mirror of
https://github.com/openai/codex.git
synced 2026-11-15 21:50:46 +00:00
Support fractional rollout budgets and configurable context reminders (#61579)
## What changed - Add a positive, finite `features.rollout_budget.limit_budget` allowance while retaining `limit_tokens` compatibility. Reject configurations that specify both. - Preserve fractional balances in reminders and compare thresholds without rounding the balance down. - Accumulate charges with compensated summation and allow spending exactly the allowance, including floating-point rounding tolerance for cases such as `0.1 + 0.2` against `0.3`. - Add `reminder_at_context_start`, defaulting to `true`. Set it to `false` to suppress reminders at thread start and after compaction; combine it with an empty threshold list to disable budget reminders entirely. ## Testing Extend configuration and rollout-budget tests to cover fractional allowances, legacy integer persistence, invalid limits, threshold-only reminders, and enforcement at decimal equality and after overspending. GitOrigin-RevId: b8e7a2faa8d39e562bcae196364de419698bc67d
This commit is contained in:
@@ -51,7 +51,8 @@ async fn first_request_item_types_roles_and_content_annotations() -> Result<()>
|
||||
});
|
||||
config.rollout_budget = Some(RolloutBudgetConfig {
|
||||
enforce_limit: true,
|
||||
limit_tokens: 100,
|
||||
limit_budget: 100.0,
|
||||
reminder_at_context_start: true,
|
||||
reminder_at_remaining_tokens: Vec::new(),
|
||||
sampling_token_weight: 1.0,
|
||||
prefill_token_weight: 1.0,
|
||||
|
||||
@@ -280,7 +280,8 @@ async fn review_respects_complete_context_budget(
|
||||
config.current_time_reminder = Some(CurrentTimeReminderConfig::default());
|
||||
config.rollout_budget = Some(RolloutBudgetConfig {
|
||||
enforce_limit: true,
|
||||
limit_tokens: 1_000_000,
|
||||
limit_budget: 1_000_000.0,
|
||||
reminder_at_context_start: true,
|
||||
reminder_at_remaining_tokens: Vec::new(),
|
||||
sampling_token_weight: 1.0,
|
||||
prefill_token_weight: 1.0,
|
||||
|
||||
@@ -108,7 +108,8 @@ impl ThreadLifecycleContributor<Config> for HostAccounting {
|
||||
pub(super) fn rollout_budget() -> RolloutBudgetConfig {
|
||||
RolloutBudgetConfig {
|
||||
enforce_limit: true,
|
||||
limit_tokens: 100,
|
||||
limit_budget: 100.0,
|
||||
reminder_at_context_start: true,
|
||||
reminder_at_remaining_tokens: vec![75, 50, 25],
|
||||
sampling_token_weight: 1.0,
|
||||
prefill_token_weight: 1.0,
|
||||
@@ -123,7 +124,7 @@ fn rollout_budget_texts(request: &ResponsesRequest) -> Vec<String> {
|
||||
.collect()
|
||||
}
|
||||
|
||||
fn rollout_budget_message(remaining_tokens: i64) -> String {
|
||||
fn rollout_budget_message(remaining_tokens: impl std::fmt::Display) -> String {
|
||||
format!(
|
||||
"<rollout_budget>\nYou have {remaining_tokens} weighted tokens left in the shared session token budget.\n</rollout_budget>"
|
||||
)
|
||||
@@ -134,10 +135,11 @@ fn wire_request_contains(request: &wiremock::Request, text: &str) -> bool {
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
#[test_case(None ; "weighted token usage")]
|
||||
#[test_case(Some(40.5) ; "provider budget units")]
|
||||
#[test_case(None, true ; "weighted tokens and initial reminder")]
|
||||
#[test_case(Some(40.5), false ; "opaque units and threshold reminders only")]
|
||||
async fn adds_weighted_initial_and_threshold_reminders(
|
||||
rollout_budget_units: Option<f64>,
|
||||
initial_reminder: bool,
|
||||
) -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
@@ -167,11 +169,18 @@ async fn adds_weighted_initial_and_threshold_reminders(
|
||||
)
|
||||
.await;
|
||||
let test = test_codex()
|
||||
.with_config(|config| {
|
||||
.with_config(move |config| {
|
||||
let mut budget = rollout_budget();
|
||||
if !initial_reminder {
|
||||
// Starting at 100.75 has not crossed the threshold at 100.
