ZIm/crates/assistant_tools/src/code_action_tool.rs
duvetfall fddaa31655
assistant_tools: Fix code_action and rename schemas for Gemini (#28634)
Closes #28475

Updates `rename` and `code_action` `input_schema` methods to use
`json_schema_for<T>()` which transforms standard JSONSchema into the
subset required by Gemini.
Also makes `input_schema` implementations consistent.
Tested tools against Gemini 2.5 Pro Preview, Zed Claude 3.7 Sonnet
Thinking, o3-mini

Release Notes:

- Agent Beta: Fixed error 400 `INVALID_ARGUMENT` when using Gemini with
`code_actions` or `rename` tools enabled.
2025-04-14 11:01:47 -04:00

390 lines
16 KiB
Rust

use anyhow::{Context as _, Result, anyhow};
use assistant_tool::{ActionLog, Tool};
use gpui::{App, Entity, Task};
use language::{self, Anchor, Buffer, ToPointUtf16};
use language_model::LanguageModelRequestMessage;
use project::{self, LspAction, Project};
use regex::Regex;
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};
use std::{ops::Range, sync::Arc};
use ui::IconName;
use crate::schema::json_schema_for;
#[derive(Debug, Serialize, Deserialize, JsonSchema)]
pub struct CodeActionToolInput {
/// The relative path to the file containing the text range.
///
/// WARNING: you MUST start this path with one of the project's root directories.
pub path: String,
/// The specific code action to execute.
///
/// If this field is provided, the tool will execute the specified action.
/// If omitted, the tool will list all available code actions for the text range.
///
/// Here are some actions that are commonly supported (but may not be for this particular
/// text range; you can omit this field to list all the actions, if you want to know
/// what your options are, or you can just try an action and if it fails I'll tell you
/// what the available actions were instead):
/// - "quickfix.all" - applies all available quick fixes in the range
/// - "source.organizeImports" - sorts and cleans up import statements
/// - "source.fixAll" - applies all available auto fixes
/// - "refactor.extract" - extracts selected code into a new function or variable
/// - "refactor.inline" - inlines a variable by replacing references with its value
/// - "refactor.rewrite" - general code rewriting operations
/// - "source.addMissingImports" - adds imports for references that lack them
/// - "source.removeUnusedImports" - removes imports that aren't being used
/// - "source.implementInterface" - generates methods required by an interface/trait
/// - "source.generateAccessors" - creates getter/setter methods
/// - "source.convertToAsyncFunction" - converts callback-style code to async/await
///
/// Also, there is a special case: if you specify exactly "textDocument/rename" as the action,
/// then this will rename the symbol to whatever string you specified for the `arguments` field.
pub action: Option<String>,
/// Optional arguments to pass to the code action.
///
/// For rename operations (when action="textDocument/rename"), this should contain the new name.
/// For other code actions, these arguments may be passed to the language server.
pub arguments: Option<serde_json::Value>,
/// The text that comes immediately before the text range in the file.
pub context_before_range: String,
/// The text range. This text must appear in the file right between `context_before_range`
/// and `context_after_range`.
///
/// The file must contain exactly one occurrence of `context_before_range` followed by
/// `text_range` followed by `context_after_range`. If the file contains zero occurrences,
/// or if it contains more than one occurrence, the tool will fail, so it is absolutely
/// critical that you verify ahead of time that the string is unique. You can search
/// the file's contents to verify this ahead of time.
///
/// To make the string more likely to be unique, include a minimum of 1 line of context
/// before the text range, as well as a minimum of 1 line of context after the text range.
/// If these lines of context are not enough to obtain a string that appears only once
/// in the file, then double the number of context lines until the string becomes unique.
/// (Start with 1 line before and 1 line after though, because too much context is
/// needlessly costly.)
///
/// Do not alter the context lines of code in any way, and make sure to preserve all
