Default Guardian v2 async classification to conversation mode (#53733)

## What changed

Use `conversation` when neither a local `async_classifier_mode` override nor a model default specifies a mode. Keep `snapshot` available as an explicit override and update the configuration documentation and schema to describe the new fallback.

## Testing

Expand configuration tests to cover all local override and model default combinations. Update app-server approval tests to exercise conversation mode without an explicit override, and explicitly select snapshot mode in tests that depend on it.

GitOrigin-RevId: a62479a3f6454bf28d589373a236d8b0ab5023b9

codex-rs/app-server/tests/suite/v2/guardian_v2.rs

@@ -919,9 +919,12 @@ async fn guardian_v2_routes_scoped_tool_approvals(
     } else {
         "prompt"
     };
-    let classifier_mode = match classifier_mode {
-        codex_protocol::openai_models::AsyncClassifierMode::Snapshot => "snapshot",
-        codex_protocol::openai_models::AsyncClassifierMode::Conversation => "conversation",
+    // Exercise the default conversation path and the explicit snapshot rollback.
+    let classifier_config = match classifier_mode {
+        codex_protocol::openai_models::AsyncClassifierMode::Snapshot => {
+            "\nasync_classifier_mode = \"snapshot\""
+        }
+        codex_protocol::openai_models::AsyncClassifierMode::Conversation => "",
     };
     let mut mock_config = MockResponsesConfig::new(&responses_url)
         .with_model(MODEL)
@@ -932,7 +935,7 @@ async fn guardian_v2_routes_scoped_tool_approvals(
             analytics_server.uri(),
         ))
         .with_extra_config(&format!(
-            "[mcp_servers.{server_name}]\nurl = \"{mcp_server_url}/mcp\"\ndefault_tools_approval_mode = \"{tool_approval_mode}\"\n\n[analytics]\nenabled = true\n\n[otel]\nmetrics_exporter = {{ otlp-http = {{ endpoint = \"{responses_url}/metrics\", protocol = \"json\" }} }}\n\n[features.guardianv2]\nenabled = true\nasync_classifier_mode = \"{classifier_mode}\"{guardian_scope_config}"
+            "[mcp_servers.{server_name}]\nurl = \"{mcp_server_url}/mcp\"\ndefault_tools_approval_mode = \"{tool_approval_mode}\"\n\n[analytics]\nenabled = true\n\n[otel]\nmetrics_exporter = {{ otlp-http = {{ endpoint = \"{responses_url}/metrics\", protocol = \"json\" }} }}\n\n[features.guardianv2]\nenabled = true{classifier_config}{guardian_scope_config}"
         ))
         .enable_feature(Feature::GuardianApproval);
     if lifecycle.has_user_input() || lifecycle.has_root_user_input() {
@@ -1259,7 +1262,8 @@ async fn guardian_v2_routes_scoped_tool_approvals(
                 | ThreadLifecycle::RootUserRestriction
                 | ThreadLifecycle::RootRestrictionDuringClassification
         ) {
-            let retention_allowed = classifier_mode == "conversation"
+            let retention_allowed = classifier_mode
+                == codex_protocol::openai_models::AsyncClassifierMode::Conversation
                 && !fail_after_classification
                 && !matches!(risk, GuardianRisk::InvalidResponse)
                 && !late_root_restriction;
@@ -1553,7 +1557,7 @@ async fn guardian_v2_routes_scoped_tool_approvals(
                 .filter_map(|item| item["id"].as_str())
                 .filter(|id| id.starts_with("luna-score-message-"))
                 .collect::<Vec<_>>();
-            if classifier_mode == "conversation"
+            if classifier_mode == codex_protocol::openai_models::AsyncClassifierMode::Conversation
                 && !fail_after_classification
                 && let Some(latest) = retained_ids.last()
             {
@@ -1907,7 +1911,7 @@ async fn guardian_v2_routes_scoped_tool_approvals(
         }
         let post_authorization_change_sample =
             wait_for_luna_request(responses_state.as_ref(), /*index*/ 2).await?;
-        if classifier_mode == "conversation"
+        if classifier_mode == codex_protocol::openai_models::AsyncClassifierMode::Conversation
             && matches!(lifecycle, ThreadLifecycle::RootUserRestriction)
         {
             let input = post_authorization_change_sample["input"]

codex-rs/core/config.schema.json

@@ -1960,7 +1960,7 @@
               "$ref": "#/definitions/AsyncClassifierMode"
             }
           ],
-          "description": "Classifier experiment override. Otherwise use the model default, then snapshots."
+          "description": "Classifier mode override. Otherwise use the model default, then conversation."
         },
         "classifier_instructions": {
           "type": "string"