|
||||
budget.limit_budget = 100.75;
|
||||
budget.reminder_at_remaining_tokens = vec![100];
|
||||
}
|
||||
config.rollout_budget = Some(RolloutBudgetConfig {
|
||||
reminder_at_context_start: initial_reminder,
|
||||
sampling_token_weight: 2.0,
|
||||
prefill_token_weight: 0.5,
|
||||
..rollout_budget()
|
||||
..budget
|
||||
});
|
||||
})
|
||||
.build(&server)
|
||||
@@ -181,22 +190,21 @@ async fn adds_weighted_initial_and_threshold_reminders(
|
||||
test.submit_turn("second turn").await?;
|
||||
|
||||
let requests = responses.requests();
|
||||
assert!(requests[0].has_content_kinds(&["rollout_budget.remaining_tokens"]));
|
||||
assert_eq!(
|
||||
rollout_budget_texts(&requests[0]),
|
||||
vec![rollout_budget_message(/*remaining_tokens*/ 100)]
|
||||
);
|
||||
assert_eq!(
|
||||
rollout_budget_texts(&requests[1]),
|
||||
vec![
|
||||
rollout_budget_message(/*remaining_tokens*/ 100),
|
||||
rollout_budget_message(if rollout_budget_units.is_some() {
|
||||
59
|
||||
} else {
|
||||
60
|
||||
}),
|
||||
]
|
||||
requests[0].has_content_kinds(&["rollout_budget.remaining_tokens"]),
|
||||
initial_reminder
|
||||
);
|
||||
let mut expected: Vec<_> = initial_reminder
|
||||
.then(|| rollout_budget_message(/*remaining_tokens*/ 100))
|
||||
.into_iter()
|
||||
.collect();
|
||||
assert_eq!(rollout_budget_texts(&requests[0]), expected);
|
||||
expected.push(rollout_budget_message(if rollout_budget_units.is_some() {
|
||||
60.25
|
||||
} else {
|
||||
60.0
|
||||
}));
|
||||
assert_eq!(rollout_budget_texts(&requests[1]), expected);
|
||||
|
||||
Ok(())
|
||||
}
|
||||
@@ -352,7 +360,7 @@ async fn subagent_usage_draws_from_the_shared_budget(host_accounting: bool) -> R
|
||||
.expect("test config should allow multi-agent v2");
|
||||
config.rollout_budget = Some(RolloutBudgetConfig {
|
||||
// Host success must not also charge the Local budget.
|
||||
limit_tokens: if host_accounting { 1 } else { 100 },
|
||||
limit_budget: if host_accounting { 1.0 } else { 100.0 },
|
||||
..rollout_budget()
|
||||
});
|
||||
})
|
||||
@@ -426,16 +434,31 @@ async fn exhausted_budget_respects_enforcement_on_current_and_later_turns(
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
let server = start_mock_server().await;
|
||||
let mut sequence = vec![sse(vec![
|
||||
ev_response_created("exhaust-budget"),
|
||||
ev_completed_with_tokens("exhaust-budget", /*total_tokens*/ 30),
|
||||
])];
|
||||
if owner == AgentRuntimeOwnership::Host || !enforce_limit {
|
||||
sequence.push(sse(vec![
|
||||
ev_response_created("already-exhausted"),
|
||||
ev_completed_with_tokens("already-exhausted", /*total_tokens*/ 1),
|
||||
]));
|
||||
}
|
||||
// Enforced: decimal equality is allowed; the next material charge exceeds it.
|
||||
// Advisory: a real overspend still permits a later request with a zero-balance reminder.