/// whitespace and indentation for all lines of code. The combined string must be exactly
/// as it appears in the file, or else this tool call will fail.
pub text_range: String,
/// The text that comes immediately after the text range in the file.
pub context_after_range: String,
}
pub struct CodeActionTool;
impl Tool for CodeActionTool {
fn name(&self) -> String {
"code_actions".into()
}
fn needs_confirmation(&self, _input: &serde_json::Value, _cx: &App) -> bool {
false
}
fn description(&self) -> String {
include_str!("./code_action_tool/description.md").into()
}
fn icon(&self) -> IconName {
IconName::Wand
}
fn input_schema(
&self,
format: language_model::LanguageModelToolSchemaFormat,
) -> serde_json::Value {
json_schema_for::<CodeActionToolInput>(format)
}
fn ui_text(&self, input: &serde_json::Value) -> String {
match serde_json::from_value::<CodeActionToolInput>(input.clone()) {
Ok(input) => {
if let Some(action) = &input.action {
if action == "textDocument/rename" {
let new_name = match &input.arguments {
Some(serde_json::Value::String(new_name)) => new_name.clone(),
Some(value) => {
if let Ok(new_name) =
serde_json::from_value::<String>(value.clone())
{
new_name
} else {
"invalid name".to_string()
}
}
None => "missing name".to_string(),
};
format!("Rename '{}' to '{}'", input.text_range, new_name)
} else {
format!(
"Execute code action '{}' for '{}'",
action, input.text_range
)
}
} else {
format!("List available code actions for '{}'", input.text_range)
}
}
Err(_) => "Perform code action".to_string(),
}
}
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::<CodeActionToolInput>(input) {
Ok(input) => input,
Err(err) => return Task::ready(Err(anyhow!(err))),
};
cx.spawn(async move |cx| {
let buffer = {
let project_path = project.read_with(cx, |project, cx| {
project
.find_project_path(&input.path, cx)
.context("Path not found in project")
})??;
project.update(cx, |project, cx| project.open_buffer(project_path, cx))?.await?
};
action_log.update(cx, |action_log, cx| {
action_log.buffer_read(buffer.clone(), cx);
})?;
let range = {
let Some(range) = buffer.read_with(cx, |buffer, _cx| {
find_text_range(&buffer, &input.context_before_range, &input.text_range, &input.context_after_range)
})? else {
return Err(anyhow!(
"Failed to locate the text specified by context_before_range, text_range, and context_after_range. Make sure context_before_range and context_after_range each match exactly once in the file."
));
};
range
};
if let Some(action_type) = &input.action {
// Special-case the `rename` operation
let response = if action_type == "textDocument/rename" {
let Some(new_name) = input.arguments.and_then(|args| serde_json::from_value::<String>(args).ok()) else {
return Err(anyhow!("For rename operations, 'arguments' must be a string containing the new name"));
};
let position = buffer.read_with(cx, |buffer, _| {
range.start.to_point_utf16(&buffer.snapshot())
})?;
project
.update(cx, |project, cx| {
project.perform_rename(buffer.clone(), position, new_name.clone(), cx)
})?
.await?;
format!("Renamed '{}' to '{}'", input.text_range, new_name)
} else {
// Get code actions for the range
let actions = project
.update(cx, |project, cx| {
project.code_actions(&buffer, range.clone(), None, cx)
})?
.await?;
if actions.is_empty() {
return Err(anyhow!("No code actions available for this range"));
}
// Find all matching actions
let regex = match Regex::new(action_type) {
Ok(regex) => regex,
Err(err) => return Err(anyhow!("Invalid regex pattern: {}", err)),
};
let mut matching_actions = actions
.into_iter()
.filter(|action| { regex.is_match(action.lsp_action.title()) });
let Some(action) = matching_actions.next() else {
return Err(anyhow!("No code actions match the pattern: {}", action_type));
};
// There should have been exactly one matching action.
if let Some(second) = matching_actions.next() {
let mut all_matches = vec![action, second];
all_matches.extend(matching_actions);
return Err(anyhow!(
"Pattern '{}' matches multiple code actions: {}",
action_type,
all_matches.into_iter().map(|action| action.lsp_action.title().to_string()).collect::<Vec<_>>().join(", ")
));
}
let title = action.lsp_action.title().to_string();
project
.update(cx, |project, cx| {
project.apply_code_action(buffer.clone(), action, true, cx)
