优化: 设置域体验完善

- 校验失败红框定位字段,保存反馈改持久化成功后
- 数值越界钳制,双提示队列化不再重叠
- 搜索防抖跳分类,托盘气泡文案进多语言

crates/df-ai/src/coordinator.rs

@@ -508,14 +508,50 @@ impl Coordinator {
                     });
                 }
 
+                // 记录最近一次 JoinError 详情:JoinError 不携带 subtask 身份,
+                // 缺失任务靠循环结束后按 id 对账补齐,错误文本由此带出
+                let mut join_err_msg: Option<String> = None;
+
                 while let Some(join_result) = join_set.join_next().await {
                     match join_result {
                         Ok(result) => {
                             results_map.insert(result.subtask_id.clone(), result);
                         }
                         Err(e) => {
+                            // panic 与取消分开留痕:panic 是执行体自身崩溃需排查,取消是外部干预
+                            if e.is_panic() {
                                 tracing::error!("[COORDINATOR] 子任务 panic: {}", e);
+                            } else {
+                                tracing::warn!("[COORDINATOR] 子任务被取消: {}", e);
                             }
+                            join_err_msg = Some(e.to_string());
+                        }
+                    }
+                }
+
+                // panic/取消的子任务不能凭空消失:返回契约是与 plan.tasks 一一对应,
+                // 缺失会被下游误读为"无此任务";此处按该任务失败语义补失败结果收敛,
+                // merge 按 success 过滤自然跳过,不扩散中断整层
+                for task in layer {
+                    if !results_map.contains_key(&task.id) {
+                        tracing::error!(
+                            subtask = %task.id,
+                            "[COORDINATOR] 子任务执行体崩溃,按失败结果收敛(不扩散中断)"
+                        );
+                        results_map.insert(
+                            task.id.clone(),
+                            ExecutionResult {
+                                subtask_id: task.id.clone(),
+                                persona_id: task_persona
+                                    .get(&task.id)
+                                    .cloned()
+                                    .unwrap_or_default(),
+                                output: join_err_msg
+                                    .clone()
+                                    .unwrap_or_else(|| "子任务 panic/被取消".to_string()),
+                                success: false,
+                            },
+                        );
                     }
                 }
             } else {
@@ -1415,6 +1451,39 @@ mod tests {
         assert_eq!(results[1].subtask_id, "b");
     }
 
+    #[tokio::test]
+    async fn par_05_panicked_task_converges_to_failed_result() {
+        // 同层并行中某子任务 panic:JoinSet 返回 JoinError(panic 变体),
+        // 应收敛为该任务的失败结果,而不是让它从结果列表凭空消失
+        let coord = make_coord();
+        let plan = Plan::from_tasks(vec![
+            SubTask::new("t1", "task1"),
+            SubTask::new("t2", "task2"),
+            SubTask::new("t3", "task3"),
+        ]);
+
+        let results = coord
+            .dispatch_with_budget(&plan, |task, persona_id: String| async move {
+                if task.id == "t2" {
+                    panic!("模拟子任务执行体崩溃");
+                }
+                ExecutionResult {
+                    subtask_id: task.id,
+                    persona_id,
+                    output: format!("done {}", task.intent),
+                    success: true,
+                }
+            }, None)
+            .await;
+
+        assert_eq!(results.len(), 3, "panic 子任务也应占一个结果位");
+        let t2 = results.iter().find(|r| r.subtask_id == "t2").expect("t2 应有失败结果");
+        assert!(!t2.success, "panic 子任务应收敛为失败而非成功");
+        assert!(!t2.output.is_empty(), "失败结果应带 JoinError 详情");
+        assert!(results.iter().filter(|r| r.subtask_id != "t2").all(|r| r.success),
+            "其余子任务不受 panic 扩散影响");
+    }
+
     // -- Plan-driven LLM 规划开关 --
 
