Allow model metadata to control multi-agent tools in code mode (#61854)

## What changed

Add optional `multi_agent_tools_in_code_mode` model metadata to control whether native multi-agent tools follow the model's `tool_mode`. When `true`, expose them according to `Direct`, `CodeMode`, or `CodeModeOnly`; when `false`, keep them direct-only. When omitted, preserve the `multi_agent_v2.non_code_mode_only` runtime setting.

## Testing

Add coverage for metadata overrides and runtime fallback across all three tool modes, plus serialization and deserialization of absent, null, and boolean metadata values. Update the namespace-prefix integration scenario to enable code-mode tools through model metadata.

GitOrigin-RevId: be90f436066226bf9971cbd17f28c396341bb83a
This commit is contained in:
jif
2026-11-16 18:54:05 +00:00
committed by copyberry
parent 35145aba21
commit 3257e1b3eb
15 changed files with 94 additions and 4 deletions
@@ -78,6 +78,7 @@ fn preset_to_info(preset: &ModelPreset, priority: i32) -> ModelInfo {
tool_mode: None,
multi_agent_version: preset.multi_agent_version,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}
}
@@ -170,6 +170,7 @@ async fn models_client_hits_models_endpoint() {
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}],
};
+7 -3
View File
@@ -1517,10 +1517,14 @@ fn add_collaboration_tools(context: &CoreToolPlanContext<'_>, registry: &mut Too
let model_messages = ResolvedModelMessages::from_model(context.model_info);
let spawn_agent_description =
model_messages.multi_agent_tool_description_override("spawn_agent");
let exposure = if turn_context.config.multi_agent_v2.non_code_mode_only {
ToolExposure::DirectModelOnly
} else {
let tools_in_code_mode = context
.model_info
.multi_agent_tools_in_code_mode
.unwrap_or(!turn_context.config.multi_agent_v2.non_code_mode_only);
let exposure = if tools_in_code_mode {
ToolExposure::Direct
} else {
ToolExposure::DirectModelOnly
};
let tool_namespace = turn_context.config.multi_agent_v2.tool_namespace.as_deref();
let native_review = if turn_context
@@ -3144,6 +3144,55 @@ async fn multi_agent_v2_bedrock_workers_only_delegate_when_model_supports_v2() {
}
}
#[tokio::test]
async fn multi_agent_catalog_code_mode_follows_tool_mode_and_preserves_runtime_fallback() {
for (catalog, non_code_mode_only, follows_tool_mode) in [
(None, true, false),
(None, false, true),
(Some(true), true, true),
(Some(false), false, false),
] {
for tool_mode in [ToolMode::Direct, ToolMode::CodeMode, ToolMode::CodeModeOnly] {
let plan = probe(|turn| {
set_feature(turn, Feature::MultiAgentV2, /*enabled*/ true);
update_config(turn, |config| {
config.multi_agent_v2.non_code_mode_only = non_code_mode_only;
config.multi_agent_v2.tool_namespace = Some("agents".to_string());
});
update_turn_settings_for_test(turn, |settings| {
let model = Arc::make_mut(&mut settings.model_info);
model.tool_mode = Some(tool_mode);
model.multi_agent_tools_in_code_mode = catalog;
});
})
.await;
for name in [
"spawn_agent",
"send_message",
"followup_task",
"wait_agent",
"interrupt_agent",
"list_agents",
] {
assert_eq!(
(
plan.namespace_function_names("agents")
.iter()
.any(|tool| tool == name),
plan.code_mode_tool_names
.contains_key(&format!("agents__{name}")),
),
(
!follows_tool_mode || tool_mode != ToolMode::CodeModeOnly,
follows_tool_mode && tool_mode != ToolMode::Direct,
),
"{name}: catalog={catalog:?}, non_code_mode_only={non_code_mode_only}, mode={tool_mode:?}",
);
}
}
}
}
#[tokio::test]
async fn code_mode_only_can_expose_namespaced_multi_agent_v2_as_normal_tools() {
let plan = probe(|turn| {
+1
View File
@@ -656,6 +656,7 @@ fn remote_model_with_auto_review_override(slug: &str, review_model: &str) -> Mod
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
@@ -135,6 +135,7 @@ fn test_model_info(
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
