Skip to content

armgit / codex

Public Imported from GitHub
0

Lightweight coding agent that runs in your terminal

Rust

Enable multi-agent V2 and Ultra reasoning on Amazon Bedrock (#49345)

Celia Chen8ffd91e42aa0
diff --git a/codex-rs/core/tests/suite/bedrock_multi_agent_tests.rs b/codex-rs/core/tests/suite/bedrock_multi_agent_tests.rs
new file mode 100644
index 0000000000..ddc5ebf1cf
--- /dev/null
+++ b/codex-rs/core/tests/suite/bedrock_multi_agent_tests.rs
@@ -0,0 +1,62 @@
+//! Verifies default Bedrock catalogs select V2 and expose its collaboration tools.
+
+use codex_features::Feature;
+use codex_login::CodexAuth;
+use codex_login::auth::BedrockApiKeyAuth;
+use codex_model_provider_info::AMAZON_BEDROCK_PROVIDER_ID;
+use codex_model_provider_info::AMAZON_BEDROCK_RUNTIME_PROVIDER_ID;
+use codex_model_provider_info::built_in_model_providers;
+use codex_protocol::protocol::MultiAgentVersion;
+use core_test_support::responses;
+use core_test_support::skip_if_no_network;
+use core_test_support::test_codex::test_codex;
+use pretty_assertions::assert_eq;
+
+#[test_case::test_case(AMAZON_BEDROCK_PROVIDER_ID, "openai.gpt-5.6-sol"; "mantle")]
+#[test_case::test_case(AMAZON_BEDROCK_RUNTIME_PROVIDER_ID, "global.openai.gpt-5.6-sol"; "runtime")]
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn bedrock_catalog_selects_v2_by_default(
+ provider_id: &'static str,
+ model: &'static str,
+) -> anyhow::Result<()> {
+ skip_if_no_network!(Ok(()));
+ let server = responses::start_mock_server().await;
+ let mock = responses::mount_sse_once(
+ &server,
+ responses::sse(vec![responses::ev_completed("done")]),
+ )
+ .await;
+ let test = test_codex()
+ .with_auth(CodexAuth::BedrockApiKey(BedrockApiKeyAuth {
+ api_key: "dummy".to_string(),
+ region: "us-east-1".to_string(),
+ }))
+ .with_model(model)
+ .with_config(move |config| {
+ let base_url = config.model_provider.base_url.clone();
+ config.model_provider_id = provider_id.to_string();
+ config.model_provider = built_in_model_providers(/*openai_base_url*/ None)
+ .remove(provider_id)
+ .expect("built-in Bedrock provider");
+ config.model_provider.base_url = base_url;
+ })
+ .build_with_auto_env(&server)
+ .await?;
+ assert!(!test.config.features.enabled(Feature::MultiAgentV2));
+ assert!(test.config.model_catalog.is_none());
+
+ test.submit_text_turn("Hello").await?;
+
+ assert_eq!(
+ test.codex.multi_agent_version(),
+ Some(MultiAgentVersion::V2)
+ );
+ let request = mock.single_request();
+ for tool in ["spawn_agent", "followup_task"] {
+ assert!(
+ request.tool_by_name("collaboration", tool).is_some(),
+ "default Bedrock catalog should expose collaboration.{tool}",
+ );
+ }
+ Ok(())
+}
diff --git a/codex-rs/core/tests/suite/mod.rs b/codex-rs/core/tests/suite/mod.rs
index 14805e25b5..1be4f9710f 100644
--- a/codex-rs/core/tests/suite/mod.rs
+++ b/codex-rs/core/tests/suite/mod.rs
@@ -54,6 +54,8 @@ mod approvals;
mod audio_truncation;
mod auth_recovery_policy;
mod auto_review;
+#[path = "bedrock_multi_agent_tests.rs"]
+mod bedrock_multi_agent;
mod catalog_permission_messages;
mod cli_stream;
mod client;
diff --git a/codex-rs/model-provider/src/amazon_bedrock/catalog.rs b/codex-rs/model-provider/src/amazon_bedrock/catalog.rs
index c7c478436b..1de3b24436 100644
--- a/codex-rs/model-provider/src/amazon_bedrock/catalog.rs
+++ b/codex-rs/model-provider/src/amazon_bedrock/catalog.rs
@@ -10,9 +10,7 @@ use codex_models_manager::bundled_models_response;
use codex_protocol::openai_models::ModelInfo;
use codex_protocol::openai_models::ModelVisibility;
use codex_protocol::openai_models::ModelsResponse;
-use codex_protocol::openai_models::ReasoningEffort;
use codex_protocol::openai_models::WebSearchToolType;
-use codex_protocol::protocol::MultiAgentVersion;
const GPT_5_BEDROCK_CONTEXT_WINDOW: i64 = 272_000;
const GPT_5_6_SOL_OPENAI_MODEL_ID: &str = "gpt-5.6-sol";
@@ -98,8 +96,6 @@ pub(crate) fn normalize_bedrock_catalog(mut catalog: ModelsResponse) -> ModelsRe
model.default_service_tier = None;
// Bedrock rejects the `search_content_types` field used by multimodal search.
model.web_search_tool_type = WebSearchToolType::Text;
- // Bedrock does not support the response items used by multi-agent V2.
- model.multi_agent_version = Some(MultiAgentVersion::V1);
}
catalog
}
@@ -137,9 +133,6 @@ fn bedrock_model(
