Enable remote compaction for Amazon Bedrock (#36981)

## What changed

- Add provider-owned remote compaction capabilities for unsupported, v1, and v2 protocols.
- Mark Amazon Bedrock as v1-only so manual and automatic compaction use `/v1/responses/compact`, even when the v2 feature is enabled.
- Preserve v2 support for OpenAI and Azure Responses providers and local compaction for unsupported providers.

## Testing

- Cover Amazon Bedrock manual and automatic remote compaction, including endpoint selection, authentication headers, model selection, and reuse of the compacted summary.

GitOrigin-RevId: d58093178605d1f8155b93d9a161408177f6ac72
This commit is contained in:
rka-oai
2026-08-04 22:07:56 +00:00
committed by copyberry
parent e1f39b5f5b
commit b9ba969f39
10 changed files with 277 additions and 156 deletions
-6
View File
@@ -50,8 +50,6 @@ use codex_utils_output_truncation::truncate_text;
use futures::prelude::*;
use tracing::error;
use codex_model_provider_info::ModelProviderInfo;
pub use codex_prompts::SUMMARIZATION_PROMPT;
pub use codex_prompts::SUMMARY_PREFIX;
const COMPACT_USER_MESSAGE_MAX_TOKENS: usize = 20_000;
@@ -105,10 +103,6 @@ pub(crate) async fn build_compaction_initial_context(
}
}
pub(crate) fn should_use_remote_compact_task(provider: &ModelProviderInfo) -> bool {
provider.supports_remote_compaction()
}
pub(crate) async fn run_inline_auto_compact_task(
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
-27
View File
@@ -1,6 +1,4 @@
use super::*;
use codex_model_provider_info::ModelProviderInfo;
use codex_model_provider_info::WireApi;
use codex_protocol::ResponseItemId;
use codex_protocol::models::DEFAULT_IMAGE_DETAIL;
use codex_protocol::models::InternalChatMessageMetadataPassthrough;
@@ -266,31 +264,6 @@ fn build_compacted_history_preserves_user_message_passthrough_metadata() {
assert_eq!(history[1].turn_id(), None);
}
#[test]
fn should_use_remote_compact_task_for_azure_provider() {
let provider = ModelProviderInfo {
name: "Azure".into(),
base_url: Some("https://example.com/openai".into()),
env_key: Some("AZURE_OPENAI_API_KEY".into()),
env_key_instructions: None,
experimental_bearer_token: None,
auth: None,
aws: None,
wire_api: WireApi::Responses,
query_params: None,
http_headers: None,
env_http_headers: None,
request_max_retries: None,
stream_max_retries: None,
stream_idle_timeout_ms: None,
websocket_connect_timeout_ms: None,
requires_openai_auth: false,
supports_websockets: false,
supports_standalone_web_search: false,
};
assert!(should_use_remote_compact_task(&provider));
}
#[tokio::test]
async fn process_compacted_history_replaces_developer_messages() {
let compacted_history = vec![
+39 -36
View File
@@ -11,7 +11,6 @@ use crate::client_common::ResponseEvent;
use crate::collect_explicit_skill_mentions;
use crate::compact::InitialContextInjection;
use crate::compact::run_inline_auto_compact_task;
use crate::compact::should_use_remote_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;
@@ -81,6 +80,7 @@ use codex_features::Feature;
use codex_file_system::FindUpErrorPolicy;
use codex_file_system::find_nearest_ancestor_with_markers;
use codex_login::CodexAuth;
use codex_model_provider::RemoteCompactionSupport;
