Yehowshua/ZIm
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Include full abs paths of worktrees in system prompt (#32725)

Browse source 
Some MCP servers expose tools that take absolute paths as arguments. To
interact with these, the agent needs to know the absolute path to the
project directories, not just their names. This PR changes the system
prompt to include the full path to each worktree, and updates some tool
descriptions to reflect this.

Todo:

* [x] Run evals, make sure assistant still understand how to specify
paths for tools, now that we include abs paths in the system prompt.

Release Notes:

- Improved the agent's ability to use MPC tools that require absolute
paths to files and directories in the project.

---------

Co-authored-by: Ben Brandt <benjamin.j.brandt@gmail.com>
This commit is contained in:
Max Brunsfeld 2025-06-15 06:45:26 -07:00 • committed by GitHub
parent 0433b8859d
commit a994666888
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 63 additions and 57 deletions
Download patch file Download diff file Expand all files Collapse all files

72
crates/assistant_tools/src/edit_agent/evals.rs
View file

@ -26,6 +26,7 @@ use std::{
cmp::Reverse,
fmt::{self, Display},
io::Write as _,
path::Path,
str::FromStr,
sync::mpsc,
};
@ -38,10 +39,11 @@ fn eval_extract_handle_command_output() {
//
// Model | Pass rate
// ----------------------------|----------
// claude-3.7-sonnet | 0.98
// gemini-2.5-pro-06-05 | 0.77
// gemini-2.5-flash | 0.11
// gpt-4.1 | 1.00
// claude-3.7-sonnet | 0.99 (2025-06-14)
// claude-sonnet-4 | 0.97 (2025-06-14)
// gemini-2.5-pro-06-05 | 0.77 (2025-05-22)
// gemini-2.5-flash | 0.11 (2025-05-22)
// gpt-4.1 | 1.00 (2025-05-22)
let input_file_path = "root/blame.rs";
let input_file_content = include_str!("evals/fixtures/extract_handle_command_output/before.rs");
@ -110,6 +112,13 @@ fn eval_extract_handle_command_output() {
#[test]
#[cfg_attr(not(feature = "eval"), ignore)]
fn eval_delete_run_git_blame() {
// Model | Pass rate
// ----------------------------|----------
// claude-3.7-sonnet | 1.0 (2025-06-14)
// claude-sonnet-4 | 0.96 (2025-06-14)
// gemini-2.5-pro-06-05 |
// gemini-2.5-flash |
// gpt-4.1 |
let input_file_path = "root/blame.rs";
let input_file_content = include_str!("evals/fixtures/delete_run_git_blame/before.rs");
let output_file_content = include_str!("evals/fixtures/delete_run_git_blame/after.rs");
@ -165,13 +174,12 @@ fn eval_delete_run_git_blame() {
#[test]
#[cfg_attr(not(feature = "eval"), ignore)]
fn eval_translate_doc_comments() {
// Results for 2025-05-22
//
// Model | Pass rate
// ============================================
//
// claude-3.7-sonnet |
// gemini-2.5-pro-preview-03-25 | 1.0
// claude-3.7-sonnet | 1.0 (2025-06-14)
// claude-sonnet-4 | 1.0 (2025-06-14)
// gemini-2.5-pro-preview-03-25 | 1.0 (2025-05-22)
// gemini-2.5-flash-preview-04-17 |
// gpt-4.1 |
let input_file_path = "root/canvas.rs";
@ -228,13 +236,12 @@ fn eval_translate_doc_comments() {
#[test]
#[cfg_attr(not(feature = "eval"), ignore)]
fn eval_use_wasi_sdk_in_compile_parser_to_wasm() {
// Results for 2025-05-22
//
// Model | Pass rate
// ============================================
//
// claude-3.7-sonnet | 0.98
// gemini-2.5-pro-preview-03-25 | 0.99
// claude-3.7-sonnet | 0.96 (2025-06-14)
// claude-sonnet-4 | 0.11 (2025-06-14)
// gemini-2.5-pro-preview-03-25 | 0.99 (2025-05-22)
// gemini-2.5-flash-preview-04-17 |
// gpt-4.1 |
let input_file_path = "root/lib.rs";
@ -354,13 +361,12 @@ fn eval_use_wasi_sdk_in_compile_parser_to_wasm() {
#[test]
#[cfg_attr(not(feature = "eval"), ignore)]
fn eval_disable_cursor_blinking() {
// Results for 2025-05-22
//
// Model | Pass rate
// ============================================
//
// claude-3.7-sonnet |
// gemini-2.5-pro-preview-03-25 | 1.0
// claude-3.7-sonnet | 0.99 (2025-06-14)
// claude-sonnet-4 | 0.85 (2025-06-14)
// gemini-2.5-pro-preview-03-25 | 1.0 (2025-05-22)
// gemini-2.5-flash-preview-04-17 |
