fix: enforce provider upstream stream policy

apps/aether-gateway/src/ai_serving/planner/common.rs

@@ -37,6 +37,52 @@ pub(crate) fn force_upstream_streaming_for_provider(
     force_upstream_streaming_for_provider_impl(provider_type, provider_api_format)
 }
 
+pub(crate) fn resolve_upstream_is_stream_for_provider(
+    endpoint_config: Option<&serde_json::Value>,
+    provider_type: &str,
+    provider_api_format: &str,
+    client_is_stream: bool,
+    hard_requires_streaming: bool,
+) -> bool {
+    let hard_requires_streaming = hard_requires_streaming
+        || force_upstream_streaming_for_provider(provider_type, provider_api_format);
+    aether_ai_formats::resolve_upstream_is_stream_from_endpoint_config(
+        endpoint_config,
+        client_is_stream,
+        hard_requires_streaming,
+    )
+}
+
+pub(crate) fn endpoint_config_forces_body_stream_field(
+    endpoint_config: Option<&serde_json::Value>,
+) -> bool {
+    aether_ai_formats::endpoint_config_forces_upstream_stream_policy(endpoint_config)
+}
+
+pub(crate) fn request_requires_body_stream_field(
+    body_json: &serde_json::Value,
+    force_body_stream_field: bool,
+) -> bool {
+    force_body_stream_field
+        || body_json
+            .as_object()
+            .is_some_and(|object| object.contains_key("stream"))
+}
+
+pub(crate) fn enforce_provider_body_stream_policy(
+    provider_request_body: &mut serde_json::Value,
+    provider_api_format: &str,
+    upstream_is_stream: bool,
+    require_body_stream_field: bool,
+) {
+    aether_ai_formats::enforce_request_body_stream_field(
+        provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        require_body_stream_field,
+    );
+}
+
 pub(crate) fn extract_standard_requested_model(body_json: &serde_json::Value) -> Option<String> {
     aether_ai_serving::extract_ai_standard_requested_model(body_json)
 }
@@ -56,8 +102,10 @@ pub(crate) fn extract_requested_model_from_request(
 #[cfg(test)]
 mod tests {
     use super::{
+        endpoint_config_forces_body_stream_field, enforce_provider_body_stream_policy,
         extract_requested_model_from_request, extract_standard_requested_model,
-        force_upstream_streaming_for_provider, RequestedModelFamily,
+        force_upstream_streaming_for_provider, resolve_upstream_is_stream_for_provider,
+        RequestedModelFamily,
     };
     use axum::http::Request;
     use serde_json::json;
@@ -90,6 +138,67 @@ mod tests {
         ));
     }
 
+    #[test]
+    fn resolves_endpoint_upstream_stream_policy_with_provider_hard_constraints() {
+        assert!(resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "force_stream"})),
+            "openai",
+            "openai:chat",
+            false,
+            false,
+        ));
+        assert!(!resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "force_non_stream"})),
+            "openai",
+            "openai:chat",
+            true,
+            false,
+        ));
+        assert!(resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "auto"})),
+            "openai",
+            "openai:chat",
+            true,
+            false,
+        ));
+        assert!(resolve_upstream_is_stream_for_provider(
+            Some(&json!({"upstream_stream_policy": "force_non_stream"})),
+            "codex",
+            "openai:responses",
+            true,
+            false,
+        ));
+    }
+
+    #[test]
+    fn enforces_provider_body_stream_policy_for_body_and_streamless_formats() {
+        let mut openai_chat = json!({"stream": true});
+        enforce_provider_body_stream_policy(&mut openai_chat, "openai:chat", false, false);
+        assert_eq!(openai_chat.get("stream"), Some(&json!(false)));
+
+        let mut ordinary_sync = json!({"messages": []});
+        enforce_provider_body_stream_policy(&mut ordinary_sync, "openai:chat", false, false);
+        assert!(ordinary_sync.get("stream").is_none());
+
+        let mut compact = json!({"stream": true});
+        enforce_provider_body_stream_policy(&mut compact, "openai:responses:compact", true, true);
+        assert!(compact.get("stream").is_none());
+    }
+
+    #[test]
+    fn detects_endpoint_configs_that_force_body_stream_field() {
+        assert!(endpoint_config_forces_body_stream_field(Some(
+            &json!({"upstream_stream_policy": "force_stream"})
+        )));
+        assert!(endpoint_config_forces_body_stream_field(Some(
+            &json!({"upstream_stream_policy": "force_non_stream"})
+        )));
+        assert!(!endpoint_config_forces_body_stream_field(Some(
+            &json!({"upstream_stream_policy": "auto"})
+        )));
+        assert!(!endpoint_config_forces_body_stream_field(None));
+    }
+
     #[test]
     fn extracts_standard_requested_model_from_request_body() {
         let requested_model =

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/request.rs

@@ -3,6 +3,9 @@ use std::sync::Arc;
 
 use serde_json::Value;
 
+use crate::ai_serving::planner::common::{
+    enforce_provider_body_stream_policy, request_requires_body_stream_field,
+};
 use crate::ai_serving::transport::antigravity::{
     build_antigravity_safe_v1internal_request, build_antigravity_static_identity_headers,
     classify_local_antigravity_request_support, AntigravityEnvelopeRequestType,
@@ -50,6 +53,7 @@ pub(crate) fn resolve_same_format_provider_transport_unsupported_reason_for_trac
         };
     let behavior = policy::classify_same_format_provider_request_behavior(
         transport,
+        provider_api_format,
         crate::ai_serving::planner::spec_metadata::LocalExecutionSurfaceSpecMetadata {
             api_format: provider_api_format,
             require_streaming: false,
@@ -131,6 +135,7 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
             prepared.transport.endpoint.body_rules.as_ref(),
             Some(&parts.headers),
             prepared.upstream_is_stream,
+            prepared.force_body_stream_field,
             prepared.kiro_auth.as_ref(),
             prepared.is_claude_code,
             enable_model_directives,
@@ -170,6 +175,18 @@ pub(crate) async fn resolve_local_same_format_provider_candidate_payload_parts(
             &mut base_provider_request_body,
             &mapping,
         );
+        // Directive mapping is a deep-merge patch and may overwrite/add `stream`;
+        // re-enforce stream-field policy afterward.
+        // Kiro behavior classification already hard-requires upstream streaming,
+        // and the Kiro envelope does not use a top-level body stream field.
+        if prepared.kiro_auth.is_none() {
+            enforce_provider_body_stream_policy(
+                &mut base_provider_request_body,
+                prepared.provider_api_format.as_str(),
+                prepared.upstream_is_stream,
+                request_requires_body_stream_field(body_json, prepared.force_body_stream_field),
+            );
+        }
     }
 
     let antigravity_auth = if prepared.is_antigravity {

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/request/policy.rs

@@ -13,12 +13,14 @@ use super::super::LocalSameFormatProviderFamily;
 
 pub(super) fn classify_same_format_provider_request_behavior(
     transport: &GatewayProviderTransportSnapshot,
+    provider_api_format: &str,
     spec_metadata: LocalExecutionSurfaceSpecMetadata,
 ) -> SameFormatProviderRequestBehavior {
     classify_same_format_provider_request_behavior_impl(
         transport,
         SameFormatProviderRequestBehaviorParams {
             require_streaming: spec_metadata.require_streaming,
+            provider_api_format,
             report_kind: spec_metadata
                 .report_kind
                 .expect("same-format provider specs should declare report kind"),

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/family/request/prepare.rs

@@ -35,6 +35,7 @@ pub(super) struct PreparedSameFormatProviderCandidate {
     pub(super) mapped_model: String,
     pub(super) report_kind: &'static str,
     pub(super) upstream_is_stream: bool,
+    pub(super) force_body_stream_field: bool,
 }
 
 pub(super) async fn prepare_local_same_format_provider_candidate(
@@ -51,7 +52,11 @@ pub(super) async fn prepare_local_same_format_provider_candidate(
     let candidate = &eligible.candidate;
     let transport = Arc::clone(&eligible.transport);
     let provider_api_format = eligible.provider_api_format.as_str();
-    let behavior = classify_same_format_provider_request_behavior(&transport, spec_metadata);
+    let behavior = classify_same_format_provider_request_behavior(
+        &transport,
+        provider_api_format,
+        spec_metadata,
+    );
 
     if !same_format_provider_transport_supported(
         &behavior,
@@ -174,5 +179,6 @@ pub(super) async fn prepare_local_same_format_provider_candidate(
         mapped_model,
         report_kind: behavior.report_kind,
         upstream_is_stream: behavior.upstream_is_stream,
+        force_body_stream_field: behavior.force_body_stream_field,
     })
 }

apps/aether-gateway/src/ai_serving/planner/passthrough/provider/request/body.rs

@@ -14,6 +14,7 @@ pub(crate) fn build_same_format_provider_request_body(
     body_rules: Option<&Value>,
     request_headers: Option<&http::HeaderMap>,
     upstream_is_stream: bool,
+    force_body_stream_field: bool,
     kiro_auth: Option<&crate::ai_serving::transport::kiro::KiroRequestAuth>,
     is_claude_code: bool,
     enable_model_directives: bool,
@@ -28,6 +29,7 @@ pub(crate) fn build_same_format_provider_request_body(
         body_rules,
         request_headers,
         upstream_is_stream,
+        force_body_stream_field,
         kiro_auth_config: kiro_auth.map(|auth| &auth.auth_config),
         is_claude_code,
         enable_model_directives,

apps/aether-gateway/src/ai_serving/planner/standard/family/request.rs

@@ -7,7 +7,10 @@ use crate::ai_serving::planner::candidate_preparation::{
     prepare_header_authenticated_candidate, prepare_header_authenticated_candidate_from_auth,
     OauthPreparationContext,
 };
-use crate::ai_serving::planner::common::force_upstream_streaming_for_provider;
+use crate::ai_serving::planner::common::{
+    endpoint_config_forces_body_stream_field, enforce_provider_body_stream_policy,
+    request_requires_body_stream_field, resolve_upstream_is_stream_for_provider,
+};
 use crate::ai_serving::planner::spec_metadata::local_standard_spec_metadata;
 use crate::ai_serving::planner::standard::{
     apply_codex_openai_responses_special_headers, request_body_build_failure_extra_data,
@@ -175,11 +178,15 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
         }
     };
 
