compressed
} else {
patch.to_string()
}
})
.collect::<Vec<_>>();
if current_size <= max_size {
return compressed_patches.join("\n\n");
}
// Compression level 2: list paths of the changed files only
let filenames = file_patches
.iter()
.map(|patch| {
let patch = diffy::Patch::from_str(patch).unwrap();
let path = patch
.modified()
.and_then(|path| path.strip_prefix("b/"))
.unwrap_or_default();
format!("- {path}\n")
})
.collect::<Vec<_>>();
filenames.join("")
}
/// Split a potentially multi-file patch into multiple single-file patches
fn split_patch(patch: &str) -> Vec<String> {
let mut result = Vec::new();
let mut current_patch = String::new();
for line in patch.lines() {
if line.starts_with("---") && !current_patch.is_empty() {
result.push(current_patch.trim_end_matches('\n').into());
current_patch = String::new();
}
current_patch.push_str(line);
current_patch.push('\n');
}
if !current_patch.is_empty() {
result.push(current_patch.trim_end_matches('\n').into());
}
result
}
fn compress_patch(patch: &str) -> anyhow::Result<String> {
let patch = diffy::Patch::from_str(patch)?;
let mut out = String::new();
writeln!(out, "--- {}", patch.original().unwrap_or("a"))?;
writeln!(out, "+++ {}", patch.modified().unwrap_or("b"))?;
for hunk in patch.hunks() {
writeln!(out, "@@ -{} +{} @@", hunk.old_range(), hunk.new_range())?;
writeln!(out, "[...skipped...]")?;
}
Ok(out)
}
#[cfg(test)]
mod tests {
use super::*;
use assistant_tool::ToolResultContent;
use gpui::{AppContext, TestAppContext};
use indoc::indoc;
use language_model::{LanguageModelRequest, fake_provider::FakeLanguageModelProvider};
use project::{FakeFs, Project};
use serde_json::json;
use settings::SettingsStore;
use std::sync::Arc;
use util::path;
#[gpui::test]
async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
init_test(cx);
let fs = FakeFs::new(cx.executor());
fs.insert_tree(
path!("/test"),
json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
)
.await;
let project = Project::test(fs, [path!("/test").as_ref()], cx).await;
let action_log = cx.new(|_| ActionLog::new(project.clone()));
let buffer_path = project
.read_with(cx, |project, cx| {
project.find_project_path("test/code.rs", cx)
})
.unwrap();
let buffer = project
.update(cx, |project, cx| {
project.open_buffer(buffer_path.clone(), cx)
})
.await
.unwrap();
// Start tracking the buffer
action_log.update(cx, |log, cx| {
log.buffer_read(buffer.clone(), cx);
});
cx.run_until_parked();
// Run the tool before any changes
let tool = Arc::new(ProjectNotificationsTool);
let provider = Arc::new(FakeLanguageModelProvider);
let model: Arc<dyn LanguageModel> = Arc::new(provider.test_model());
let request = Arc::new(LanguageModelRequest::default());
let tool_input = json!({});
let result = cx.update(|cx| {
tool.clone().run(
tool_input.clone(),
request.clone(),
project.clone(),
action_log.clone(),
model.clone(),
None,
cx,
)
});
cx.run_until_parked();
let response = result.output.await.unwrap();
let response_text = match &response.content {
ToolResultContent::Text(text) => text.clone(),
_ => panic!("Expected text response"),
};
assert_eq!(
response_text.as_str(),
"No new notifications",
"Tool should return 'No new notifications' when no stale buffers"
);
// Modify the buffer (makes it stale)
buffer.update(cx, |buffer, cx| {
buffer.edit([(1..1, "\nChange!\n")], None, cx);
});
cx.run_until_parked();
// Run the tool again
let result = cx.update(|cx| {
tool.clone().run(
tool_input.clone(),
request.clone(),
project.clone(),
action_log.clone(),
model.clone(),
None,
cx,
)
});
cx.run_until_parked();
// This time the buffer is stale, so the tool should return a notification
let response = result.output.await.unwrap();
let response_text = match &response.content {
ToolResultContent::Text(text) => text.clone(),
_ => panic!("Expected text response"),
};
assert!(
response_text.contains("These files have changed"),
"Tool should return the stale buffer notification"
);
assert!(
response_text.contains("test/code.rs"),
"Tool should return the stale buffer notification"
);
// Run the tool once more without any changes - should get no new notifications
let result = cx.update(|cx| {
tool.run(
tool_input.clone(),
request.clone(),
project.clone(),
action_log,
model.clone(),
None,
cx,
)
});
cx.run_until_parked();
let response = result.output.await.unwrap();
let response_text = match &response.content {
ToolResultContent::Text(text) => text.clone(),
_ => panic!("Expected text response"),
};
assert_eq!(
response_text.as_str(),
"No new notifications",
"Tool should return 'No new notifications' when running again without changes"
);
}
#[test]
fn test_patch_compression() {
// Given a patch that doesn't fit into the size budget
let patch = indoc! {"
--- a/dir/test.txt
+++ b/dir/test.txt
@@ -1,3 +1,3 @@
line 1
-line 2
+CHANGED
line 3
@@ -10,2 +10,2 @@
line 10
-line 11
+line eleven
--- a/dir/another.txt
+++ b/dir/another.txt
@@ -100,1 +1,1 @@
-before
+after
"};
// When the size deficit can be compensated by dropping the body,
// then the body should be trimmed for larger files first
let limit = patch.len() - 10;
let compressed = fit_patch_to_size(patch, limit);
let expected = indoc! {"
--- a/dir/test.txt
+++ b/dir/test.txt
@@ -1,3 +1,3 @@
[...skipped...]
@@ -10,2 +10,2 @@
[...skipped...]
--- a/dir/another.txt
+++ b/dir/another.txt
@@ -100,1 +1,1 @@
-before
+after"};
assert_eq!(compressed, expected);
// When the size deficit is too large, then only file paths
// should be returned
let limit = 10;
let compressed = fit_patch_to_size(patch, limit);
let expected = indoc! {"
- dir/another.txt
- dir/test.txt
"};
assert_eq!(compressed, expected);
}
fn init_test(cx: &mut TestAppContext) {
cx.update(|cx| {
let settings_store = SettingsStore::test(cx);
cx.set_global(settings_store);
language::init(cx);
Project::init_settings(cx);
assistant_tool::init(cx);
});
}
}