diff --git a/Cargo.lock b/Cargo.lock
index a785f9a..366017a 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -869,6 +869,7 @@ dependencies = [
  "reqwest 0.12.28",
  "serde",
  "serde_json",
+ "sys-locale",
  "tauri",
  "tauri-build",
  "tauri-plugin-dialog",
@@ -4422,6 +4423,15 @@ dependencies = [
  "syn 2.0.117",
 ]
 
+[[package]]
+name = "sys-locale"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8eab9a99a024a169fe8a903cf9d4a3b3601109bcc13bd9e3c6fff259138626c4"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "system-configuration"
 version = "0.7.0"
diff --git a/crates/df-ai-core/src/model.rs b/crates/df-ai-core/src/model.rs
index 795749e..8c0cc2e 100644
--- a/crates/df-ai-core/src/model.rs
+++ b/crates/df-ai-core/src/model.rs
@@ -138,6 +138,14 @@ pub struct ModelConfig {
     #[serde(default = "default_context_window")]
     pub context_window: usize,
 
+    /// 是否支持 temperature 参数(默认 true)。
+    ///
+    /// 部分模型仅接受固定温度值(如商汤 sensenova 系列仅允许 1.0),透传其它温度(标题 0.3/
+    /// 压缩 0.2 等)→ 上游 HTTP 400,实测 9 次。false 时协议转换层省略 temperature 字段
+    /// (交由上游默认),避免无效往返;true 保持原样透传(零回归)。
+    #[serde(default = "default_true")]
+    pub temperature_supported: bool,
+
     // ── 探测元数据(只读,由 ModelProbe 填充,Phase 2)──
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub probe_source: Option<ProbeSource>,
@@ -173,6 +181,9 @@ impl ModelConfig {
             intelligence: IntelligenceTier::Standard,
             weight: default_weight(),
             context_window: default_context_window(),
+            // 默认支持 temperature(老格式升级/未知模型保持原样透传,零回归);
+            // 明确不支持的模型由 presets 表 temperature_supported=false 标注。
+            temperature_supported: true,
             probe_source: None,
         }
     }
@@ -339,6 +350,7 @@ mod tests {
             intelligence: IntelligenceTier::Plus,
             weight: 70,
             context_window: 128_000,
+            temperature_supported: true,
             probe_source: Some(ProbeSource::PresetTable),
         };
         let json = serde_json::to_string(&mc).unwrap();
diff --git a/crates/df-ai-core/src/provider.rs b/crates/df-ai-core/src/provider.rs
index f953a9b..5d92f6e 100644
--- a/crates/df-ai-core/src/provider.rs
+++ b/crates/df-ai-core/src/provider.rs
@@ -472,6 +472,7 @@ mod tests {
             usage: None,
             error: None,
             reasoning_content: Some("thought".to_string()),
+            truncated_by_max_tokens: false,
         };
         let json = serde_json::to_string(&chunk).unwrap();
         assert!(json.contains("reasoning_content"));
diff --git a/crates/df-ai-core/src/types.rs b/crates/df-ai-core/src/types.rs
index f7b6e4f..99f96c9 100644
--- a/crates/df-ai-core/src/types.rs
+++ b/crates/df-ai-core/src/types.rs
@@ -350,6 +350,11 @@ pub struct StreamChunk {
     /// 非空表示流中途出错，不应视为正常完成（finished 路径），由 stream_llm 转 AiError。
     #[serde(skip)]
     pub error: Option<String>,
+    /// 本流因 provider 输出上限截断收尾(openai finish_reason=length /
+    /// anthropic stop_reason=max_tokens)。finished 仍为 true(正常终止非断连),但文本/
+    /// tool_calls 可能不完整;消费方跨帧 OR 累积该标志,轮产出据此标记截断,不冒充正常完成。
+    #[serde(default)]
+    pub truncated_by_max_tokens: bool,
     /// DeepSeek thinking 模式推理内容增量
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub reasoning_content: Option<String>,
diff --git a/crates/df-ai/presets/models.json b/crates/df-ai/presets/models.json
index 771f17e..ef308e0 100644
--- a/crates/df-ai/presets/models.json
+++ b/crates/df-ai/presets/models.json
@@ -133,7 +133,8 @@
     "modalities": ["text", "vision"],
     "capabilities": ["tool_use"],
     "weight": 85,
-    "context_window": 262144
+    "context_window": 262144,
+    "temperature_supported": false
   },
   {
     "model_id": "deepseek-v4-flash",
diff --git a/crates/df-ai/src/anthropic_compat.rs b/crates/df-ai/src/anthropic_compat.rs
index 4c56148..9d23bb2 100644
--- a/crates/df-ai/src/anthropic_compat.rs
+++ b/crates/df-ai/src/anthropic_compat.rs
@@ -262,12 +262,23 @@ impl AnthropicCompatProvider {
                 .collect()
         });
 
+        // 模型不支持 temperature(如商汤 sensenova 仅接受 1.0)→ 省略该字段,避免上游 HTTP 400。
+        // 判定走 model_probe(命中 presets 表;未命中默认支持,保持原样透传零回归)。
+        let temperature = if req.temperature.is_some()
+            && !crate::model_probe::probe(&model).temperature_supported
+        {
+            tracing::debug!(model = %model, "模型不支持 temperature,已省略该参数");
+            None
+        } else {
+            req.temperature
+        };
+
         AnthropicRequest {
             model,
             messages,
             max_tokens: req.max_tokens.unwrap_or(DEFAULT_MAX_TOKENS),
             system,
-            temperature: req.temperature,
+            temperature,
             stream: req.stream,
             tools,
             tool_choice: req.tool_choice,
@@ -927,6 +938,21 @@ mod tests {
         assert_eq!(a.total_tokens, 15);
     }
 