use codex_protocol::ResponseItemId;
use codex_protocol::config_types::AutoCompactTokenLimitScope;
use codex_protocol::config_types::ModeKind;
@@ -1171,11 +1171,12 @@ async fn run_auto_compact(
return Ok(());
}
if should_use_remote_compact_task(turn_context.provider.info()) {
if turn_context
.config
.features
.enabled(Feature::RemoteCompactionV2)
match turn_context.provider.capabilities().remote_compaction {
RemoteCompactionSupport::V2
if turn_context
.config
.features
.enabled(Feature::RemoteCompactionV2) =>
{
emit_compact_metric(
&sess.services.session_telemetry,
@@ -1192,37 +1193,39 @@ async fn run_auto_compact(
phase,
)
.await?;
return Ok(());
}
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?;
} else {
emit_compact_metric(
&sess.services.session_telemetry,
"local",
/*manual*/ false,
);
run_inline_auto_compact_task(
Arc::clone(sess),
Arc::clone(turn_context),
initial_context_injection,
reason,
phase,
)
.await?;
RemoteCompactionSupport::V1 | 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,
"local",
/*manual*/ false,
);
run_inline_auto_compact_task(
Arc::clone(sess),
Arc::clone(turn_context),
initial_context_injection,
reason,
phase,
)
.await?;
}
}
Ok(())
}
+24 -23
View File
@@ -8,6 +8,7 @@ use crate::session::session::Session;
use crate::session::turn_context::TurnContext;
use crate::state::TaskKind;
use codex_features::Feature;
use codex_model_provider::RemoteCompactionSupport;
use codex_protocol::error::CodexErrorDetails;
use codex_protocol::user_input::UserInput;
use tokio_util::sync::CancellationToken;
@@ -37,11 +38,9 @@ impl SessionTask for CompactTask {
return Ok(None);
}
let result = if crate::compact::should_use_remote_compact_task(ctx.provider.info()) {
if ctx
.config
.features
.enabled(codex_features::Feature::RemoteCompactionV2)
let result = match ctx.provider.capabilities().remote_compaction {
RemoteCompactionSupport::V2
if ctx.config.features.enabled(Feature::RemoteCompactionV2) =>
{
emit_compact_metric(
&session.services.session_telemetry,
@@ -49,7 +48,8 @@ impl SessionTask for CompactTask {
/*manual*/ true,
);
crate::compact_remote_v2::run_remote_compact_task(session.clone(), ctx).await
} else {
}
RemoteCompactionSupport::V1 | RemoteCompactionSupport::V2 => {
emit_compact_metric(
&session.services.session_telemetry,
"remote",
@@ -57,23 +57,24 @@ impl SessionTask for CompactTask {
);
crate::compact_remote::run_remote_compact_task(session.clone(), ctx).await
}
} else {
emit_compact_metric(
&session.services.session_telemetry,
"local",
/*manual*/ true,
);
let input = vec![UserInput::Text {
text: ctx
.config
.compact_prompt
.as_deref()
.unwrap_or(crate::compact::SUMMARIZATION_PROMPT)
.to_string(),
// Compaction prompt is synthesized; no UI element ranges to preserve.
text_elements: Vec::new(),
}];
crate::compact::run_compact_task(session.clone(), ctx, input).await
RemoteCompactionSupport::Unsupported => {
emit_compact_metric(
&session.services.session_telemetry,
"local",
/*manual*/ true,
);
let input = vec![UserInput::Text {
text: ctx
.config
.compact_prompt
.as_deref()
.unwrap_or(crate::compact::SUMMARIZATION_PROMPT)
.to_string(),
// Compaction prompt is synthesized; no UI element ranges to preserve.