codex-rs/core/tests/suite/scenarios_guardian_orchestrator_connector.rs

@@ -94,7 +94,7 @@ async fn host_owned_orchestrator_connector_reaches_guardian_context() -> Result<
         .with_pre_build_hook(|home| {
             std::fs::write(
                 home.join("config.toml"),
-                "[features.guardianv2]\nenabled = true\n\n[features.guardianv2.review_scope]\ncomputer_use_only = false\n\n[apps._default]\ndefault_tools_approval_mode = \"prompt\"\n",
+                "[features.guardianv2]\nenabled = true\nasync_classifier_mode = \"snapshot\"\n\n[features.guardianv2.review_scope]\ncomputer_use_only = false\n\n[apps._default]\ndefault_tools_approval_mode = \"prompt\"\n",
             )
             .expect("write Guardian configuration");
         })

codex-rs/ext/guardian-v2/src/async_scorer/config_tests.rs

@@ -340,17 +340,23 @@ fn classifier_mode_uses_local_override_then_model_default() {
             ..Default::default()
         })
         .unwrap();
-        let defaults = GuardianV2ModelConfig {
-            async_classifier_mode: Some(AsyncClassifierMode::Conversation),
-            ..Default::default()
-        };
-        assert_eq!(
-            config.classifier_mode(Some(&defaults)),
-            local.unwrap_or(AsyncClassifierMode::Conversation)
-        );
+        for model in [
+            None,
+            Some(AsyncClassifierMode::Snapshot),
+            Some(AsyncClassifierMode::Conversation),
+        ] {
+            let defaults = GuardianV2ModelConfig {
+                async_classifier_mode: model,
+                ..Default::default()
+            };
+            assert_eq!(
+                config.classifier_mode(Some(&defaults)),
+                local.or(model).unwrap_or(AsyncClassifierMode::Conversation)
+            );
+        }
         assert_eq!(
             config.classifier_mode(/*model_defaults*/ None),
-            local.unwrap_or_default()
+            local.unwrap_or(AsyncClassifierMode::Conversation)
         );
     }
 }

codex-rs/ext/guardian-v2/src/async_scorer/decisions_lifecycle_tests.rs

@@ -17,7 +17,7 @@ async fn decisions_disagreement_failures_and_disabled_flag_preserve_baseline() -
     ] {
         let server = responses::start_mock_server().await;
         let fixture = GuardianFailureFixture::with_config(&format!(
-            "[features]\nguardianv2_decisions_comparison = {enabled}\n"
+            "[features]\nguardianv2_decisions_comparison = {enabled}\n[features.guardianv2]\nasync_classifier_mode = \"snapshot\"\n"
         ))
         .await?;
         let test = &fixture.test;
@@ -181,7 +181,11 @@ async fn decisions_disagreement_failures_and_disabled_flag_preserve_baseline() -
 #[tokio::test]
 async fn guardian_budget_rejects_both_backends_before_sending() -> Result<()> {
     skip_if_no_network!(Ok(()));
-    let fixture = GuardianFailureFixture::new().await?;
+    // This test replaces the snapshot sampler to force the Responses budget failure.
+    let fixture = GuardianFailureFixture::with_config(
+        "[features.guardianv2]\nenabled = true\nasync_classifier_mode = \"snapshot\"\n",
+    )
+    .await?;
     let server = responses::start_mock_server().await;
     let store = fixture.test.codex.thread_extension_data();
     let mut config = fixture.test.config.clone();

codex-rs/ext/guardian-v2/src/async_scorer/extension_tests.rs

@@ -2092,7 +2092,7 @@ async fn contributor_samples_tool_calls_with_the_existing_luna_pool() -> Result<
         conversation_history,
         r#"{"path":"README.md"}"#,
         Some(TEST_GUARDIAN_POLICY),
-        "[features.guardianv2.transcript]\nmax_message_entry_tokens = 100\nmax_message_transcript_tokens = 100\n",
+        "[features.guardianv2]\nasync_classifier_mode = \"snapshot\"\n[features.guardianv2.transcript]\nmax_message_entry_tokens = 100\nmax_message_transcript_tokens = 100\n",
         /*model_defaults*/ None,
     )
     .await?;

codex-rs/features/src/feature_configs.rs

@@ -152,7 +152,7 @@ pub struct GuardianV2ConfigToml {
     /// Transcript encoding for synchronous Guardian review and asynchronous scoring. Defaults to line.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub transcript_mode: Option<codex_protocol::TranscriptFormat>,
-    /// Classifier experiment override. Otherwise use the model default, then snapshots.
+    /// Classifier mode override. Otherwise use the model default, then conversation.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub async_classifier_mode: Option<codex_protocol::openai_models::AsyncClassifierMode>,
     /// Reset retained async history when the next request exceeds this token estimate.

codex-rs/protocol/src/openai_models/guardian_v2.rs

@@ -10,8 +10,8 @@ use super::ReasoningEffort;
 #[serde(rename_all = "snake_case")]
 #[ts(rename_all = "snake_case")]
 pub enum AsyncClassifierMode {
-    #[default]
     Snapshot,
+    #[default]
     Conversation,
 }