+    /// stop_reason=max_tokens(message_delta 帧携带)必须置截断标志;end_turn 不置。
+    /// 终态 finished 在后续 message_stop 帧,消费方跨帧 OR 累积标志。
+    #[test]
+    fn anthropic_max_tokens_stop_reason_sets_truncation_flag() {
+        let mut acc: Option<TokenUsage> = None;
+        let data = r#"{"type":"message_delta","delta":{"stop_reason":"max_tokens"},"usage":{"output_tokens":9}}"#;
+        let c = apply_anthropic_event(data, &mut acc);
+        assert!(!c.finished, "message_delta 非终态帧");
+        assert!(c.truncated_by_max_tokens, "max_tokens 必须置截断标志");
+
+        let mut acc2: Option<TokenUsage> = None;
+        let c2 = apply_anthropic_event(&message_delta(5), &mut acc2);
+        assert!(!c2.truncated_by_max_tokens, "end_turn 不置截断标志");
+    }
+
     /// message_delta 的 output_tokens 必须是累计覆盖（非累加）：连续两个 delta 5 和 8，结果应是 8 不是 13
     #[test]
     fn anthropic_message_delta_output_is_cumulative_not_incremental() {
diff --git a/crates/df-ai/src/anthropic_helpers.rs b/crates/df-ai/src/anthropic_helpers.rs
index c3fe395..0e6ca03 100644
--- a/crates/df-ai/src/anthropic_helpers.rs
+++ b/crates/df-ai/src/anthropic_helpers.rs
@@ -137,7 +137,7 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
     let v: serde_json::Value = match serde_json::from_str(data) {
         Ok(v) => v,
         Err(_) => {
-            return StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None }
+            return StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None, truncated_by_max_tokens: false }
         }
     };
     let ty = v.get("type").and_then(|t| t.as_str()).unwrap_or("");
@@ -169,7 +169,7 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
                     reasoning_tokens: 0,
                 });
             }
-            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: usage_accum.clone(), error: None, reasoning_content: None }
+            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: usage_accum.clone(), error: None, reasoning_content: None, truncated_by_max_tokens: false }
         }
         // 消息增量：output_tokens 是累计值（非增量），直接覆盖 completion + 重算 total
         "message_delta" => {
@@ -179,9 +179,12 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
                 acc.completion_tokens = out as u32;
                 acc.total_tokens = acc.prompt_tokens.saturating_add(acc.completion_tokens);
             }
+            // stop_reason=max_tokens(message_delta 帧携带):输出被 max_tokens 截断。
+            // 终态 finished 在后续 message_stop 帧,截断标志挂本帧,消费方跨帧 OR 累积。
+            let truncated_by_max_tokens = v.get("delta").and_then(|d| d.get("stop_reason")).and_then(|s| s.as_str()) == Some("max_tokens");
             // 加固:usage 挂到本帧(而非仅 message_stop 带出),中途断连/端点不发 message_stop
             // 时仍能拿到真实 usage,对称 openai_helpers 的修复。
-            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: usage_accum.clone(), error: None, reasoning_content: None }
+            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: usage_accum.clone(), error: None, reasoning_content: None, truncated_by_max_tokens }
         }
         // 文本增量
         "content_block_delta" => {
@@ -190,16 +193,16 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
                 // 不混入正文 delta。本侧请求未启用 thinking，此分支只为不丢端点侧自行产出的思考。
                 if delta.get("type").and_then(|t| t.as_str()) == Some("thinking_delta") {
                     let text = delta.get("thinking").and_then(|t| t.as_str()).unwrap_or("").to_string();
-                    return StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: Some(text) };
+                    return StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: Some(text), truncated_by_max_tokens: false };
                 }
                 // signature_delta：thinking 块签名字段，仅回传校验用。DevFlow 从不把 thinking 回传
                 // anthropic 历史（丢弃 signature 后伪造回传必 400），签名无处消费，安静忽略。
                 if delta.get("type").and_then(|t| t.as_str()) == Some("signature_delta") {
-                    return StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None };
+                    return StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None, truncated_by_max_tokens: false };
                 }
                 if delta.get("type").and_then(|t| t.as_str()) == Some("text_delta") {
                     let text = delta.get("text").and_then(|t| t.as_str()).unwrap_or("").to_string();
-                    return StreamChunk { delta: text, finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None };
+                    return StreamChunk { delta: text, finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None, truncated_by_max_tokens: false };
                 }
                 // 工具入参增量
                 if delta.get("type").and_then(|t| t.as_str()) == Some("input_json_delta") {
@@ -217,10 +220,11 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
                         usage: None,
                         error: None,
                         reasoning_content: None,
+                        truncated_by_max_tokens: false,
                     };
                 }
             }
-            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None }
+            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None, truncated_by_max_tokens: false }
         }
         // 工具块开始：带 id + name
         "content_block_start" => {
@@ -253,10 +257,11 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
                         usage: None,
                         error: None,
                         reasoning_content: None,
+                        truncated_by_max_tokens: false,
                     };
                 }
             }
-            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None }
+            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None, truncated_by_max_tokens: false }
         }
         // 消息结束：带出累积 usage
         "message_stop" => StreamChunk {
@@ -266,15 +271,16 @@ pub(crate) fn apply_anthropic_event(data: &str, usage_accum: &mut Option<TokenUs
             usage: usage_accum.take(),
             error: None,
             reasoning_content: None,
+            truncated_by_max_tokens: false,
         },
         // 错误事件：流中途出错。不走 finished 完成路径（避免残缺响应被当正常完成入库），
         // 改由 stream_llm 识别 error 非空 → 发 AiError + 丢弃残缺（与 OpenAI 路径 Err 一致）。
         "error" => {
             let msg = v.get("error").and_then(|e| e.get("message")).and_then(|m| m.as_str()).unwrap_or("stream error").to_string();
             error!(%msg, "Anthropic 流式错误事件");
-            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: Some(msg), reasoning_content: None }
+            StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: Some(msg), reasoning_content: None, truncated_by_max_tokens: false }
         }
         // content_block_stop / ping 等不产出 chunk
-        _ => StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None },
+        _ => StreamChunk { delta: String::new(), finished: false, tool_calls: None, usage: None, error: None, reasoning_content: None, truncated_by_max_tokens: false },
     }
 }
diff --git a/crates/df-ai/src/context/manager_tests.rs b/crates/df-ai/src/context/manager_tests.rs
index 2e0ffce..4ba62fe 100644
--- a/crates/df-ai/src/context/manager_tests.rs
+++ b/crates/df-ai/src/context/manager_tests.rs
@@ -841,12 +841,12 @@ fn phase1_push_only_keeps_clean_no_dirty_flag() {
 
 #[test]
 fn phase1_advance_persisted_count_sets_baseline_and_clears_dirty() {
-    // advance_persisted_count(new_count):推进基线 + 清 needs_full_rewrite(DB 已与内存一致)
+    // advance_persisted_count(new_count, full):推进基线;全量重写才清 needs_full_rewrite
     let mut mgr = ContextManager::new(cfg(100_000));
     mgr.push(ChatMessage::user("x"));
     mgr.push(ChatMessage::user("y"));
-    // 模拟 save 成功后推进
-    mgr.advance_persisted_count(2);
+    // 模拟 save(全量重写)成功后推进
+    mgr.advance_persisted_count(2, true);
     assert_eq!(mgr.persisted_msg_count(), 2);
     assert!(!mgr.needs_full_rewrite());
 }
@@ -855,11 +855,26 @@ fn phase1_advance_persisted_count_sets_baseline_and_clears_dirty() {
 fn phase1_advance_persisted_count_max_no_regression() {
     // 防回退:advance 传更小值不回退基线(防误传 / 并发致基线丢失 → 下次增量漏写)
     let mut mgr = ContextManager::new(cfg(100_000));
-    mgr.advance_persisted_count(5);
-    mgr.advance_persisted_count(3); // 误传更小
+    mgr.advance_persisted_count(5, true);
+    mgr.advance_persisted_count(3, true); // 误传更小
     assert_eq!(mgr.persisted_msg_count(), 5, "基线不可回退");
 }
 
+#[test]
+fn phase1_advance_incremental_keeps_dirty() {
+    // 增量/无变化路径(full=false)不得清 dirty:写库期间新置的 dirty(如压缩链 mark)
+    // 与本次 append 写库无关,误清会致改动滞留内存永不落库
+    let mut mgr = ContextManager::new(cfg(100_000));
+    mgr.push(ChatMessage::user("a"));
+    mgr.advance_persisted_count(1, false); // 增量 save 成功
+    assert!(!mgr.needs_full_rewrite());
+    mgr.mark_needs_full_rewrite(); // 模拟写库期间压缩 mark
+    mgr.advance_persisted_count(1, false); // 又一次增量 save
+    assert!(mgr.needs_full_rewrite(), "增量路径 advance 不得清 dirty");
+    mgr.advance_persisted_count(1, true); // 直到全量重写才清
+    assert!(!mgr.needs_full_rewrite());
+}
+
 #[test]
 fn phase1_compress_old_messages_sets_dirty() {
     // compress 改既有消息 status(active→compressed) → 必须全量重写收敛
@@ -881,7 +896,7 @@ fn phase1_compress_no_op_does_not_set_dirty() {
     mgr.push(ChatMessage::user("a"));
     let _ = mgr.compress_old_messages(1);
     assert!(mgr.needs_full_rewrite(), "首次压缩应置 dirty");
-    mgr.advance_persisted_count(1); // 模拟 save 后清 dirty
+    mgr.advance_persisted_count(1, true); // 模拟 save(全量)后清 dirty
     // 二次压缩同范围:全已 compressed,无新标 → 不应再置 dirty
     let compressed = mgr.compress_old_messages(1);
     assert!(compressed.is_empty());
@@ -974,7 +989,7 @@ fn phase1_clear_resets_baseline_and_sets_dirty() {
     // clear:基线重置 0 + 置 dirty(下次 save 全量重写空列表清 DB)
     let mut mgr = ContextManager::new(cfg(100_000));
     mgr.push(ChatMessage::user("a"));
-    mgr.advance_persisted_count(1);
+    mgr.advance_persisted_count(1, true);
     assert_eq!(mgr.persisted_msg_count(), 1);
     mgr.clear();
     assert_eq!(mgr.persisted_msg_count(), 0, "clear 应重置基线到 0");
@@ -1021,7 +1036,7 @@ fn phase1_append_after_dirty_stays_dirty_until_save() {
     mgr.push(ChatMessage::user("b")); // 追加
     assert!(mgr.needs_full_rewrite(), "dirty 后 push 不应清 dirty");
     // 模拟 save(全量重写)成功后推进 + 清
-    mgr.advance_persisted_count(2);
+    mgr.advance_persisted_count(2, true);
     assert!(!mgr.needs_full_rewrite());
 }
 
diff --git a/crates/df-ai/src/context/mod.rs b/crates/df-ai/src/context/mod.rs
index 2f33431..7e1bab9 100644
--- a/crates/df-ai/src/context/mod.rs
+++ b/crates/df-ai/src/context/mod.rs
@@ -528,11 +528,15 @@ impl ContextManager {
     }
 