     #[test]

crates/df-ai/src/lib.rs

@@ -134,6 +134,10 @@ pub(crate) fn send_error_class(p: SendErrPreds) -> &'static str {
 /// 其余（openai / glm / deepseek 等 OpenAI 兼容）走 OpenAICompatProvider。
 /// 调用方（AI Chat 的 run_agentic_loop、df-nodes 的 AiNode）统一引用此工厂，
 /// 新增 provider 只改这一处。
+///
+/// 未知 provider_type 无法走显式报错（工厂签名不动，调用方横跨 src-tauri/df-nodes），
+/// 兜底 OpenAI 兼容协议但 warn 留痕，不再静默——协议错配的请求失败可溯源到配置项。
+/// anthropic_family 口径与 prompt.rs cache_split / chat.rs needs_url_prefetch 对齐。
 pub fn build_provider(
     provider_type: &str,
     base_url: &str,
@@ -141,11 +145,22 @@ pub fn build_provider(
     model: &str,
 ) -> Box<dyn LlmProvider> {
     match provider_type {
-        "anthropic" => Box::new(anthropic_compat::AnthropicCompatProvider::new(
+        "anthropic" | "anthropic_compat" => Box::new(
+            anthropic_compat::AnthropicCompatProvider::new(base_url, api_key, model),
+        ),
+        "openai_compat" | "openai" | "glm" | "deepseek" => {
+            Box::new(openai_compat::OpenAICompatProvider::new(
                 base_url, api_key, model,
-        )),
-        _ => Box::new(openai_compat::OpenAICompatProvider::new(
+            ))
+        }
+        unknown => {
+            tracing::warn!(
+                "未知 provider_type {:?}，按 OpenAI 兼容协议兜底；请检查 ai_providers 配置的 provider_type",
+                unknown
+            );
+            Box::new(openai_compat::OpenAICompatProvider::new(
                 base_url, api_key, model,
-        )),
+            ))
+        }
     }
 }

crates/df-ai/src/model_probe.rs

@@ -333,6 +333,36 @@ mod tests {
         assert_eq!(m.cost_tier, CostTier::Medium);
     }
 
+    // ── 启发式:版本号后缀不误判 Vision ──
+
+    #[test]
+    fn heuristic_version_suffix_no_vision() {
+        // "-v" 后接数字属版本号(deepseek-v3 为纯文本模型),不得误判 Vision
+        for name in ["deepseek-v3", "acme-chat-v2", "acme-v12"] {
+            let m = probe(name);
+            assert_eq!(m.probe_source, Some(ProbeSource::Heuristic));
+            assert!(
+                !m.modalities.contains(&Modality::Vision),
+                "{name} 版本号后缀不应误判 Vision: {:?}",
+                m.modalities
+            );
+        }
+    }
+
+    #[test]
+    fn heuristic_v_morpheme_tail_still_vision() {
+        // v 词素尾仍判视觉:独立 v 段(qwen-v-plus 形态)与段尾 v(glm-4v 形态,假名避开预设表)
+        for name in ["acme-chat-v", "acme-4v-plus"] {
+            let m = probe(name);
+            assert_eq!(m.probe_source, Some(ProbeSource::Heuristic));
+            assert!(
+                m.modalities.contains(&Modality::Vision),
+                "{name} 段尾 v 词素应判 Vision: {:?}",
+                m.modalities
+            );
+        }
+    }
+
     // ── 启发式:默认(Standard/Medium/Text/ToolUse) ──
 
     #[test]

crates/df-ai/src/model_probe_helpers.rs

@@ -112,8 +112,11 @@ pub(super) fn heuristic_infer(model_id: &str) -> ModelConfig {
 
 // —— 词素判定助手(命名约定式启发式,集中维护便于增删) ——
 
-/// Vision 词素:`-v`(尾缀如 glm-4v)、以 `v` 结尾、含 `vl` / `vision`。
-/// 不直接 `contains('v')`(避免误伤 review/verbal 等)。
+/// Vision 词素:`vision` / `vl` 子串、`v` 词素尾(任一分隔段以 v 结尾:glm-4v / qwen-v-plus)。
+/// 裸 "v" 必须是词素尾:`contains("-v")` 会把版本号误判成视觉(deepseek-v3 纯文本,
+/// "-v" 后接数字属版本号而非视觉词素),按段尾判定后 v3/v2 不命中、4v/v 段仍命中。
 pub(super) fn has_vision_token(name: &str) -> bool {
-    name.contains("vision") || name.contains("vl") || name.ends_with('v') || name.contains("-v")
+    name.contains("vision")
+        || name.contains("vl")
+        || name.split(['-', '.', '_']).any(|seg| seg.ends_with('v'))
 }

