diff --git a/server-rs/crates/api-server/src/editor_screen_background_decision.rs b/server-rs/crates/api-server/src/editor_screen_background_decision.rs
index 12f2bb8a1..28fa38d23 100644
--- a/server-rs/crates/api-server/src/editor_screen_background_decision.rs
+++ b/server-rs/crates/api-server/src/editor_screen_background_decision.rs
@@ -94,25 +94,28 @@ pub(crate) async fn resolve_editor_screen_background_color(
         .map(|report| report.allowed.clone())
         .unwrap_or_else(|| EDITOR_SCREEN_BACKGROUND_COLORS.to_vec());
 
-    // 默认 llm_client 是纯文本模型（Ark），收到图片分片会被上游 400 拒绝；
-    // 带图决策必须走 VectorEngine 视觉客户端，没有视觉客户端时降级为纯文本决策。
-    let (llm_client, vision_model) = if has_source_image {
-        match vision_llm_client {
-            Some(client) => (
-                Some(client),
-                Some(crate::llm_model_routing::EDITOR_SCREEN_BACKGROUND_VISION_LLM_MODEL),
-            ),
-            None => {
+    // 决策统一走 VectorEngine gpt-5-mini：有图用视觉档、无图用文本档（两个独立常量），
+    // 都不继承 Ark 默认文本模型（豆包，选色能力弱）。仅当 gpt5 客户端未配置时才降级回默认
+    // llm_client；该默认客户端是纯文本模型（Ark），收到图片分片会被上游 400 拒绝，故先丢弃图片分片。
+    let (llm_client, decision_model) = match vision_llm_client {
+        Some(client) if has_source_image => (
+            Some(client),
+            Some(crate::llm_model_routing::EDITOR_SCREEN_BACKGROUND_VISION_LLM_MODEL),
+        ),
+        Some(client) => (
+            Some(client),
+            Some(crate::llm_model_routing::EDITOR_SCREEN_BACKGROUND_TEXT_LLM_MODEL),
+        ),
+        None => {
+            if has_source_image {
                 warn!(
                     kind = input.kind.label(),
                     "editor_screen_background_vision_client_missing_downgrade_to_text"
                 );
                 input.source_image_data_url = None;
-                (llm_client, None)
             }
+            (llm_client, None)
         }
-    } else {
-        (llm_client, None)
     };
 
     // 默认兜底色被过滤掉时，改用危险度最小的候选兜底。
@@ -152,15 +155,15 @@ pub(crate) async fn resolve_editor_screen_background_color(
             None => LlmMessage::user(user_prompt.as_str()),
         };
         // 预算要够推理模型（如 gpt-5-mini）先花几百 token 推理、再吐 JSON 答案；
-        // 实测 low 档推理约 320~384 token，取 1024 留足余量。非推理模型遇 stop 提前结束，不会多花。
+        // 实测 low 档推理约 320~384 token，取 1024 留足余量。降级客户端遇 stop 提前结束，不会多花。
         let mut request = LlmTextRequest::new(vec![LlmMessage::system(system_prompt), user_message])
             .with_max_tokens(1024)
             .with_request_timeout_ms(EDITOR_SCREEN_BACKGROUND_DECISION_TIMEOUT_MS);
-        if let Some(vision_model) = vision_model {
-            // 视觉模型 gpt-5-mini 是推理模型：走 Responses 协议并压到 low 推理档，
-            // 否则默认档会把预算全烧在推理上、返回空答案。
+        if let Some(decision_model) = decision_model {
+            // gpt-5-mini 是推理模型（有图视觉档 / 无图文本档均适用）：走 Responses 协议并压到 low
+            // 推理档，否则默认档会把预算全烧在推理上、返回空答案。
             request = request
-                .with_model(vision_model)
+                .with_model(decision_model)
                 .with_responses_api()
                 .with_response_reasoning_effort(LlmResponseReasoningEffort::Low);
         }
@@ -526,4 +529,143 @@ mod tests {
             "默认兜底色与前景撞色时应改用危险度最小的候选"
         );
     }
+
+    // 真机联调：按 build_creative_agent_gpt5_client 的方式组 VectorEngine 客户端，直接跑
+    // resolve_editor_screen_background_color 的完整代码路径（无图文本档 + 有图视觉档），
+    // 验证决策请求真的打到 VectorEngine 并被解析成候选色（decision.fallback == false）。
+    // 凭证从仓库根 .env.local / .env.secrets.local 读，需要真实 VECTOR_ENGINE_* 才有意义。
+    // 运行：cargo test -p api-server --manifest-path server-rs/Cargo.toml \
+    //         editor_screen_background_decision::tests::live -- --ignored --nocapture
+    fn read_live_env(key: &str) -> Option<String> {
+        use std::collections::BTreeMap;
+        use std::path::PathBuf;
+
+        let repo_root = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+            .join("..")
