assistant: Use tool interface for scripts (#26377)

We decided to expose scripting as tools again. We are aware of the UX
downsides of doing so, but we want to focus on getting it working well
first, and the model seems to make better use of it as an actual tool.

In the future, the tools API might support streaming. If it doesn't and
we need to ship, we can consider reverting this.

Release Notes:

- N/A
This commit is contained in:
Agus Zubiaga 2025-03-10 13:59:31 -03:00 committed by GitHub
parent 3891381d3e
commit 2fc4dec58f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
17 changed files with 163 additions and 652 deletions

View file

@ -0,0 +1,74 @@
mod session;
use project::Project;
use session::*;
use assistant_tool::{Tool, ToolRegistry};
use gpui::{App, AppContext as _, Entity, Task};
use schemars::JsonSchema;
use serde::Deserialize;
use std::sync::Arc;
pub fn init(cx: &App) {
let registry = ToolRegistry::global(cx);
registry.register_tool(ScriptingTool);
}
#[derive(Debug, Deserialize, JsonSchema)]
struct ScriptingToolInput {
lua_script: String,
}
struct ScriptingTool;
impl Tool for ScriptingTool {
fn name(&self) -> String {
"lua-interpreter".into()
}
fn description(&self) -> String {
include_str!("scripting_tool_description.txt").into()
}
fn input_schema(&self) -> serde_json::Value {
let schema = schemars::schema_for!(ScriptingToolInput);
serde_json::to_value(&schema).unwrap()
}
fn run(
self: Arc<Self>,
input: serde_json::Value,
project: Entity<Project>,
cx: &mut App,
) -> Task<anyhow::Result<String>> {
let input = match serde_json::from_value::<ScriptingToolInput>(input) {
Err(err) => return Task::ready(Err(err.into())),
Ok(input) => input,
};
// TODO: Store a session per thread
let session = cx.new(|cx| ScriptSession::new(project, cx));
let lua_script = input.lua_script;
let (script_id, script_task) =
session.update(cx, |session, cx| session.run_script(lua_script, cx));
cx.spawn(|cx| async move {
script_task.await;
let message = session.read_with(&cx, |session, _cx| {
// Using a id to get the script output seems impractical.
// Why not just include it in the Task result?
// This is because we'll later report the script state as it runs,
// currently not supported by the `Tool` interface.
session
.get(script_id)
.output_message_for_llm()
.expect("Script shouldn't still be running")
})?;
drop(session);
Ok(message)
})
}
}