     /// msg-split-phase1:推进已落库消息条数(insert_batch/replace_conversation 成功后调用)。
-    /// 同时清 needs_full_rewrite(DB 已与内存一致)。供 save_conversation 在写库成功后调用。
-    pub fn advance_persisted_count(&mut self, new_count: usize) {
+    /// `full_rewrite_applied`=本次写库是否走了全量重写(快照时刻 needs_full_rewrite 快照)。
+    /// 仅全量重写覆盖了既有行才清 needs_full_rewrite;增量/无变化路径写库期间新置的 dirty
+    /// (如压缩链 mark)与本次 append 写库无关,误清会致改动滞留内存永不落库。供 save_conversation 写库成功后调用。
+    pub fn advance_persisted_count(&mut self, new_count: usize, full_rewrite_applied: bool) {
         // 防回退:若调用方误传更小值(理论上不应发生),取 max 保不丢基线(下次 save 仍能收敛)。
         self.persisted_msg_count = self.persisted_msg_count.max(new_count);
-        self.needs_full_rewrite = false;
+        if full_rewrite_applied {
+            self.needs_full_rewrite = false;
+        }
     }
 
     /// msg-split-phase1:是否需要全量重写(改过既有消息 / 缩短 / 中段插入 / 清空)。
diff --git a/crates/df-ai/src/model_probe.rs b/crates/df-ai/src/model_probe.rs
index 92722cc..6eaf7cc 100644
--- a/crates/df-ai/src/model_probe.rs
+++ b/crates/df-ai/src/model_probe.rs
@@ -159,6 +159,24 @@ mod tests {
         assert!(m.modalities.contains(&Modality::Text));
         // context_window 应为预设的 256K
         assert_eq!(m.context_window, 262144);
+        // P1 温度:商汤 sensenova 仅接受 temperature=1.0,透传其它值(标题 0.3/压缩 0.2)
+        // → 上游 HTTP 400(实测 9 次)。预设表须标 temperature_supported=false,
+        // 协议转换层据此省略该字段。
+        assert!(
+            !m.temperature_supported,
+            "sensenova 应标 temperature_supported=false(仅接受 1.0)"
+        );
+    }
+
+    /// 温度支持判定:仅 sensenova 系列为 false,其余模型默认 true(零回归)。
+    #[test]
+    fn probe_temperature_supported_defaults_true_except_sensenova() {
+        assert!(!probe("sensenova-6.7-flash-lite").temperature_supported);
+        // 启发式路径(非预设表):sensenova 前缀仍判 false
+        assert!(!probe("sensenova-unknown-model").temperature_supported);
+        // 常规模型保持 true(原样透传)
+        assert!(probe("deepseek-v4-flash").temperature_supported);
+        assert!(probe("glm-4v").temperature_supported);
     }
 
     #[test]
diff --git a/crates/df-ai/src/model_probe_helpers.rs b/crates/df-ai/src/model_probe_helpers.rs
index ab38ee0..158e0f3 100644
--- a/crates/df-ai/src/model_probe_helpers.rs
+++ b/crates/df-ai/src/model_probe_helpers.rs
@@ -74,6 +74,7 @@ pub(super) fn heuristic_infer(model_id: &str) -> ModelConfig {
             intelligence,
             weight: 50,
             context_window: 8192,
+            temperature_supported: temperature_supported_by_name(&name),
             probe_source: None,
         };
     }
@@ -106,6 +107,7 @@ pub(super) fn heuristic_infer(model_id: &str) -> ModelConfig {
         intelligence,
         weight: 50,
         context_window: 8192,
+        temperature_supported: temperature_supported_by_name(&name),
         probe_source: None,
     }
 }
@@ -120,3 +122,11 @@ pub(super) fn has_vision_token(name: &str) -> bool {
         || name.contains("vl")
         || name.split(['-', '.', '_']).any(|seg| seg.ends_with('v'))
 }
+
+/// 温度参数支持判定(命名约定式):商汤 sensenova 系列仅接受 temperature=1.0,
+/// 透传其它温度值(标题 0.3/压缩 0.2 等)→ 上游 HTTP 400(实测 9 次)。
+/// 返回 false 时协议转换层省略 temperature 字段(交上游默认),避免无效往返;
+/// 其余模型默认 true(原样透传,零回归)。
+pub(super) fn temperature_supported_by_name(name: &str) -> bool {
+    !name.starts_with("sensenova")
+}
diff --git a/crates/df-ai/src/openai_compat.rs b/crates/df-ai/src/openai_compat.rs
index e723065..93fcdf5 100644
--- a/crates/df-ai/src/openai_compat.rs
+++ b/crates/df-ai/src/openai_compat.rs
@@ -209,10 +209,21 @@ impl OpenAICompatProvider {
                 .collect()
         });
 
+        // 模型不支持 temperature(如商汤 sensenova 仅接受 1.0)→ 省略该字段,避免上游 HTTP 400。
+        // 判定走 model_probe(命中 presets 表;未命中默认支持,保持原样透传零回归)。
+        let temperature = if req.temperature.is_some()
+            && !crate::model_probe::probe(&model).temperature_supported
+        {
+            tracing::debug!(model = %model, "模型不支持 temperature,已省略该参数");
+            None
+        } else {
+            req.temperature
+        };
+
         OpenAiRequest {
             model,
             messages,
-            temperature: req.temperature,
+            temperature,
             max_tokens: req.max_tokens,
             stream: req.stream,
             tools,
diff --git a/crates/df-ai/src/openai_helpers.rs b/crates/df-ai/src/openai_helpers.rs
index b77a19b..2cb14a6 100644
--- a/crates/df-ai/src/openai_helpers.rs
+++ b/crates/df-ai/src/openai_helpers.rs
@@ -274,6 +274,7 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
             usage: usage_accum.take(),
             error: None,
             reasoning_content: None,
+            truncated_by_max_tokens: false,
         };
     }
 
@@ -296,6 +297,7 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
                     usage: None,
                     error: Some(msg),
                     reasoning_content: None,
+                    truncated_by_max_tokens: false,
                 };
             }
             // 提取 usage（带 include_usage 时末段 chunk 携带，覆盖累积）
@@ -334,6 +336,9 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
                     choice.finish_reason.as_deref(),
                     Some("stop") | Some("tool_calls") | Some("length") | Some("content_filter")
                 );
+                // length = 输出被 max_tokens 截断:仍按正常终止收敛(finished=true),
+                // 但截断标志随 chunk 带出,消费方据此把该轮产出标记截断,不冒充完整。
+                let truncated_by_max_tokens = choice.finish_reason.as_deref() == Some("length");
 