model.use_responses_lite = false;
model.tool_mode = None;
model
- .supported_reasoning_levels
- .retain(|level| level.effort != ReasoningEffort::Ultra);
- model
}
fn bundled_openai_model(slug: &str) -> ModelInfo {
@@ -154,6 +147,7 @@ fn bundled_openai_model(slug: &str) -> ModelInfo {
#[cfg(test)]
mod tests {
use codex_protocol::config_types::SERVICE_TIER_DEFAULT_REQUEST_VALUE;
+ use codex_protocol::protocol::MultiAgentVersion;
use pretty_assertions::assert_eq;
use super::*;
@@ -251,21 +245,30 @@ mod tests {
#[test]
fn configured_bedrock_catalogs_normalize_unsupported_model_capabilities() {
- let model = bundled_openai_model(GPT_5_5_OPENAI_MODEL_ID);
- let mut expected = model.clone();
- expected.additional_speed_tiers.clear();
- expected.service_tiers.clear();
- expected.default_service_tier = None;
- expected.web_search_tool_type = WebSearchToolType::Text;
- expected.multi_agent_version = Some(MultiAgentVersion::V1);
+ let models = [
+ Some(MultiAgentVersion::V2),
+ Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::Disabled),
+ None,
+ ]
+ .into_iter()
+ .map(|version| {
+ let mut model = bundled_openai_model(GPT_5_5_OPENAI_MODEL_ID);
+ model.multi_agent_version = version;
+ model
+ })
+ .collect::<Vec<_>>();
+ let mut expected = models.clone();
+ for model in &mut expected {
+ model.additional_speed_tiers.clear();
+ model.service_tiers.clear();
+ model.default_service_tier = None;
+ model.web_search_tool_type = WebSearchToolType::Text;
+ }
assert_eq!(
- normalize_bedrock_catalog(ModelsResponse {
- models: vec![model],
- }),
- ModelsResponse {
- models: vec![expected],
- }
+ normalize_bedrock_catalog(ModelsResponse { models }),
+ ModelsResponse { models: expected }
);
}
@@ -343,14 +346,10 @@ mod tests {
expected.upgrade = None;
expected.use_responses_lite = false;
expected.tool_mode = None;
- expected
- .supported_reasoning_levels
- .retain(|level| level.effort != ReasoningEffort::Ultra);
expected.additional_speed_tiers.clear();
expected.service_tiers.clear();
expected.default_service_tier = None;
expected.web_search_tool_type = WebSearchToolType::Text;
- expected.multi_agent_version = Some(MultiAgentVersion::V1);
assert_eq!(
catalog.models.iter().find(|model| model.slug == slug),
diff --git a/codex-rs/model-provider/src/amazon_bedrock/runtime_catalog_tests.rs b/codex-rs/model-provider/src/amazon_bedrock/runtime_catalog_tests.rs
index 9f6f6ad8b7..f769d03481 100644
--- a/codex-rs/model-provider/src/amazon_bedrock/runtime_catalog_tests.rs
+++ b/codex-rs/model-provider/src/amazon_bedrock/runtime_catalog_tests.rs
@@ -64,13 +64,13 @@ fn runtime_catalog_disables_web_search_without_overriding_review_models() {
"global.openai.gpt-6.1-sol",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"global.openai.gpt-6-astra",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"global.openai.gpt-6-sol",
@@ -82,19 +82,19 @@ fn runtime_catalog_disables_web_search_without_overriding_review_models() {
"global.openai.gpt-6-luna",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"global.openai.gpt-5.6-sol",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"global.openai.gpt-5.6-terra",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"global.openai.gpt-5.6-luna",
@@ -106,13 +106,13 @@ fn runtime_catalog_disables_web_search_without_overriding_review_models() {
"us.openai.gpt-6.1-sol",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"us.openai.gpt-6-astra",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"us.openai.gpt-6-sol",
@@ -124,19 +124,19 @@ fn runtime_catalog_disables_web_search_without_overriding_review_models() {
"us.openai.gpt-6-luna",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"us.openai.gpt-5.6-sol",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"us.openai.gpt-5.6-terra",
None,
false,
- Some(MultiAgentVersion::V1),
+ Some(MultiAgentVersion::V2),
),
(
"us.openai.gpt-5.6-luna",
diff --git a/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_advanced_reasoning.snap b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_advanced_reasoning.snap
new file mode 100644
index 0000000000..8518e8551c
--- /dev/null
+++ b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_advanced_reasoning.snap
@@ -0,0 +1,12 @@
+---
+source: tui/src/chatwidget/tests/bedrock_catalog_tests.rs
+expression: "render_bottom_popup(&chat, 100)"
+---
+ Advanced Reasoning