-    let upstream_is_stream = spec_metadata.require_streaming
+    let upstream_is_stream = resolve_upstream_is_stream_for_provider(
-        || force_upstream_streaming_for_provider(
+        transport.endpoint.config.as_ref(),
-            transport.provider.provider_type.as_str(),
+        transport.provider.provider_type.as_str(),
-            provider_api_format,
+        provider_api_format,
-        );
+        spec_metadata.require_streaming,
+        is_kiro_claude_cli,
+    );
+    let force_body_stream_field =
+        endpoint_config_forces_body_stream_field(transport.endpoint.config.as_ref());
     let enable_model_directives =
         crate::system_features::reasoning_model_directive_enabled_for_api_format_and_model(
             state,
@@ -225,6 +232,12 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
                 return None;
             }
         };
+    enforce_provider_body_stream_policy(
+        &mut provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        request_requires_body_stream_field(body_json, force_body_stream_field),
+    );
     if let Some(mapping) =
         crate::system_features::reasoning_model_directive_mapping_for_api_format_and_model(
             state,
@@ -237,6 +250,14 @@ pub(crate) async fn resolve_local_standard_candidate_payload_parts(
             &mut provider_request_body,
             &mapping,
         );
+        // Directive mapping is a deep-merge patch and may overwrite/add `stream`;
+        // re-enforce stream-field policy afterward.
+        enforce_provider_body_stream_policy(
+            &mut provider_request_body,
+            provider_api_format,
+            upstream_is_stream,
+            request_requires_body_stream_field(body_json, force_body_stream_field),
+        );
     }
 
     if let Some(kiro_auth) = kiro_auth.as_ref() {

apps/aether-gateway/src/ai_serving/planner/standard/gemini/plan_builders.rs

@@ -116,9 +116,9 @@ pub(crate) fn build_gemini_stream_plan_from_decision(
             content_type: payload.content_type.as_deref(),
             provider_api_format: core.provider_api_format.as_str(),
             client_api_format: core.client_api_format.as_str(),
-            upstream_is_stream: true,
+            upstream_is_stream: payload.upstream_is_stream,
             build_from_request_when_empty: false,
-            accept_policy: StandardPlanFallbackAcceptPolicy::TextEventStreamRequired,
+            accept_policy: StandardPlanFallbackAcceptPolicy::TextEventStreamIfStreaming,
         });
     let content_type = payload
         .content_type

apps/aether-gateway/src/ai_serving/planner/standard/normalize.rs

@@ -15,3 +15,6 @@ pub(crate) use self::responses::{
     build_cross_format_openai_responses_upstream_url, build_local_openai_responses_request_body,
     build_local_openai_responses_upstream_url,
 };
+pub(super) use crate::ai_serving::planner::common::{
+    enforce_provider_body_stream_policy, request_requires_body_stream_field,
+};

apps/aether-gateway/src/ai_serving/planner/standard/normalize/chat.rs

@@ -9,10 +9,13 @@ use crate::ai_serving::{
     GatewayProviderTransportSnapshot,
 };
 
+use super::{enforce_provider_body_stream_policy, request_requires_body_stream_field};
+
 pub(crate) fn build_local_openai_chat_request_body(
     body_json: &Value,
     mapped_model: &str,
     upstream_is_stream: bool,
+    force_body_stream_field: bool,
     body_rules: Option<&Value>,
     request_headers: &http::HeaderMap,
     enable_model_directives: bool,
@@ -23,12 +26,20 @@ pub(crate) fn build_local_openai_chat_request_body(
         upstream_is_stream,
         enable_model_directives,
     )?;
-    apply_standard_provider_request_body_rules_with_request_headers(
+    let mut provider_request_body =
-        provider_request_body,
+        apply_standard_provider_request_body_rules_with_request_headers(
-        body_rules,
+            provider_request_body,
-        body_json,
+            body_rules,
-        request_headers,
+            body_json,
-    )
+            request_headers,
+        )?;
+    enforce_provider_body_stream_policy(
+        &mut provider_request_body,
+        "openai:chat",
+        upstream_is_stream,
+        request_requires_body_stream_field(body_json, force_body_stream_field),
+    );
+    Some(provider_request_body)
 }
 
 pub(crate) fn build_local_openai_chat_upstream_url(
@@ -44,6 +55,7 @@ pub(crate) fn build_cross_format_openai_chat_request_body(
     provider_type: &str,
     provider_api_format: &str,
     upstream_is_stream: bool,
+    force_body_stream_field: bool,
     body_rules: Option<&Value>,
     user_api_key_id: Option<&str>,
     request_headers: &http::HeaderMap,
@@ -74,6 +86,12 @@ pub(crate) fn build_cross_format_openai_chat_request_body(
         &mut provider_request_body,
         provider_api_format,
     );
+    enforce_provider_body_stream_policy(
+        &mut provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        request_requires_body_stream_field(body_json, force_body_stream_field),
+    );
     Some(provider_request_body)
 }
 

apps/aether-gateway/src/ai_serving/planner/standard/normalize/responses.rs

@@ -9,10 +9,13 @@ use crate::ai_serving::{
     GatewayProviderTransportSnapshot,
 };
 
+use super::{enforce_provider_body_stream_policy, request_requires_body_stream_field};
+
 pub(crate) fn build_local_openai_responses_request_body(
     body_json: &Value,
     mapped_model: &str,
     require_streaming: bool,
+    force_body_stream_field: bool,
     provider_type: &str,
     provider_api_format: &str,
     body_rules: Option<&Value>,
@@ -44,6 +47,12 @@ pub(crate) fn build_local_openai_responses_request_body(
         &mut provider_request_body,
         provider_api_format,
     );
+    enforce_provider_body_stream_policy(
+        &mut provider_request_body,
+        provider_api_format,
+        require_streaming,
+        request_requires_body_stream_field(body_json, force_body_stream_field),
+    );
     Some(provider_request_body)
 }
 
@@ -53,6 +62,7 @@ pub(crate) fn build_cross_format_openai_responses_request_body(
     client_api_format: &str,
     provider_api_format: &str,
     upstream_is_stream: bool,
+    force_body_stream_field: bool,
     provider_type: &str,
     body_rules: Option<&Value>,
     user_api_key_id: Option<&str>,
@@ -85,6 +95,12 @@ pub(crate) fn build_cross_format_openai_responses_request_body(
         &mut provider_request_body,
         provider_api_format,
     );
+    enforce_provider_body_stream_policy(
+        &mut provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        request_requires_body_stream_field(body_json, force_body_stream_field),
+    );
     Some(provider_request_body)
 }
 

apps/aether-gateway/src/ai_serving/planner/standard/normalize/tests.rs

@@ -87,6 +87,7 @@ fn builds_openai_chat_cross_format_request_body_from_openai_responses_source() {
         "openai:responses",
         "openai:chat",
         false,
+        false,
         "openai",
         None,
         None,
@@ -121,6 +122,7 @@ fn local_openai_responses_wrapper_preserves_body_order_after_edits() {
         &body_json,
         "gpt-5.4",
         true,
+        false,
         "codex",
         "openai:responses",
         None,
@@ -160,6 +162,7 @@ fn local_openai_responses_compact_wrapper_strips_store_for_same_format_requests(
         &body_json,
         "gpt-5.4",
         false,
+        false,
         "openai",
         "openai:responses:compact",
         None,
@@ -170,6 +173,7 @@ fn local_openai_responses_compact_wrapper_strips_store_for_same_format_requests(
     .expect("local openai compact body should build");
 
     assert!(provider_request_body.get("store").is_none());
+    assert!(provider_request_body.get("stream").is_none());
 }
 
 #[test]
@@ -187,6 +191,7 @@ fn local_openai_responses_wrapper_applies_model_directive_before_body_rules() {
         &body_json,
         "gpt-5.4",
         false,
+        false,
         "openai",
         "openai:responses",
         Some(&body_rules),
@@ -237,6 +242,7 @@ fn strips_metadata_for_codex_openai_responses_requests() {
         "claude:messages",
         "openai:responses",
         true,
+        false,
         "codex",
         None,
         None,
@@ -271,6 +277,7 @@ fn applies_codex_defaults_unless_body_rules_handle_the_field() {
         "claude:messages",
         "openai:responses",
         true,
+        false,
         "codex",
         Some(&body_rules),
         None,
@@ -300,6 +307,7 @@ fn injects_codex_prompt_cache_key_for_openai_responses_cross_format_requests() {
         "claude:messages",
         "openai:responses",
         true,
+        false,
         "codex",
         None,
         Some("key-123"),
@@ -330,6 +338,7 @@ fn injects_codex_prompt_cache_key_for_openai_chat_cross_format_requests() {
         "codex",
         "openai:responses",
         false,
+        false,
         None,
         Some("key-123"),
         &http::HeaderMap::new(),

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/decision/payload.rs

@@ -1,4 +1,5 @@
 use crate::ai_serving::build_request_trace_proxy_value;
+use crate::ai_serving::planner::common::OPENAI_CHAT_STREAM_PLAN_KIND;
 use crate::ai_serving::planner::report_context::{
     build_local_execution_report_context, insert_provider_stream_event_api_format,
     LocalExecutionReportContextParts,
@@ -29,6 +30,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
     report_kind: &str,
     upstream_is_stream: bool,
 ) -> Option<AiExecutionDecision> {
+    let decision_is_stream = decision_kind == OPENAI_CHAT_STREAM_PLAN_KIND;
     let attempt_identity = attempt.attempt_identity();
     let LocalOpenAiChatCandidateAttempt {
         eligible,
@@ -149,7 +151,7 @@ pub(crate) async fn maybe_build_local_openai_chat_decision_payload_for_candidate
 
     Some(build_ai_execution_decision_response(
         AiExecutionDecisionResponseParts {
-            decision_is_stream: upstream_is_stream,
+            decision_is_stream,
             decision_kind: decision_kind.to_string(),
             execution_strategy,
             conversion_mode,

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/decision/request.rs

@@ -8,7 +8,10 @@ use crate::ai_serving::planner::candidate_preparation::{
     OauthPreparationContext,
 };
 use crate::ai_serving::planner::candidate_resolution::EligibleLocalExecutionCandidate;
-use crate::ai_serving::planner::common::OPENAI_CHAT_STREAM_PLAN_KIND;
+use crate::ai_serving::planner::common::{
+    endpoint_config_forces_body_stream_field, enforce_provider_body_stream_policy,
+    request_requires_body_stream_field, OPENAI_CHAT_STREAM_PLAN_KIND,
+};
 use crate::ai_serving::planner::standard::{
     apply_codex_openai_responses_special_headers, build_cross_format_openai_chat_request_body,
     build_cross_format_openai_chat_upstream_url, build_local_openai_chat_request_body,
@@ -72,6 +75,8 @@ pub(crate) async fn resolve_local_openai_chat_candidate_payload_parts(
     let candidate = &eligible.candidate;
     let provider_api_format = eligible.provider_api_format.as_str();
     let transport = &eligible.transport;
+    let force_body_stream_field =
+        endpoint_config_forces_body_stream_field(transport.endpoint.config.as_ref());
     let enable_model_directives =
         crate::system_features::reasoning_model_directive_enabled_for_api_format_and_model(
             state,
@@ -128,6 +133,7 @@ pub(crate) async fn resolve_local_openai_chat_candidate_payload_parts(
             body_json,
             &prepared_candidate.mapped_model,
             upstream_is_stream,
+            force_body_stream_field,
             transport.endpoint.body_rules.as_ref(),
             &parts.headers,
             enable_model_directives,
@@ -214,6 +220,13 @@ pub(crate) async fn resolve_local_openai_chat_candidate_payload_parts(
 
         let (execution_strategy, conversion_mode) =
             ai_local_execution_contract_for_formats("openai:chat", "openai:chat");
+        let resolved_report_kind =
+            if decision_kind == OPENAI_CHAT_STREAM_PLAN_KIND || !upstream_is_stream {
+                report_kind.to_string()
+            } else {
+                "openai_chat_sync_finalize".to_string()
+            };
+
         return Some(LocalOpenAiChatCandidatePayloadParts {
             auth_header: resolved_headers.auth_header,
             auth_value: resolved_headers.auth_value,
@@ -224,7 +237,7 @@ pub(crate) async fn resolve_local_openai_chat_candidate_payload_parts(
             upstream_url,
             execution_strategy,
             conversion_mode,
-            report_kind: report_kind.to_string(),
+            report_kind: resolved_report_kind,
             envelope_name: None,
             transport: Arc::clone(transport),
         });
@@ -349,6 +362,7 @@ pub(crate) async fn resolve_local_openai_chat_candidate_payload_parts(
         transport.provider.provider_type.as_str(),
         provider_api_format.as_str(),
         upstream_is_stream,
+        force_body_stream_field,
         if is_kiro_claude_cli {
             None
         } else {
@@ -387,6 +401,14 @@ pub(crate) async fn resolve_local_openai_chat_candidate_payload_parts(
             &mut provider_request_body,
             &mapping,
         );
+        // Directive mapping is a deep-merge patch and may overwrite/add `stream`;
+        // re-enforce stream-field policy afterward.
+        enforce_provider_body_stream_policy(
+            &mut provider_request_body,
+            provider_api_format.as_str(),
+            upstream_is_stream,
+            request_requires_body_stream_field(body_json, force_body_stream_field),
+        );
     }
 
     if let Some(kiro_auth) = kiro_auth.as_ref() {

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/plans.rs

@@ -9,6 +9,9 @@ mod stream;
 #[path = "plans/sync.rs"]
 mod sync;
 
