Initial public release
Luxe 0.1.0: a coding agent harness for the terminal. See README.md to get started and CHANGELOG.md for releases.
This commit is contained in:
480
crates/luxe-core/tests/integration_usage_ledger.rs
Normal file
480
crates/luxe-core/tests/integration_usage_ledger.rs
Normal file
@@ -0,0 +1,480 @@
|
||||
//! Regression: a session's reported spend is the sum of every provider call it made.
|
||||
//!
|
||||
//! The defect this covers cost real money to find. `Event::TurnUsage` is emitted **per provider
|
||||
//! call**, and a turn that runs tools makes many; the status reducer read them as cumulative-per-
|
||||
//! turn and added only the positive delta, so a forty-seven-call turn reported roughly its largest
|
||||
//! single call. A real session billed at $0.47 displayed $0.09.
|
||||
//!
|
||||
//! What is asserted here:
|
||||
//!
|
||||
//! - the running total equals the sum of the per-call events, for a turn with several calls;
|
||||
//! - the spend is written to the transcript and a resumed session adopts it;
|
||||
//! - work that is not the agent turn (the compaction summarizer) is billed to its own source
|
||||
//! rather than vanishing.
|
||||
|
||||
#![allow(clippy::unwrap_used)]
|
||||
|
||||
use std::collections::HashMap;
|
||||
use std::sync::atomic::{AtomicUsize, Ordering};
|
||||
use std::sync::Arc;
|
||||
|
||||
use futures::stream;
|
||||
use futures::stream::BoxStream;
|
||||
use luxe_core::engine::{Action, Engine, EngineConfig};
|
||||
use luxe_core::event::Event;
|
||||
use luxe_core::message::ContentBlock;
|
||||
use luxe_core::provider::{
|
||||
CompletionRequest, Cost, ModelInfo, ModelRef, Provider, ProviderError, ProviderEvent,
|
||||
StopReason, ToolSchema, Usage,
|
||||
};
|
||||
use luxe_core::session::{EntryKind, JsonlSessionStore, SessionId, SessionStore};
|
||||
use luxe_core::tool::{Tool, ToolContext, ToolRegistry, ToolResultData};
|
||||
use luxe_core::usage::{SessionUsage, UsageSource};
|
||||
use luxe_core::HookRegistry;
|
||||
|
||||
/// `claude-sonnet-5`'s published rates.
|
||||
fn sonnet() -> Cost {
|
||||
Cost {
|
||||
input: 2.0,
|
||||
output: 10.0,
|
||||
cache_read: 0.2,
|
||||
cache_write: 2.5,
|
||||
cache_write_1h: 0.0,
|
||||
tiers: Vec::new(),
|
||||
}
|
||||
}
|
||||
|
||||
fn priced_model() -> ModelInfo {
|
||||
ModelInfo {
|
||||
r#ref: ModelRef::new("anthropic", "claude-sonnet-5"),
|
||||
context_window: 1_000_000,
|
||||
max_tokens: 64_000,
|
||||
cost: sonnet(),
|
||||
..Default::default()
|
||||
}
|
||||
}
|
||||
|
||||
/// Calls a no-op tool `tool_calls` times, then answers — one provider call per step, each with its
|
||||
/// own usage over a context that grows the way a real turn's does.
