fix(ws): harden Responses continuation state

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

@@ -50,6 +50,69 @@ pub(crate) fn build_local_openai_responses_request_body_with_codex_model_capabil
     request_headers: &http::HeaderMap,
     model_capabilities: Option<&crate::ai_serving::CodexResponsesModelCapabilities>,
     enable_model_directives: bool,
+) -> Option<Value> {
+    build_local_openai_responses_request_body_with_codex_model_capabilities_and_websocket_mode(
+        body_json,
+        mapped_model,
+        require_streaming,
+        force_body_stream_field,
+        provider_type,
+        provider_api_format,
+        body_rules,
+        request_headers,
+        model_capabilities,
+        enable_model_directives,
+        false,
+    )
+}
+
+/// Builds a Responses body for a pinned WebSocket continuation.
+///
+/// This is intentionally an additive variant of the ordinary HTTP builder.
+/// The WebSocket framing layer, rather than a JSON-body heuristic, tells the
+/// Codex compatibility pass that `previous_response_id` is transport state and
+/// that Responses Lite `tools`/`instructions` must not be materialized into a
+/// second historical input prefix.
+pub(crate) fn build_local_openai_responses_request_body_with_codex_model_capabilities_for_websocket_continuation(
+    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>,
+    request_headers: &http::HeaderMap,
+    model_capabilities: Option<&crate::ai_serving::CodexResponsesModelCapabilities>,
+    enable_model_directives: bool,
+) -> Option<Value> {
+    build_local_openai_responses_request_body_with_codex_model_capabilities_and_websocket_mode(
+        body_json,
+        mapped_model,
+        require_streaming,
+        force_body_stream_field,
+        provider_type,
+        provider_api_format,
+        body_rules,
+        request_headers,
+        model_capabilities,
+        enable_model_directives,
+        true,
+    )
+}
+
+#[allow(clippy::too_many_arguments)]
+fn build_local_openai_responses_request_body_with_codex_model_capabilities_and_websocket_mode(
+    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>,
+    request_headers: &http::HeaderMap,
+    model_capabilities: Option<&crate::ai_serving::CodexResponsesModelCapabilities>,
+    enable_model_directives: bool,
+    websocket_continuation: bool,
 ) -> Option<Value> {
     let provider_request_body = surface_build_local_openai_responses_request_body(
         body_json,
@@ -68,15 +131,27 @@ pub(crate) fn build_local_openai_responses_request_body_with_codex_model_capabil
         .get("model")
         .and_then(Value::as_str)
         .unwrap_or(mapped_model);
-    crate::ai_serving::apply_codex_openai_responses_special_body_edits_with_source_model_and_capabilities(
-        &mut provider_request_body,
-        provider_type,
-        provider_api_format,
-        mapped_model,
-        source_model,
-        model_capabilities,
-        body_rules,
-    );
+    if websocket_continuation {
+        crate::ai_serving::apply_codex_openai_responses_websocket_continuation_body_edits_with_source_model_and_capabilities(
+            &mut provider_request_body,
+            provider_type,
+            provider_api_format,
+            mapped_model,
+            source_model,
+            model_capabilities,
+            body_rules,
+        );
+    } else {
+        crate::ai_serving::apply_codex_openai_responses_special_body_edits_with_source_model_and_capabilities(
+            &mut provider_request_body,
+            provider_type,
+            provider_api_format,
+            mapped_model,
+            source_model,
+            model_capabilities,
+            body_rules,
+        );
+    }
     apply_openai_responses_compact_special_body_edits(
         &mut provider_request_body,
         provider_api_format,

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

@@ -153,7 +153,7 @@ fn local_openai_responses_wrapper_preserves_body_order_after_edits() {
 }
 
 #[test]
-fn local_openai_responses_wrapper_strips_foreign_reasoning_item_ids() {
+fn local_openai_responses_wrapper_defers_reasoning_replay_filtering() {
     let body_json = json!({
         "model": "gpt-5.4",
         "input": [
@@ -184,9 +184,14 @@ fn local_openai_responses_wrapper_strips_foreign_reasoning_item_ids() {
     let input = provider_request_body["input"]
         .as_array()
         .expect("input array");
-    assert_eq!(input.len(), 2);
+    // This provider-agnostic normalization layer cannot decide whether an
+    // id-less/foreign reasoning item is opaque state required by DeepSeek.
+    // The provider-aware finalization pass applies the strict or DeepSeek
+    // replay policy once the selected upstream base URL is known.
+    assert_eq!(input.len(), 3);
     assert_eq!(input[0]["id"], "rs_provider_123");
-    assert_eq!(input[1]["type"], "message");
+    assert_eq!(input[1]["id"], "item_72d3bd8d367d01977ace23f1");
+    assert_eq!(input[2]["type"], "message");
 }
 
 #[test]