|
||||
let (limit_budget, reminder_at_remaining_tokens, charges): (f64, Vec<i64>, &[f64]) =
|
||||
if enforce_limit {
|
||||
(0.3, vec![], &[0.1, 0.2, 0.000001, 0.0])
|
||||
} else {
|
||||
(30.0, vec![20, 10], &[31.0, 1.0])
|
||||
};
|
||||
let crossing_turn = if enforce_limit { 2 } else { 0 };
|
||||
// The later user turn still runs, but an enforced Core budget refuses its model send.
|
||||
let model_charges = if enforce_limit && owner == AgentRuntimeOwnership::Controller {
|
||||
&charges[..=crossing_turn]
|
||||
} else {
|
||||
charges
|
||||
};
|
||||
let sequence = model_charges
|
||||
.iter()
|
||||
.enumerate()
|
||||
.map(|(turn, units)| {
|
||||
let id = format!("budget-{turn}");
|
||||
let mut completed = ev_completed_with_tokens(&id, /*total_tokens*/ 1);
|
||||
completed["response"]["usage"]["codex_rollout_budget_units"] = json!(units);
|
||||
sse(vec![ev_response_created(&id), completed])
|
||||
})
|
||||
.collect();
|
||||
let responses = mount_sse_sequence(&server, sequence).await;
|
||||
let mut extensions = ExtensionRegistryBuilder::new();
|
||||
if owner == AgentRuntimeOwnership::Host {
|
||||
@@ -446,8 +469,8 @@ async fn exhausted_budget_respects_enforcement_on_current_and_later_turns(
|
||||
.with_config(move |config| {
|
||||
config.rollout_budget = Some(RolloutBudgetConfig {
|
||||
enforce_limit,
|
||||
limit_tokens: 30,
|
||||
reminder_at_remaining_tokens: vec![20, 10],
|
||||
limit_budget,
|
||||
reminder_at_remaining_tokens,
|
||||
..rollout_budget()
|
||||
});
|
||||
})
|
||||
@@ -461,10 +484,10 @@ async fn exhausted_budget_respects_enforcement_on_current_and_later_turns(
|
||||
data.insert(late.clone() as Arc<dyn AgentUsageAccounting>);
|
||||
}
|
||||
|
||||
for prompt in ["exhaust the budget", "try another turn"] {
|
||||
for turn in 0..charges.len() {
|
||||
test.codex
|
||||
.start_or_steer_turn(TurnInputRequest::user_input(vec![UserInput::Text {
|
||||
text: prompt.to_string(),
|
||||
text: format!("Budget turn {turn}"),
|
||||
text_elements: Vec::new(),
|
||||
}]))
|
||||
.await?;
|
||||
@@ -473,7 +496,7 @@ async fn exhausted_budget_respects_enforcement_on_current_and_later_turns(
|
||||
matches!(event, EventMsg::Error(_) | EventMsg::TurnComplete(_))
|
||||
})
|
||||
.await;
|
||||
if owner == AgentRuntimeOwnership::Host || !enforce_limit {
|
||||
if turn < crossing_turn || owner == AgentRuntimeOwnership::Host || !enforce_limit {
|
||||
assert!(matches!(event, EventMsg::TurnComplete(done) if done.error.is_none()));
|
||||
continue;
|
||||
}
|
||||
@@ -490,7 +513,7 @@ async fn exhausted_budget_respects_enforcement_on_current_and_later_turns(
|
||||
|
||||
if owner == AgentRuntimeOwnership::Host {
|
||||
let requests = responses.requests();
|
||||
assert_eq!(requests.len(), 2);
|
||||
assert_eq!(requests.len(), charges.len());
|
||||
for request in &requests {
|
||||
assert!(rollout_budget_texts(request).is_empty());
|
||||
}
|
||||
@@ -557,7 +580,7 @@ async fn compaction_accounting_errors_fail_without_retry(
|
||||
.with_config(move |config| {
|
||||
config.model_provider.stream_max_retries = Some(2);
|
||||
config.rollout_budget = Some(RolloutBudgetConfig {
|
||||
limit_tokens: 10,
|
||||
limit_budget: 9.0,
|
||||
reminder_at_remaining_tokens: vec![5],
|
||||
..rollout_budget()
|
||||
});
|
||||
|
||||
Reference in New Issue
Block a user