+use crate::ai_serving::planner::common::resolve_upstream_is_stream_for_provider;
+use crate::ai_serving::GatewayProviderTransportSnapshot;
+
 pub(super) use self::candidates::list_local_openai_chat_candidates;
 pub(super) use self::diagnostic::set_local_openai_chat_miss_diagnostic;
 pub(super) use self::resolve::resolve_local_openai_chat_decision_input;
@@ -18,3 +21,147 @@ pub(super) use self::stream::{
 pub(super) use self::sync::{
     build_local_openai_chat_sync_attempt_source, build_local_openai_chat_sync_plan_and_reports,
 };
+
+fn openai_chat_upstream_is_stream_for_candidate(
+    transport: &GatewayProviderTransportSnapshot,
+    provider_api_format: &str,
+    client_is_stream: bool,
+) -> bool {
+    let hard_requires_streaming =
+        crate::ai_serving::transport::kiro::is_kiro_claude_messages_transport(
+            transport,
+            provider_api_format,
+        );
+    resolve_upstream_is_stream_for_provider(
+        transport.endpoint.config.as_ref(),
+        transport.provider.provider_type.as_str(),
+        provider_api_format,
+        client_is_stream,
+        hard_requires_streaming,
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::openai_chat_upstream_is_stream_for_candidate;
+    use aether_provider_transport::snapshot::{
+        GatewayProviderTransportEndpoint, GatewayProviderTransportKey,
+        GatewayProviderTransportProvider, GatewayProviderTransportSnapshot,
+    };
+    use serde_json::{json, Value};
+
+    fn sample_transport(
+        provider_type: &str,
+        api_format: &str,
+        endpoint_config: Option<Value>,
+    ) -> GatewayProviderTransportSnapshot {
+        GatewayProviderTransportSnapshot {
+            provider: GatewayProviderTransportProvider {
+                id: "provider-1".to_string(),
+                name: "Provider".to_string(),
+                provider_type: provider_type.to_string(),
+                website: None,
+                is_active: true,
+                keep_priority_on_conversion: false,
+                enable_format_conversion: true,
+                concurrent_limit: None,
+                max_retries: None,
+                proxy: None,
+                request_timeout_secs: None,
+                stream_first_byte_timeout_secs: None,
+                config: None,
+            },
+            endpoint: GatewayProviderTransportEndpoint {
+                id: "endpoint-1".to_string(),
+                provider_id: "provider-1".to_string(),
+                api_format: api_format.to_string(),
+                api_family: None,
+                endpoint_kind: None,
+                is_active: true,
+                base_url: "https://api.example.test".to_string(),
+                header_rules: None,
+                body_rules: None,
+                max_retries: None,
+                custom_path: None,
+                config: endpoint_config,
+                format_acceptance_config: None,
+                proxy: None,
+            },
+            key: GatewayProviderTransportKey {
+                id: "key-1".to_string(),
+                provider_id: "provider-1".to_string(),
+                name: "key".to_string(),
+                auth_type: "api_key".to_string(),
+                is_active: true,
+                api_formats: None,
+                auth_type_by_format: None,
+                allow_auth_channel_mismatch_formats: None,
+                allowed_models: None,
+                capabilities: None,
+                rate_multipliers: None,
+                global_priority_by_format: None,
+                expires_at_unix_secs: None,
+                proxy: None,
+                fingerprint: None,
+                decrypted_api_key: "secret".to_string(),
+                decrypted_auth_config: None,
+            },
+        }
+    }
+
+    #[test]
+    fn openai_chat_policy_resolver_supports_force_stream_force_non_stream_and_auto() {
+        let force_stream = sample_transport(
+            "openai",
+            "openai:chat",
+            Some(json!({"upstream_stream_policy": "force_stream"})),
+        );
+        assert!(openai_chat_upstream_is_stream_for_candidate(
+            &force_stream,
+            "openai:chat",
+            false,
+        ));
+
+        let force_non_stream = sample_transport(
+            "openai",
+            "openai:chat",
+            Some(json!({"upstream_stream_policy": "force_non_stream"})),
+        );
+        assert!(!openai_chat_upstream_is_stream_for_candidate(
+            &force_non_stream,
+            "openai:chat",
+            true,
+        ));
+
+        let auto = sample_transport(
+            "openai",
+            "openai:chat",
+            Some(json!({"upstream_stream_policy": "auto"})),
+        );
+        assert!(openai_chat_upstream_is_stream_for_candidate(
+            &auto,
+            "openai:chat",
+            true,
+        ));
+        assert!(!openai_chat_upstream_is_stream_for_candidate(
+            &auto,
+            "openai:chat",
+            false,
+        ));
+    }
+
+    #[test]
+    fn openai_chat_policy_resolver_preserves_provider_hard_streaming() {
+        let codex = sample_transport(
+            "codex",
+            "openai:responses",
+            Some(json!({"upstream_stream_policy": "force_non_stream"})),
+        );
+
+        assert!(openai_chat_upstream_is_stream_for_candidate(
+            &codex,
+            "openai:responses",
+            false,
+        ));
+    }
+}

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/plans/stream.rs

@@ -13,6 +13,7 @@ use super::candidates::list_local_openai_chat_candidates;
 use super::diagnostic::{
     set_local_openai_chat_candidate_evaluation_diagnostic, set_local_openai_chat_miss_diagnostic,
 };
+use super::openai_chat_upstream_is_stream_for_candidate;
 use super::resolve::resolve_local_openai_chat_decision_input;
 use crate::ai_serving::planner::candidate_materialization::LocalExecutionAttemptSource;
 use crate::ai_serving::planner::common::OPENAI_CHAT_STREAM_PLAN_KIND;
@@ -120,6 +121,11 @@ impl LocalOpenAiChatStreamAttemptSource<'_> {
         &self,
         attempt: LocalOpenAiChatCandidateAttempt,
     ) -> Result<Option<AiStreamAttempt>, GatewayError> {
+        let upstream_is_stream = openai_chat_upstream_is_stream_for_candidate(
+            &attempt.eligible.transport,
+            attempt.eligible.provider_api_format.as_str(),
+            true,
+        );
         let Some(payload) = maybe_build_local_openai_chat_decision_payload_for_candidate(
             self.state,
             self.parts,
@@ -129,7 +135,7 @@ impl LocalOpenAiChatStreamAttemptSource<'_> {
             attempt,
             OPENAI_CHAT_STREAM_PLAN_KIND,
             "openai_chat_stream_success",
-            true,
+            upstream_is_stream,
         )
         .await
         else {
@@ -222,6 +228,11 @@ pub(crate) async fn build_local_openai_chat_stream_plan_and_reports(
 
     let mut plans = Vec::new();
     for attempt in attempts {
+        let upstream_is_stream = openai_chat_upstream_is_stream_for_candidate(
+            &attempt.eligible.transport,
+            attempt.eligible.provider_api_format.as_str(),
+            true,
+        );
         let Some(payload) = maybe_build_local_openai_chat_decision_payload_for_candidate(
             state,
             parts,
@@ -231,7 +242,7 @@ pub(crate) async fn build_local_openai_chat_stream_plan_and_reports(
             attempt,
             OPENAI_CHAT_STREAM_PLAN_KIND,
             "openai_chat_stream_success",
-            true,
+            upstream_is_stream,
         )
         .await
         else {

apps/aether-gateway/src/ai_serving/planner/standard/openai/chat/plans/sync.rs

@@ -13,11 +13,10 @@ use super::candidates::list_local_openai_chat_candidates;
 use super::diagnostic::{
     set_local_openai_chat_candidate_evaluation_diagnostic, set_local_openai_chat_miss_diagnostic,
 };
+use super::openai_chat_upstream_is_stream_for_candidate;
 use super::resolve::resolve_local_openai_chat_decision_input;
 use crate::ai_serving::planner::candidate_materialization::LocalExecutionAttemptSource;
-use crate::ai_serving::planner::common::{
+use crate::ai_serving::planner::common::OPENAI_CHAT_SYNC_PLAN_KIND;
-    force_upstream_streaming_for_provider, OPENAI_CHAT_SYNC_PLAN_KIND,
-};
 use crate::ai_serving::planner::plan_builders::{
     build_openai_chat_sync_plan_from_decision, AiSyncAttempt,
 };
@@ -32,13 +31,6 @@ pub(crate) struct LocalOpenAiChatSyncAttemptSource<'a> {
     candidates: LocalOpenAiChatCandidateAttemptSource<'a>,
 }
 
-fn openai_chat_sync_upstream_is_stream_for_candidate(
-    provider_type: &str,
-    provider_api_format: &str,
-) -> bool {
-    force_upstream_streaming_for_provider(provider_type, provider_api_format)
-}
-
 pub(crate) async fn build_local_openai_chat_sync_attempt_source<'a>(
     state: &'a AppState,
     parts: &'a http::request::Parts,
@@ -129,9 +121,10 @@ impl LocalOpenAiChatSyncAttemptSource<'_> {
         &self,
         attempt: LocalOpenAiChatCandidateAttempt,
     ) -> Result<Option<AiSyncAttempt>, GatewayError> {
-        let upstream_is_stream = openai_chat_sync_upstream_is_stream_for_candidate(
+        let upstream_is_stream = openai_chat_upstream_is_stream_for_candidate(
-            attempt.eligible.transport.provider.provider_type.as_str(),
+            &attempt.eligible.transport,
             attempt.eligible.provider_api_format.as_str(),
+            false,
         );
         let Some(payload) = maybe_build_local_openai_chat_decision_payload_for_candidate(
             self.state,
@@ -235,9 +228,10 @@ pub(crate) async fn build_local_openai_chat_sync_plan_and_reports(
 
     let mut plans = Vec::new();
     for attempt in attempts {
-        let upstream_is_stream = openai_chat_sync_upstream_is_stream_for_candidate(
+        let upstream_is_stream = openai_chat_upstream_is_stream_for_candidate(
-            attempt.eligible.transport.provider.provider_type.as_str(),
+            &attempt.eligible.transport,
             attempt.eligible.provider_api_format.as_str(),
+            false,
         );
         let Some(payload) = maybe_build_local_openai_chat_decision_payload_for_candidate(
             state,
@@ -272,28 +266,3 @@ pub(crate) async fn build_local_openai_chat_sync_plan_and_reports(
 