crates/df-ai/src/namespace_store.rs

@@ -81,6 +81,10 @@ impl NamespaceStore {
     }
 
     /// 存入 namespace。返回引用路径字符串。
+    ///
+    /// 超限行为分两档：常规超限 LRU 淘汰旧条目腾位（既有语义）；单条本身超过
+    /// max_bytes 时淘汰清空也放不下，拒绝入库只返回引用路径——展开点对
+    /// read_only()=None 统一落 EVICTED_PLACEHOLDER，与 LRU 淘汰后的不可解析引用同语义。
     pub fn store(&mut self, tool_name: &str, content: &str) -> String {
         let key = simple_hash(content);
         let path = format!("{}{}/{}", NAMESPACE_REF_PREFIX, tool_name, key);
@@ -93,8 +97,15 @@ impl NamespaceStore {
             return path;
         }
 
-        // 检查容量，超限淘汰
         let new_bytes = content.len();
+
+        // 单条超过总容量：硬入库会把 current_bytes 永久顶破上限（evict 清空后仍无条件
+        // insert 的旧路径），拒绝入库保住 current_bytes <= max_bytes 不变式
+        if new_bytes > self.max_bytes {
+            return path;
+        }
+
+        // 检查容量，超限淘汰
         self.evict_if_needed(new_bytes);
 
         let entry = NamespaceEntry {
@@ -268,6 +279,23 @@ mod tests {
         }
     }
 
+    #[test]
+    fn store_oversized_single_entry_rejected_cap_intact() {
+        // 回归:单条超过 max_bytes 时旧路径淘汰清空后仍无条件入库,current_bytes
+        // 永久突破上限。现拒绝入库,仅返回引用路径,展开点落 EVICTED_PLACEHOLDER。
+        let mut ns = NamespaceStore::new(100);
+        let path = ns.store("t1", &"x".repeat(200));
+        assert_eq!(ns.len(), 0, "超容量单条应拒绝入库");
+        assert_eq!(ns.current_bytes, 0, "字节账目不得突破上限");
+        assert!(is_namespace_ref(&path), "返回值仍是引用路径");
+        assert_eq!(ns.read_only(&path), None, "不可解析 → 既有淘汰占位语义");
+
+        // 拒绝后不污染状态,常规条目仍可正常入
+        let p2 = ns.store("t2", "ok");
+        assert_eq!(ns.read(&p2).map(str::to_owned), Some("ok".to_string()));
+        assert_eq!(ns.current_bytes, 2);
+    }
+
     #[test]
     fn clear_resets_all() {
         let mut ns = NamespaceStore::new(100_000);

crates/df-ai/src/openai_compat.rs

@@ -310,7 +310,15 @@ impl OpenAICompatProvider {
             .into_iter()
             .enumerate()
             .map(|(i, c)| {
+                let missing_id = c.id.is_empty();
                 let id = tool_call_id_or_fallback(&c.id, i, "gen_tool");
+                if missing_id {
+                    tracing::warn!(
+                        function = %c.function.name,
+                        fallback_id = %id,
+                        "[OpenAI] 同步响应 tool_call 缺/空 id,已生成 fallback id(防结果路由串号)"
+                    );
+                }
                 ToolCall::new(id, c.function.name, c.function.arguments)
             })
             .collect()
@@ -1293,6 +1301,26 @@ mod tests {
         assert_eq!(ids.len(), unique.len(), "id 应全部唯一，实际 {:?}", ids);
     }
 
