mirror of
https://github.com/openai/codex.git
synced 2026-11-15 21:50:46 +00:00
Reserve item ID overhead before elastic context allocation (#61347)
## Why Elastic request allocation did not account for item IDs added after evidence selection. Their token cost could push the completed request over its input limit instead of allowing optional evidence to be dropped. ## What changed Reserve IDs for the possible new request items before allocating evidence, including their JSON escaping and field overhead in the global request budget. Assign the same reserved IDs to the selected items while preserving existing IDs and committed history. ## Testing Add coverage for escaped IDs and native tool call outputs, and a regression test with long IDs that verifies optional images are omitted while committed history and new transcript text are retained within budget. Update budget boundary tests to distinguish fixed rejection from elastic evidence reduction. GitOrigin-RevId: 621d6d83e3cc95efb9eef4976dff9b7799a6e7c8
This commit is contained in:
@@ -7,6 +7,7 @@
|
||||
use codex_context_fragments::ContextualUserFragment;
|
||||
use codex_core::context::GuardianBudgetOmission;
|
||||
use codex_guardian_context::RequestBudget;
|
||||
use codex_guardian_context::ReservedElasticItemIds;
|
||||
use codex_guardian_context::TranscriptCursor;
|
||||
use codex_guardian_context::TruncationObservation;
|
||||
use codex_guardian_context::estimate_input_tokens;
|
||||
@@ -64,17 +65,20 @@ impl LunaSamplingRequest {
|
||||
.map(|item| estimate_input_tokens(&item.item))
|
||||
.fold(0usize, usize::saturating_add);
|
||||
let maximum = max_input_tokens.saturating_sub(INPUT_TOKEN_MARGIN);
|
||||
if let Some(policy) = &evidence.elastic {
|
||||
context = context
|
||||
let ids = if let Some(policy) = &evidence.elastic {
|
||||
let prepared = context
|
||||
.prepare_elastic_entries(policy)
|
||||
.ok()?
|
||||
.prepare_delivery(history)
|
||||
.ok()?
|
||||
.ok()?;
|
||||
let ids = ReservedElasticItemIds::reserve(&prepared, &self.item_id_generator);
|
||||
context = prepared
|
||||
.enforce_budget(
|
||||
RequestBudget {
|
||||
max_input_tokens: maximum,
|
||||
existing_context_tokens: existing_context_tokens
|
||||
.saturating_add(setup_tokens),
|
||||
.saturating_add(setup_tokens)
|
||||
.saturating_add(ids.tokens()),
|
||||
},
|
||||
GuardianBudgetOmission
|
||||
.render()
|
||||
@@ -82,11 +86,17 @@ impl LunaSamplingRequest {
|
||||
.unwrap_or_default(),
|
||||
)
|
||||
.ok()?;
|
||||
}
|
||||
Some(ids)
|
||||
} else {
|
||||
None
|
||||
};
|
||||
let truncations = std::mem::take(&mut context.truncations);
|
||||
let section_costs = context.section_costs().collect();
|
||||
let (mut input, new_input_tokens) =
|
||||
LunaSampler::finish_input(self, setup, context.into_annotated_messages());
|
||||
let mut suffix = context.into_annotated_messages();
|
||||
if let Some(ids) = ids {
|
||||
ids.assign(&mut suffix).ok()?;
|
||||
}
|
||||
let (mut input, new_input_tokens) = LunaSampler::finish_input(self, setup, suffix);
|
||||
let input_tokens = existing_context_tokens.saturating_add(new_input_tokens);
|
||||
if input_tokens > maximum {
|
||||
return None;
|
||||
|
||||
@@ -1,4 +1,5 @@
|
||||
use super::*;
|
||||
use crate::async_scorer::sampler::INPUT_TOKEN_MARGIN;
|
||||
use crate::async_scorer::transcript::CollectedTranscript;
|
||||
use crate::async_scorer::transcript::ContextInput;
|
||||
use crate::async_scorer::transcript::TranscriptConfig;
|
||||
@@ -235,6 +236,7 @@ fn preparation_preserves_committed_prefix_and_selects_only_new_entries_and_image
|
||||
#[test_case::test_case(ContextBudgeting::Fixed; "fixed")]
|
||||
#[test_case::test_case(ContextBudgeting::Elastic; "elastic")]
|
||||
fn rejected_preparation_preserves_committed_history_and_cursor(mode: ContextBudgeting) {
|
||||
const FRESH_REQUEST_HEADROOM: usize = 128;
|
||||
let mut history = History {
|
||||
items: vec![responses::user_message_item("Inspect README.md.")],
|
||||
version: 1,
|
||||
@@ -244,7 +246,8 @@ fn rejected_preparation_preserves_committed_history_and_cursor(mode: ContextBudg
|
||||
let first = config
|
||||
.prepare_retained(&collect(&history, &[], mode), &mut state, INPUT_BUDGET)
|
||||
.unwrap();
|
||||
let limit = first.input_tokens + 256;
|
||||
// Leave room for the fresh request's allocation estimate and ID reservation.