+            .join("..")
+            .join("..");
+        let mut map: BTreeMap<String, String> = BTreeMap::new();
+        for name in [".env.local", ".env.secrets.local", ".env"] {
+            let path = repo_root.join(name);
+            let Ok(content) = std::fs::read_to_string(&path) else {
+                continue;
+            };
+            for line in content.lines() {
+                let trimmed = line.trim();
+                if trimmed.is_empty() || trimmed.starts_with('#') || !trimmed.contains('=') {
+                    continue;
+                }
+                let (k, v) = trimmed.split_once('=').unwrap();
+                let v = v.trim().trim_matches('"').trim_matches('\'');
+                // 先出现的文件优先（.env.local > .env.secrets.local > .env），与服务端 dotenv 顺序一致。
+                map.entry(k.trim().to_string()).or_insert_with(|| v.to_string());
+            }
+        }
+        std::env::var(key).ok().or_else(|| map.get(key).cloned())
+    }
+
+    fn build_live_vector_engine_client() -> Option<LlmClient> {
+        use platform_llm::{LlmConfig, LlmProvider};
+
+        let base_url = read_live_env("VECTOR_ENGINE_BASE_URL")?;
+        let api_key = read_live_env("VECTOR_ENGINE_API_KEY")?;
+        // 与 state.rs build_creative_agent_gpt5_client 一致：规整到以 /v1 结尾。
+        let base_url = if base_url.trim_end_matches('/').ends_with("/v1") {
+            base_url.trim_end_matches('/').to_string()
+        } else {
+            format!("{}/v1", base_url.trim_end_matches('/'))
+        };
+        let config = LlmConfig::new(
+            LlmProvider::OpenAiCompatible,
+            base_url,
+            api_key,
+            crate::llm_model_routing::EDITOR_SCREEN_BACKGROUND_TEXT_LLM_MODEL.to_string(),
+            60_000,
+            0,
+            500,
+        )
+        .expect("live VectorEngine LlmConfig should build");
+        Some(LlmClient::new(config).expect("live VectorEngine LlmClient should build"))
+    }
+
+    fn solid_source_image_data_url() -> String {
+        use base64::Engine as _;
+        use image::{Rgba, RgbaImage};
+
+        let image = RgbaImage::from_pixel(64, 64, Rgba([120, 180, 120, 255]));
+        let mut bytes = Vec::new();
+        image::DynamicImage::ImageRgba8(image)
+            .write_to(&mut std::io::Cursor::new(&mut bytes), image::ImageFormat::Png)
+            .expect("test image should encode");
+        format!(
+            "data:image/png;base64,{}",
+            base64::engine::general_purpose::STANDARD.encode(&bytes)
+        )
+    }
+
+    #[tokio::test]
+    #[ignore = "真机联调：需要 .env.local / .env.secrets.local 中真实 VECTOR_ENGINE_* 凭证"]
+    async fn live_screen_background_decision_hits_vector_engine_without_image() {
+        let Some(client) = build_live_vector_engine_client() else {
+            panic!("缺少 VECTOR_ENGINE_BASE_URL / VECTOR_ENGINE_API_KEY，无法真机联调");
+        };
+        let decision = resolve_editor_screen_background_color(
+            None,
+            Some(&client),
+            EditorScreenBackgroundDecisionInput {
+                kind: EditorScreenBackgroundDecisionKind::Character,