                 let tool_calls = choice.delta.tool_calls.map(|tcs| {
                     // 缺 index 的增量缺省 0(单工具调用语义等价);但同帧多 call 且 index
@@ -379,6 +384,7 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
                     usage: usage_accum.clone(),
                     error: None,
                     reasoning_content: choice.delta.reasoning_content,
+                    truncated_by_max_tokens,
                 }
             } else {
                 // choices 为空 = usage-only chunk，不输出文本（usage 已累积），usage 一并带出
@@ -389,6 +395,7 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
                     usage: usage_accum.clone(),
                     error: None,
                     reasoning_content: None,
+                    truncated_by_max_tokens: false,
                 }
             }
         }
@@ -401,6 +408,7 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
                 usage: None,
                 error: None,
                 reasoning_content: None,
+                truncated_by_max_tokens: false,
             }
         }
     }
@@ -503,4 +511,32 @@ mod tests {
         );
         assert!(!chunk2.finished, "无终态 finish_reason 的普通 delta 不应终结");
     }
+
+    /// finish_reason=length:max_tokens 截断仍按正常终止收敛(finished=true),
+    /// 但截断标志必须随 chunk 带出(零信号会让腰斩报告冒充完整完成);
+    /// stop 等正常终态不置标志。
+    #[test]
+    fn finish_reason_length_carries_truncation_flag() {
+        let mut accum = None;
+        let chunk = apply_openai_sse(
+            r#"{"choices":[{"delta":{"content":"报告写到一半"},"finish_reason":"length"}]}"#,
+            &mut accum,
+        );
+        assert!(chunk.finished, "length 属正常终止");
+        assert!(chunk.truncated_by_max_tokens, "length 必须带截断标志");
+
+        let mut accum2 = None;
+        let chunk2 = apply_openai_sse(
+            r#"{"choices":[{"delta":{"content":"ok"},"finish_reason":"stop"}]}"#,
+            &mut accum2,
+        );
+        assert!(chunk2.finished && !chunk2.truncated_by_max_tokens, "stop 不置截断标志");
+
+        let mut accum3 = None;
+        let chunk3 = apply_openai_sse(
+            r#"{"choices":[{"delta":{"content":"x"},"finish_reason":null}]}"#,
+            &mut accum3,
+        );
+        assert!(!chunk3.truncated_by_max_tokens, "普通 delta 不置截断标志");
+    }
 }
diff --git a/crates/df-ai/src/retry.rs b/crates/df-ai/src/retry.rs
index c502359..762773f 100644
--- a/crates/df-ai/src/retry.rs
+++ b/crates/df-ai/src/retry.rs
@@ -5,7 +5,7 @@
 //! 但重试能在不侵入上层的前提下吸收瞬态错误，对低频后台调用零风险。
 //!
 //! 策略:
-//! - 最多 3 次尝试(初次 + 2 次重试)，指数退避 1s → 2s → 4s + ±20% jitter
+//! - 最多 3 次尝试(初次 + 2 次重试)，指数退避 1s → 2s → 4s + ±50% jitter
 //! - 总挂钟预算 30s(含 sleep + 各次请求耗时)，超预算直接放弃重试返回最后错误
 //! - 仅对可重试错误重试: timeout / connect / 5xx / 429; 不重试 4xx(非429) / 参数错 / 鉴权错
 //! - 不引入 reqwest-retry / tower: complete 路径简单，手写 loop 零依赖、可调试、不与 SSE/双层 Semaphore 耦合
@@ -34,6 +34,11 @@ const BASE_BACKOFF_SECS: u64 = 1;
 /// 与单请求 60s timeout 配合: 最坏 3 次 × 60s = 180s，30s 预算会在此之前主动止损。
 const MAX_TOTAL_BUDGET: Duration = Duration::from_secs(30);
 
+/// 单次退避基准上限(毫秒)，与 [`MAX_TOTAL_BUDGET`] 同口径: 超预算的退避没有意义，
+/// attempt 走深时不再指数膨胀。与总预算的一致性由测试 `backoff_cap_matches_budget` 守住。
+/// (同构先例: df-tunnel 的 backoff_delay 也以 RECONNECT_MAX 封顶。)
+const MAX_BACKOFF_MS: u64 = 30_000;
+
 /// jitter 上限(相对 base 的 ±比例)。避免重试风暴对齐。
 /// CR-XX: 范围扩到 ±50%(由 gen_range(-0.5..0.5) × JITTER_RATIO=1.0 合成)。
 const JITTER_RATIO: f64 = 1.0;
@@ -65,20 +70,45 @@ pub fn is_status_retryable(status: u16) -> bool {
     status == 429 || (500..600).contains(&status)
 }
 
+/// 退避倍率 `2^(attempt-1)`，位移溢出安全。
+///
+/// 替代裸 `1u64 << (attempt - 1)`: attempt > 64 时位移越界(debug panic / release 环绕)、
+/// attempt = 0 时下溢 panic。现 attempt = 0 视作 1(倍率 1)，attempt - 1 ≥ 64 时
+/// checked_shl 返回 None，倍率封顶 u64::MAX，最终时长由 [`MAX_BACKOFF_MS`] 兜底。
+/// 全输入无 panic，序列单调不减。
+fn backoff_multiplier(attempt: u32) -> u64 {
+    match 1u64.checked_shl(attempt.saturating_sub(1)) {
+        Some(m) => m,
+        None => u64::MAX,
+    }
+}
+
+/// 退避基准毫秒(未加 jitter)，`BASE_BACKOFF_SECS * 2^(attempt-1) * 1000`，封顶 [`MAX_BACKOFF_MS`]。
+///
+/// 乘法全程 saturating: 倍率封顶 u64::MAX 后 `* 1000` 也会二次溢出，不能裸乘。
+/// 单一事实源——backoff_delay 与单测的 base 断言都走这里，不再各写一份公式。
+fn base_backoff_ms(attempt: u32) -> u64 {
+    BASE_BACKOFF_SECS
+        .saturating_mul(backoff_multiplier(attempt))
+        .saturating_mul(1000)
+        .min(MAX_BACKOFF_MS)
+}
+
 /// 指数退避 + jitter: 返回第 `attempt`(1-based)次重试前应 sleep 的时长。
 /// attempt=1 → ~1s, attempt=2 → ~2s, attempt=3 → ~4s，各 ±50% jitter。
 ///
 /// jitter 用 `rand::thread_rng().gen_range(-0.5..0.5)` 生成 ±50% 比例，避免多客户端同步重试风暴。
 /// 以毫秒粒度计算后向下取整(避免秒级截断把 0.9s 砍成 0)。
+/// 全输入安全: attempt 任意 u32(含 0 / >64)不 panic，结果 ≤ [`MAX_BACKOFF_MS`]。
 ///
 /// 暴露 pub 供 src-tauri/agentic.rs 流前重试复用
 /// ("复用 retry.rs backoff_delay 退避 1s→2s→4s+jitter"),避免重写退避逻辑。
 pub fn backoff_delay(attempt: u32) -> Duration {
-    let base_ms = BASE_BACKOFF_SECS.saturating_mul(1u64 << (attempt - 1)) * 1000;
+    let base_ms = base_backoff_ms(attempt) as f64;
     // ±50% jitter，相对 base 时长的浮动比例
     let jitter_ratio = rand::thread_rng().gen_range(-0.5..0.5); // [-0.5, 0.5)
     let factor = 1.0 + jitter_ratio * JITTER_RATIO;
-    let ms = (base_ms as f64 * factor).max(0.0) as u64;
+    let ms = (base_ms * factor).clamp(0.0, MAX_BACKOFF_MS as f64) as u64;
     Duration::from_millis(ms)
 }
 
@@ -160,12 +190,12 @@ mod tests {
 
     #[test]
     fn backoff_is_monotonic_within_jitter() {
-        // base 序列: 1, 2, 4 秒; jitter ±20% 后仍在 [0.8,4.8] 范围
+        // base 序列: 1, 2, 4 秒; jitter ±50%(JITTER_RATIO=1.0) 后仍在 [0.5x, 1.5x] 范围
         for attempt in 1..=3 {
             let d = backoff_delay(attempt);
-            let base_ms = (BASE_BACKOFF_SECS * (1u64 << (attempt - 1))) * 1000;
-            let lo = (base_ms as f64 * (1.0 - JITTER_RATIO)) as u64;
-            let hi = (base_ms as f64 * (1.0 + JITTER_RATIO)) as u64;
+            let base_ms = base_backoff_ms(attempt);
+            let lo = (base_ms as f64 * (1.0 - JITTER_RATIO / 2.0)) as u64;
+            let hi = (base_ms as f64 * (1.0 + JITTER_RATIO / 2.0)) as u64;
             assert!(
                 d.as_millis() as u64 >= lo && d.as_millis() as u64 <= hi,
                 "attempt={} delay={:?} 应在 [{},{}]ms",
@@ -177,6 +207,64 @@ mod tests {
         }
     }
 