     Ok(plans)
 }
-
-#[cfg(test)]
-mod tests {
-    use super::openai_chat_sync_upstream_is_stream_for_candidate;
-
-    #[test]
-    fn openai_chat_sync_forces_streaming_for_codex_openai_responses_candidates() {
-        assert!(openai_chat_sync_upstream_is_stream_for_candidate(
-            "codex",
-            "openai:responses"
-        ));
-        assert!(!openai_chat_sync_upstream_is_stream_for_candidate(
-            "codex",
-            "openai:responses:compact"
-        ));
-        assert!(!openai_chat_sync_upstream_is_stream_for_candidate(
-            "openai",
-            "openai:responses"
-        ));
-        assert!(!openai_chat_sync_upstream_is_stream_for_candidate(
-            "codex",
-            "openai:chat"
-        ));
-    }
-}

apps/aether-gateway/src/ai_serving/planner/standard/openai/plan_builders/stream.rs

@@ -9,6 +9,7 @@ use super::super::{
     take_ai_upstream_auth_pair, take_non_empty_string, AiExecutionPlanFromDecisionParts,
     AiStreamAttempt,
 };
+use crate::ai_serving::planner::common::enforce_provider_body_stream_policy;
 use crate::ai_serving::provider_adaptation_requires_eventstream_accept;
 use crate::ai_serving::transport::{
     build_standard_plan_fallback_headers, build_standard_plan_fallback_openai_chat_url,
@@ -53,21 +54,31 @@ pub(crate) fn build_openai_chat_stream_plan_from_decision(
             provider_request_body
                 .insert("model".to_string(), serde_json::Value::String(mapped_model));
         }
-        provider_request_body.insert("stream".to_string(), serde_json::Value::Bool(true));
+        let require_body_stream_field = provider_request_body.contains_key("stream");
+        let mut provider_request_body = serde_json::Value::Object(provider_request_body);
+        enforce_provider_body_stream_policy(
+            &mut provider_request_body,
+            core.provider_api_format.as_str(),
+            payload.upstream_is_stream,
+            require_body_stream_field,
+        );
+        let Some(provider_request_object) = provider_request_body.as_object_mut() else {
+            return Ok(None);
+        };
         if let Some(prompt_cache_key) = take_non_empty_string(&mut payload.prompt_cache_key) {
-            let existing = provider_request_body
+            let existing = provider_request_object
                 .get("prompt_cache_key")
                 .and_then(|value| value.as_str())
                 .map(str::trim)
                 .unwrap_or_default();
             if existing.is_empty() {
-                provider_request_body.insert(
+                provider_request_object.insert(
                     "prompt_cache_key".to_string(),
                     serde_json::Value::String(prompt_cache_key),
                 );
             }
         }
-        serde_json::Value::Object(provider_request_body)
+        provider_request_body
     };
     let extra_headers = std::mem::take(&mut payload.extra_headers);
     let mut provider_request_headers =
@@ -82,9 +93,9 @@ pub(crate) fn build_openai_chat_stream_plan_from_decision(
             content_type: payload.content_type.as_deref(),
             provider_api_format: core.provider_api_format.as_str(),
             client_api_format: core.client_api_format.as_str(),
-            upstream_is_stream: true,
+            upstream_is_stream: payload.upstream_is_stream,
             build_from_request_when_empty: true,
-            accept_policy: StandardPlanFallbackAcceptPolicy::TextEventStreamRequired,
+            accept_policy: StandardPlanFallbackAcceptPolicy::TextEventStreamIfStreaming,
         });
     let content_type = payload
         .content_type
@@ -157,13 +168,14 @@ pub(crate) fn build_openai_responses_stream_plan_from_decision(
         .as_ref()
         .and_then(|context| context.get("envelope_name"))
         .and_then(serde_json::Value::as_str);
-    let accept_policy = if provider_adaptation_requires_eventstream_accept(
+    let accept_policy = if payload.upstream_is_stream
-        envelope_name,
+        && provider_adaptation_requires_eventstream_accept(
-        core.provider_api_format.as_str(),
+            envelope_name,
-    ) {
+            core.provider_api_format.as_str(),
+        ) {
         StandardPlanFallbackAcceptPolicy::ProviderEventStreamIfMissing
     } else {
-        StandardPlanFallbackAcceptPolicy::TextEventStreamRequired
+        StandardPlanFallbackAcceptPolicy::TextEventStreamIfStreaming
     };
     let mut provider_request_headers =
         build_standard_plan_fallback_headers(StandardPlanFallbackHeadersInput {
@@ -177,7 +189,7 @@ pub(crate) fn build_openai_responses_stream_plan_from_decision(
             content_type: payload.content_type.as_deref(),
             provider_api_format: core.provider_api_format.as_str(),
             client_api_format: core.client_api_format.as_str(),
-            upstream_is_stream: true,
+            upstream_is_stream: payload.upstream_is_stream,
             build_from_request_when_empty: false,
             accept_policy,
         });
@@ -427,6 +439,104 @@ mod tests {
         );
     }
 
+    #[test]
+    fn build_openai_chat_stream_plan_keeps_downstream_stream_for_force_non_stream_upstream() {
+        fn force_non_stream_payload(provider_request_body: Option<Value>) -> AiExecutionDecision {
+            AiExecutionDecision {
+                action: "stream".to_string(),
+                decision_kind: Some("openai_chat_stream".to_string()),
+                execution_strategy: None,
+                conversion_mode: None,
+                request_id: Some("req_force_non_stream".to_string()),
+                candidate_id: Some("cand_force_non_stream".to_string()),
+                provider_name: Some("OpenAI".to_string()),
+                provider_id: Some("prov_force_non_stream".to_string()),
+                endpoint_id: Some("ep_force_non_stream".to_string()),
+                key_id: Some("key_force_non_stream".to_string()),
+                upstream_base_url: Some("https://example.com".to_string()),
+                upstream_url: Some("https://example.com/v1/chat/completions".to_string()),
+                provider_request_method: None,
+                auth_header: Some("authorization".to_string()),
+                auth_value: Some("Bearer upstream-token".to_string()),
+                provider_api_format: Some("openai:chat".to_string()),
+                client_api_format: Some("openai:chat".to_string()),
+                provider_contract: Some("openai:chat".to_string()),
+                client_contract: Some("openai:chat".to_string()),
+                model_name: Some("gpt-5.4".to_string()),
+                mapped_model: Some("gpt-5.4".to_string()),
+                prompt_cache_key: None,
+                extra_headers: BTreeMap::new(),
+                provider_request_headers: BTreeMap::new(),
+                provider_request_body,
+                provider_request_body_base64: None,
+                content_type: Some("application/json".to_string()),
+                proxy: None,
+                transport_profile: None,
+                timeouts: None,
+                upstream_is_stream: false,
+                report_kind: Some("openai_chat_stream_success".to_string()),
+                report_context: Some(json!({})),
+                auth_context: None,
+            }
+        }
+
+        let parts = http::Request::builder()
+            .uri("http://localhost/v1/chat/completions")
+            .body(())
+            .expect("request should build")
+            .into_parts()
+            .0;
+
+        let built = build_openai_chat_stream_plan_from_decision(
+            &parts,
+            &json!({}),
+            force_non_stream_payload(Some(json!({
+                "model": "gpt-5.4",
+                "messages": [],
+                "stream": false
+            }))),
+        )
+        .expect("plan build should succeed")
+        .expect("plan should be produced");
+
+        assert!(built.plan.stream);
+        assert_eq!(
+            built
+                .plan
+                .body
+                .json_body
+                .as_ref()
+                .and_then(|body| body.get("stream"))
+                .and_then(Value::as_bool),
+            Some(false)
+        );
+
+        let fallback_body = json!({
+            "model": "client-model",
+            "messages": [],
+            "stream": true
+        });
+        let built = build_openai_chat_stream_plan_from_decision(
+            &parts,
+            &fallback_body,
+            force_non_stream_payload(None),
+        )
+        .expect("fallback plan build should succeed")
+        .expect("fallback plan should be produced");
+
+        assert!(built.plan.stream);
+        assert_eq!(
+            built
+                .plan
+                .body
+                .json_body
+                .as_ref()
+                .and_then(|body| body.get("stream"))
+                .and_then(Value::as_bool),
+            Some(false)
+        );
+    }
+
     #[test]
     fn build_openai_chat_stream_plan_fallback_restores_claude_headers_for_cross_format() {
         let parts = http::Request::builder()

apps/aether-gateway/src/ai_serving/planner/standard/openai/plan_builders/sync.rs

@@ -9,6 +9,7 @@ use super::super::{
     take_ai_upstream_auth_pair, take_non_empty_string, AiExecutionPlanFromDecisionParts,
     AiSyncAttempt,
 };
+use crate::ai_serving::planner::common::enforce_provider_body_stream_policy;
 use crate::ai_serving::transport::{
     build_standard_plan_fallback_headers, build_standard_plan_fallback_openai_chat_url,
     build_standard_plan_fallback_openai_responses_url, StandardPlanFallbackAcceptPolicy,
@@ -51,23 +52,31 @@ pub(crate) fn build_openai_chat_sync_plan_from_decision(
             provider_request_body
                 .insert("model".to_string(), serde_json::Value::String(mapped_model));
         }
-        if payload.upstream_is_stream {
+        let require_body_stream_field = provider_request_body.contains_key("stream");
-            provider_request_body.insert("stream".to_string(), serde_json::Value::Bool(true));
+        let mut provider_request_body = serde_json::Value::Object(provider_request_body);
-        }
+        enforce_provider_body_stream_policy(
+            &mut provider_request_body,
+            core.provider_api_format.as_str(),
+            payload.upstream_is_stream,
+            require_body_stream_field,
+        );
+        let Some(provider_request_object) = provider_request_body.as_object_mut() else {
+            return Ok(None);
+        };
         if let Some(prompt_cache_key) = take_non_empty_string(&mut payload.prompt_cache_key) {
-            let existing = provider_request_body
+            let existing = provider_request_object
                 .get("prompt_cache_key")
                 .and_then(|value| value.as_str())
                 .map(str::trim)
                 .unwrap_or_default();
             if existing.is_empty() {
-                provider_request_body.insert(
+                provider_request_object.insert(
                     "prompt_cache_key".to_string(),
                     serde_json::Value::String(prompt_cache_key),
                 );
             }
         }
-        serde_json::Value::Object(provider_request_body)
+        provider_request_body
     };
     let extra_headers = std::mem::take(&mut payload.extra_headers);
     let mut provider_request_headers =

apps/aether-gateway/src/ai_serving/planner/standard/openai/responses/decision/request.rs