+                screen_color: Some("auto".to_string()),
+                prompt: "赛博朋克风格的机械猫，霓虹蓝紫主体".to_string(),
+                icon_descriptions: Vec::new(),
+                reference_count: 1,
+                source_image_data_url: None,
+            },
+        )
+        .await
+        .expect("live 无图决策应成功");
+
+        eprintln!(
+            "[live 无图] mode={:?} hex={} label={} attempts={} fallback={}",
+            decision.mode, decision.color.hex, decision.color.label, decision.attempts, decision.fallback
+        );
+        assert_eq!(decision.mode, EditorScreenBackgroundDecisionMode::Auto);
+        assert!(
+            !decision.fallback,
+            "若走到兜底说明 LLM 没答复（VectorEngine 未被成功调用或响应解析失败）"
+        );
+        assert!(decision.attempts >= 1);
+    }
+
+    #[tokio::test]
+    #[ignore = "真机联调：需要 .env.local / .env.secrets.local 中真实 VECTOR_ENGINE_* 凭证"]
+    async fn live_screen_background_decision_hits_vector_engine_with_image() {
+        let Some(client) = build_live_vector_engine_client() else {
+            panic!("缺少 VECTOR_ENGINE_BASE_URL / VECTOR_ENGINE_API_KEY，无法真机联调");
+        };
+        let decision = resolve_editor_screen_background_color(
+            None,
+            Some(&client),
+            EditorScreenBackgroundDecisionInput {
+                kind: EditorScreenBackgroundDecisionKind::CharacterAnimation,
+                screen_color: Some("auto".to_string()),
+                prompt: "角色起跳动作".to_string(),
+                icon_descriptions: Vec::new(),
+                reference_count: 1,
+                source_image_data_url: Some(solid_source_image_data_url()),
+            },
+        )
+        .await
+        .expect("live 有图决策应成功");
+
+        eprintln!(
+            "[live 有图] mode={:?} hex={} label={} attempts={} fallback={}",
+            decision.mode, decision.color.hex, decision.color.label, decision.attempts, decision.fallback
+        );
+        assert_eq!(decision.mode, EditorScreenBackgroundDecisionMode::Auto);
+        assert!(
+            !decision.fallback,
+            "若走到兜底说明视觉 LLM 没答复（VectorEngine 未被成功调用或响应解析失败）"
+        );
+        assert!(decision.attempts >= 1);
+    }
 }
diff --git a/server-rs/crates/api-server/src/llm_model_routing.rs b/server-rs/crates/api-server/src/llm_model_routing.rs
index 49f8e6740..e3485fe09 100644
--- a/server-rs/crates/api-server/src/llm_model_routing.rs
+++ b/server-rs/crates/api-server/src/llm_model_routing.rs
@@ -1,6 +1,9 @@
 pub(crate) const RPG_STORY_LLM_MODEL: &str = "doubao-seed-character-251128";
 pub(crate) const CREATION_TEMPLATE_LLM_MODEL: &str = "deepseek-v3-2-251201";
 pub(crate) const PUZZLE_LEVEL_NAME_VISION_LLM_MODEL: &str = "gpt-4o-mini";
-// 抠图背景色决策的视觉模型。gpt-5-mini 是推理模型，会先花若干 token 做推理，
+// 抠图背景色决策的视觉模型（有源图）。gpt-5-mini 是推理模型，会先花若干 token 做推理，
 // 故决策请求的 max_tokens 需留足推理开销（见 editor_screen_background_decision）。
 pub(crate) const EDITOR_SCREEN_BACKGROUND_VISION_LLM_MODEL: &str = "gpt-5-mini";
+// 抠图背景色决策的文本模型（无源图）。独立于视觉档，便于两条路径分别调参；
+// 硬编码而非继承 Ark 默认文本模型（豆包，选色能力弱）。同为 gpt-5-mini 推理模型，同样留足推理预算。
+pub(crate) const EDITOR_SCREEN_BACKGROUND_TEXT_LLM_MODEL: &str = "gpt-5-mini";