+    /// 倍率溢出安全: attempt=0 不下溢，≥65 封顶 u64::MAX，序列单调不减
+    #[test]
+    fn backoff_multiplier_extremes_safe() {
+        assert_eq!(backoff_multiplier(0), 1, "attempt=0 视作 1，不下溢 panic");
+        assert_eq!(backoff_multiplier(1), 1);
+        assert_eq!(backoff_multiplier(2), 2);
+        assert_eq!(backoff_multiplier(3), 4);
+        assert_eq!(backoff_multiplier(63), 1u64 << 62);
+        assert_eq!(backoff_multiplier(64), 1u64 << 63, "u64 可表达的最大位移");
+        assert_eq!(backoff_multiplier(65), u64::MAX, "checked_shl 越界封顶");
+        assert_eq!(backoff_multiplier(100), u64::MAX);
+        assert_eq!(backoff_multiplier(u32::MAX), u64::MAX);
+        let seq: Vec<u64> = [0u32, 1, 2, 3, 10, 63, 64, 65, 100, u32::MAX]
+            .iter()
+            .map(|&a| backoff_multiplier(a))
+            .collect();
+        assert!(seq.windows(2).all(|w| w[0] <= w[1]), "倍率应单调不减: {:?}", seq);
+    }
+
+    /// 基准毫秒单调不减、封顶值正确(2^5×1s=32s 已超 30s 预算，从 attempt=6 起封顶)
+    #[test]
+    fn base_backoff_ms_monotonic_and_capped() {
+        let mut prev = 0u64;
+        for attempt in [0u32, 1, 2, 3, 4, 5, 6, 10, 63, 64, 65, 100, u32::MAX] {
+            let b = base_backoff_ms(attempt);
+            assert!(b >= prev, "attempt={} base_ms={} 应 ≥ 前值 {}", attempt, b, prev);
+            assert!(b <= MAX_BACKOFF_MS, "attempt={} base_ms={} 应 ≤ {}", attempt, b, MAX_BACKOFF_MS);
+            prev = b;
+        }
+        assert_eq!(base_backoff_ms(1), 1000);
+        assert_eq!(base_backoff_ms(2), 2000);
+        assert_eq!(base_backoff_ms(3), 4000);
+        assert_eq!(base_backoff_ms(5), 16_000, "未触及封顶");
+        assert_eq!(base_backoff_ms(6), MAX_BACKOFF_MS, "32s 超预算封顶 30s");
+        assert_eq!(base_backoff_ms(u32::MAX), MAX_BACKOFF_MS);
+    }
+
+    /// 极端 attempt 下 backoff_delay 不 panic(修复: 1u64<<(attempt-1) 曾 debug panic/release 环绕)
+    #[test]
+    fn backoff_delay_extreme_attempts_no_panic_and_capped() {
+        for attempt in [0u32, 1, 10, 63, 64, 65, 100, u32::MAX] {
+            let ms = backoff_delay(attempt).as_millis() as u64;
+            assert!(
+                ms <= MAX_BACKOFF_MS,
+                "attempt={} 实际 {}ms 应 ≤ {}ms(含 jitter 后仍不超)",
+                attempt,
+                ms,
+                MAX_BACKOFF_MS
+            );
+        }
+    }
+
+    /// 封顶常量与总预算同口径(改预算时两处同步)
+    #[test]
+    fn backoff_cap_matches_budget() {
+        assert_eq!(MAX_BACKOFF_MS, MAX_TOTAL_BUDGET.as_millis() as u64);
+    }
+
     #[test]
     fn status_classification() {
         assert!(is_status_retryable(429));
diff --git a/crates/df-ai/src/router.rs b/crates/df-ai/src/router.rs
index c3d9ae8..df03dcc 100644
--- a/crates/df-ai/src/router.rs
+++ b/crates/df-ai/src/router.rs
@@ -163,6 +163,7 @@ mod tests {
             intelligence: IntelligenceTier::Standard,
             weight: 50,
             context_window: 8192,
+            temperature_supported: true,
             probe_source: None,
         }
     }
diff --git a/src-tauri/Cargo.toml b/src-tauri/Cargo.toml
index a55483a..b13cbab 100644
--- a/src-tauri/Cargo.toml
+++ b/src-tauri/Cargo.toml
@@ -39,6 +39,9 @@ async-trait = { workspace = true }
 # 零新依赖树)。rust-work 改代理端口只改网关映射表,工具代码不感知端口。
 axum = { version = "0.7" }
 
+# 托盘菜单 i18n:系统 locale 判定(零传递依赖,判定一次缓存)
+sys-locale = "0.3"
+
 # 后端 crate
 df-types = { path = "../crates/df-types" }
 df-storage = { path = "../crates/df-storage" }
diff --git a/src-tauri/src/commands/ai/remote_bridge.rs b/src-tauri/src/commands/ai/remote_bridge.rs
index e7854d6..a20215c 100644
--- a/src-tauri/src/commands/ai/remote_bridge.rs
+++ b/src-tauri/src/commands/ai/remote_bridge.rs
@@ -51,7 +51,7 @@
 
 use serde::Deserialize;
 use serde_json::Value;
-use tauri::{AppHandle, Emitter, State};
+use tauri::{AppHandle, State};
 
 use crate::state::AppState;
 
diff --git a/src-tauri/src/commands/ai/stream_recv.rs b/src-tauri/src/commands/ai/stream_recv.rs
index bb7f340..b5e4546 100644
--- a/src-tauri/src/commands/ai/stream_recv.rs
+++ b/src-tauri/src/commands/ai/stream_recv.rs
@@ -102,6 +102,9 @@ pub(crate) enum StreamResult {
         usage: df_ai::provider::TokenUsage,
         /// DeepSeek thinking 模式推理内容（需回传到下一轮请求）
         reasoning_content: Option<String>,
+        /// 流因 provider max_tokens 上限截断(finish_reason=length / stop_reason=max_tokens)。
+        /// true=文本/tool_calls 可能不完整,调用方据此标记该轮产出截断,不冒充正常完成。
+        truncated_by_max_tokens: bool,
     },
     /// 流中途断（MidStream chunk Err / idle timeout / provider stream-error / 有 partial_text
     /// 但流尽未收到 finished）。已 emit 过 AiTextDelta（前端 currentText 已累积），
@@ -282,6 +285,9 @@ pub(crate) async fn stream_llm(
     let mut finished_received = false;
     let mut stopped = false;
     let mut final_usage: Option<df_ai::provider::TokenUsage> = None;
+    // max_tokens 截断标志跨帧 OR 累积(openai finish_reason=length / anthropic
+    // stop_reason=max_tokens 的帧先于终态帧到达),Complete 随旗标带出供上层标记截断。
+    let mut truncated_by_max_tokens = false;
     // DeepSeek thinking 模式推理内容累积（多轮需回传）
     let mut reasoning_content_acc: Option<String> = None;
     // per-conv delta 合批累加器。pending_delta 累积未 flush 的文本,
@@ -419,6 +425,9 @@ pub(crate) async fn stream_llm(
                         if let Some(u) = &chunk.usage {
                             final_usage = Some(u.clone());
                         }
+                        if chunk.truncated_by_max_tokens {
+                            truncated_by_max_tokens = true;
+                        }
                         if let Some(ref rc) = chunk.reasoning_content {
                             reasoning_content_acc.get_or_insert_with(String::new).push_str(rc);
                         }
@@ -540,6 +549,7 @@ pub(crate) async fn stream_llm(
             tool_calls: tool_calls_acc,
             usage: final_usage.unwrap_or_default(),
             reasoning_content: reasoning_content_acc,
+            truncated_by_max_tokens,
         };
     }
 
@@ -573,6 +583,7 @@ pub(crate) async fn stream_llm(
         tool_calls: tool_calls_acc,
         usage: final_usage.unwrap_or_default(),
         reasoning_content: reasoning_content_acc,
+        truncated_by_max_tokens,
     }
 }
 
