409 lines
13 KiB
Rust
409 lines
13 KiB
Rust
mod edit_action;
|
|
pub mod log;
|
|
mod replace;
|
|
|
|
use anyhow::{anyhow, Context, Result};
|
|
use assistant_tool::{ActionLog, Tool};
|
|
use collections::HashSet;
|
|
use edit_action::{EditAction, EditActionParser};
|
|
use futures::StreamExt;
|
|
use gpui::{App, AsyncApp, Entity, Task};
|
|
use language_model::{
|
|
LanguageModelRegistry, LanguageModelRequest, LanguageModelRequestMessage, MessageContent, Role,
|
|
};
|
|
use log::{EditToolLog, EditToolRequestId};
|
|
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::ResultExt;
|
|
|
|
#[derive(Debug, Serialize, Deserialize, JsonSchema)]
|
|
pub struct EditFilesToolInput {
|
|
/// High-level edit instructions. These will be interpreted by a smaller
|
|
/// model, so explain the changes you want that model to make and which
|
|
/// file paths need changing.
|
|
///
|
|
/// The description should be concise and clear. We will show this
|
|
/// description to the user as well.
|
|
///
|
|
/// WARNING: When specifying which file paths need changing, you MUST
|
|
/// start each path with one of the project's root directories.
|
|
///
|
|
/// WARNING: NEVER include code blocks or snippets in edit instructions.
|
|
/// Only provide natural language descriptions of the changes needed! The tool will
|
|
/// reject any instructions that contain code blocks or snippets.
|
|
///
|
|
/// The following examples assume we have two root directories in the project:
|
|
/// - root-1
|
|
/// - root-2
|
|
///
|
|
/// <example>
|
|
/// If you want to introduce a new quit function to kill the process, your
|
|
/// instructions should be: "Add a new `quit` function to
|
|
/// `root-1/src/main.rs` to kill the process".
|
|
///
|
|
/// Notice how the file path starts with root-1. Without that, the path
|
|
/// would be ambiguous and the call would fail!
|
|
/// </example>
|
|
///
|
|
/// <example>
|
|
/// If you want to change documentation to always start with a capital
|
|
/// letter, your instructions should be: "In `root-2/db.js`,
|
|
/// `root-2/inMemory.js` and `root-2/sql.js`, change all the documentation
|
|
/// to start with a capital letter".
|
|
///
|
|
/// Notice how we never specify code snippets in the instructions!
|
|
/// </example>
|
|
pub edit_instructions: String,
|
|
}
|
|
|
|
pub struct EditFilesTool;
|
|
|
|
impl Tool for EditFilesTool {
|
|
fn name(&self) -> String {
|
|
"edit-files".into()
|
|
}
|
|
|
|
fn description(&self) -> String {
|
|
include_str!("./edit_files_tool/description.md").into()
|
|
}
|
|
|
|
fn input_schema(&self) -> serde_json::Value {
|
|
let schema = schemars::schema_for!(EditFilesToolInput);
|
|
serde_json::to_value(&schema).unwrap()
|
|
}
|
|
|
|
fn run(
|
|
self: Arc<Self>,
|
|
input: serde_json::Value,
|
|
messages: &[LanguageModelRequestMessage],
|
|
project: Entity<Project>,
|
|
action_log: Entity<ActionLog>,
|
|
cx: &mut App,
|
|
) -> Task<Result<String>> {
|
|
let input = match serde_json::from_value::<EditFilesToolInput>(input) {
|
|
Ok(input) => input,
|
|
Err(err) => return Task::ready(Err(anyhow!(err))),
|
|
};
|
|
|
|
match EditToolLog::try_global(cx) {
|
|
Some(log) => {
|
|
let req_id = log.update(cx, |log, cx| {
|
|
log.new_request(input.edit_instructions.clone(), cx)
|
|
});
|
|
|
|
let task = EditToolRequest::new(
|
|
input,
|
|
messages,
|
|
project,
|
|
action_log,
|
|
Some((log.clone(), req_id)),
|
|
cx,
|
|
);
|
|
|
|
cx.spawn(async move |cx| {
|
|
let result = task.await;
|
|
|
|
let str_result = match &result {
|
|
Ok(out) => Ok(out.clone()),
|
|
Err(err) => Err(err.to_string()),
|
|
};
|
|
|
|
log.update(cx, |log, cx| log.set_tool_output(req_id, str_result, cx))
|
|
.log_err();
|
|
|
|
result
|
|
})
|
|
}
|
|
|
|
None => EditToolRequest::new(input, messages, project, action_log, None, cx),
|
|
}
|
|
}
|
|
}
|
|
|
|
struct EditToolRequest {
|