|
||||
let limit = first.input_tokens + INPUT_TOKEN_MARGIN + FRESH_REQUEST_HEADROOM;
|
||||
let mut committed = first.input;
|
||||
committed.push(responses::user_message_item(&"output ".repeat(/*n*/ 400)).into());
|
||||
state.complete_review(first.cursor);
|
||||
@@ -278,6 +281,62 @@ fn rejected_preparation_preserves_committed_history_and_cursor(mode: ContextBudg
|
||||
assert_eq!(state.snapshot().unwrap().history(), &committed);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn elastic_fits_ids_without_resetting_committed_history() {
|
||||
const ID_PADDING_BYTES: usize = 1_024;
|
||||
let mode = ContextBudgeting::Elastic;
|
||||
let mut config = sample_request("parent");
|
||||
let original_generator = std::sync::Arc::clone(&config.item_id_generator);
|
||||
let padding = "x".repeat(ID_PADDING_BYTES);
|
||||
config.item_id_generator = std::sync::Arc::new(move |prefix| {
|
||||
ResponseItemId::from_server(format!("{}{padding}", original_generator(prefix)))
|
||||
});
|
||||
let mut history = History {
|
||||
items: vec![responses::user_message_item("Inspect README.md.")],
|
||||
version: 1,
|
||||
};
|
||||
let mut state = ConversationState::default();
|
||||
let first = config
|
||||
.prepare_retained(&collect(&history, &[], mode), &mut state, INPUT_BUDGET)
|
||||
.unwrap();
|
||||
state.complete_review(first.cursor);
|
||||
state.commit_snapshot(first.input.clone());
|
||||
history
|
||||
.items
|
||||
.push(responses::user_message_item("Inspect LICENSE next."));
|
||||
let image = ContentItem::InputImage {
|
||||
image: ImageReference::Inline {
|
||||
image_url: "data:image/png;base64,b3B0aW9uYWw=".to_owned(),
|
||||
},
|
||||
detail: None,
|
||||
};
|
||||
let evidence = collect(&history, std::slice::from_ref(&image), mode);
|
||||
let mut generous = ConversationState::from(state.snapshot().unwrap().clone());
|
||||
let full = config
|
||||
.prepare_retained(&evidence, &mut generous, INPUT_BUDGET)
|
||||
.unwrap();
|
||||
assert!(full.input.iter().any(|item| matches!(
|
||||
&item.item, ResponseItem::Message { content, .. } if content.contains(&image)
|
||||
)));
|
||||
let maximum = full.input_tokens - 1;
|
||||
let bounded = config
|
||||
.prepare_retained(&evidence, &mut state, maximum + INPUT_TOKEN_MARGIN)
|
||||
.expect("allocate evidence with room for IDs");
|
||||
assert_eq!(&bounded.input[..first.input.len()], first.input);
|
||||
assert!(bounded.input_tokens <= maximum);
|
||||
assert!(bounded.input.iter().all(|item| item.id().is_some()));
|
||||
let suffix = serde_json::to_string(
|
||||
&bounded.input[first.input.len()..]
|
||||
.iter()
|
||||
.map(|item| &item.item)
|
||||
.collect::<Vec<_>>(),
|
||||
)
|
||||
.unwrap();
|
||||
assert!(suffix.contains("Inspect LICENSE next."));
|
||||
assert!(!suffix.contains("data:image/png;base64,b3B0aW9uYWw="));
|
||||
assert!(suffix.contains("<guardian_context_omission>"));
|
||||
}
|
||||
|
||||
#[test_case::test_case(ContextBudgeting::Fixed; "fixed")]
|
||||
#[test_case::test_case(ContextBudgeting::Elastic; "elastic")]
|
||||
fn complete_budget_counts_setup_checkpoint_images_and_trusted_evidence(mode: ContextBudgeting) {
|
||||
@@ -307,13 +366,32 @@ fn complete_budget_counts_setup_checkpoint_images_and_trusted_evidence(mode: Con
|
||||
config.parent_compaction.as_ref()
|
||||
);
|
||||
assert!(request.input_tokens > 10_000);
|
||||
assert!(
|
||||
config
|
||||
.prepare_retained(&evidence, &mut state, request.input_tokens + 255)
|
||||
.is_none()
|
||||
);
|
||||
// The complete input fits exactly at the reserved-output boundary.
|
||||
config
|
||||
.prepare_retained(&evidence, &mut state, request.input_tokens + 256)
|
||||
.unwrap();
|
||||
let maximum = request.input_tokens - 1;
|
||||
let smaller = config.prepare_retained(&evidence, &mut state, maximum + INPUT_TOKEN_MARGIN);
|
||||
match mode {
|
||||
ContextBudgeting::Fixed => {
|
||||
assert!(smaller.is_none());
|
||||
let exact = config
|
||||
.prepare_retained(
|
||||
&evidence,
|
||||
&mut state,
|
||||
request.input_tokens + INPUT_TOKEN_MARGIN,
|
||||
)
|
||||
.unwrap();
|
||||
assert_eq!(exact.input_tokens, request.input_tokens);
|
||||
assert!(exact.input.iter().any(|item| matches!(
|
||||
&item.item, ResponseItem::Message { content, .. } if content.contains(&images[0])
|
||||
)));
|
||||
}
|
||||
ContextBudgeting::Elastic => {
|
||||
let smaller = smaller.expect("Elastic can shed optional evidence");
|
||||
assert!(smaller.input_tokens <= maximum);
|
||||
assert!(
|
||||
smaller
|
||||
.input
|
||||
.iter()
|
||||
.any(|item| Some(&item.item) == config.parent_compaction.as_ref())
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user