diff --git a/src-tauri/src/commands/ai/tools/git.rs b/src-tauri/src/commands/ai/tools/git.rs
index 05e9e3d..e9784bc 100644
--- a/src-tauri/src/commands/ai/tools/git.rs
+++ b/src-tauri/src/commands/ai/tools/git.rs
@@ -156,9 +156,13 @@ async fn run_git_diff(working_dir: &str, staged: bool) -> serde_json::Value {
     let mut patch_args = vec!["diff"];
     if staged { patch_args.push("--cached"); }
     let patch_raw = exec_git_exit(working_dir, &patch_args).await.0;
-    // 截断到 8000 字符(防大体量 diff)
+    // 截断到 8000 字符(防大体量 diff);中文字节切片撞 UTF-8 边界会 panic,回退到字符边界
     let patch_truncated = if patch_raw.len() > 8000 {
-        format!("{}\n... (diff 截断,共 {} 字符)", &patch_raw[..8000], patch_raw.len())
+        let mut end = 8000;
+        while !patch_raw.is_char_boundary(end) {
+            end -= 1;
+        }
+        format!("{}\n... (diff 截断,共 {} 字符)", &patch_raw[..end], patch_raw.len())
     } else {
         patch_raw
     };
diff --git a/src-tauri/src/commands/idea.rs b/src-tauri/src/commands/idea.rs
index 49cd5e9..ade336a 100644
--- a/src-tauri/src/commands/idea.rs
+++ b/src-tauri/src/commands/idea.rs
@@ -313,7 +313,8 @@ pub async fn evaluate_idea(
 ///
 /// - 成功的灵感返回更新后的记录，失败的灵感记录错误信息，不中断后续评估。
 /// - provider 构造仅一次（批量场景减少重复初始化开销）。
-/// - 返回 BatchEvalResult { success, errors }，前端据此展示部分成功/失败。
+// 2026-10-04: 未在 lib.rs 注册 IPC(dead_code),保留作批量评估参考实现;复用下方 evaluate_single 共用内部函数。
+#[allow(dead_code)]
 #[tauri::command]
 pub async fn evaluate_ideas_batch(
     state: State<'_, AppState>,
diff --git a/src-tauri/src/tray.rs b/src-tauri/src/tray.rs
index 96c745a..bdbe334 100644
--- a/src-tauri/src/tray.rs
+++ b/src-tauri/src/tray.rs
@@ -4,6 +4,9 @@
 //! - 托盘菜单「显示主窗口」唤起 / 「退出」app.exit(0) 真正结束
 //! - TrayHandle 经 app.manage 保活,防托盘图标被 GC 移除
 //! - P1 bg 化:运行态菜单项 + tooltip 随 ConvState 迁移刷新(guard.transition 挂钩 update_tray_runtime)
+//! - i18n:文案按系统 locale 取中/英双语静态映射(locale 判定一次缓存),无 Rust i18n 机制的最小实现
+use std::sync::OnceLock;
+
 use tauri::{
     AppHandle, Manager,
     menu::{Menu, MenuItem},
@@ -12,6 +15,45 @@ use tauri::{
 
 use crate::state::AppState;
 
+/// 托盘文案双语映射(系统 locale 判定一次,进程内缓存;判定失败回退中文——应用主语言)。
+/// 菜单项与运行态文本分两处消费,映射同源防两处漂移。
+struct TrayTexts {
+    status_idle: &'static str,
+    status_running: &'static str,
+    show: &'static str,
+    quit: &'static str,
+    tooltip_running: &'static str,
+}
+
+/// 按系统 locale 取托盘文案(判定一次缓存)。zh* → 中文,其余(含判定失败)→ 中文/英文
+/// 各取其一:locale 明确非 zh 用英文,判定失败(None)回退中文(应用主语言)。
+fn tray_texts() -> &'static TrayTexts {
+    static TEXTS: OnceLock<TrayTexts> = OnceLock::new();
+    TEXTS.get_or_init(|| {
+        let zh = sys_locale::get_locale()
+            .map(|l| l.to_ascii_lowercase().starts_with("zh"))
+            // None = 系统locale不可得,回退中文(应用默认语言),与首次安装体验一致
+            .unwrap_or(true);
+        if zh {
+            TrayTexts {
+                status_idle: "无运行中会话",
+                status_running: "运行中 {} 个会话",
+                show: "显示主窗口",
+                quit: "退出",
+                tooltip_running: "DevFlow (运行中 {} 个会话)",
+            }
+        } else {
+            TrayTexts {
+                status_idle: "No running sessions",
+                status_running: "{} session(s) running",
+                show: "Show Main Window",
+                quit: "Quit",
+                tooltip_running: "DevFlow ({} session(s) running)",
+            }
+        }
+    })
+}
+
 /// managed state 保活，防 GC 移除托盘图标。
 /// tray 持 TrayIcon 本体(保活为主,update_tray_runtime 借用刷 tooltip);
 /// status 为运行态纯展示项句柄,供 update_tray_runtime 刷文本。
@@ -24,10 +66,11 @@ pub struct TrayHandle {
 
 /// 创建托盘:图标 + 菜单(运行态/显示主窗口/退出)。
 pub fn setup_tray(app: &AppHandle) -> tauri::Result<()> {
+    let t = tray_texts();
     // 运行态纯展示项:enabled=false 置灰不响应点击,文本由 update_tray_runtime 维护
-    let status = MenuItem::with_id(app, "status", "无运行中会话", false, None::<&str>)?;
-    let show = MenuItem::with_id(app, "show", "显示主窗口", true, None::<&str>)?;
-    let quit = MenuItem::with_id(app, "quit", "退出", true, None::<&str>)?;
+    let status = MenuItem::with_id(app, "status", t.status_idle, false, None::<&str>)?;
+    let show = MenuItem::with_id(app, "show", t.show, true, None::<&str>)?;
+    let quit = MenuItem::with_id(app, "quit", t.quit, true, None::<&str>)?;
     let menu = Menu::with_items(app, &[&status, &show, &quit])?;
     // 复用窗口图标(tauri.conf.json bundle.icon 已配置),缺则 panic 提示配置
     let icon = app
@@ -60,14 +103,15 @@ pub fn update_tray_runtime(app: &AppHandle) {
     else {
         return;
     };
+    let t = tray_texts();
     let n = state.conv_states.active_convs().len();
     let text = if n > 0 {
-        format!("运行中 {} 个会话", n)
+        t.status_running.replace("{}", &n.to_string())
     } else {
-        "无运行中会话".to_string()
+        t.status_idle.to_string()
     };
     let tooltip = if n > 0 {
-        format!("DevFlow (运行中 {} 个会话)", n)
+        t.tooltip_running.replace("{}", &n.to_string())
     } else {
         "DevFlow (MCP server running)".to_string()
     };
diff --git a/src/App.vue b/src/App.vue
index dcbdc81..d6ca16b 100644
--- a/src/App.vue
+++ b/src/App.vue
@@ -157,6 +157,10 @@ function initTheme() {
 // UI 按钮/审批激活仍用 togglePanel 二态开关,语义互不干扰
 function handleKeydown(e: KeyboardEvent) {
   if ((e.ctrlKey || e.metaKey) && e.key === 'i') {
+    // 输入焦点守卫(2026-10-04 回归修复):输入控件内 Ctrl+I 让路(打字/组合键不切面板),
+    // 对齐 AiChat onGlobalKeydown Esc 分支的 closest target 守卫写法
+    const el = e.target as HTMLElement | null
+    if (el?.closest?.('input, textarea, [contenteditable]')) return
     e.preventDefault()
     aiStore.cyclePanelMode()
   }
diff --git a/src/components/ai/ChatInput.vue b/src/components/ai/ChatInput.vue
index bc5dcf5..f2d20f7 100644
--- a/src/components/ai/ChatInput.vue
+++ b/src/components/ai/ChatInput.vue
@@ -15,6 +15,8 @@
       />
       <!-- F-260614-05 Phase 2b: 待发送图片预览 -->
       <ImageInput :pending-images="pendingImages" @remove="removePendingImage" />
+      <!-- B1: 当前模型不看图时贴图警示(发送仍允许,后端温和降级丢图;chip 持续可见防误判网络问题) -->
+      <div v-if="showImageIgnoreWarn" class="ai-img-warn-chip">{{ t('aiChat.imageIgnoredWarn') }}</div>
       <textarea
         ref="inputEl"
         v-model="inputText"
@@ -304,6 +306,22 @@ const stopBtnState = computed<'idle' | 'stop' | 'stopping' | 'retry'>(() => {
   return textActive ? 'stop' : 'idle'
 })
 