|
parser: EditActionParser,
|
|
output: String,
|
|
changed_buffers: HashSet<Entity<language::Buffer>>,
|
|
bad_searches: Vec<BadSearch>,
|
|
project: Entity<Project>,
|
|
action_log: Entity<ActionLog>,
|
|
tool_log: Option<(Entity<EditToolLog>, EditToolRequestId)>,
|
|
}
|
|
|
|
#[derive(Debug)]
|
|
enum DiffResult {
|
|
BadSearch(BadSearch),
|
|
Diff(language::Diff),
|
|
}
|
|
|
|
#[derive(Debug)]
|
|
struct BadSearch {
|
|
file_path: String,
|
|
search: String,
|
|
}
|
|
|
|
impl EditToolRequest {
|
|
fn new(
|
|
input: EditFilesToolInput,
|
|
messages: &[LanguageModelRequestMessage],
|
|
project: Entity<Project>,
|
|
action_log: Entity<ActionLog>,
|
|
tool_log: Option<(Entity<EditToolLog>, EditToolRequestId)>,
|
|
cx: &mut App,
|
|
) -> Task<Result<String>> {
|
|
let model_registry = LanguageModelRegistry::read_global(cx);
|
|
let Some(model) = model_registry.editor_model() else {
|
|
return Task::ready(Err(anyhow!("No editor model configured")));
|
|
};
|
|
|
|
let mut messages = messages.to_vec();
|
|
// Remove the last tool use (this run) to prevent an invalid request
|
|
'outer: for message in messages.iter_mut().rev() {
|
|
for (index, content) in message.content.iter().enumerate().rev() {
|
|
match content {
|
|
MessageContent::ToolUse(_) => {
|
|
message.content.remove(index);
|
|
break 'outer;
|
|
}
|
|
MessageContent::ToolResult(_) => {
|
|
// If we find any tool results before a tool use, the request is already valid
|
|
break 'outer;
|
|
}
|
|
MessageContent::Text(_) | MessageContent::Image(_) => {}
|
|
}
|
|
}
|
|
}
|
|
|
|
messages.push(LanguageModelRequestMessage {
|
|
role: Role::User,
|
|
content: vec![
|
|
include_str!("./edit_files_tool/edit_prompt.md").into(),
|
|
input.edit_instructions.into(),
|
|
],
|
|
cache: false,
|
|
});
|
|
|
|
cx.spawn(async move |cx| {
|
|
let llm_request = LanguageModelRequest {
|
|
messages,
|
|
tools: vec![],
|
|
stop: vec![],
|
|
temperature: Some(0.0),
|
|
};
|
|
|
|
let stream = model.stream_completion_text(llm_request, &cx);
|
|
let mut chunks = stream.await?;
|
|
|
|
let mut request = Self {
|
|
parser: EditActionParser::new(),
|
|
// we start with the success header so we don't need to shift the output in the common case
|
|
output: Self::SUCCESS_OUTPUT_HEADER.to_string(),
|
|
changed_buffers: HashSet::default(),
|
|
bad_searches: Vec::new(),
|
|
action_log,
|
|
project,
|
|
tool_log,
|
|
};
|
|
|
|
while let Some(chunk) = chunks.stream.next().await {
|
|
request.process_response_chunk(&chunk?, cx).await?;
|
|
}
|
|
|
|
request.finalize(cx).await
|
|
})
|
|
}
|
|
|
|
async fn process_response_chunk(&mut self, chunk: &str, cx: &mut AsyncApp) -> Result<()> {
|
|
let new_actions = self.parser.parse_chunk(chunk);
|
|
|
|
if let Some((ref log, req_id)) = self.tool_log {
|
|
log.update(cx, |log, cx| {
|
|
log.push_editor_response_chunk(req_id, chunk, &new_actions, cx)
|
|
})
|
|
.log_err();
|
|
}
|
|
|
|
for action in new_actions {
|
|
self.apply_action(action, cx).await?;
|
|
}
|
|
|
|
Ok(())
|
|
}
|
|
|
|
async fn apply_action(
|
|
&mut self,
|
|
(action, source): (EditAction, String),
|
|
cx: &mut AsyncApp,
|
|
) -> Result<()> {
|
|
let project_path = self.project.read_with(cx, |project, cx| {
|
|
project
|
|
.find_project_path(action.file_path(), cx)
|
|
.context("Path not found in project")
|
|
})??;
|
|
|
|
let buffer = self
|
|
.project
|
|
.update(cx, |project, cx| project.open_buffer(project_path, cx))?
|
|
.await?;
|
|
|
|
let result = match action {
|
|
EditAction::Replace {
|
|
old,
|
|
new,
|
|
file_path,
|
|
} => {
|
|
let snapshot = buffer.read_with(cx, |buffer, _cx| buffer.snapshot())?;
|
|
|
|
cx.background_executor()
|
|
.spawn(Self::replace_diff(old, new, file_path, snapshot))
|
|
.await
|
|
}
|
|
EditAction::Write { content, .. } => Ok(DiffResult::Diff(
|
|
buffer
|
|
.read_with(cx, |buffer, cx| buffer.diff(content, cx))?