})?
.await?;
format!("Completed code action: {}", title)
};
project
.update(cx, |project, cx| project.save_buffer(buffer.clone(), cx))?
.await?;
action_log.update(cx, |log, cx| {
log.buffer_edited(buffer.clone(), cx)
})?;
Ok(response)
} else {
// No action specified, so list the available ones.
let (position_start, position_end) = buffer.read_with(cx, |buffer, _| {
let snapshot = buffer.snapshot();
(
range.start.to_point_utf16(&snapshot),
range.end.to_point_utf16(&snapshot)
)
})?;
// Convert position to display coordinates (1-based)
let position_start_display = language::Point {
row: position_start.row + 1,
column: position_start.column + 1,
};
let position_end_display = language::Point {
row: position_end.row + 1,
column: position_end.column + 1,
};
// Get code actions for the range
let actions = project
.update(cx, |project, cx| {
project.code_actions(&buffer, range.clone(), None, cx)
})?
.await?;
let mut response = format!(
"Available code actions for text range '{}' at position {}:{} to {}:{} (UTF-16 coordinates):\n\n",
input.text_range,
position_start_display.row, position_start_display.column,
position_end_display.row, position_end_display.column
);
if actions.is_empty() {
response.push_str("No code actions available for this range.");
} else {
for (i, action) in actions.iter().enumerate() {
let title = match &action.lsp_action {
LspAction::Action(code_action) => code_action.title.as_str(),
LspAction::Command(command) => command.title.as_str(),
LspAction::CodeLens(code_lens) => {
if let Some(cmd) = &code_lens.command {
cmd.title.as_str()
} else {
"Unknown code lens"
}
},
};
let kind = match &action.lsp_action {
LspAction::Action(code_action) => {
if let Some(kind) = &code_action.kind {
kind.as_str()
} else {
"unknown"
}
},
LspAction::Command(_) => "command",
LspAction::CodeLens(_) => "code_lens",
};
response.push_str(&format!("{}. {title} ({kind})\n", i + 1));
}
}
Ok(response)
}
})
}
}
/// Finds the range of the text in the buffer, if it appears between context_before_range
/// and context_after_range, and if that combined string has one unique result in the buffer.
///
/// If an exact match fails, it tries adding a newline to the end of context_before_range and
/// to the beginning of context_after_range to accommodate line-based context matching.
fn find_text_range(
buffer: &Buffer,
context_before_range: &str,
text_range: &str,
context_after_range: &str,
) -> Option<Range<Anchor>> {
let snapshot = buffer.snapshot();
let text = snapshot.text();
// First try with exact match
let search_string = format!("{context_before_range}{text_range}{context_after_range}");
let mut positions = text.match_indices(&search_string);
let position_result = positions.next();
if let Some(position) = position_result {
// Check if the matched string is unique
if positions.next().is_none() {
let range_start = position.0 + context_before_range.len();
let range_end = range_start + text_range.len();
let range_start_anchor = snapshot.anchor_before(snapshot.offset_to_point(range_start));
let range_end_anchor = snapshot.anchor_before(snapshot.offset_to_point(range_end));
return Some(range_start_anchor..range_end_anchor);
}
}
// If exact match fails or is not unique, try with line-based context
// Add a newline to the end of before context and beginning of after context
let line_based_before = if context_before_range.ends_with('\n') {
context_before_range.to_string()
} else {
format!("{context_before_range}\n")
};
let line_based_after = if context_after_range.starts_with('\n') {
context_after_range.to_string()
} else {
format!("\n{context_after_range}")
};
let line_search_string = format!("{line_based_before}{text_range}{line_based_after}");
let mut line_positions = text.match_indices(&line_search_string);
let line_position = line_positions.next()?;
// The line-based search string must also appear exactly once
if line_positions.next().is_some() {
return None;
}
let line_range_start = line_position.0 + line_based_before.len();
let line_range_end = line_range_start + text_range.len();
let line_range_start_anchor =
snapshot.anchor_before(snapshot.offset_to_point(line_range_start));
let line_range_end_anchor = snapshot.anchor_before(snapshot.offset_to_point(line_range_end));
Some(line_range_start_anchor..line_range_end_anchor)
}