Always use streamed remote compaction for supported providers (#44255)

## What changed

- Route manual and automatic remote compaction through the v2 implementation.
- Retire the `remote_compaction_v2` feature toggle while continuing to advertise it in the model client beta features header.
- Migrate compaction tests from `/responses/compact` to streamed `/responses` requests with `compaction_trigger` items and remove legacy parity coverage.

## Testing

Add streamed compaction coverage for tool search output trimming and session base instruction token estimates. Update app-server tests to verify started/completed notifications for both local and streamed remote compaction.

GitOrigin-RevId: cea392a6166a9b116ff83708fd9126d5546476e4
This commit is contained in:
rka-oai authored and copyberry committed 2026-09-09 18:04:04 +00:00
1 parent 2617ed2e1c
commit 3dc1e2a584
25 files changed
+675 -5635

No files matched your search

+13 -21
View File
@@ -1088,28 +1088,20 @@ impl Session {
///
/// `ModelClient` is session-scoped and intentionally does not depend on the full `Config`, so
/// we precompute the comma-separated list of enabled experimental feature keys at session
/// creation time and thread it into the client.
/// creation time and thread it into the client. Remote compaction stays advertised unconditionally.
fn build_model_client_beta_features_header(config: &Config) -> Option<String> {
let beta_features_header = FEATURES
.iter()
.filter_map(|spec| {
let advertise_in_model_client_header =
spec.stage.experimental_menu_description().is_some()
|| spec.id == Feature::RemoteCompactionV2;
if advertise_in_model_client_header && config.features.enabled(spec.id) {
Some(spec.key)
} else {
None
}
})
.collect::<Vec<_>>()
.join(",");
if beta_features_header.is_empty() {
None
} else {
Some(beta_features_header)
}
Some(
FEATURES
.iter()
.filter(|spec| {
spec.id == Feature::RemoteCompactionV2
|| (spec.stage.experimental_menu_description().is_some()
&& config.features.enabled(spec.id))
})
.map(|spec| spec.key)
.collect::<Vec<_>>()
.join(","),
)
}
async fn start_managed_network_proxy(
+20 -7
View File
@@ -11314,7 +11314,7 @@ async fn make_remote_compaction_session(
#[test_case(FirstAttempt::Succeeds; "primary succeeds")]
#[test_case(FirstAttempt::Retries; "fallback executes")]
#[tokio::test]
async fn legacy_compaction_retains_only_the_selected_step(first_attempt: FirstAttempt) {
async fn remote_compaction_v2_retains_only_the_selected_step(first_attempt: FirstAttempt) {
let server = responses::start_mock_server().await;
let (session, turn, events) = make_remote_compaction_session(&server.uri()).await;
session
@@ -11360,9 +11360,15 @@ async fn legacy_compaction_retains_only_the_selected_step(first_attempt: FirstAt
Some(Arc::as_ptr(&primary)),
);
let success = ResponseTemplate::new(/*status*/ 200).set_body_json(json!({
"output": [{ "type": "compaction", "encrypted_content": "summary" }]
}));
let success = ResponseTemplate::new(/*status*/ 200)
.insert_header("content-type", "text/event-stream")
.set_body_string(responses::sse(vec![
json!({
"type": "response.output_item.done",
"item": { "type": "compaction", "encrypted_content": "summary" },
}),
responses::ev_completed("compact-response"),
]));
let replies = match first_attempt {
FirstAttempt::Succeeds => vec![success],
FirstAttempt::Retries => vec![
@@ -11371,12 +11377,13 @@ async fn legacy_compaction_retains_only_the_selected_step(first_attempt: FirstAt
success,
],
};
let requests = responses::mount_compact_response_sequence(&server, replies).await;
crate::compact_remote::run_inline_remote_auto_compact_task(
let requests = responses::mount_response_sequence(&server, replies).await;
let mut client_session = session.services.model_client.new_session();
crate::compact_remote_v2::run_inline_remote_auto_compact_task(
Arc::clone(&session),
Arc::clone(&primary),
Some(Arc::clone(&fallback)),
Arc::new(OnceLock::new()),
&mut client_session,
InitialContextInjection::DoNotInject,
CompactionReason::ModelDownshift,
CompactionPhase::PreTurn,