|
||||
struct MultiCallProvider {
|
||||
tool_calls: usize,
|
||||
seen: AtomicUsize,
|
||||
}
|
||||
|
||||
#[async_trait::async_trait]
|
||||
impl Provider for MultiCallProvider {
|
||||
fn name(&self) -> &str {
|
||||
"multi-call"
|
||||
}
|
||||
|
||||
async fn stream(
|
||||
&self,
|
||||
_request: CompletionRequest,
|
||||
_cancel: tokio::sync::watch::Receiver<bool>,
|
||||
) -> Result<BoxStream<'static, ProviderEvent>, ProviderError> {
|
||||
let step = self.seen.fetch_add(1, Ordering::SeqCst);
|
||||
let prompt = 10_000 + (step as u32) * 700;
|
||||
let usage = Usage {
|
||||
prompt_tokens: prompt,
|
||||
completion_tokens: 250,
|
||||
total_tokens: prompt + 250,
|
||||
cache_read_tokens: prompt.saturating_sub(600),
|
||||
cache_write_tokens: 600,
|
||||
reasoning_tokens: 0,
|
||||
cache_write_1h_tokens: 0,
|
||||
};
|
||||
let mut events = Vec::new();
|
||||
if step < self.tool_calls {
|
||||
events.push(ProviderEvent::ToolCallStart {
|
||||
index: 0,
|
||||
id: format!("call-{step}"),
|
||||
name: "noop".into(),
|
||||
});
|
||||
events.push(ProviderEvent::ToolCallDelta { index: 0, json_delta: "{}".into() });
|
||||
events.push(ProviderEvent::Usage(usage));
|
||||
events.push(ProviderEvent::Done { reason: StopReason::ToolCall });
|
||||
} else {
|
||||
events.push(ProviderEvent::TextDelta { index: 0, delta: "done".into() });
|
||||
events.push(ProviderEvent::Usage(usage));
|
||||
events.push(ProviderEvent::Done { reason: StopReason::EndTurn });
|
||||
}
|
||||
Ok(Box::pin(stream::iter(events)))
|
||||
}
|
||||
}
|
||||
|
||||
struct NoopTool;
|
||||
|
||||
#[async_trait::async_trait]
|
||||
impl Tool for NoopTool {
|
||||
fn name(&self) -> &str {
|
||||
"noop"
|
||||
}
|
||||
|
||||
fn schema(&self) -> ToolSchema {
|
||||
ToolSchema {
|
||||
name: "noop".into(),
|
||||
description: "Does nothing.".into(),
|
||||
parameters: serde_json::json!({"type": "object", "properties": {}}),
|
||||
}
|
||||
}
|
||||
|
||||
async fn run(&self, _args: serde_json::Value, _cx: &ToolContext) -> ToolResultData {
|
||||
ToolResultData { content: vec![ContentBlock::Text("ok".into())], is_error: false }
|
||||
}
|
||||
}
|
||||
|
||||
struct Harness {
|
||||
action_tx: tokio::sync::mpsc::Sender<Action>,
|
||||
event_rx: tokio::sync::mpsc::Receiver<Event>,
|
||||
root: std::path::PathBuf,
|
||||
session: SessionId,
|
||||
_tmp: tempfile::TempDir,
|
||||
}
|
||||
|
||||
fn harness(provider: Arc<dyn Provider + Send + Sync>, model: ModelInfo) -> Harness {
|
||||
harness_with(provider, model, luxe_core::compaction::CompactionSettings::default())
|
||||
}
|
||||
|
||||
fn harness_with(
|
||||
provider: Arc<dyn Provider + Send + Sync>,
|
||||
model: ModelInfo,
|
||||
compaction_settings: luxe_core::compaction::CompactionSettings,
|
||||
) -> Harness {
|
||||
let tmp = tempfile::tempdir().unwrap();
|
||||
let root = tmp.path().join("sessions");
|
||||
let mut store = JsonlSessionStore::new(root.clone());
|
||||
let session = store.create("/tmp/test").unwrap();
|
||||
let mut tools = ToolRegistry::new();
|
||||
tools.register(Arc::new(NoopTool));
|
||||
let config = EngineConfig {
|
||||
session_store: std::sync::RwLock::new(Box::new(store)),
|
||||
provider,
|
||||
providers: HashMap::new(),
|
||||
model,
|
||||
tool_registry: tools,
|
||||
hook_registry: HookRegistry::new(),
|
||||
session_id: session.clone(),
|
||||
system_prompt: None,
|
||||
todo_store: None,
|
||||
compaction_settings,
|
||||
permission: Default::default(),
|
||||
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_runtime_state: false,
|
||||
cache_warming: luxe_core::cache_warming::CacheWarming::Off,
|
||||
review_insight_provider: None,
|
||||
github: None,
|
||||
permissions_file: None,
|
||||
max_tool_iterations: luxe_core::engine::DEFAULT_MAX_TOOL_ITERATIONS,
|
||||
startup_notices: Vec::new(),
|
||||
action_tx: None,
|
||||
};
|
||||
let (action_tx, event_rx) = Engine::spawn(config);
|
||||
Harness { action_tx, event_rx, root, session, _tmp: tmp }
|
||||
}
|
||||
|
||||
/// Everything one turn reported: each call's cost, and the last running total published.