text_elements: Vec::new(),
}];
crate::compact::run_compact_task(session.clone(), ctx, input).await
}
};
if let Err(err) = result
&& matches!(err.details(), CodexErrorDetails::TurnAborted)
+146
View File
@@ -10,6 +10,10 @@ use codex_features::Feature;
use codex_login::CodexAuth;
use codex_login::auth::AgentIdentityAuth;
use codex_login::auth::AgentIdentityAuthRecord;
use codex_login::auth::BedrockApiKeyAuth;
use codex_model_provider_info::AMAZON_BEDROCK_GPT_5_5_MODEL_ID;
use codex_model_provider_info::AMAZON_BEDROCK_PROVIDER_ID;
use codex_model_provider_info::ModelProviderInfo;
use codex_protocol::AgentPath;
use codex_protocol::account::PlanType as AccountPlanType;
use codex_protocol::config_types::ServiceTier;
@@ -346,6 +350,148 @@ async fn wait_for_turn_complete(codex: &codex_core::CodexThread) {
.await;
}
fn amazon_bedrock_test_codex() -> TestCodexBuilder {
let auth = CodexAuth::BedrockApiKey(BedrockApiKeyAuth {
api_key: "bedrock-test-api-key".to_string(),
region: "us-east-1".to_string(),
});
test_codex()
.with_auth(auth)
.with_model(AMAZON_BEDROCK_GPT_5_5_MODEL_ID)
.with_config(|config| {
let _ = config.features.enable(Feature::RemoteCompactionV2);
config.model_provider = ModelProviderInfo {
base_url: config.model_provider.base_url.clone(),
..ModelProviderInfo::create_amazon_bedrock_provider(/*aws*/ None)
};
config.model_provider_id = AMAZON_BEDROCK_PROVIDER_ID.to_string();
})
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn amazon_bedrock_uses_remote_compaction_endpoint() -> Result<()> {
skip_if_no_network!(Ok(()));
let harness = TestCodexHarness::with_auto_env_builder(amazon_bedrock_test_codex()).await?;
let response_mock = responses::mount_sse_sequence(
harness.server(),
vec![
sse(vec![
responses::ev_assistant_message("message-1", "before compaction"),
responses::ev_completed("response-1"),
]),
sse(vec![
responses::ev_assistant_message("message-2", "after compaction"),
responses::ev_completed("response-2"),
]),
],
)
.await;
let compact_mock = responses::mount_compact_user_history_with_summary_once(
harness.server(),
"BEDROCK_REMOTE_COMPACTED_SUMMARY",
)
.await;
harness.test().submit_turn("before compact").await?;
harness.test().codex.submit(Op::Compact).await?;
wait_for_turn_complete(&harness.test().codex).await;
harness.test().submit_turn("after compact").await?;
let compact_request = compact_mock.single_request();
assert_eq!(compact_request.path(), "/v1/responses/compact");
assert_eq!(
compact_request.header("authorization").as_deref(),
Some("Bearer bedrock-test-api-key")
);
assert_eq!(
compact_request
.header("x-amzn-mantle-client-agent")
.as_deref(),
Some("codex")
);
assert_eq!(
compact_request.body_json()["model"],
AMAZON_BEDROCK_GPT_5_5_MODEL_ID
);
let response_requests = response_mock.requests();
assert_eq!(response_requests.len(), 2);
assert!(
response_requests
.iter()
.all(|request| request.inputs_of_type("compaction_trigger").is_empty())
);
assert!(response_requests[1].input().iter().any(|item| {
item["type"] == "compaction"
&& item["encrypted_content"] == "BEDROCK_REMOTE_COMPACTED_SUMMARY"
}));
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn amazon_bedrock_automatic_compaction_uses_v1_endpoint_when_v2_is_enabled() -> Result<()> {
skip_if_no_network!(Ok(()));
let harness = TestCodexHarness::with_auto_env_builder(amazon_bedrock_test_codex().with_config(
|config| {
config.model_auto_compact_token_limit = Some(200);
},
))
.await?;
let response_mock = responses::mount_sse_sequence(
harness.server(),
vec![
sse(vec![
responses::ev_assistant_message("message-1", "before automatic compaction"),
responses::ev_completed_with_tokens("response-1", /*total_tokens*/ 500),
]),
sse(vec![
responses::ev_assistant_message("message-2", "after automatic compaction"),
responses::ev_completed("response-2"),
]),
],
)
.await;