+    /// 同步响应 tool_call 缺 id/arguments 字段:serde default 兜底(空串/空串),
+    /// 缺 id 经 parse_tool_calls 生成 fallback。旧行为:id 必填,单条缺字段令整个
+    /// 响应 serde 失败 → complete 报"响应解析失败"整批丢弃。
+    #[test]
+    fn openai_sync_tool_call_missing_fields_default_instead_of_failing() {
+        let calls: Vec<OpenAiToolCallResp> = serde_json::from_str(
+            r#"[{"type":"function","function":{"name":"list_dir"}}]"#,
+        )
+        .expect("缺 id/arguments 的 tool_call 应可反序列化,不再整响应解析失败");
+        let parsed = OpenAICompatProvider::parse_tool_calls(calls);
+        assert_eq!(parsed.len(), 1);
+        assert!(
+            parsed[0].id.starts_with("gen_tool_"),
+            "缺 id → fallback id, got: {}",
+            parsed[0].id
+        );
+        assert_eq!(parsed[0].function.name, "list_dir");
+        assert_eq!(parsed[0].function.arguments, "", "缺 arguments 缺省空串");
+    }
+
     /// CR-空 id 流式：SSE chunk 携带 `"id":""`（SenseNova 兼容缺陷）→ ToolCallDelta.id
     /// 转为 `gen_stream_{n}` fallback（n 取自全局计数器，跨轮跨 assistant 递增，非 None），
     /// 保证下游 accumulate_tool_calls 写入 draft.id 非空。chunk 完全无 id 字段（None）保持

crates/df-ai/src/openai_helpers.rs

@@ -102,10 +102,14 @@ pub(crate) struct OpenAiMessageResp {
 
 #[derive(Debug, Deserialize)]
 pub(crate) struct OpenAiToolCallResp {
+    /// 缺省容错:个别兼容端点同步响应省略 id → 缺省空串,由 parse_tool_calls 的
+    /// tool_call_id_or_fallback 生成唯一 fallback。此前必填,单条 tool_call 缺 id 令
+    /// 整个响应 serde 失败 → complete 报"响应解析失败"整批丢弃(流式路径早已容错)。
+    #[serde(default)]
     pub id: String,
     // 反序列化 #[serde(rename="type")] 字段,保留以对齐 OpenAI 响应结构,标注意图消除 dead_code warning
     #[allow(dead_code)]
-    #[serde(rename = "type")]
+    #[serde(rename = "type", default)]
     pub call_type: String,
     pub function: OpenAiFunctionResp,
 }
@@ -113,6 +117,9 @@ pub(crate) struct OpenAiToolCallResp {
 #[derive(Debug, Deserialize)]
 pub(crate) struct OpenAiFunctionResp {
     pub name: String,
+    /// 缺省容错:arguments 缺失 → 空串(对齐 anthropic 同步 tool_use input 缺失 args="" 口径),
+    /// 防单条缺字段令整个响应解析失败;name 保持必填(无名 tool_call 无法派发,无安全缺省)。
+    #[serde(default)]
     pub arguments: String,
 }
 
@@ -130,8 +137,13 @@ pub(crate) struct PromptTokensDetails {
 /// 嵌套 cached_tokens 回填,消费方一律读归一后的扁平字段。
 #[derive(Debug, Deserialize)]
 pub(crate) struct OpenAiUsageRaw {
+    /// 三个计数缺省 0:部分兼容网关 usage 只报部分字段(如仅 total_tokens),必填会让
+    /// 整帧 serde 失败走空 chunk 分支 — 同帧 content 一并蒸发、usage 归零无痕。
+    #[serde(default)]
     pub prompt_tokens: u32,
+    #[serde(default)]
     pub completion_tokens: u32,
+    #[serde(default)]
     pub total_tokens: u32,
     /// DeepSeek 扩展:缓存命中 token(低价,deepseek-chat/reasoner)。无此字段 → default 0。
     #[serde(default)]
@@ -288,6 +300,14 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
             }
             // 提取 usage（带 include_usage 时末段 chunk 携带，覆盖累积）
             if let Some(u) = chunk.usage {
+                // 缺省计数的留痕:两个分项均 0 基本可断定 provider 上报不完整(真实完成
+                // 必有 completion 计数),非流路径下游拿到零 usage 无从察觉。
+                if u.prompt_tokens == 0 && u.completion_tokens == 0 {
+                    tracing::warn!(
+                        total = u.total_tokens,
+                        "[OpenAI] usage 缺 prompt/completion 分项(按 0 兜底),provider 上报不完整"
+                    );
+                }
                 tracing::info!(
                     prompt = u.prompt_tokens,
                     completion = u.completion_tokens,
@@ -316,6 +336,21 @@ pub(crate) fn apply_openai_sse(data: &str, usage_accum: &mut Option<TokenUsage>)
                 );
 