|
||||
#[derive(Default)]
|
||||
struct TurnReport {
|
||||
call_costs: Vec<f64>,
|
||||
calls: usize,
|
||||
session: Option<SessionUsage>,
|
||||
}
|
||||
|
||||
async fn run_turn(harness: &mut Harness) -> TurnReport {
|
||||
harness
|
||||
.action_tx
|
||||
.send(Action::SubmitUserMessage {
|
||||
content: "go".into(),
|
||||
images: Vec::new(),
|
||||
remediation: None,
|
||||
})
|
||||
.await
|
||||
.unwrap();
|
||||
|
||||
let mut report = TurnReport::default();
|
||||
let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(20);
|
||||
while tokio::time::Instant::now() < deadline {
|
||||
match tokio::time::timeout(std::time::Duration::from_millis(500), harness.event_rx.recv())
|
||||
.await
|
||||
{
|
||||
Ok(Some(Event::TurnUsage { cost_usd, .. })) => {
|
||||
report.calls += 1;
|
||||
if let Some(cost) = cost_usd {
|
||||
report.call_costs.push(cost);
|
||||
}
|
||||
}
|
||||
Ok(Some(Event::SessionUsage(usage))) => report.session = Some(usage),
|
||||
Ok(Some(Event::TurnEnded { .. })) => break,
|
||||
Ok(Some(_)) | Err(_) => {}
|
||||
Ok(None) => break,
|
||||
}
|
||||
}
|
||||
report
|
||||
}
|
||||
|
||||
async fn drain(harness: &mut Harness) {
|
||||
while let Ok(Some(_)) =
|
||||
tokio::time::timeout(std::time::Duration::from_millis(100), harness.event_rx.recv()).await
|
||||
{
|
||||
}
|
||||
}
|
||||
|
||||
/// The headline regression. A turn with eleven provider calls must report all eleven.
|
||||
#[tokio::test]
|
||||
async fn a_turn_with_many_tool_calls_reports_the_sum_of_its_calls() {
|
||||
let provider = Arc::new(MultiCallProvider { tool_calls: 10, seen: AtomicUsize::new(0) });
|
||||
let mut harness = harness(provider, priced_model());
|
||||
assert!(matches!(harness.event_rx.recv().await, Some(Event::SessionLoaded { .. })));
|
||||
drain(&mut harness).await;
|
||||
|
||||
let report = run_turn(&mut harness).await;
|
||||
assert_eq!(report.calls, 11, "ten tool steps plus the final answer");
|
||||
|
||||
let expected: f64 = report.call_costs.iter().sum();
|
||||
let session = report.session.expect("a running total was published");
|
||||
assert_eq!(session.total.calls, 11);
|
||||
assert!(
|
||||
(session.total.cost_usd - expected).abs() < 1e-9,
|
||||
"session total must equal the sum of its calls: {} vs {expected}",
|
||||
session.total.cost_usd
|
||||
);
|
||||
|
||||
// And the old arithmetic — treating each event as a cumulative restatement — really did
|
||||
// collapse to roughly one call. Guard the shape of the bug, not just the number.
|
||||
let mut delta_accumulated = 0.0;
|
||||
let mut previous = 0.0;
|
||||
for cost in &report.call_costs {
|
||||
delta_accumulated += (cost - previous).max(0.0);
|
||||
previous = *cost;
|
||||
}
|
||||
assert!(
|
||||
session.total.cost_usd > delta_accumulated * 3.0,
|
||||
"the regression would report ${delta_accumulated} instead of ${}",
|
||||
session.total.cost_usd
|
||||
);
|
||||
}
|
||||
|
||||
/// Spend is written to the transcript, and a session resumed from it continues the same tally.