let compact_mock = responses::mount_compact_user_history_with_summary_once(
harness.server(),
"BEDROCK_AUTOMATIC_REMOTE_COMPACTED_SUMMARY",
)
.await;
harness
.test()
.submit_turn("before automatic compact")
.await?;
harness
.test()
.submit_turn("after automatic compact")
.await?;
let compact_request = compact_mock.single_request();
assert_eq!(compact_request.path(), "/v1/responses/compact");
assert_eq!(
compact_request.header("authorization").as_deref(),
Some("Bearer bedrock-test-api-key")
);
let response_requests = response_mock.requests();
assert_eq!(response_requests.len(), 2);
assert!(
response_requests
.iter()
.all(|request| request.inputs_of_type("compaction_trigger").is_empty())
);
assert!(response_requests[1].input().iter().any(|item| {
item["type"] == "compaction"
&& item["encrypted_content"] == "BEDROCK_AUTOMATIC_REMOTE_COMPACTED_SUMMARY"
}));
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn remote_compact_replaces_history_for_followups() -> Result<()> {
skip_if_no_network!(Ok(()));
-5
View File
@@ -7,7 +7,6 @@
use codex_api::Provider as ApiProvider;
use codex_api::RetryConfig as ApiRetryConfig;
use codex_api::is_azure_responses_provider;
use codex_protocol::auth::AuthMode;
use codex_protocol::config_types::ModelProviderAuthInfo;
use codex_protocol::error::CodexErr;
@@ -419,10 +418,6 @@ impl ModelProviderInfo {
self.name == AMAZON_BEDROCK_PROVIDER_NAME
}
pub fn supports_remote_compaction(&self) -> bool {
self.is_openai() || is_azure_responses_provider(&self.name, self.base_url.as_deref())
}
pub fn has_command_auth(&self) -> bool {
self.auth.is_some()
}
@@ -136,13 +136,6 @@ supports_websockets = true
assert_eq!(provider.websocket_connect_timeout_ms, Some(15_000));
}
#[test]
fn test_supports_remote_compaction_for_openai() {
let provider = ModelProviderInfo::create_openai_provider(/*base_url*/ None);
assert!(provider.supports_remote_compaction());
}
#[test]
fn test_personal_access_token_uses_chatgpt_codex_base_url() {
let api_provider = ModelProviderInfo::create_openai_provider(/*base_url*/ None)
@@ -161,58 +154,6 @@ fn test_header_auth_uses_chatgpt_codex_base_url() {
assert_eq!(api_provider.base_url, CHATGPT_CODEX_BASE_URL);
}
#[test]
fn test_supports_remote_compaction_for_azure_name() {
let provider = ModelProviderInfo {
name: "Azure".into(),
base_url: Some("https://example.com/openai".into()),
env_key: Some("AZURE_OPENAI_API_KEY".into()),
env_key_instructions: None,
experimental_bearer_token: None,
auth: None,
aws: None,
wire_api: WireApi::Responses,
query_params: None,
http_headers: None,
env_http_headers: None,
request_max_retries: None,
stream_max_retries: None,
stream_idle_timeout_ms: None,
websocket_connect_timeout_ms: None,
requires_openai_auth: false,
supports_websockets: false,
supports_standalone_web_search: false,
};
assert!(provider.supports_remote_compaction());
}
#[test]
fn test_supports_remote_compaction_for_non_openai_non_azure_provider() {
let provider = ModelProviderInfo {
name: "Example".into(),
base_url: Some("https://example.com/v1".into()),
env_key: Some("API_KEY".into()),
env_key_instructions: None,
experimental_bearer_token: None,
auth: None,
aws: None,
wire_api: WireApi::Responses,
query_params: None,
http_headers: None,
env_http_headers: None,
request_max_retries: None,
stream_max_retries: None,
stream_idle_timeout_ms: None,
websocket_connect_timeout_ms: None,
requires_openai_auth: false,
supports_websockets: false,
supports_standalone_web_search: false,
};
assert!(!provider.supports_remote_compaction());
}
#[test]
fn test_uses_openai_actor_authorization() {
let mut provider = ModelProviderInfo {
@@ -29,6 +29,7 @@ use crate::provider::ModelProviderFuture;
use crate::provider::ProviderAccountResult;
use crate::provider::ProviderAccountState;
use crate::provider::ProviderCapabilities;