@@ -11405,6 +11412,12 @@ async fn legacy_compaction_retains_only_the_selected_step(first_attempt: FirstAt
.collect::<Vec<_>>(),
models,
);
assert!(
requests
.requests()
.iter()
.all(|request| request.inputs_of_type("compaction_trigger").len() == 1),
);
session.abort_all_tasks(TurnAbortReason::Interrupted).await;
recv_terminal_event(&events, TerminalEventKind::TurnAborted).await;
}
@@ -39,11 +39,9 @@ use codex_network_proxy::NetworkProtocol;
use codex_protocol::config_types::ApprovalsReviewer;
use codex_protocol::error::SandboxErr;
use codex_protocol::models::AdditionalPermissionProfile as PermissionProfile;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ContentItemKind;
use codex_protocol::models::NetworkPermissions;
use codex_protocol::models::ResponseInputItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::protocol::AskForApproval;
use codex_protocol::request_permissions::PermissionGrantScope;
use codex_protocol::request_permissions::RequestPermissionProfile;
@@ -1053,7 +1051,7 @@ async fn guardian_allows_unified_exec_additional_permissions_requests_past_polic
}
#[tokio::test]
async fn process_compacted_history_preserves_separate_guardian_developer_message() {
async fn compaction_initial_context_preserves_separate_guardian_developer_message() {
let (session, mut turn_context) = make_session_and_context().await;
update_turn_settings_for_test(&mut turn_context, |settings| {
update_selected_settings_for_test(settings, |selected| {
@@ -1084,40 +1082,20 @@ async fn process_compacted_history_preserves_separate_guardian_developer_message
step_context,
};
let (refreshed, _) = crate::compact_remote::process_compacted_history(
&session,
vec![
ResponseItem::Message {
id: None,
role: "developer".to_string(),
content: vec![ContentItem::InputText {
text: "stale developer message".to_string(),
}],
phase: None,
internal_chat_message_metadata_passthrough: None,
},
ResponseItem::Message {
id: None,
role: "user".to_string(),
content: vec![ContentItem::InputText {
text: "summary".to_string(),
}],
phase: None,
internal_chat_message_metadata_passthrough: None,
},
],
&initial_context_injection,
)
.await;
let (refreshed, _) =
crate::compact::build_compaction_initial_context(&session, &initial_context_injection)
.await;
let developer_messages = refreshed
.iter()
.filter_map(|item| match item {
.filter_map(|envelope| match &envelope.item {
ResponseItem::Message { role, content, .. } if role == "developer" => {
crate::content_items_to_text(content).map(|text| {
(
text,
item.executed_tool_call_metadata()
envelope
.item
.executed_tool_call_metadata()
.and_then(|metadata| metadata.content_item_kinds.clone()),
)
})
@@ -1126,11 +1104,6 @@ async fn process_compacted_history_preserves_separate_guardian_developer_message
})
.collect::<Vec<_>>();
assert!(
!developer_messages
.iter()
.any(|(message, _)| message.contains("stale developer message"))
);
assert!(
!developer_messages
.iter()
+1 -24
View File
@@ -9,7 +9,6 @@ use crate::client_common::Prompt;
use crate::client_common::ResponseEvent;
use crate::compact::InitialContextInjection;
use crate::compact::run_inline_auto_compact_task;
use crate::compact_remote::run_inline_remote_auto_compact_task;
use crate::compact_remote_v2::run_inline_remote_auto_compact_task as run_inline_remote_auto_compact_task_v2;
use crate::connectors;
use crate::context::ContextualUserFragment;
@@ -1315,12 +1314,7 @@ async fn run_auto_compact(
}
match turn_context.provider.capabilities().remote_compaction {
RemoteCompactionSupport::V2
if turn_context
.config
.features
.enabled(Feature::RemoteCompactionV2) =>
{
RemoteCompactionSupport::V2 => {
emit_compact_metric(
&sess.services.session_telemetry,
"remote_v2",
@@ -1337,23 +1331,6 @@ async fn run_auto_compact(
)
.await?;
}
RemoteCompactionSupport::V2 => {
emit_compact_metric(
&sess.services.session_telemetry,
"remote",
/*manual*/ false,
);
run_inline_remote_auto_compact_task(
Arc::clone(sess),
step_context,
fallback_step_context,
client_session.turn_state(),
initial_context_injection,
reason,
phase,
)
.await?;
}
RemoteCompactionSupport::Unsupported => {
emit_compact_metric(
&sess.services.session_telemetry,