|
||||
#[tokio::test]
|
||||
async fn spend_is_persisted_and_a_resumed_session_adopts_it() {
|
||||
let provider = Arc::new(MultiCallProvider { tool_calls: 3, seen: AtomicUsize::new(0) });
|
||||
let mut harness = harness(provider, priced_model());
|
||||
assert!(matches!(harness.event_rx.recv().await, Some(Event::SessionLoaded { .. })));
|
||||
drain(&mut harness).await;
|
||||
|
||||
let report = run_turn(&mut harness).await;
|
||||
let live = report.session.expect("a running total was published");
|
||||
assert_eq!(live.total.calls, 4);
|
||||
|
||||
// The transcript carries it.
|
||||
let store = JsonlSessionStore::new(harness.root.clone());
|
||||
let entries = store.load(&harness.session).unwrap();
|
||||
let persisted: Vec<_> = entries
|
||||
.iter()
|
||||
.filter_map(|entry| match &entry.kind {
|
||||
EntryKind::Usage { source, model, totals } => Some((*source, model.clone(), *totals)),
|
||||
_ => None,
|
||||
})
|
||||
.collect();
|
||||
assert!(!persisted.is_empty(), "usage must reach the transcript");
|
||||
let persisted_calls: u64 = persisted.iter().map(|(_, _, totals)| totals.calls).sum();
|
||||
let persisted_cost: f64 = persisted.iter().map(|(_, _, totals)| totals.cost_usd).sum();
|
||||
assert_eq!(persisted_calls, live.total.calls, "every call is on disk");
|
||||
assert!((persisted_cost - live.total.cost_usd).abs() < 1e-9);
|
||||
assert!(persisted.iter().all(|(_, model, _)| model == "claude-sonnet-5"), "priced by model");
|
||||
|
||||
// The conversation still resumes: usage entries are chained, so the leaf is still the tip of
|
||||
// the conversation and the path back to the root is intact.
|
||||
let leaf = luxe_core::session::latest_leaf(&entries).expect("a leaf");
|
||||
let path = store.active_path(&harness.session, &leaf).unwrap();
|
||||
let messages = path.iter().filter(|entry| matches!(entry.kind, EntryKind::Message(_))).count();
|
||||
assert!(messages >= 5, "user + 4 assistant replies survive on the active path: {messages}");
|
||||
|
||||
// Resuming adopts the tally rather than starting at zero.
|
||||
harness.action_tx.send(Action::LoadSession(harness.session.to_string())).await.unwrap();
|
||||
let mut resumed: Option<SessionUsage> = None;
|
||||
let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(5);
|
||||
while tokio::time::Instant::now() < deadline {
|
||||
match tokio::time::timeout(std::time::Duration::from_millis(300), harness.event_rx.recv())
|
||||
.await
|
||||
{
|
||||
Ok(Some(Event::SessionUsage(usage))) => {
|
||||
resumed = Some(usage);
|
||||
break;
|
||||
}
|
||||
Ok(Some(_)) | Err(_) => {}
|
||||
Ok(None) => break,
|
||||
}
|
||||
}
|
||||
let resumed = resumed.expect("a resumed session republishes its tally");
|
||||
assert_eq!(resumed.total.calls, live.total.calls, "resume continues the same tally");
|
||||
assert!((resumed.total.cost_usd - live.total.cost_usd).abs() < 1e-9);
|
||||
}
|
||||
|
||||
/// A cache refresh is persisted as its own entry kind, so an older Luxe can skip it, and a resumed
|
||||
/// session still counts it, under its own source rather than as conversation.