                 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
+                    // 冲突时缺省会互相踩(下游按 index 累积 → 参数拼错位),warn 留痕定位
+                    // provider 兼容缺陷,不做静默改写。
+                    if tcs.len() > 1 {
+                        let mut idxs: Vec<u32> = tcs.iter().map(|tc| tc.index).collect();
+                        idxs.sort_unstable();
+                        idxs.dedup();
+                        if idxs.len() < tcs.len() {
+                            tracing::warn!(
+                                calls = tcs.len(),
+                                unique = idxs.len(),
+                                "[OpenAI] 同帧多 tool_call 增量 index 冲突(缺省 0 互踩),按 index 累积将错位"
+                            );
+                        }
+                    }
                     tcs.into_iter()
                         .map(|tc| {
                             // CR-空 id：流式 chunk 的 id 可能为 Some("")（SenseNova 兼容缺陷）。
@@ -423,6 +458,35 @@ mod tests {
         assert_eq!(usage.prompt_cache_miss_tokens, 36);
     }
 
+    /// 部分兼容网关 usage 只报部分字段:缺省 0 入累加器,同帧 content 不再因整帧
+    /// serde 失败被吞(旧行为:必填字段缺失 → 空 chunk → 文本蒸发 + usage 归零无痕)
+    #[test]
+    fn sse_partial_usage_frame_parses_with_defaults() {
+        let mut accum = None;
+        let data = r#"{"choices":[{"delta":{"content":"hi"}}],"usage":{"total_tokens":10}}"#;
+        let c = apply_openai_sse(data, &mut accum);
+        assert_eq!(c.delta, "hi", "usage 缺分项不应令整帧解析失败丢 content");
+        let u = accum.expect("部分 usage 应写入累加器");
+        assert_eq!(u.total_tokens, 10);
+        assert_eq!(u.prompt_tokens, 0, "缺省分项按 0 兜底");
+        assert_eq!(u.completion_tokens, 0);
+    }
+
+    /// 同帧多 tool_call 增量缺 index:缺省 0 触发 index 冲突路径(仅 warn 留痕,不改写
+    /// 缺省策略);解析本身不失败,增量仍逐条带出。
+    #[test]
+    fn sse_tool_calls_missing_index_in_multi_call_frame_still_parse() {
+        let mut accum = None;
+        let data = r#"{"choices":[{"delta":{"tool_calls":[
+            {"id":"call_a","function":{"name":"ta","arguments":"{"}},
+            {"id":"call_b","function":{"name":"tb","arguments":"}"}}
+        ]}}]}"#;
+        let c = apply_openai_sse(data, &mut accum);
+        let tcs = c.tool_calls.expect("缺 index 的多 call 帧应可解析");
+        assert_eq!(tcs.len(), 2);
+        assert!(tcs.iter().all(|tc| tc.index == 0), "缺 index 缺省 0");
+    }
+
     /// content_filter 属终态:敏感内容拦截的流部分 provider 不发 [DONE] 帧,
     /// 必须靠 finish_reason 收敛,否则悬挂;未列入终态的 finish_reason 保持非终态
     #[test]

crates/df-ai/src/plan_executor.rs

@@ -2,7 +2,8 @@
 //!
 //! 按 Plan::to_layers 输出的层级顺序执行子任务：
 //! - 层间串行：上层全部完成才进下一层（deps 保证）
-//! - 层内并行：同层 SubTask 用 tokio::JoinSet 并发执行
+//! - 层内并行：同层 SubTask 用 tokio::JoinSet 并发执行（设计目标；
+//!   当前为串行循环模拟，见 execute 内 TODO，本执行器未接生产链路）
 //!
 //! **当前状态**：骨架实现，PLAN_EXECUTION_ENABLED 门控（默认关）。
 //! 未接入 agentic loop 主路径，翻 true 启用后仍需与 run_agentic_loop 对接。
@@ -16,7 +17,9 @@ use crate::planner::{Plan, SubTask};
 /// Plan 执行总开关。false = 不启用（走单链 ReAct 旧行为）。
 ///
 /// 运行时原子门控,支持热切换(通过 IPC 或前端设置开关)。
-/// 默认关,翻 true 后 process_tool_calls 内以 JoinSet 并行执行同层工具。
+/// 默认关。现状:PlanExecutor::execute 零生产调用点(仅测试),本执行器未接入主链;
+/// agentic loop 读本开关走的是 coordinator::dispatch(Phase 1 串行桩),与本文件无关。
+/// JoinSet 真并行留 Phase 3 对接时接通(见 execute 内 TODO)。
 /// 与 PLANNING_ENABLED 解耦:plan_hint 独立使能,编排可见性始终可开。
 use std::sync::atomic::{AtomicBool, Ordering};
 pub static PLAN_EXECUTION_ENABLED: AtomicBool = AtomicBool::new(false);
@@ -87,16 +90,16 @@ impl PlanExecutor {
                 layer.len()
             );
 