+// ── B1: 非视觉模型贴图前置感知 ──
+// 有效模型判定镜像 TopBar 口径(active provider → enabled 池 → override 命中优先,否则权重最高)。
+// 池未知(未拉取/空)时保守视为支持(不误伤;后端 model_supports_vision 有 model_probe 兜底,
+// 发图照旧温和降级)。发送仍允许(与现状兼容),仅以警示 chip 持续可见防「静默丢图」误判网络问题。
+const currentModelSupportsVision = computed<boolean>(() => {
+  const p = store.state.providers.find(x => x.id === store.state.activeProvider)
+    || store.state.providers.find(x => x.is_default)
+    || store.state.providers[0]
+  if (!p?.model_configs?.length) return true
+  const models = p.model_configs.filter(m => m.enabled)
+  const chosen = (store.modelOverride.value ? models.find(m => m.model_id === store.modelOverride.value) : undefined) || models[0]
+  if (!chosen) return true
+  return (chosen.modalities ?? []).includes('vision')
+})
+const showImageIgnoreWarn = computed(() => pendingImages.value.length > 0 && !currentModelSupportsVision.value)
+
 // ── 技能 `/` 联想 ──
 const skillOpen = ref(false)
 const skillIndex = ref(0)
@@ -849,6 +867,12 @@ async function handleSend() {
   const skill = pendingSkill.value
   // 无文本、无图片、未选技能时忽略；选了技能允许空文本（纯技能调用）
   if (!text && !skill && pendingImages.value.length === 0) return
+  // B2: 纯发图(无文本无技能)+ 非视觉模型 → 后端会拒绝(占位串不喂模型不浪费一轮),
+  // 前端先行拦截给 i18n 提示省一轮 IPC;图文混合(有文本)不受限,维持占位后缀现状
+  if (!text && !skill && pendingImages.value.length > 0 && !currentModelSupportsVision.value) {
+    emit('error', { msg: t('aiChat.imageOnlyBlocked'), type: 'warning' })
+    return
+  }
   // F-260614-05 Phase 2b: 快照待发送图片,构造 ContentPart[]([Text?, Image...])。
   // 后端 ai_chat_send IPC 当前不接 parts(Phase 2c 接入);本轮 parts 仅写本地 push 的
   // user 消息渲染多模态图,后端请求仍走 content 文本路径。
@@ -1181,4 +1205,11 @@ defineExpose({
 }
 
 /* ── ⑥.4 Phase 4: @项目展开摘要样式已随 EnrichmentPanel.vue 提取 ── */
+
+/* B1: 非视觉模型贴图警示条(待发送图片预览行下方,warning 色轻量提示,不抢输入框注意力) */
+.ai-img-warn-chip {
+  font-size: 10.5px;
+  color: var(--df-warning);
+  padding: 1px 4px;
+}
 </style>
diff --git a/src/components/settings/ConnectionPanel.vue b/src/components/settings/ConnectionPanel.vue
index 9dd8564..4fb4102 100644
--- a/src/components/settings/ConnectionPanel.vue
+++ b/src/components/settings/ConnectionPanel.vue
@@ -34,7 +34,8 @@
   <section class="panel" v-if="connForm.visible">
     <div class="panel-header">
       <h2>{{ connForm.editId ? $t('settings.editConnectionTitle') : $t('settings.addConnectionTitle') }}</h2>
-      <button class="btn btn-ghost btn-sm" @click="connForm.visible = false">{{ $t('common.cancel') }}</button>
+      <!-- L19 对齐: 取消走 closeConnForm,脏表单二次确认后再关(防误弃改动) -->
+      <button class="btn btn-ghost btn-sm" @click="closeConnForm">{{ $t('common.cancel') }}</button>
     </div>
     <div class="form-grid">
       <div class="form-field">
@@ -77,7 +78,7 @@
 </template>
 
 <script setup lang="ts">
-import { reactive, ref, onMounted } from 'vue'
+import { reactive, ref, computed, onMounted } from 'vue'
 import { useI18n } from 'vue-i18n'
 import { useAppSettingsStore } from '@/stores/appSettings'
 
@@ -120,6 +121,8 @@ const connForm = reactive({
   port: DEFAULT_PORTS.mysql,
   user: '',
   password: '',
+  // L19 对齐: 表单打开/保存成功时的「已落库快照」JSON 串,用于 isConnDirty 比对(同 ProviderPanel 范式)
+  savedSnapshot: '' as string,
 })
 
 // 异步操作禁用态(防双击重复提交):saveConn 保存连接期间禁用按钮
@@ -146,6 +149,34 @@ function typeLabel(type: string) {
   return map[type] || type
 }
 
+// L19 对齐: 脏表单二次确认——关闭前与已落库快照比对,有改动先 confirmDialog 防误关丢配置
+// (password 计入快照:输入了未保存的密钥同样算脏)
+function connBusinessSnapshot(): string {
+  return JSON.stringify({
+    name: connForm.name,
+    type: connForm.type,
+    host: connForm.host,
+    port: connForm.port,
+    user: connForm.user,
+    password: connForm.password,
+  })
+}
+
+const isConnDirty = computed(() => {
+  if (!connForm.visible) return false
+  return connBusinessSnapshot() !== connForm.savedSnapshot
+})
+
+/** 关闭表单入口(模板「取消」用):脏表单二次确认后再关,防误弃改动 */
+async function closeConnForm() {
+  if (isConnDirty.value) {
+    const ok = await props.confirmDialog(t('settings.confirmDiscardUnsaved'))
+    if (!ok) return  // 用户取消关闭,继续编辑
+  }
+  connForm.visible = false
+  connForm.savedSnapshot = ''
+}
+
 function openConnForm(conn?: ConnRecord) {
   if (conn) {
     connForm.editId = conn.id
@@ -165,6 +196,8 @@ function openConnForm(conn?: ConnRecord) {
     connForm.password = ''
   }
   connForm.visible = true
+  // 记初始快照(visible=true、字段回填完成后),isConnDirty 据此比对
+  connForm.savedSnapshot = connBusinessSnapshot()
 }
 