|
||||
#[tokio::test]
|
||||
async fn a_resumed_session_counts_its_cache_refreshes_under_their_own_source() {
|
||||
let provider = Arc::new(MultiCallProvider { tool_calls: 0, seen: AtomicUsize::new(0) });
|
||||
let mut harness = harness(provider, priced_model());
|
||||
assert!(matches!(harness.event_rx.recv().await, Some(Event::SessionLoaded { .. })));
|
||||
drain(&mut harness).await;
|
||||
let live = run_turn(&mut harness).await.session.expect("a running total was published");
|
||||
|
||||
let mut store = JsonlSessionStore::new(harness.root.clone());
|
||||
let leaf = luxe_core::session::latest_leaf(&store.load(&harness.session).unwrap());
|
||||
let refresh = luxe_core::usage::UsageTotals {
|
||||
calls: 1,
|
||||
prompt_tokens: 11_512,
|
||||
completion_tokens: 1,
|
||||
cache_read_tokens: 11_510,
|
||||
cost_usd: 0.00579,
|
||||
..Default::default()
|
||||
};
|
||||
store
|
||||
.append(
|
||||
&harness.session,
|
||||
leaf,
|
||||
EntryKind::CacheWarm { model: "claude-sonnet-5".into(), totals: refresh },
|
||||
)
|
||||
.unwrap();
|
||||
|
||||
harness.action_tx.send(Action::LoadSession(harness.session.to_string())).await.unwrap();
|
||||
let mut resumed: Option<SessionUsage> = None;
|
||||
let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(5);
|
||||
while resumed.is_none() && tokio::time::Instant::now() < deadline {
|
||||
if let Ok(Some(Event::SessionUsage(usage))) =
|
||||
tokio::time::timeout(std::time::Duration::from_millis(300), harness.event_rx.recv())
|
||||
.await
|
||||
{
|
||||
resumed = Some(usage);
|
||||
}
|
||||
}
|
||||
let resumed = resumed.expect("a resumed session republishes its tally");
|
||||
let warming = resumed
|
||||
.by_source
|
||||
.iter()
|
||||
.find(|entry| entry.source == UsageSource::CacheWarm)
|
||||
.expect("the refresh is in the breakdown");
|
||||
assert_eq!(warming.totals.calls, 1);
|
||||
assert_eq!(resumed.total.calls, live.total.calls + 1, "and in the total");
|
||||
let agent = resumed.by_source.iter().find(|entry| entry.source == UsageSource::Agent);
|
||||
assert_eq!(
|
||||
agent.map(|entry| entry.totals.calls),
|
||||
Some(live.total.calls),
|
||||
"not as conversation"
|
||||
);
|
||||
}
|
||||
|
||||
/// A model with no published rates reports tokens but claims no cost.
|
||||
#[tokio::test]
|
||||
async fn an_unpriced_model_reports_tokens_without_inventing_a_cost() {
|
||||
let provider = Arc::new(MultiCallProvider { tool_calls: 1, seen: AtomicUsize::new(0) });
|
||||
let unpriced = ModelInfo {
|
||||
r#ref: ModelRef::new("lmstudio", "local"),
|
||||
context_window: 200_000,
|
||||
..Default::default()
|
||||
};
|
||||
let mut harness = harness(provider, unpriced);
|
||||
assert!(matches!(harness.event_rx.recv().await, Some(Event::SessionLoaded { .. })));
|
||||
drain(&mut harness).await;
|
||||
|
||||
let report = run_turn(&mut harness).await;
|
||||
let session = report.session.expect("a running total was published");
|
||||
assert!(report.call_costs.is_empty(), "no call claims a price");
|
||||
assert_eq!(session.total.unpriced_calls, session.total.calls);
|
||||
assert!(session.total.prompt_tokens > 0, "tokens are still counted");
|
||||
assert_eq!(session.total.cost_label(), None, "and no cost is shown at all");
|
||||
}
|
||||
|
||||
/// Compaction is the one provider call the user never asked for. Its input is the whole slice of
|
||||
/// context being summarized, which on a long session makes it the single most expensive call — and
|
||||
/// it used to be spent entirely off the books.