-            // 层内并行：JoinSet 收集
-            let _join_set: tokio::task::JoinSet<SubTaskResult> = tokio::task::JoinSet::new();
             let summary_clone = parent_summary.clone();
 
             for task in layer {
                 let task = Arc::new(task.clone());
                 let summary = summary_clone.clone();
-                // 注意:run_subtask 是 FnMut,不能直接 move 进 JoinSet(多任务并发调用)。
-                // Phase 3 对接时,run_subtask 改为 Arc<Fn> 或 channel 模式。
-                // 当前骨架用串行执行模拟(层内不并行),验证分层逻辑正确性。
+                // TODO(Phase 3 真并行): 当前为串行循环模拟并行(假并行),仅验证分层逻辑。
+                // run_subtask 是 FnMut,不能直接 move 进 JoinSet(多任务并发调用);
+                // 接真并行的改法:run_subtask 换 Arc<Fn> 或 channel 模式,
+                // 本层全部 spawn 进 JoinSet 统一收集后再汇总 parent_summary,替换此串行 await。
+                // 现状无用户可见影响:本执行器未接生产链路(execute 零调用点)。
                 let result = run_subtask(task.clone(), summary).await;
                 let sr = match result {
                     Ok(output) => SubTaskResult {

src-tauri/src/commands/ai/audit/record.rs

@@ -10,11 +10,12 @@
 //! 依赖 audit/utils.rs 的 `truncate_chars` 做参数/结果截断，通过 `super::truncate_chars` 引用。
 
 use serde::{Deserialize, Serialize};
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use tauri::State;
 
 use df_ai::ai_tools::RiskLevel;
-use df_storage::crud::{AiToolExecutionRepo, AuditQuery};
+use df_storage::crud::{AiConversationRepo, AiToolExecutionRepo, AuditQuery};
+use df_storage::db::Database;
 use df_storage::models::AiToolExecutionRecord;
 use df_types::types::new_id;
 