use crate::provider::RemoteCompactionSupport;
use auth::resolve_provider_auth as resolve_bedrock_provider_auth;
use catalog::normalize_bedrock_catalog;
pub(crate) use catalog::static_model_catalog;
@@ -126,6 +127,7 @@ impl ModelProvider for AmazonBedrockModelProvider {
image_generation: false,
web_search: true,
external_web_access: false,
remote_compaction: RemoteCompactionSupport::V1,
}
}
@@ -368,6 +370,7 @@ mod tests {
image_generation: false,
web_search: true,
external_web_access: false,
remote_compaction: RemoteCompactionSupport::V1,
}
);
}
+1
View File
@@ -22,5 +22,6 @@ pub use provider::ProviderAccountError;
pub use provider::ProviderAccountResult;
pub use provider::ProviderAccountState;
pub use provider::ProviderCapabilities;
pub use provider::RemoteCompactionSupport;
pub use provider::SharedModelProvider;
pub use provider::create_model_provider;
+64
View File
@@ -7,6 +7,7 @@ use std::sync::Arc;
use codex_api::ApiError;
use codex_api::Provider;
use codex_api::SharedAuthProvider;
use codex_api::is_azure_responses_provider;
use codex_login::AuthManager;
use codex_login::CodexAuth;
use codex_model_provider_info::ModelProviderInfo;
@@ -25,6 +26,17 @@ use crate::auth::resolve_provider_auth;
use crate::auth::resolve_provider_auth_for_scope;
use crate::models_endpoint::OpenAiModelsEndpoint;
/// Remote context-compaction protocols supported by a model provider.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum RemoteCompactionSupport {
/// The provider does not support remote compaction.
Unsupported,
/// The provider supports only the dedicated `/v1/responses/compact` endpoint.
V1,
/// The provider supports both the dedicated endpoint and `compaction_trigger` items.
V2,
}
/// Optional provider-backed features that Codex may expose at runtime.
///
/// These capabilities are a provider-owned upper bound. Callers can disable
@@ -36,6 +48,7 @@ pub struct ProviderCapabilities {
pub image_generation: bool,
pub web_search: bool,
pub external_web_access: bool,
pub remote_compaction: RemoteCompactionSupport,
}
impl Default for ProviderCapabilities {
@@ -45,6 +58,7 @@ impl Default for ProviderCapabilities {
image_generation: true,
web_search: true,
external_web_access: true,
remote_compaction: RemoteCompactionSupport::V2,
}
}
}
@@ -264,6 +278,21 @@ impl ModelProvider for ConfiguredModelProvider {
&self.info
}
fn capabilities(&self) -> ProviderCapabilities {
let remote_compaction = if self.info.is_openai()
|| is_azure_responses_provider(&self.info.name, self.info.base_url.as_deref())
{
RemoteCompactionSupport::V2
} else {
RemoteCompactionSupport::Unsupported
};
ProviderCapabilities {
remote_compaction,
..ProviderCapabilities::default()
}
}
fn auth_manager(&self) -> Option<Arc<AuthManager>> {
self.auth_manager.clone()
}
@@ -508,6 +537,41 @@ mod tests {
assert_eq!(provider.capabilities(), ProviderCapabilities::default());
}
#[test]
fn configured_provider_remote_compaction_matches_provider_support() {
let cases = [
(
ModelProviderInfo::create_openai_provider(/*base_url*/ None),
RemoteCompactionSupport::V2,
),
(
ModelProviderInfo {
name: "Azure".to_string(),
base_url: Some("https://example.com/openai".to_string()),
..ModelProviderInfo::default()
},
RemoteCompactionSupport::V2,
),
(
ModelProviderInfo {
name: "Custom".to_string(),
base_url: Some("https://example.openai.azure.com/openai/v1".to_string()),
..ModelProviderInfo::default()
},
RemoteCompactionSupport::V2,
),
(
provider_for("https://example.test/v1".to_string()),
RemoteCompactionSupport::Unsupported,
),
];
for (provider_info, expected) in cases {
let provider = create_model_provider(provider_info, /*auth_manager*/ None);
assert_eq!(provider.capabilities().remote_compaction, expected);
}
}
#[test]
fn configured_provider_uses_default_approval_review_preferred_model() {
let provider = create_model_provider(