Files
luxe/crates/luxe-core/tests/integration_branch_rerun.rs

298 lines
11 KiB
Rust
Raw Normal View History

//! Integration: edit-and-resend, retry and regenerate put the right conversation in front of the
//! model, and run as ordinary turns.
//!
//! Each test reads what the provider was actually sent, because that is the whole contract: an
//! edit replaces the message it edits, a regeneration replaces the answer, and nothing is ever
//! sent twice in a row. The last test is the reason these three run through the same door as a
//! submitted message: run inline, a permission prompt inside one waited on an answer that nothing
//! read, and an interrupt never arrived either.
#![allow(clippy::unwrap_used)]
use std::sync::{Arc, Mutex};
use futures::stream;
use futures::stream::BoxStream;
use luxe_core::compaction::CompactionSettings;
use luxe_core::engine::{Action, Engine, EngineConfig};
use luxe_core::event::Event;
use luxe_core::message::{ContentBlock, Role};
use luxe_core::permission::{
AskResolution, PermValue, PermissionAction, PermissionConfig, PermissionEngine,
PermissionReply, PermissionSettings,
};
use luxe_core::provider::{
CompletionRequest, ModelInfo, Provider, ProviderError, ProviderEvent, StopReason,
};
use luxe_core::session::{JsonlSessionStore, SessionStore};
use luxe_core::tool::{Tool, ToolContext, ToolRegistry, ToolResultData};
use luxe_core::HookRegistry;
use tokio::sync::mpsc;
/// Every conversation the model was sent, each as its messages' roles and text.
type Sent = Arc<Mutex<Vec<Vec<(Role, String)>>>>;
/// One scripted reply: text, or a call to the `bash` tool.
#[derive(Clone)]
enum Reply {
Text(&'static str),
Bash,
Fail,
}
/// Answers from a script, in order, and keeps every conversation it was sent.
struct Scripted {
script: Mutex<Vec<Reply>>,
sent: Sent,
}
#[async_trait::async_trait]
impl Provider for Scripted {
fn name(&self) -> &str {
"scripted"
}
async fn stream(
&self,
req: CompletionRequest,
_cancel: tokio::sync::watch::Receiver<bool>,
) -> Result<BoxStream<'static, ProviderEvent>, ProviderError> {
let conversation = req
.messages
.iter()
.map(|message| {
let text = message
.content
.iter()
.map(|block| match block {
ContentBlock::Text(text) => text.clone(),
ContentBlock::ToolCall { name, .. } => format!("<call {name}>"),
ContentBlock::ToolResult { .. } => "<result>".to_string(),
_ => String::new(),
})
.collect::<String>();
(message.role.clone(), text)
})
.collect();
self.sent.lock().unwrap().push(conversation);
let next = {
let mut script = self.script.lock().unwrap();
if script.is_empty() {
Reply::Text("done")
} else {
script.remove(0)
}
};
let events = match next {
Reply::Text(text) => vec![
ProviderEvent::TextDelta { index: 0, delta: text.to_string() },
ProviderEvent::Done { reason: StopReason::EndTurn },
],
Reply::Bash => vec![
ProviderEvent::ToolCallStart {
index: 0,
id: "call_bash".to_string(),
name: "bash".to_string(),
},
ProviderEvent::ToolCallDelta {
index: 0,
json_delta: r#"{"command":"git commit -m more"}"#.to_string(),
},
ProviderEvent::ToolCallEnd { index: 0 },
ProviderEvent::Done { reason: StopReason::ToolCall },
],
Reply::Fail => return Err(ProviderError::Auth("the key was refused".to_string())),
};
Ok(Box::pin(stream::iter(events)))
}
}
/// A `bash`-named tool that never needs to run in these tests.
struct Bash;
#[async_trait::async_trait]
impl Tool for Bash {
fn name(&self) -> &str {
"bash"
}
fn schema(&self) -> luxe_core::provider::ToolSchema {
luxe_core::provider::ToolSchema {
name: "bash".to_string(),
description: "Run a command.".to_string(),
parameters: serde_json::json!({
"type": "object",
"properties": { "command": { "type": "string" } },
}),
}
}
async fn run(&self, _args: serde_json::Value, _cx: &ToolContext) -> ToolResultData {
ToolResultData { content: vec![ContentBlock::Text("ran".to_string())], is_error: false }
}
}
struct Session {
actions: mpsc::Sender<Action>,
events: mpsc::Receiver<Event>,
sent: Sent,
}
/// An engine whose model answers from `script`, asking before every `bash` call.