|
|
.await,
|
|
)),
|
|
}?;
|
|
|
|
match result {
|
|
DiffResult::BadSearch(invalid_replace) => {
|
|
self.bad_searches.push(invalid_replace);
|
|
}
|
|
DiffResult::Diff(diff) => {
|
|
let _clock = buffer.update(cx, |buffer, cx| buffer.apply_diff(diff, cx))?;
|
|
|
|
write!(&mut self.output, "\n\n{}", source)?;
|
|
self.changed_buffers.insert(buffer);
|
|
}
|
|
}
|
|
|
|
Ok(())
|
|
}
|
|
|
|
async fn replace_diff(
|
|
old: String,
|
|
new: String,
|
|
file_path: std::path::PathBuf,
|
|
snapshot: language::BufferSnapshot,
|
|
) -> Result<DiffResult> {
|
|
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 Some(diff) = result else {
|
|
return anyhow::Ok(DiffResult::BadSearch(BadSearch {
|
|
search: old,
|
|
file_path: file_path.display().to_string(),
|
|
}));
|
|
};
|
|
|
|
anyhow::Ok(DiffResult::Diff(diff))
|
|
}
|
|
|
|
const SUCCESS_OUTPUT_HEADER: &str = "Successfully applied. Here's a list of changes:";
|
|
const ERROR_OUTPUT_HEADER_NO_EDITS: &str = "I couldn't apply any edits!";
|
|
const ERROR_OUTPUT_HEADER_WITH_EDITS: &str =
|
|
"Errors occurred. First, here's a list of the edits we managed to apply:";
|
|
|
|
async fn finalize(self, cx: &mut AsyncApp) -> Result<String> {
|
|
let changed_buffer_count = self.changed_buffers.len();
|
|
|
|
// Save each buffer once at the end
|
|
for buffer in &self.changed_buffers {
|
|
self.project
|
|
.update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))?
|
|
.await?;
|
|
}
|
|
|
|
self.action_log
|
|
.update(cx, |log, cx| log.buffer_edited(self.changed_buffers, cx))
|
|
.log_err();
|
|
|
|
let errors = self.parser.errors();
|
|
|
|
if errors.is_empty() && self.bad_searches.is_empty() {
|
|
if changed_buffer_count == 0 {
|
|
return Err(anyhow!(
|
|
"The instructions didn't lead to any changes. You might need to consult the file contents first."
|
|
));
|
|
}
|
|
|
|
Ok(self.output)
|
|
} else {
|
|
let mut output = self.output;
|
|
|
|
if output.is_empty() {
|
|
output.replace_range(
|
|
0..Self::SUCCESS_OUTPUT_HEADER.len(),
|
|
Self::ERROR_OUTPUT_HEADER_NO_EDITS,
|
|
);
|
|
} else {
|
|
output.replace_range(
|
|
0..Self::SUCCESS_OUTPUT_HEADER.len(),
|
|
Self::ERROR_OUTPUT_HEADER_WITH_EDITS,
|
|
);
|
|
}
|
|
|
|
if !self.bad_searches.is_empty() {
|
|
writeln!(
|
|
&mut output,
|
|
"\n\n# {} SEARCH/REPLACE block(s) failed to match:\n",
|
|
self.bad_searches.len()
|
|
)?;
|
|
|
|
for replace in self.bad_searches {
|
|
writeln!(
|
|
&mut output,
|
|
"## No exact match in: {}\n```\n{}\n```\n",
|
|
replace.file_path, replace.search,
|
|
)?;
|
|
}
|
|
|
|
write!(&mut output,
|
|
"The SEARCH section must exactly match an existing block of lines including all white \
|
|
space, comments, indentation, docstrings, etc."
|
|
)?;
|
|
}
|
|
|
|
if !errors.is_empty() {
|
|
writeln!(
|
|
&mut output,
|
|
"\n\n# {} SEARCH/REPLACE blocks failed to parse:",
|
|
errors.len()
|
|
)?;
|
|
|
|
for error in errors {
|
|
writeln!(&mut output, "- {}", error)?;
|
|
}
|
|
}
|
|
|
|
if changed_buffer_count > 0 {
|
|
writeln!(
|
|
&mut output,
|
|
"\n\nThe other SEARCH/REPLACE blocks were applied successfully. Do not re-send them!",
|
|
)?;
|
|
}
|
|
|
|
writeln!(
|
|
&mut output,
|
|
"{}You can fix errors by running the tool again. You can include instructions, \
|
|
but errors are part of the conversation so you don't need to repeat them.",
|
|
if changed_buffer_count == 0 {
|
|
"\n\n"
|
|
} else {
|
|
""
|
|
}
|
|
)?;
|
|
|
|
Err(anyhow!(output))
|
|
}
|
|
}
|
|
}
|