@@ -9,7 +9,10 @@ use crate::ai_serving::planner::candidate_preparation::{
     OauthPreparationContext,
 };
 use crate::ai_serving::planner::candidate_resolution::EligibleLocalExecutionCandidate;
-use crate::ai_serving::planner::common::force_upstream_streaming_for_provider;
+use crate::ai_serving::planner::common::{
+    endpoint_config_forces_body_stream_field, enforce_provider_body_stream_policy,
+    request_requires_body_stream_field, resolve_upstream_is_stream_for_provider,
+};
 use crate::ai_serving::planner::spec_metadata::local_openai_responses_spec_metadata;
 use crate::ai_serving::planner::standard::{
     apply_codex_openai_responses_special_headers, build_cross_format_openai_responses_request_body,
@@ -224,12 +227,15 @@ pub(crate) async fn resolve_local_openai_responses_candidate_payload_parts(
         .await;
 
     let needs_bidirectional_conversion = !same_format && conversion_kind.is_some();
-    let upstream_is_stream = spec_metadata.require_streaming
+    let upstream_is_stream = resolve_upstream_is_stream_for_provider(
-        || is_antigravity
+        transport.endpoint.config.as_ref(),
-        || force_upstream_streaming_for_provider(
+        transport.provider.provider_type.as_str(),
-            transport.provider.provider_type.as_str(),
+        provider_api_format,
-            provider_api_format,
+        spec_metadata.require_streaming,
-        );
+        is_antigravity || is_kiro_claude_cli,
+    );
+    let force_body_stream_field =
+        endpoint_config_forces_body_stream_field(transport.endpoint.config.as_ref());
     let Some(mut base_provider_request_body) = (if needs_bidirectional_conversion {
         build_cross_format_openai_responses_request_body(
             body_json,
@@ -237,6 +243,7 @@ pub(crate) async fn resolve_local_openai_responses_candidate_payload_parts(
             spec_metadata.api_format,
             provider_api_format,
             upstream_is_stream,
+            force_body_stream_field,
             transport.provider.provider_type.as_str(),
             if is_kiro_claude_cli {
                 None
@@ -252,6 +259,7 @@ pub(crate) async fn resolve_local_openai_responses_candidate_payload_parts(
             body_json,
             &mapped_model,
             upstream_is_stream,
+            force_body_stream_field,
             transport.provider.provider_type.as_str(),
             provider_api_format,
             if is_kiro_claude_cli {
@@ -293,6 +301,14 @@ pub(crate) async fn resolve_local_openai_responses_candidate_payload_parts(
             &mut base_provider_request_body,
             &mapping,
         );
+        // Directive mapping is a deep-merge patch and may overwrite/add `stream`;
+        // re-enforce stream-field policy afterward.
+        enforce_provider_body_stream_policy(
+            &mut base_provider_request_body,
+            provider_api_format,
+            upstream_is_stream,
+            request_requires_body_stream_field(body_json, force_body_stream_field),
+        );
     }
     let antigravity_auth = if is_antigravity {
         match classify_local_antigravity_request_support(

apps/aether-gateway/src/ai_serving/planner/standard/plan_builders.rs

@@ -112,13 +112,14 @@ pub(crate) fn build_standard_stream_plan_from_decision(
         .as_ref()
         .and_then(|context| context.get("envelope_name"))
         .and_then(serde_json::Value::as_str);
-    let accept_policy = if provider_adaptation_requires_eventstream_accept(
+    let accept_policy = if payload.upstream_is_stream
-        envelope_name,
+        && provider_adaptation_requires_eventstream_accept(
-        core.provider_api_format.as_str(),
+            envelope_name,
-    ) {
+            core.provider_api_format.as_str(),
+        ) {
         StandardPlanFallbackAcceptPolicy::ProviderEventStreamIfMissing
     } else {
-        StandardPlanFallbackAcceptPolicy::TextEventStreamRequired
+        StandardPlanFallbackAcceptPolicy::TextEventStreamIfStreaming
     };
     let mut provider_request_headers =
         build_standard_plan_fallback_headers(StandardPlanFallbackHeadersInput {
@@ -132,7 +133,7 @@ pub(crate) fn build_standard_stream_plan_from_decision(
             content_type: payload.content_type.as_deref(),
             provider_api_format: core.provider_api_format.as_str(),
             client_api_format: core.client_api_format.as_str(),
-            upstream_is_stream: true,
+            upstream_is_stream: payload.upstream_is_stream,
             build_from_request_when_empty: false,
             accept_policy,
         });

crates/aether-ai-formats/src/api.rs

@@ -82,7 +82,9 @@ pub use crate::formats::shared::passthrough::{
     LocalSameFormatProviderSpec,
 };
 pub use crate::formats::shared::request::{
+    endpoint_config_forces_upstream_stream_policy, enforce_request_body_stream_field,
     force_upstream_streaming_for_provider, parse_direct_request_body,
+    resolve_upstream_is_stream_from_endpoint_config,
 };
 pub use crate::formats::shared::request_matrix::{
     build_standard_request_body_from_canonical,

crates/aether-ai-formats/src/formats/id.rs

@@ -135,10 +135,18 @@ pub fn is_openai_responses_family_format(value: &str) -> bool {
     )
 }
 
+pub fn api_format_uses_body_stream_field(value: &str) -> bool {
+    matches!(
+        FormatId::parse(value).map(FormatId::canonical),
+        Some(FormatId::OpenAiChat | FormatId::OpenAiResponses | FormatId::ClaudeMessages)
+    )
+}
+
 #[cfg(test)]
 mod tests {
     use super::{
-        api_format_alias_matches, api_format_storage_aliases, normalize_api_format_alias, FormatId,
+        api_format_alias_matches, api_format_storage_aliases, api_format_uses_body_stream_field,
+        normalize_api_format_alias, FormatId,
     };
 
     #[test]
@@ -304,4 +312,22 @@ mod tests {
             vec!["doubao:embedding".to_string()]
         );
     }
+
+    #[test]
+    fn body_stream_field_support_matches_provider_wire_formats() {
+        assert!(api_format_uses_body_stream_field("openai:chat"));
+        assert!(api_format_uses_body_stream_field("/v1/chat/completions"));
+        assert!(api_format_uses_body_stream_field("openai:responses"));
+        assert!(api_format_uses_body_stream_field("/v1/responses"));
+        assert!(api_format_uses_body_stream_field("claude:messages"));
+        assert!(api_format_uses_body_stream_field("/v1/messages"));
+        assert!(!api_format_uses_body_stream_field(
+            "openai:responses:compact"
+        ));
+        assert!(!api_format_uses_body_stream_field("/v1/responses/compact"));
+        assert!(!api_format_uses_body_stream_field(
+            "gemini:generate_content"
+        ));
+        assert!(!api_format_uses_body_stream_field("openai:embedding"));
+    }
 }

crates/aether-ai-formats/src/formats/openai/responses/codex.rs

@@ -281,8 +281,9 @@ pub fn apply_openai_responses_compact_special_body_edits(
         return;
     };
 
-    // `/v1/responses/compact` does not accept `store`.
+    // `/v1/responses/compact` does not accept `store` or body-level `stream`.
     body_object.remove("store");
+    body_object.remove("stream");
 }
 
 pub fn apply_codex_openai_responses_special_body_edits(

crates/aether-ai-formats/src/formats/openai/responses/request.rs

@@ -188,6 +188,9 @@ pub fn to_raw(
         OPENAI_RESPONSES_LEGACY_EXTENSION_NAMESPACE,
         &output,
     ));
+    if compact {
+        output.remove("stream");
+    }
     output.remove("verbosity");
     Some(Value::Object(output))
 }

crates/aether-ai-formats/src/formats/shared/request.rs

@@ -1,5 +1,14 @@
 use base64::Engine as _;
 
+use crate::formats::id::api_format_uses_body_stream_field;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum UpstreamStreamPolicy {
+    Auto,
+    ForceStream,
+    ForceNonStream,
+}
+
 pub fn parse_direct_request_body(
     is_json_request: bool,
     body_bytes: &[u8],
@@ -29,9 +38,130 @@ pub fn force_upstream_streaming_for_provider(
         && aether_ai_formats::is_openai_responses_format(provider_api_format)
 }
 
+pub(crate) fn parse_upstream_stream_policy(
+    value: Option<&serde_json::Value>,
+) -> UpstreamStreamPolicy {
+    let Some(value) = value else {
+        return UpstreamStreamPolicy::Auto;
+    };
+    if let Some(value) = value.as_bool() {
+        return if value {
+            UpstreamStreamPolicy::ForceStream
+        } else {
+            UpstreamStreamPolicy::ForceNonStream
+        };
+    }
+
+    let serde_json::Value::String(value) = value else {
+        return UpstreamStreamPolicy::Auto;
+    };
+    let raw = value.trim().to_ascii_lowercase();
+    match raw.as_str() {
+        "" | "auto" | "follow" | "client" | "default" => UpstreamStreamPolicy::Auto,
+        "force_stream" | "stream" | "sse" | "true" | "1" | "yes" => {
+            UpstreamStreamPolicy::ForceStream
+        }
+        "force_non_stream" | "force_sync" | "non_stream" | "sync" | "false" | "0" | "no" => {
+            UpstreamStreamPolicy::ForceNonStream
+        }
+        _ => UpstreamStreamPolicy::Auto,
+    }
+}
+
+pub(crate) fn upstream_stream_policy_from_endpoint_config(
+    endpoint_config: Option<&serde_json::Value>,
+) -> UpstreamStreamPolicy {
+    let Some(config) = endpoint_config.and_then(serde_json::Value::as_object) else {
+        return UpstreamStreamPolicy::Auto;
+    };
+    for key in [
+        "upstream_stream_policy",
+        "upstreamStreamPolicy",
+        "upstream_stream",
+    ] {
+        if let Some(value) = config.get(key) {
+            return parse_upstream_stream_policy(Some(value));
+        }
+    }
+    UpstreamStreamPolicy::Auto
+}
+
+pub fn endpoint_config_forces_upstream_stream_policy(
+    endpoint_config: Option<&serde_json::Value>,
+) -> bool {
+    matches!(
+        upstream_stream_policy_from_endpoint_config(endpoint_config),
+        UpstreamStreamPolicy::ForceStream | UpstreamStreamPolicy::ForceNonStream
+    )
+}
+
+/// Resolves the upstream provider transport mode.
+///
+/// `client_is_stream` means the request landed on a streaming surface or should
+/// be treated as streaming; the original JSON body may not have had
+/// `"stream": true`.
+pub(crate) fn resolve_upstream_is_stream(
+    client_is_stream: bool,
+    hard_requires_streaming: bool,
+    policy: UpstreamStreamPolicy,
+) -> bool {
+    // ForceStream is unconditional, while ForceNonStream yields to hard
+    // stream-only constraints such as Kiro or Codex OpenAI Responses.
+    match policy {
+        UpstreamStreamPolicy::ForceStream => true,
+        UpstreamStreamPolicy::ForceNonStream => hard_requires_streaming,
+        UpstreamStreamPolicy::Auto => hard_requires_streaming || client_is_stream,
+    }
+}
+
+pub fn enforce_request_body_stream_field(
+    body: &mut serde_json::Value,
+    provider_api_format: &str,
+    upstream_is_stream: bool,
+    require_body_stream_field: bool,
+) {
+    let Some(body_object) = body.as_object_mut() else {
+        return;
+    };
+    if !api_format_uses_body_stream_field(provider_api_format) {
+        body_object.remove("stream");
+        return;
+    }
+
+    // Final-body fallback catches body rules, directive patches, and other
+    // provider-body mutations that introduce `stream`.
+    if upstream_is_stream || require_body_stream_field || body_object.contains_key("stream") {
+        body_object.insert(
+            "stream".to_string(),
+            serde_json::Value::Bool(upstream_is_stream),
+        );
+    } else {
+        body_object.remove("stream");
+    }
+}
+
+pub fn resolve_upstream_is_stream_from_endpoint_config(
+    endpoint_config: Option<&serde_json::Value>,
+    client_is_stream: bool,
+    hard_requires_streaming: bool,
+) -> bool {
+    resolve_upstream_is_stream(
+        client_is_stream,
+        hard_requires_streaming,
+        upstream_stream_policy_from_endpoint_config(endpoint_config),
+    )
+}
+
 #[cfg(test)]
 mod tests {
-    use super::{force_upstream_streaming_for_provider, parse_direct_request_body};
+    use super::{
+        endpoint_config_forces_upstream_stream_policy, enforce_request_body_stream_field,
+        force_upstream_streaming_for_provider, parse_direct_request_body,
+        parse_upstream_stream_policy, resolve_upstream_is_stream,
+        resolve_upstream_is_stream_from_endpoint_config,
+        upstream_stream_policy_from_endpoint_config, UpstreamStreamPolicy,
+    };
+    use serde_json::json;
 