fn session(script: Vec<Reply>) -> Session {
let dir = tempfile::tempdir().unwrap().keep();
let mut store = JsonlSessionStore::new(dir.join("sessions"));
let session_id = store.create("/tmp/branch-rerun").unwrap();
let sent = Arc::new(Mutex::new(Vec::new()));
let provider = Arc::new(Scripted { script: Mutex::new(script), sent: sent.clone() });
let mut tools = ToolRegistry::new();
tools.register(Arc::new(Bash));
let asks = PermissionConfig::Keyed(
[("bash".to_string(), PermValue::Scalar(PermissionAction::Ask))].into_iter().collect(),
)
.to_ruleset();
let config = EngineConfig {
session_store: std::sync::RwLock::new(Box::new(store)),
provider,
providers: std::collections::HashMap::new(),
model: ModelInfo::default(),
tool_registry: tools,
hook_registry: HookRegistry::new(),
session_id,
system_prompt: None,
todo_store: None,
compaction_settings: CompactionSettings::default(),
permission: PermissionSettings {
engine: PermissionEngine::new(asks, vec![], false),
ask: AskResolution::Prompt,
},
modes: Default::default(),
reviewer: None,
formatter_registry: None,
spend_sink: None,
guided_review_project_root: None,
invoked_from: None,
external_work: None,
review_context_files: Vec::new(),
review_exclude: Vec::new(),
remember_choices_in: None,
cache_warming: luxe_core::cache_warming::CacheWarming::Off,
review_insight_provider: None,
github: None,
permissions_file: None,
max_tool_iterations: Some(8),
startup_notices: Vec::new(),
action_tx: None,
};
let (actions, events) = Engine::spawn(config);
Session { actions, events, sent }
}
impl Session {
/// Sends `action` and waits for the turn it starts to end, answering any prompt with `reply`.
async fn turn(&mut self, action: Action, reply: PermissionReply) {
self.actions.send(action).await.unwrap();
loop {
let event = tokio::time::timeout(std::time::Duration::from_secs(5), self.events.recv())
.await
.expect("the turn ended rather than hanging")
.expect("the engine is running");
match event {
Event::PermissionRequested { id, .. } => {
self.actions.send(Action::RespondPermission { id, reply }).await.unwrap();
}
Event::TurnEnded { .. } => return,
_ => {}
}
}
}
async fn say(&mut self, text: &str) {
let submit = Action::SubmitUserMessage {
content: text.to_string(),
images: Vec::new(),
remediation: None,
};
self.turn(submit, PermissionReply::Reject).await;
}
/// The conversation the last provider call carried, user and assistant text only.
fn last_sent(&self) -> Vec<(Role, String)> {
self.sent.lock().unwrap().last().cloned().unwrap_or_default()
}
}
fn user(text: &str) -> (Role, String) {
(Role::User, text.to_string())
}
fn assistant(text: &str) -> (Role, String) {
(Role::Assistant, text.to_string())
}
/// An edit replaces the message it edits: the model sees the edited message and not the original
/// or its reply, which it used to answer as a pair.
#[tokio::test]
async fn an_edit_replaces_the_message_it_edits() {
let mut session = session(vec![Reply::Text("first answer"), Reply::Text("second answer")]);
session.say("teh question").await;
session
.turn(Action::EditResend { content: "the question".into() }, PermissionReply::Reject)
.await;
assert_eq!(session.last_sent(), [user("the question")]);
// And the conversation goes on from the edit
session.say("and then").await;
assert_eq!(
session.last_sent(),
[user("the question"), assistant("second answer"), user("and then")]
);
}
/// Regenerating replaces the answer, and the question goes in once.
#[tokio::test]
async fn a_regeneration_replaces_the_answer_and_asks_once() {
let mut session = session(vec![Reply::Text("first answer")]);
session.say("the question").await;
session.turn(Action::Regenerate, PermissionReply::Reject).await;
assert_eq!(session.last_sent(), [user("the question")]);
}
/// Asking again after an answer puts the question after it, once.
#[tokio::test]
async fn asking_again_follows_the_answer() {
let mut session = session(vec![Reply::Text("first answer")]);
session.say("the question").await;
session.turn(Action::Retry, PermissionReply::Reject).await;
assert_eq!(
session.last_sent(),
[user("the question"), assistant("first answer"), user("the question")]
);
}
/// Asking again after a turn that failed asks in place, rather than sending the question twice in
/// a row.
#[tokio::test]
async fn asking_again_after_a_failure_asks_in_place() {
let mut session = session(vec![Reply::Fail]);
session.say("the question").await;
session.turn(Action::Retry, PermissionReply::Reject).await;
assert_eq!(session.last_sent(), [user("the question")]);
}
/// A permission prompt inside an edited turn is answered, and the turn ends. Run inline, the
/// engine read no actions until the turn was over, so the answer never arrived and the turn waited
/// on it for good.
#[tokio::test]
async fn a_prompt_inside_an_edited_turn_is_answered() {
let mut session =
session(vec![Reply::Text("first answer"), Reply::Bash, Reply::Text("left it")]);
session.say("commit it").await;
session
.turn(Action::EditResend { content: "commit it now".into() }, PermissionReply::Reject)
.await;
let sent = session.last_sent();
assert_eq!(sent.first(), Some(&user("commit it now")));
assert!(sent.iter().any(|(_, text)| text == "<result>"), "the refusal reached the model");
}