|
||||
#[tokio::test]
|
||||
async fn the_compaction_summarizer_bills_to_its_own_source() {
|
||||
let provider = Arc::new(MultiCallProvider { tool_calls: 0, seen: AtomicUsize::new(0) });
|
||||
// A window small enough that the very first reply puts the session over the threshold.
|
||||
let model = ModelInfo { context_window: 12_000, max_tokens: 1_000, ..priced_model() };
|
||||
let settings = luxe_core::compaction::CompactionSettings {
|
||||
enabled: true,
|
||||
compact_at_fraction: 0.1,
|
||||
keep_recent_tokens: 1,
|
||||
reserve_tokens: 500,
|
||||
..Default::default()
|
||||
};
|
||||
let mut harness = harness_with(provider, model, settings);
|
||||
assert!(matches!(harness.event_rx.recv().await, Some(Event::SessionLoaded { .. })));
|
||||
drain(&mut harness).await;
|
||||
|
||||
// The threshold is checked at the *start* of a turn against the previous turn's occupancy, so
|
||||
// the second submission is the one that compacts.
|
||||
run_turn(&mut harness).await;
|
||||
harness
|
||||
.action_tx
|
||||
.send(Action::SubmitUserMessage {
|
||||
content: "go again".into(),
|
||||
images: Vec::new(),
|
||||
remediation: None,
|
||||
})
|
||||
.await
|
||||
.unwrap();
|
||||
|
||||
let mut sources = Vec::new();
|
||||
let mut session = None;
|
||||
let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(10);
|
||||
let mut ended = false;
|
||||
while tokio::time::Instant::now() < deadline {
|
||||
match tokio::time::timeout(std::time::Duration::from_millis(400), harness.event_rx.recv())
|
||||
.await
|
||||
{
|
||||
Ok(Some(Event::TurnUsage { source, .. })) => sources.push(source),
|
||||
Ok(Some(Event::SessionUsage(usage))) => session = Some(usage),
|
||||
Ok(Some(Event::TurnEnded { .. })) => ended = true,
|
||||
// Compaction runs after the turn ends, so keep listening past it.
|
||||
Ok(Some(_)) => {}
|
||||
Ok(None) => break,
|
||||
Err(_) if ended => break,
|
||||
Err(_) => {}
|
||||
}
|
||||
}
|
||||
|
||||
assert!(
|
||||
sources.contains(&UsageSource::Compaction),
|
||||
"the summarizer's call must be reported: {sources:?}"
|
||||
);
|
||||
let session = session.expect("a running total");
|
||||
let compaction = session.source(UsageSource::Compaction).expect("compaction has its own share");
|
||||
assert!(compaction.calls > 0);
|
||||
assert!(compaction.cost_usd > 0.0, "and it is priced");
|
||||
assert!(
|
||||
session.total.cost_usd > session.source(UsageSource::Agent).unwrap().cost_usd,
|
||||
"the roll-up includes work the turn did not do"
|
||||
);
|
||||
}
|
||||
|
||||
/// Every source the ledger knows about is reachable through the same path, and the roll-up is the
|
||||
/// sum of its parts — so a future call site cannot report into a total that nobody displays.
|
||||
#[tokio::test]
|
||||
async fn the_breakdown_sums_to_the_total() {
|
||||
let provider = Arc::new(MultiCallProvider { tool_calls: 2, seen: AtomicUsize::new(0) });
|
||||
let mut harness = harness(provider, priced_model());
|
||||
assert!(matches!(harness.event_rx.recv().await, Some(Event::SessionLoaded { .. })));
|
||||
drain(&mut harness).await;
|
||||
|
||||
let session = run_turn(&mut harness).await.session.expect("a running total");
|
||||
let summed: f64 = session.by_source.iter().map(|entry| entry.totals.cost_usd).sum();
|
||||
assert!((summed - session.total.cost_usd).abs() < 1e-9);
|
||||
assert_eq!(
|
||||
session.source(UsageSource::Agent).map(|totals| totals.calls),
|
||||
Some(session.total.calls),
|
||||
"an ordinary turn bills entirely to the agent"
|
||||
);
|
||||
}
|
||||
Reference in New Issue
Block a user