     #[test]
     fn parses_empty_json_body_as_empty_object() {
@@ -77,4 +207,177 @@ mod tests {
             "openai:responses"
         ));
     }
+
+    #[test]
+    fn parses_python_compatible_upstream_stream_policy_values() {
+        assert_eq!(
+            parse_upstream_stream_policy(None),
+            UpstreamStreamPolicy::Auto
+        );
+        for value in [
+            json!(""),
+            json!("auto"),
+            json!("follow"),
+            json!("client"),
+            json!("default"),
+            json!("unknown"),
+        ] {
+            assert_eq!(
+                parse_upstream_stream_policy(Some(&value)),
+                UpstreamStreamPolicy::Auto
+            );
+        }
+        for value in [
+            json!(true),
+            json!("force_stream"),
+            json!("stream"),
+            json!("sse"),
+            json!("true"),
+            json!("1"),
+            json!("yes"),
+        ] {
+            assert_eq!(
+                parse_upstream_stream_policy(Some(&value)),
+                UpstreamStreamPolicy::ForceStream
+            );
+        }
+        for value in [
+            json!(false),
+            json!("force_non_stream"),
+            json!("force_sync"),
+            json!("non_stream"),
+            json!("sync"),
+            json!("false"),
+            json!("0"),
+            json!("no"),
+        ] {
+            assert_eq!(
+                parse_upstream_stream_policy(Some(&value)),
+                UpstreamStreamPolicy::ForceNonStream
+            );
+        }
+    }
+
+    #[test]
+    fn parses_non_string_non_bool_policy_values_as_auto() {
+        for value in [json!(1), json!(0), json!(null), json!({}), json!([])] {
+            assert_eq!(
+                parse_upstream_stream_policy(Some(&value)),
+                UpstreamStreamPolicy::Auto
+            );
+        }
+    }
+
+    #[test]
+    fn enforces_request_body_stream_field_for_stream_and_streamless_formats() {
+        let mut openai_chat = json!({"stream": true});
+        enforce_request_body_stream_field(&mut openai_chat, "openai:chat", false, false);
+        assert_eq!(openai_chat.get("stream"), Some(&json!(false)));
+
+        let mut ordinary_sync = json!({"messages": []});
+        enforce_request_body_stream_field(&mut ordinary_sync, "openai:chat", false, false);
+        assert!(ordinary_sync.get("stream").is_none());
+
+        let mut forced_sync = json!({"messages": []});
+        enforce_request_body_stream_field(&mut forced_sync, "openai:chat", false, true);
+        assert_eq!(forced_sync.get("stream"), Some(&json!(false)));
+
+        let mut compact = json!({"stream": true});
+        enforce_request_body_stream_field(&mut compact, "openai:responses:compact", true, true);
+        assert!(compact.get("stream").is_none());
+    }
+
+    #[test]
+    fn reads_endpoint_policy_keys_in_python_compatible_order() {
+        assert_eq!(
+            upstream_stream_policy_from_endpoint_config(Some(&json!({
+                "upstream_stream_policy": "force_non_stream",
+                "upstreamStreamPolicy": "force_stream",
+                "upstream_stream": "force_stream"
+            }))),
+            UpstreamStreamPolicy::ForceNonStream
+        );
+        assert_eq!(
+            upstream_stream_policy_from_endpoint_config(Some(&json!({
+                "upstreamStreamPolicy": "force_stream"
+            }))),
+            UpstreamStreamPolicy::ForceStream
+        );
+        assert_eq!(
+            upstream_stream_policy_from_endpoint_config(Some(&json!({
+                "upstream_stream": false
+            }))),
+            UpstreamStreamPolicy::ForceNonStream
+        );
+    }
+
+    #[test]
+    fn detects_forced_endpoint_policy_values() {
+        assert!(endpoint_config_forces_upstream_stream_policy(Some(
+            &json!({"upstream_stream_policy": "force_stream"})
+        )));
+        assert!(endpoint_config_forces_upstream_stream_policy(Some(
+            &json!({"upstream_stream_policy": "force_non_stream"})
+        )));
+        assert!(!endpoint_config_forces_upstream_stream_policy(Some(
+            &json!({"upstream_stream_policy": "auto"})
+        )));
+        assert!(!endpoint_config_forces_upstream_stream_policy(None));
+    }
+
+    #[test]
+    fn resolves_upstream_stream_policy_against_client_mode_and_hard_constraints() {
+        assert!(resolve_upstream_is_stream(
+            false,
+            false,
+            UpstreamStreamPolicy::ForceStream
+        ));
+        assert!(!resolve_upstream_is_stream(
+            true,
+            false,
+            UpstreamStreamPolicy::ForceNonStream
+        ));
+        assert!(resolve_upstream_is_stream(
+            true,
+            true,
+            UpstreamStreamPolicy::ForceNonStream
+        ));
+        assert!(!resolve_upstream_is_stream(
+            false,
+            false,
+            UpstreamStreamPolicy::Auto
+        ));
+        assert!(resolve_upstream_is_stream(
+            true,
+            false,
+            UpstreamStreamPolicy::Auto
+        ));
+        assert!(resolve_upstream_is_stream(
+            false,
+            true,
+            UpstreamStreamPolicy::Auto
+        ));
+    }
+
+    #[test]
+    fn resolves_endpoint_policy_config_to_upstream_mode() {
+        assert!(resolve_upstream_is_stream_from_endpoint_config(
+            Some(&json!({"upstream_stream_policy": "force_stream"})),
+            false,
+            false,
+        ));
+        assert!(!resolve_upstream_is_stream_from_endpoint_config(
+            Some(&json!({"upstream_stream_policy": "force_non_stream"})),
+            true,
+            false,
+        ));
+        assert!(resolve_upstream_is_stream_from_endpoint_config(
+            Some(&json!({"upstream_stream_policy": "auto"})),
+            true,
+            false,
+        ));
+        assert!(!resolve_upstream_is_stream_from_endpoint_config(
+            None, false, false,
+        ));
+    }
 }

crates/aether-ai-formats/src/formats/shared/standard_matrix.rs

@@ -128,6 +128,18 @@ pub fn build_standard_request_body_with_model_directives_and_request_headers(
         &mut provider_request_body,
         provider_api_format,
     );
+    let require_body_stream_field = body_json
+        .as_object()
+        .is_some_and(|object| object.contains_key("stream"))
+        || provider_request_body
+            .as_object()
+            .is_some_and(|object| object.contains_key("stream"));
+    crate::formats::shared::request::enforce_request_body_stream_field(
+        &mut provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        require_body_stream_field,
+    );
     Some(provider_request_body)
 }
 
@@ -302,13 +314,23 @@ mod tests {
     fn assert_stream_flag(provider_api_format: &str, upstream_is_stream: bool, converted: &Value) {
         match provider_api_format {
             "openai:chat" | "openai:responses" | "claude:messages" => {
-                assert_eq!(
+                if upstream_is_stream {
-                    converted
+                    assert_eq!(
-                        .get("stream")
+                        converted.get("stream").and_then(Value::as_bool),
-                        .and_then(Value::as_bool)
+                        Some(true),
-                        .unwrap_or(false),
+                        "{provider_api_format} stream flag should be true for upstream streaming"
-                    upstream_is_stream,
+                    );
-                    "{provider_api_format} stream flag should follow upstream_is_stream"
+                } else {
+                    assert!(
+                        converted.get("stream").is_none(),
+                        "{provider_api_format} should not gain stream:false for ordinary sync requests"
+                    );
+                }
+            }
+            "openai:responses:compact" => {
+                assert!(
+                    converted.get("stream").is_none(),
+                    "openai responses compact keeps stream out of the request body"
                 );
             }
             "gemini:generate_content" => {
@@ -321,6 +343,93 @@ mod tests {
         }
     }
 
+    fn assert_explicit_stream_flag(
+        provider_api_format: &str,
+        upstream_is_stream: bool,
+        converted: &Value,
+    ) {
+        match provider_api_format {
+            "openai:chat" | "openai:responses" | "claude:messages" => {
+                assert_eq!(
+                    converted.get("stream").and_then(Value::as_bool),
+                    Some(upstream_is_stream),
+                    "{provider_api_format} stream flag should follow upstream_is_stream"
+                );
+            }
+            "openai:responses:compact" | "gemini:generate_content" => {
+                assert!(converted.get("stream").is_none());
+            }
+            other => panic!("unexpected provider api format: {other}"),
+        }
+    }
+
+    #[test]
+    fn standard_request_body_overrides_client_stream_true_for_non_stream_upstream() {
+        let request = json!({
+            "model": "source-model",
+            "messages": [{"role": "user", "content": "hello"}],
+            "stream": true
+        });
+
+        for provider_api_format in [
+            "openai:chat",
+            "openai:responses",
+            "openai:responses:compact",
+            "claude:messages",
+            "gemini:generate_content",
+        ] {
+            let converted = build_standard_request_body(
+                &request,
+                "openai:chat",
+                "mapped-model",
+                "custom",
+                provider_api_format,
+                "/v1/chat/completions",
+                false,
+                None,
+                None,
+            )
+            .unwrap_or_else(|| panic!("openai:chat -> {provider_api_format} should build"));
+
+            assert_explicit_stream_flag(provider_api_format, false, &converted);
+        }
+    }
+
+    #[test]
+    fn standard_request_body_stream_policy_wins_after_body_rules() {
+        let request = json!({
+            "model": "source-model",
+            "messages": [{"role": "user", "content": "hello"}],
+            "stream": true
+        });
+        let body_rules = json!([
+            {"action":"set","path":"stream","value":true}
+        ]);
+
+        for provider_api_format in [
+            "openai:chat",
+            "openai:responses",
+            "openai:responses:compact",
+            "claude:messages",
+            "gemini:generate_content",
+        ] {
+            let converted = build_standard_request_body(
+                &request,
+                "openai:chat",
+                "mapped-model",
+                "custom",
+                provider_api_format,
+                "/v1/chat/completions",
+                false,
+                Some(&body_rules),
+                None,
+            )
+            .unwrap_or_else(|| panic!("openai:chat -> {provider_api_format} should build"));
+
+            assert_explicit_stream_flag(provider_api_format, false, &converted);
+        }
+    }
+
     fn codex_default_body_rules() -> Value {
         json!([
             {"action":"drop","path":"max_output_tokens"},

crates/aether-ai-formats/src/formats/shared/standard_normalize.rs

@@ -50,14 +50,24 @@ pub fn build_local_openai_chat_request_body_with_model_directives(
             }
         }
     }
-    Some(with_model_directive_overrides(
+    let mut provider_request_body = with_model_directive_overrides(
         Value::Object(provider_request_body),
         "openai:chat",
         mapped_model,
         body_json,
         None,
         enable_model_directives,
-    ))
+    );
+    let require_body_stream_field = body_json
+        .as_object()
+        .is_some_and(|object| object.contains_key("stream"));
+    crate::formats::shared::request::enforce_request_body_stream_field(
+        &mut provider_request_body,
+        "openai:chat",
+        upstream_is_stream,
+        require_body_stream_field,
+    );
+    Some(provider_request_body)
 }
 