@@ -1282,6 +1283,7 @@ async fn model_switch_to_smaller_model_updates_token_context_window() -> Result<
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
@@ -748,5 +748,6 @@ fn test_remote_model(slug: &str, priority: i32) -> ModelInfo {
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}
}
@@ -837,6 +837,7 @@ async fn remote_models_remote_model_uses_unified_exec() -> Result<()> {
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
@@ -1112,6 +1113,7 @@ async fn remote_models_apply_legacy_instructions(auth: CodexAuth) -> Result<()>
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
@@ -1701,6 +1703,7 @@ fn test_remote_model_with_policy(
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
+1
View File
@@ -3704,6 +3704,7 @@ async fn stdio_image_responses_are_sanitized_for_text_only_model() -> anyhow::Re
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}],
},
)
@@ -46,7 +46,6 @@ async fn catalog_namespace_prefixes_follow_the_selected_model(
.features
.enable(Feature::MultiAgentV2)
.expect("enable MAv2");
config.multi_agent_v2.non_code_mode_only = false;
config.multi_agent_v2.root_agent_usage_hint_text = Some("Coordinate sub-agents.".into());
config
.mcp_servers
@@ -86,6 +85,7 @@ async fn catalog_namespace_prefixes_follow_the_selected_model(
] {
builder = builder.with_model_info_override(slug, move |model| {
model.tool_mode = Some(tool_mode);
model.multi_agent_tools_in_code_mode = Some(true);
model.supports_search_tool = include_search;
model.model_messages.as_mut().expect("model messages").tools = Some(
serde_json::from_value(json!({
@@ -102,6 +102,7 @@ fn test_model_info(
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers,
+1
View File
@@ -1638,6 +1638,7 @@ async fn view_image_tool_returns_unsupported_message_for_text_only_model() -> an
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
priority: 1,
additional_speed_tiers: Vec::new(),
service_tiers: Vec::new(),
@@ -145,6 +145,7 @@ pub fn model_info_from_slug(slug: &str) -> ModelInfo {
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}
}
@@ -56,6 +56,7 @@ fn test_model(model_messages: Option<ModelMessages>) -> ModelInfo {
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}
}
+23
View File
@@ -529,6 +529,10 @@ pub struct ModelInfo {
/// Reasoning effort used for multi-agent work when the user selects Ultra.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub multi_agent_reasoning_effort: Option<ReasoningEffort>,
/// Whether native multi-agent tools follow the model's tool mode instead of staying direct-only.
/// When absent, preserve the runtime's `multi_agent_v2.non_code_mode_only` setting.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub multi_agent_tools_in_code_mode: Option<bool>,
}
/// Where model metadata was obtained during local catalog resolution.
@@ -1246,6 +1250,7 @@ mod tests {
tool_mode: None,
multi_agent_version: None,
multi_agent_reasoning_effort: None,
multi_agent_tools_in_code_mode: None,
}
}
#[test]
@@ -1948,6 +1953,24 @@ mod tests {
assert_eq!(model.multi_agent_reasoning_effort, None);
}
#[test]
fn model_info_preserves_multi_agent_code_mode_presence() {
let base = test_model(/*spec*/ None);
let base_json = serde_json::to_value(&base).expect("serialize model");
assert!(base_json.get("multi_agent_tools_in_code_mode").is_none());
assert_eq!(
serde_json::from_value::<ModelInfo>(base_json.clone()).unwrap(),
base
);
for value in [None, Some(false), Some(true)] {
let mut json = base_json.clone();
json["multi_agent_tools_in_code_mode"] = serde_json::to_value(value).unwrap();
let mut expected = base.clone();
expected.multi_agent_tools_in_code_mode = value;
assert_eq!(serde_json::from_value::<ModelInfo>(json).unwrap(), expected);
}
}
#[test]
fn model_info_deserializes_multi_agent_reasoning_effort() {
let mut value =