+ ⚠ Consumes usage limits faster
+
+
+› 1. Max For difficult problems when quality matters more than speed · higher usage
+ 2. Ultra For demanding work using multiple agents · highest usage
+
+ enter default · s session · esc back
diff --git a/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_reasoning.snap b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_reasoning.snap
index 5defe22380..07d5728a69 100644
--- a/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_reasoning.snap
+++ b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_mantle_astra_reasoning.snap
@@ -9,6 +9,6 @@ expression: "render_bottom_popup(&chat, 100)"
2. Medium Balances speed and reasoning depth for everyday tasks
3. High Greater reasoning depth for complex problems
4. Extra high Extra high reasoning depth for complex problems
- 5. More reasoning… Max consumes usage limits faster
+ 5. More reasoning… Max and Ultra consume usage limits faster
enter default · s session · esc back
diff --git a/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_advanced_reasoning.snap b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_advanced_reasoning.snap
new file mode 100644
index 0000000000..8518e8551c
--- /dev/null
+++ b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_advanced_reasoning.snap
@@ -0,0 +1,12 @@
+---
+source: tui/src/chatwidget/tests/bedrock_catalog_tests.rs
+expression: "render_bottom_popup(&chat, 100)"
+---
+ Advanced Reasoning
+ ⚠ Consumes usage limits faster
+
+
+› 1. Max For difficult problems when quality matters more than speed · higher usage
+ 2. Ultra For demanding work using multiple agents · highest usage
+
+ enter default · s session · esc back
diff --git a/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_reasoning.snap b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_reasoning.snap
index f74507aa3c..d220a75654 100644
--- a/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_reasoning.snap
+++ b/codex-rs/tui/src/chatwidget/snapshots/codex_tui__chatwidget__tests__bedrock_runtime_astra_reasoning.snap
@@ -9,6 +9,6 @@ expression: "render_bottom_popup(&chat, 100)"
2. Medium Balances speed and reasoning depth for everyday tasks
3. High Greater reasoning depth for complex problems
4. Extra high Extra high reasoning depth for complex problems
- 5. More reasoning… Max consumes usage limits faster
+ 5. More reasoning… Max and Ultra consume usage limits faster
enter default · s session · esc back
diff --git a/codex-rs/tui/src/chatwidget/tests/bedrock_catalog_tests.rs b/codex-rs/tui/src/chatwidget/tests/bedrock_catalog_tests.rs
index ca29f275dd..5f0cb91520 100644
--- a/codex-rs/tui/src/chatwidget/tests/bedrock_catalog_tests.rs
+++ b/codex-rs/tui/src/chatwidget/tests/bedrock_catalog_tests.rs
@@ -42,7 +42,7 @@ async fn bedrock_astra_model_and_reasoning_pickers() {
.find(|model| model.model == astra_model)
.expect("Astra preset")
.clone();
- let (mut chat, _events, _ops) = make_chatwidget_manual(Some(default_model)).await;
+ let (mut chat, mut events, _ops) = make_chatwidget_manual(Some(default_model)).await;
chat.thread_id = Some(ThreadId::new());
chat.model_catalog = Arc::new(ModelCatalog::new(presets));
chat.open_model_popup();
@@ -56,6 +56,35 @@ async fn bedrock_astra_model_and_reasoning_pickers() {
format!("bedrock_{name}_astra_reasoning"),
render_bottom_popup(&chat, /*width*/ 100)
);
+ chat.handle_key_event(KeyEvent::from(KeyCode::Char('5')));
+ let advanced =
+ std::iter::from_fn(|| events.try_recv().ok()).find_map(|event| match event {
+ AppEvent::OpenAdvancedReasoningPopup { model } => Some(model),
+ _ => None,
+ });
+ chat.open_advanced_reasoning_popup(advanced.expect("advanced reasoning popup"));
+ assert_chatwidget_snapshot!(
+ format!("bedrock_{name}_astra_advanced_reasoning"),
+ render_bottom_popup(&chat, /*width*/ 100)
+ );
+ chat.handle_key_event(KeyEvent::from(KeyCode::Char('2')));
+ let selected =
+ std::iter::from_fn(|| events.try_recv().ok()).find_map(|event| match event {
+ AppEvent::AstraSelectedFromModelPicker {
+ thread_id,
+ model,
+ action: AstraModelPickerAction::ApplyAdvancedReasoning { effort },
+ } => Some((thread_id, model, effort)),
+ _ => None,
+ });
+ assert_eq!(
+ selected,
+ Some((
+ chat.thread_id.expect("active thread"),
+ astra_model.to_string(),
+ ReasoningEffortConfig::Ultra,
+ ))
+ );
}
}