 pub fn build_cross_format_openai_chat_request_body(
@@ -104,14 +114,24 @@ pub fn build_cross_format_openai_chat_request_body_with_model_directives(
         }
         _ => return None,
     };
-    Some(with_model_directive_overrides(
+    let mut provider_request_body = with_model_directive_overrides(
         provider_request_body,
         provider_api_format,
         mapped_model,
         body_json,
         None,
         enable_model_directives,
-    ))
+    );
+    let require_body_stream_field = body_json
+        .as_object()
+        .is_some_and(|object| object.contains_key("stream"));
+    crate::formats::shared::request::enforce_request_body_stream_field(
+        &mut provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        require_body_stream_field,
+    );
+    Some(provider_request_body)
 }
 
 pub fn build_local_openai_responses_request_body(
@@ -143,14 +163,24 @@ pub fn build_local_openai_responses_request_body_with_model_directives(
     if require_streaming {
         provider_request_body.insert("stream".to_string(), Value::Bool(true));
     }
-    Some(with_model_directive_overrides(
+    let mut provider_request_body = with_model_directive_overrides(
         Value::Object(provider_request_body),
         "openai:responses",
         mapped_model,
         body_json,
         None,
         enable_model_directives,
-    ))
+    );
+    let require_body_stream_field = body_json
+        .as_object()
+        .is_some_and(|object| object.contains_key("stream"));
+    crate::formats::shared::request::enforce_request_body_stream_field(
+        &mut provider_request_body,
+        "openai:responses",
+        require_streaming,
+        require_body_stream_field,
+    );
+    Some(provider_request_body)
 }
 
 pub fn build_cross_format_openai_responses_request_body(
@@ -208,14 +238,24 @@ pub fn build_cross_format_openai_responses_request_body_with_model_directives(
             upstream_is_stream,
         )?,
     };
-    Some(with_model_directive_overrides(
+    let mut provider_request_body = with_model_directive_overrides(
         provider_request_body,
         provider_api_format,
         mapped_model,
         body_json,
         None,
         enable_model_directives,
-    ))
+    );
+    let require_body_stream_field = body_json
+        .as_object()
+        .is_some_and(|object| object.contains_key("stream"));
+    crate::formats::shared::request::enforce_request_body_stream_field(
+        &mut provider_request_body,
+        provider_api_format,
+        upstream_is_stream,
+        require_body_stream_field,
+    );
+    Some(provider_request_body)
 }
 
 fn with_model_directive_overrides(
@@ -324,6 +364,108 @@ mod tests {
         );
     }
 
+    #[test]
+    fn local_openai_chat_request_body_overrides_client_stream_for_non_stream_upstream() {
+        let body_json = json!({
+            "model": "gpt-5",
+            "messages": [{
+                "role": "user",
+                "content": "hello"
+            }],
+            "stream": true
+        });
+
+        let provider_request_body =
+            build_local_openai_chat_request_body(&body_json, "gpt-5-upstream", false)
+                .expect("openai chat body should build");
+
+        assert_eq!(provider_request_body["model"], "gpt-5-upstream");
+        assert_eq!(provider_request_body["stream"], false);
+    }
+
+    #[test]
+    fn local_openai_responses_request_body_overrides_client_stream_for_non_stream_upstream() {
+        let body_json = json!({
+            "model": "gpt-5",
+            "input": "hello",
+            "stream": true
+        });
+
+        let provider_request_body =
+            build_local_openai_responses_request_body(&body_json, "gpt-5-upstream", false)
+                .expect("openai responses body should build");
+
+        assert_eq!(provider_request_body["model"], "gpt-5-upstream");
+        assert_eq!(provider_request_body["stream"], false);
+    }
+
+    #[test]
+    fn cross_format_openai_chat_request_body_overrides_client_stream_for_non_stream_upstream() {
+        let body_json = json!({
+            "model": "gpt-5",
+            "messages": [{"role": "user", "content": "hello"}],
+            "stream": true
+        });
+
+        let claude = build_cross_format_openai_chat_request_body_with_model_directives(
+            &body_json,
+            "claude-sonnet-4-5",
+            "claude:messages",
+            false,
+            false,
+        )
+        .expect("claude body should build");
+        assert_eq!(claude["stream"], false);
+
+        let responses = build_cross_format_openai_chat_request_body_with_model_directives(
+            &body_json,
+            "gpt-5-upstream",
+            "openai:responses",
+            false,
+            false,
+        )
+        .expect("responses body should build");
+        assert_eq!(responses["stream"], false);
+    }
+
+    #[test]
+    fn cross_format_openai_chat_request_body_does_not_add_stream_false_for_plain_sync_body() {
+        let body_json = json!({
+            "model": "gpt-5",
+            "messages": [{"role": "user", "content": "hello"}]
+        });
+
+        let claude = build_cross_format_openai_chat_request_body_with_model_directives(
+            &body_json,
+            "claude-sonnet-4-5",
+            "claude:messages",
+            false,
+            false,
+        )
+        .expect("claude body should build");
+        assert!(claude.get("stream").is_none());
+    }
+
+    #[test]
+    fn cross_format_openai_responses_body_overrides_client_stream_for_non_stream_upstream() {
+        let body_json = json!({
+            "model": "gpt-5",
+            "input": "hello",
+            "stream": true
+        });
+
+        let provider_request_body = build_cross_format_openai_responses_request_body(
+            &body_json,
+            "claude-sonnet-4-5",
+            "openai:responses",
+            "claude:messages",
+            false,
+        )
+        .expect("claude body should build");
+
+        assert_eq!(provider_request_body["stream"], false);
+    }
+
     #[test]
     fn local_openai_chat_request_body_applies_reasoning_effort_suffix() {
         let body_json = json!({

crates/aether-ai-formats/src/lib.rs

@@ -8,9 +8,9 @@ pub mod provider_compat;
 
 pub use formats::context::{FormatContext, FormatError};
 pub use formats::id::{
-    api_format_alias_matches, api_format_storage_aliases, is_openai_responses_compact_format,
+    api_format_alias_matches, api_format_storage_aliases, api_format_uses_body_stream_field,
-    is_openai_responses_family_format, is_openai_responses_format, normalize_api_format_alias,
+    is_openai_responses_compact_format, is_openai_responses_family_format,
-    FormatFamily, FormatId, FormatProfile,
+    is_openai_responses_format, normalize_api_format_alias, FormatFamily, FormatId, FormatProfile,
 };
 pub use formats::matrix::{
     is_embedding_api_format, is_rerank_api_format, request_candidate_api_format_preference,
@@ -27,6 +27,10 @@ pub use formats::shared::model_directives::{
     normalize_model_directive_model, parse_model_directive, ModelDirective, ModelOverride,
     ReasoningEffort,
 };
+pub use formats::shared::request::{
+    endpoint_config_forces_upstream_stream_policy, enforce_request_body_stream_field,
+    resolve_upstream_is_stream_from_endpoint_config,
+};
 pub use protocol::canonical::{
     canonical_request_unknown_block_count, canonical_response_unknown_block_count,
     canonical_to_claude_request, canonical_to_claude_response, canonical_to_embedding_response,

crates/aether-provider-transport/src/same_format_provider/mod.rs

@@ -35,8 +35,9 @@ pub enum SameFormatProviderFamily {
 }
 
 #[derive(Debug, Clone, Copy)]
-pub struct SameFormatProviderRequestBehaviorParams {
+pub struct SameFormatProviderRequestBehaviorParams<'a> {
     pub require_streaming: bool,
+    pub provider_api_format: &'a str,
     pub report_kind: &'static str,
 }
 
@@ -47,6 +48,7 @@ pub struct SameFormatProviderRequestBehavior {
     pub is_vertex: bool,
     pub is_kiro: bool,
     pub upstream_is_stream: bool,
+    pub force_body_stream_field: bool,
     pub report_kind: &'static str,
 }
 
@@ -61,6 +63,7 @@ pub struct SameFormatProviderRequestBodyInput<'a> {
     pub body_rules: Option<&'a Value>,
     pub request_headers: Option<&'a http::HeaderMap>,
     pub upstream_is_stream: bool,
+    pub force_body_stream_field: bool,
     pub kiro_auth_config: Option<&'a KiroAuthConfig>,
     pub is_claude_code: bool,
     pub enable_model_directives: bool,
@@ -92,7 +95,7 @@ pub struct SameFormatProviderHeadersInput<'a> {
 
 pub fn classify_same_format_provider_request_behavior(
     transport: &GatewayProviderTransportSnapshot,
-    params: SameFormatProviderRequestBehaviorParams,
+    params: SameFormatProviderRequestBehaviorParams<'_>,
 ) -> SameFormatProviderRequestBehavior {
     let is_antigravity = is_antigravity_provider_transport(transport);
     let is_claude_code = transport
@@ -102,7 +105,19 @@ pub fn classify_same_format_provider_request_behavior(
         .eq_ignore_ascii_case("claude_code");
     let is_vertex = is_vertex_api_key_transport_context(transport);
     let is_kiro = is_kiro_provider_transport(transport);
-    let upstream_is_stream = is_kiro || is_antigravity || params.require_streaming;
+    let upstream_is_stream = aether_ai_formats::resolve_upstream_is_stream_from_endpoint_config(
+        transport.endpoint.config.as_ref(),
+        params.require_streaming,
+        is_kiro
+            || is_antigravity
+            || aether_ai_formats::api::force_upstream_streaming_for_provider(
+                transport.provider.provider_type.as_str(),
+                params.provider_api_format,
+            ),
+    );
+    let force_body_stream_field = aether_ai_formats::endpoint_config_forces_upstream_stream_policy(
+        transport.endpoint.config.as_ref(),
+    );
     let report_kind = if is_kiro && !params.require_streaming {
         "claude_cli_sync_finalize"
     } else if is_antigravity && !params.require_streaming {
@@ -121,6 +136,7 @@ pub fn classify_same_format_provider_request_behavior(
         is_vertex,
         is_kiro,
         upstream_is_stream,
+        force_body_stream_field,
         report_kind,
     }
 }
@@ -165,9 +181,6 @@ pub fn build_same_format_provider_request_body(
                 "model".to_string(),
                 Value::String(input.mapped_model.to_string()),
             );
-            if input.upstream_is_stream {
-                provider_request_body.insert("stream".to_string(), Value::Bool(true));
-            }
         }
         SameFormatProviderFamily::Gemini => {
             provider_request_body.remove("model");
@@ -195,6 +208,17 @@ pub fn build_same_format_provider_request_body(
     ) {
         return None;
     }
+    let require_body_stream_field = input.force_body_stream_field
+        || input
+            .body_json
+            .as_object()
+            .is_some_and(|object| object.contains_key("stream"));
+    aether_ai_formats::enforce_request_body_stream_field(
+        &mut provider_request_body,
+        input.provider_api_format,
+        input.upstream_is_stream,
+        require_body_stream_field,
+    );
     Some(provider_request_body)
 }
 