// gpt-4.1 |
let input_file_path = "root/editor.rs";
@ -462,7 +468,7 @@ fn eval_from_pixels_constructor() {
0.95,
// For whatever reason, this eval produces more mismatched tags.
// Increasing for now, let's see if we can bring this down.
0.2,
0.25,
EvalInput::from_conversation(
vec![
message(
@ -648,15 +654,14 @@ fn eval_from_pixels_constructor() {
#[test]
#[cfg_attr(not(feature = "eval"), ignore)]
fn eval_zode() {
// Results for 2025-05-22
//
// Model | Pass rate
// ============================================
//
// claude-3.7-sonnet | 1.0
// gemini-2.5-pro-preview-03-25 | 1.0
// gemini-2.5-flash-preview-04-17 | 1.0
// gpt-4.1 | 1.0
// claude-3.7-sonnet | 1.0 (2025-06-14)
// claude-sonnet-4 | 1.0 (2025-06-14)
// gemini-2.5-pro-preview-03-25 | 1.0 (2025-05-22)
// gemini-2.5-flash-preview-04-17 | 1.0 (2025-05-22)
// gpt-4.1 | 1.0 (2025-05-22)
let input_file_path = "root/zode.py";
let input_content = None;
let edit_description = "Create the main Zode CLI script";
@ -755,13 +760,12 @@ fn eval_zode() {
#[test]
#[cfg_attr(not(feature = "eval"), ignore)]
fn eval_add_overwrite_test() {
// Results for 2025-05-22
//
// Model | Pass rate
// ============================================
//
// claude-3.7-sonnet | 0.16
// gemini-2.5-pro-preview-03-25 | 0.35
// claude-3.7-sonnet | 0.65 (2025-06-14)
// claude-sonnet-4 | 0.07 (2025-06-14)
// gemini-2.5-pro-preview-03-25 | 0.35 (2025-05-22)
// gemini-2.5-flash-preview-04-17 |
// gpt-4.1 |
let input_file_path = "root/action_log.rs";
@ -991,15 +995,14 @@ fn eval_create_empty_file() {
// thoughts into it. This issue is not specific to empty files, but
// it's easier to reproduce with them.
//
// Results for 2025-05-21:
//
// Model | Pass rate
// ============================================
//
// claude-3.7-sonnet | 1.00
// gemini-2.5-pro-preview-03-25 | 1.00
// gemini-2.5-flash-preview-04-17 | 1.00
// gpt-4.1 | 1.00
// claude-3.7-sonnet | 1.00 (2025-06-14)
// claude-sonnet-4 | 1.00 (2025-06-14)
// gemini-2.5-pro-preview-03-25 | 1.00 (2025-05-21)
// gemini-2.5-flash-preview-04-17 | 1.00 (2025-05-21)
// gpt-4.1 | 1.00 (2025-05-21)
//
//
// TODO: gpt-4.1-mini errored 38 times:
@ -1556,6 +1559,7 @@ impl EditAgentTest {
.collect::<Vec<_>>();
let worktrees = vec![WorktreeContext {
root_name: "root".to_string(),
abs_path: Path::new("/path/to/root").into(),
rules_file: None,
}];
let prompt_builder = PromptBuilder::new(None)?;
@ -1650,7 +1654,7 @@ async fn retry_on_rate_limit<R>(mut request: impl AsyncFnMut() -> Result<R>) ->
Ok(err) => match err {
LanguageModelCompletionError::RateLimit(duration) => {
// Wait for the duration supplied, with some jitter to avoid all requests being made at the same time.
let jitter = duration.mul_f64(rand::thread_rng().gen_range(0.0..0.5));
let jitter = duration.mul_f64(rand::thread_rng().gen_range(0.0..1.0));
eprintln!(
"Attempt #{attempt}: Rate limit exceeded. Retry after {duration:?} + jitter of {jitter:?}"
);

6
crates/assistant_tools/src/edit_file_tool.rs
View file

@ -69,13 +69,13 @@ pub struct EditFileToolInput {
/// start each path with one of the project's root directories.
///
/// The following examples assume we have two root directories in the project:
/// - backend
/// - frontend
/// - /a/b/backend
/// - /c/d/frontend
///
/// <example>
/// `backend/src/main.rs`
///
/// Notice how the file path starts with root-1. Without that, the path
/// Notice how the file path starts with `backend`. Without that, the path
/// would be ambiguous and the call would fail!
/// </example>
///

4
crates/assistant_tools/src/read_file_tool.rs
View file

@ -31,8 +31,8 @@ pub struct ReadFileToolInput {
/// <example>
/// If the project has the following root directories:
///
/// - directory1
/// - directory2
/// - /a/b/directory1
/// - /c/d/directory2
///
/// If you want to access `file.txt` in `directory1`, you should use the path `directory1/file.txt`.
/// If you want to access `file.txt` in `directory2`, you should use the path `directory2/file.txt`.