 /** 用户手动切换连接类型时,端口自动重置为该类型默认端口。
@@ -202,7 +235,10 @@ function saveConn() {
       connections.value.push(record)
     }
     persistConnections()
+    // 成功明确反馈:关表单 + 清快照(刚落库无脏,直接关不走二次确认)+ 绿色「已保存」toast(对齐 ProviderPanel)
     connForm.visible = false
+    connForm.savedSnapshot = ''
+    emit('toast', t('settings.toastSaved'), 'success')
   } finally {
     submitting.value = false
   }
diff --git a/src/components/settings/ProviderPanel.vue b/src/components/settings/ProviderPanel.vue
index 2555ceb..7f21a0c 100644
--- a/src/components/settings/ProviderPanel.vue
+++ b/src/components/settings/ProviderPanel.vue
@@ -521,6 +521,12 @@ async function fetchModels() {
       apiKey: providerForm.apiKey,
       providerId: providerForm.editId || null,
     })
+    // 空列表守卫:拉到 0 个模型时保留现有模型表格(不覆盖),仅提示——
+    // 与后端 ai_fetch_models 空守卫同语义(空结果不清用户配置),防误清后顺手保存落库
+    if (models.length === 0) {
+      emit('toast', t('settings.toastFetchEmpty'), 'warning')
+      return
+    }
     // 手动优先:探测结果合并已存在的 UserSet 标注——用户手动改过的能力不被「测试连接」覆盖
     // (对齐后端 ai_fetch_models 合并逻辑:probe_source==UserSet 保留手动能力维度)。
     // enabled/weight 同理保留旧值:重拉只是刷新能力维度,用户已关掉的模型/调过的权重
@@ -599,6 +605,20 @@ async function saveProvider() {
   } else {
     providerForm.temperature = null
   }
+  // 模型权重钳制 [0,100](同温度范式:原生 min/max 只约束步进按钮,手输越界照存)。
+  // 越界/小数 → clamp+取整;清空残留非数值(v-model.number 得 '')→ 归一 50(对齐负载池 NaN 兜底)。
+  // 多行聚合一次 toast,逐行报会刷屏。
+  let weightClamped = false
+  for (const m of providerForm.models) {
+    if (typeof m.weight === 'number' && Number.isFinite(m.weight)) {
+      const c = Math.min(100, Math.max(0, Math.trunc(m.weight)))
+      if (c !== m.weight) { m.weight = c; weightClamped = true }
+    } else {
+      m.weight = 50
+      weightClamped = true
+    }
+  }
+  if (weightClamped) emit('toast', t('settings.toastWeightClamped'), 'warning')
   providerForm.saving = true
   try {
     await aiApi.saveProvider({
@@ -672,7 +692,9 @@ async function onPoolToggle(p: AiProviderConfig, enabled: boolean) {
   const prev = p.enabled
   p.enabled = enabled
   try {
-    await aiApi.updateProviderPool(p.id, enabled, p.weight ?? 50)
+    // 防御钳制:p.weight 理论上已被 onPoolWeightChange 钳过,此处兜底防脏数据直传 IPC
+    const w = Math.min(100, Math.max(0, p.weight ?? 50))
+    await aiApi.updateProviderPool(p.id, enabled, w)
     emit('toast', t('settings.toastPoolUpdateOk'), 'info')
   } catch (e) {
     p.enabled = prev  // revert
diff --git a/src/i18n/en/settings.ts b/src/i18n/en/settings.ts
index f69c4b3..6606eaa 100644
--- a/src/i18n/en/settings.ts
+++ b/src/i18n/en/settings.ts
@@ -23,6 +23,7 @@ export default {
     // Out-of-range clamp notices (concurrency 1-64 + per-conv ≤ global; temperature reports the corrected value)
     toastConcurrencyClamped: 'Value out of range, corrected automatically (concurrency 1-{max}; per-conversation capped at global)',
     toastTempClamped: 'Sampling temperature must be within 0-2, corrected to {v}',
+    toastWeightClamped: 'Model weight must be within 0-100; out-of-range values corrected (cleared inputs default to 50)',
     // Prompt caching toggle save feedback (previously silent)
     toastAnthropicCacheSaved: 'Prompt caching setting saved',
     toastAnthropicCacheFail: 'Failed to save the prompt caching setting',
@@ -88,6 +89,7 @@ export default {
     saveModels: 'Save model config',
     modelListUnsavedHint: 'Model list has unsaved changes (enabled / weight) — closing or navigating away will discard them',
     toastFetchOk: 'Fetched {count} models',
+    toastFetchEmpty: 'Fetched 0 models — existing config kept',
     toastFetchFail: 'Fetch failed: {msg}',
     fetchFailedTimeout: 'Fetch timed out, check that the Base URL is reachable: {msg}',
     fetchFailedAuth: 'Authentication failed, check your API Key: {msg}',
diff --git a/src/i18n/zh-CN/settings.ts b/src/i18n/zh-CN/settings.ts
index 36e5819..ebb8121 100644
--- a/src/i18n/zh-CN/settings.ts
+++ b/src/i18n/zh-CN/settings.ts
@@ -23,6 +23,7 @@ export default {
     // 走查 P3-14 越界钳制提示(并发 1-64 + 单对话不超全局;采样温度单独提示修正值)
     toastConcurrencyClamped: '输入超出有效范围，已自动修正（并发 1-{max}，单对话不超过全局）',
     toastTempClamped: '采样温度须在 0-2 范围内，已修正为 {v}',
+    toastWeightClamped: '模型权重须在 0-100 范围内，越界值已修正（清空视为 50）',
     // 走查 P3-11:Prompt 缓存开关保存反馈(原切换零反馈)
     toastAnthropicCacheSaved: 'Prompt 缓存设置已保存',
     toastAnthropicCacheFail: 'Prompt 缓存设置保存失败',
@@ -87,6 +88,7 @@ export default {
     saveModels: '保存模型配置',
     modelListUnsavedHint: '模型列表有未保存的调整（启用 / 权重），关闭或切走将丢弃',
     toastFetchOk: '已拉取 {count} 个模型',
+    toastFetchEmpty: '拉取到 0 个模型，已保留现有配置',
     toastFetchFail: '拉取模型失败：{msg}',
     fetchFailedTimeout: '拉取超时，请检查 Base URL 是否可达：{msg}',
     fetchFailedAuth: '鉴权失败，请检查 API Key：{msg}',
diff --git a/src/stores/appSettings.ts b/src/stores/appSettings.ts
index f139045..6dc3d86 100644
--- a/src/stores/appSettings.ts
+++ b/src/stores/appSettings.ts
@@ -3,6 +3,7 @@
 //! 后端 `app_settings` 表的 value 永远是 JSON 字符串。本层职责:
 //! - 缓存已解析(反序列化)值,按 key 索引,组件同步读
 //! - 写操作立即更新缓存(乐观),并按 key debounce ~300ms 合并连续快速写
+//! - 落库失败保留 pending 脏值待下次写入重试,经 lastWriteError 透出(见 set/flush)
 //! - 组合式 `useSetting` 让组件像用 ref 一样用 v-model 绑定某个 key
 //!
 //! 注意:此为「KV 偏好持久化」层,与 AI 提供商/连接配置(store 层)是两件事 ——
@@ -24,6 +25,11 @@ const pendingTimers = new Map<string, ReturnType<typeof setTimeout>>()
 /// 载荷队列:debounce 触发时取最新值(而非定时器创建时的快照)
 const pendingValues = new Map<string, unknown>()
 
+/// 最近一次落库失败(key + 错误),成功落库同一 key 后清 null。
+/// store 层无组件上下文,不直接弹 toast,对齐 knowledge.ts 的 state.error 范式,
+/// 上层 watch 此状态做行内红字/toast 通报。
+const lastWriteError = ref<{ key: string; error: unknown } | null>(null)
+
 /** 启动时一次性拉回全部 key/value 并解析填充缓存(App.vue 调用) */
 async function loadAll(): Promise<void> {
   const all = await settingsApi.getAll()
@@ -41,6 +47,9 @@ function get<T>(key: string, defaultValue: T): T {
 /**
  * 写入某 key:立即更新缓存(乐观,视图无延迟),再 debounce 后落库。
  * 同 key 连续快速写(如拖拽/输入)在窗口内合并,只把最后一次值 JSON.stringify 发后端。
+ *
+ * promise 契约:resolve 仅代表"已调度落库",不代表已落库;本 promise 永不
+ * reject(调用方 void/await 均安全),落库成败经 lastWriteError 透出。
  */
 async function set(key: string, value: unknown): Promise<void> {
   cache[key] = value
@@ -51,18 +60,35 @@ async function set(key: string, value: unknown): Promise<void> {
 
   const timer = setTimeout(() => {
     pendingTimers.delete(key)
-    const latest = pendingValues.get(key)
-    pendingValues.delete(key)
-    // 落库失败仅记日志:缓存已是最新值,降级为内存态,避免回滚造成 UI 与库不一致
-    settingsApi
-      .set(key, JSON.stringify(latest))
-      .catch((e) => console.error(`[appSettings] 写入 ${key} 失败:`, e))
+    void flush(key)
   }, SET_DEBOUNCE_MS)
   pendingTimers.set(key, timer)
 }
 
-/** 删除某 key:清缓存 + 取消 pending 写 + 调后端删除 */
+/**
+ * 把 pending 中该 key 的最新值落库。失败时保留 pending 脏值不清除,
+ * 下次该 key 任何 set 会合并进同一队列自然重试;cache 保持乐观值不回跳
+ * (避免落库失败把用户已输入的视图打回去),失败经 lastWriteError 与
+ * console.error 双通道透出,状态不再"假成功"。
+ */
+async function flush(key: string): Promise<void> {
+  const latest = pendingValues.get(key)
+  if (latest === undefined) return
+  try {
+    await settingsApi.set(key, JSON.stringify(latest))
+    // 成功才清 pending;若落库期间又有新写合并进来,保留新值交由新定时器落库
+    if (Object.is(pendingValues.get(key), latest)) pendingValues.delete(key)
+    if (lastWriteError.value?.key === key) lastWriteError.value = null
+  } catch (e) {
+    console.error(`[appSettings] 写入 ${key} 失败,已标记待重试(下次改动该 key 时重试):`, e)
+    lastWriteError.value = { key, error: e }
+  }
+}
+
+/** 删除某 key:清缓存 + 取消 pending 写 + 调后端删除;删除失败恢复内存态后原样上抛 */
 async function remove(key: string): Promise<void> {
+  const prevCache = cache[key]
+  const prevPending = pendingValues.get(key)
   delete cache[key]
   pendingValues.delete(key)
   const t = pendingTimers.get(key)
@@ -70,7 +96,15 @@ async function remove(key: string): Promise<void> {
     clearTimeout(t)
     pendingTimers.delete(key)
   }
-  await settingsApi.delete(key)
+  try {
+    await settingsApi.delete(key)
+  } catch (e) {
+    // 删除失败:磁盘仍有该值,恢复内存态保持两端一致,避免重启后该 key 幽灵复活而视图已丢
+    if (prevCache !== undefined) cache[key] = prevCache
+    if (prevPending !== undefined) pendingValues.set(key, prevPending)
+    lastWriteError.value = { key, error: e }
+    throw e
+  }
 }
 
 /**
@@ -149,6 +183,7 @@ function parse(raw: string): unknown {
 export function useAppSettingsStore() {
   return {
     cache,
+    lastWriteError,
     loadAll,
     get,
     set,