@@ -334,6 +358,7 @@ pub fn same_format_provider_transport_unsupported_reason_for_trace(
         transport,
         SameFormatProviderRequestBehaviorParams {
             require_streaming: false,
+            provider_api_format: normalized_api_format,
             report_kind: "trace_candidate_metadata",
         },
     );
@@ -459,6 +484,7 @@ mod tests {
             &kiro,
             SameFormatProviderRequestBehaviorParams {
                 require_streaming: false,
+                provider_api_format: "claude:messages",
                 report_kind: "claude_chat_sync_success",
             },
         );
@@ -472,6 +498,7 @@ mod tests {
             &antigravity,
             SameFormatProviderRequestBehaviorParams {
                 require_streaming: false,
+                provider_api_format: "gemini:generate_content",
                 report_kind: "gemini_chat_sync_success",
             },
         );
@@ -481,6 +508,131 @@ mod tests {
         assert_eq!(behavior.report_kind, "gemini_chat_sync_finalize");
     }
 
+    #[test]
+    fn same_format_behavior_resolves_endpoint_stream_policy() {
+        let mut force_stream = sample_transport("openai");
+        force_stream.endpoint.config = Some(json!({
+            "upstream_stream_policy": "force_stream"
+        }));
+        let behavior = classify_same_format_provider_request_behavior(
+            &force_stream,
+            SameFormatProviderRequestBehaviorParams {
+                require_streaming: false,
+                provider_api_format: "openai:chat",
+                report_kind: "openai_chat_sync_success",
+            },
+        );
+        assert!(behavior.upstream_is_stream);
+
+        let mut force_non_stream = sample_transport("openai");
+        force_non_stream.endpoint.config = Some(json!({
+            "upstreamStreamPolicy": "force_non_stream"
+        }));
+        let behavior = classify_same_format_provider_request_behavior(
+            &force_non_stream,
+            SameFormatProviderRequestBehaviorParams {
+                require_streaming: true,
+                provider_api_format: "openai:chat",
+                report_kind: "openai_chat_stream_success",
+            },
+        );
+        assert!(!behavior.upstream_is_stream);
+
+        let mut auto = sample_transport("openai");
+        auto.endpoint.config = Some(json!({
+            "upstream_stream": "auto"
+        }));
+        let stream_behavior = classify_same_format_provider_request_behavior(
+            &auto,
+            SameFormatProviderRequestBehaviorParams {
+                require_streaming: true,
+                provider_api_format: "openai:chat",
+                report_kind: "openai_chat_stream_success",
+            },
+        );
+        assert!(stream_behavior.upstream_is_stream);
+        let sync_behavior = classify_same_format_provider_request_behavior(
+            &auto,
+            SameFormatProviderRequestBehaviorParams {
+                require_streaming: false,
+                provider_api_format: "openai:chat",
+                report_kind: "openai_chat_sync_success",
+            },
+        );
+        assert!(!sync_behavior.upstream_is_stream);
+    }
+
+    #[test]
+    fn same_format_behavior_preserves_hard_streaming_constraint() {
+        let mut kiro = sample_transport("kiro");
+        kiro.endpoint.config = Some(json!({
+            "upstream_stream_policy": "force_non_stream"
+        }));
+
+        let behavior = classify_same_format_provider_request_behavior(
+            &kiro,
+            SameFormatProviderRequestBehaviorParams {
+                require_streaming: true,
+                provider_api_format: "claude:messages",
+                report_kind: "claude_chat_stream_success",
+            },
+        );
+
+        assert!(behavior.upstream_is_stream);
+    }
+
+    #[test]
+    fn same_format_policy_resolution_drives_standard_body_stream_field() {
+        for (endpoint_config, client_is_stream, expected_stream) in [
+            (
+                json!({"upstream_stream_policy": "force_stream"}),
+                false,
+                true,
+            ),
+            (
+                json!({"upstreamStreamPolicy": "force_non_stream"}),
+                true,
+                false,
+            ),
+            (json!({"upstream_stream": "auto"}), true, true),
+            (json!({"upstream_stream": "auto"}), false, false),
+        ] {
+            let mut transport = sample_transport("openai");
+            transport.endpoint.config = Some(endpoint_config);
+            let behavior = classify_same_format_provider_request_behavior(
+                &transport,
+                SameFormatProviderRequestBehaviorParams {
+                    require_streaming: client_is_stream,
+                    provider_api_format: "openai:chat",
+                    report_kind: "openai_chat_policy_test",
+                },
+            );
+            let body =
+                build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+                    body_json: &json!({
+                        "model": "client-model",
+                        "messages": [{"role": "user", "content": "hello"}],
+                        "stream": client_is_stream
+                    }),
+                    mapped_model: "upstream-model",
+                    client_api_format: "openai:chat",
+                    provider_api_format: "openai:chat",
+                    source_model: Some("client-model"),
+                    family: SameFormatProviderFamily::Standard,
+                    body_rules: None,
+                    request_headers: None,
+                    upstream_is_stream: behavior.upstream_is_stream,
+                    force_body_stream_field: behavior.force_body_stream_field,
+                    kiro_auth_config: None,
+                    is_claude_code: false,
+                    enable_model_directives: false,
+                })
+                .expect("body should build");
+
+            assert_eq!(body.get("stream"), Some(&json!(expected_stream)));
+        }
+    }
+
     #[test]
     fn resolves_direct_auth_except_vertex() {
         let transport = sample_transport("openai");
@@ -488,6 +640,7 @@ mod tests {
             &transport,
             SameFormatProviderRequestBehaviorParams {
                 require_streaming: false,
+                provider_api_format: "openai:chat",
                 report_kind: "openai_chat_sync_success",
             },
         );
@@ -517,6 +670,7 @@ mod tests {
             body_rules: None,
             request_headers: None,
             upstream_is_stream: true,
+            force_body_stream_field: false,
             kiro_auth_config: None,
             is_claude_code: false,
             enable_model_directives: false,
@@ -527,6 +681,170 @@ mod tests {
         assert_eq!(body.get("stream"), Some(&json!(true)));
     }
 
+    #[test]
+    fn same_format_standard_body_overrides_client_stream_for_non_stream_upstream() {
+        let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+            body_json: &json!({
+                "model": "client-model",
+                "messages": [{"role": "user", "content": "hello"}],
+                "stream": true
+            }),
+            mapped_model: "upstream-model",
+            client_api_format: "openai:chat",
+            provider_api_format: "openai:chat",
+            source_model: Some("client-model"),
+            family: SameFormatProviderFamily::Standard,
+            body_rules: None,
+            request_headers: None,
+            upstream_is_stream: false,
+            force_body_stream_field: false,
+            kiro_auth_config: None,
+            is_claude_code: false,
+            enable_model_directives: false,
+        })
+        .expect("body should build");
+
+        assert_eq!(body.get("model"), Some(&json!("upstream-model")));
+        assert_eq!(body.get("stream"), Some(&json!(false)));
+    }
+
+    #[test]
+    fn same_format_standard_body_does_not_add_stream_false_for_plain_sync_body() {
+        let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+            body_json: &json!({
+                "model": "client-model",
+                "messages": [{"role": "user", "content": "hello"}]
+            }),
+            mapped_model: "upstream-model",
+            client_api_format: "openai:chat",
+            provider_api_format: "openai:chat",
+            source_model: Some("client-model"),
+            family: SameFormatProviderFamily::Standard,
+            body_rules: None,
+            request_headers: None,
+            upstream_is_stream: false,
+            force_body_stream_field: false,
+            kiro_auth_config: None,
+            is_claude_code: false,
+            enable_model_directives: false,
+        })
+        .expect("body should build");
+
+        assert_eq!(body.get("model"), Some(&json!("upstream-model")));
+        assert!(body.get("stream").is_none());
+    }
+
+    #[test]
+    fn same_format_standard_body_forced_policy_adds_stream_false() {
+        let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+            body_json: &json!({
+                "model": "client-model",
+                "messages": [{"role": "user", "content": "hello"}]
+            }),
+            mapped_model: "upstream-model",
+            client_api_format: "openai:chat",
+            provider_api_format: "openai:chat",
+            source_model: Some("client-model"),
+            family: SameFormatProviderFamily::Standard,
+            body_rules: None,
+            request_headers: None,
+            upstream_is_stream: false,
+            force_body_stream_field: true,
+            kiro_auth_config: None,
+            is_claude_code: false,
+            enable_model_directives: false,
+        })
+        .expect("body should build");
+
+        assert_eq!(body.get("model"), Some(&json!("upstream-model")));
+        assert_eq!(body.get("stream"), Some(&json!(false)));
+    }
+
+    #[test]
+    fn same_format_gemini_body_removes_leaked_client_stream_field() {
+        let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+            body_json: &json!({
+                "contents": [{"role": "user", "parts": [{"text": "hello"}]}],
+                "stream": true
+            }),
+            mapped_model: "gemini-upstream",
+            client_api_format: "gemini:generate_content",
+            provider_api_format: "gemini:generate_content",
+            source_model: None,
+            family: SameFormatProviderFamily::Gemini,
+            body_rules: None,
+            request_headers: None,
+            upstream_is_stream: false,
+            force_body_stream_field: false,
+            kiro_auth_config: None,
+            is_claude_code: false,
+            enable_model_directives: false,
+        })
+        .expect("body should build");
+
+        assert!(body.get("stream").is_none());
+    }
+
+    #[test]
+    fn same_format_stream_policy_wins_after_body_rules() {
+        let body_rules = json!([
+            {"action":"set","path":"stream","value":true}
+        ]);
+
+        let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+            body_json: &json!({
+                "model": "client-model",
+                "messages": [{"role": "user", "content": "hello"}],
+                "stream": true
+            }),
+            mapped_model: "upstream-model",
+            client_api_format: "openai:chat",
+            provider_api_format: "openai:chat",
+            source_model: Some("client-model"),
+            family: SameFormatProviderFamily::Standard,
+            body_rules: Some(&body_rules),
+            request_headers: None,
+            upstream_is_stream: false,
+            force_body_stream_field: false,
+            kiro_auth_config: None,
+            is_claude_code: false,
+            enable_model_directives: false,
+        })
+        .expect("body should build");
+
+        assert_eq!(body.get("stream"), Some(&json!(false)));
+    }
+
+    #[test]
+    fn same_format_compact_stream_policy_wins_after_body_rules() {
+        let body_rules = json!([
+            {"action":"set","path":"stream","value":true}
+        ]);
+
+        let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
+            body_json: &json!({
+                "model": "client-model",
+                "input": "hello",
+                "stream": true
+            }),
+            mapped_model: "upstream-model",
+            client_api_format: "openai:responses:compact",
+            provider_api_format: "openai:responses:compact",
+            source_model: Some("client-model"),
+            family: SameFormatProviderFamily::Standard,
+            body_rules: Some(&body_rules),
+            request_headers: None,
+            upstream_is_stream: false,
+            force_body_stream_field: false,
+            kiro_auth_config: None,
+            is_claude_code: false,
+            enable_model_directives: false,
+        })
+        .expect("body should build");
+
+        assert!(body.get("stream").is_none());
+    }
+
     #[test]
     fn same_format_body_applies_model_directive_before_body_rules() {
         let body = build_same_format_provider_request_body(SameFormatProviderRequestBodyInput {
@@ -545,6 +863,7 @@ mod tests {
             ])),
             request_headers: None,
             upstream_is_stream: false,
+            force_body_stream_field: false,
             kiro_auth_config: None,
             is_claude_code: false,
             enable_model_directives: true,
@@ -571,6 +890,7 @@ mod tests {
                 is_vertex: false,
                 is_kiro: false,
                 upstream_is_stream: true,
+                force_body_stream_field: false,
                 report_kind: "openai_chat_stream_success",
             },
             auth_header: Some("x-api-key"),