@@ -250,6 +251,128 @@ pub async fn list_tool_executions(
     })
 }
 
+// ============================================================
+// S4 TaskDrawer 关联会话反查(按 task_id 推导发起会话,只读)
+// ============================================================
+
+/// 任务工具精确名单(与 tools/task.rs 声明的三个任务工具一一对应)。
+///
+/// 审计表无"task_id → conversation"映射列,主链路靠本名单行 + result JSON 反推;
+/// SQL 与本名单必须同步维护(见下方 SELECT 注释)。
+const TASK_CONV_TOOLS: [&str; 3] = ["create_task", "update_task", "advance_task"];
+
+/// 关联会话跳转条目 DTO(传给前端 TaskDrawer「关联会话」info-item)。
+#[derive(Debug, Clone, Serialize)]
+pub struct TaskConversationRef {
+    pub conversation_id: String,
+    /// 会话标题(ai_conversations.title,懒创建未命名时 None,前端回退显 id)
+    pub title: Option<String>,
+    /// 最近一次命中命中的工具名(create_task/update_task/advance_task)
+    pub tool_name: String,
+    /// 最近一次命中时间(毫秒字符串,同库 requested_at 口径)
+    pub requested_at: String,
+}
+
+/// 单行审计记录命中判定(纯函数):工具在名单内 且 result JSON 顶层 id 与 task_id 精确判等。
+///
+/// 禁 LIKE:任务 id 做子串匹配会误命中(如 "t1" 命中 "t12"),必须解析 JSON 取 id 字段判等;
+/// result 非 JSON / 缺 id 字段 / id 不等 / 工具不在名单 → 一律不命中(failed 行的
+/// `{"error":...}` 与 rejected 行的人话文本由此自然被排除)。
+fn match_task_audit_row(tool_name: &str, result: Option<&str>, task_id: &str) -> bool {
+    if !TASK_CONV_TOOLS.contains(&tool_name) {
+        return false;
+    }
+    result
+        .and_then(|r| serde_json::from_str::<serde_json::Value>(r).ok())
+        .and_then(|v| v.get("id").and_then(|id| id.as_str()).map(|s| s.to_string()))
+        .map(|id| id == task_id)
+        .unwrap_or(false)
+}
+
+/// 按 task_id 反查发起过任务工具的会话(供 IPC 与单测共用,不依赖 AppState)。
+///
+/// 数据源 ai_tool_executions 审计表(V9):每笔任务工具执行都记 conversation_id + result。
+/// repo 现有查询只支持 status/risk/tool_keyword(LIKE)维度,表达不了 `tool_name IN 精确名单`,
+/// 故走专用 SELECT(与 state.rs 清理孤儿 pending 消息同款 app 层直查先例);SQL 里的工具名单
+/// 是 TASK_CONV_TOOLS 的硬拷贝,单测覆盖三个名字防漂移。
+///
+/// 过滤:① 命中行按会话去重(行按 requested_at 倒序,首见=最近一次);② 子代理会话
+/// (origin='sub_agent',V44)剔除,v1 只列主会话;③ 会话已不在库的 id 一并剔除
+/// (switchConversation 对不存在 id 会走"新建兜底",跳过去是空会话,不如不列)。
+/// LIMIT 1000 扫描上界:抽屉每次打开都查,封顶防超大审计表全表拖慢;取最近 1000 行,
+/// 更早的关联对"跳回最近上下文"场景无意义。
+pub(crate) async fn find_task_conversations(
+    db: &Database,
+    task_id: &str,
+) -> Result<Vec<TaskConversationRef>, String> {
+    let conn = db.conn();
+    let rows: Vec<(String, String, Option<String>, String)> =
+        tokio::task::spawn_blocking(move || -> Result<Vec<(String, String, Option<String>, String)>, String> {
+        let guard = conn.blocking_lock();
+        let mut stmt = guard
+            .prepare(
+                // 工具名单硬拷贝自 TASK_CONV_TOOLS(三个名字),改动须两处同步
+                "SELECT conversation_id, tool_name, result, requested_at FROM ai_tool_executions \
+                 WHERE tool_name IN ('create_task','update_task','advance_task') \
+                 AND conversation_id IS NOT NULL \
+                 ORDER BY CAST(requested_at AS INTEGER) DESC LIMIT 1000",
+            )
+            .map_err(|e| e.to_string())?;
+        let mapped = stmt
+            .query_map([], |row| {
+                Ok((
+                    row.get::<_, String>(0)?,
+                    row.get::<_, String>(1)?,
+                    row.get::<_, Option<String>>(2)?,
+                    row.get::<_, String>(3)?,
+                ))
+            })
+            .map_err(|e| e.to_string())?;
+        let mut out = Vec::new();
+        for r in mapped {
+            out.push(r.map_err(|e| e.to_string())?);
+        }
+        Ok(out)
+    })
+    .await
+    .map_err(|e| format!("find_task_conversations join 失败: {}", e))??;
+
+    let conv_repo = AiConversationRepo::new(db);
+    let mut visited: HashSet<String> = HashSet::new();
+    let mut out = Vec::new();
+    for (conv_id, tool_name, result, requested_at) in rows {
+        if !match_task_audit_row(&tool_name, result.as_deref(), task_id) {
+            continue;
+        }
+        // 去重 + 防同会话重复查 origin(首个可见行即该会话最近一次命中)
+        if !visited.insert(conv_id.clone()) {
+            continue;
+        }
+        if let Ok(Some(rec)) = conv_repo.get_by_id(&conv_id).await {
+            // origin=None(主会话)才列;查无记录的会话同样不列(见函数注释③)
+            if rec.origin.is_none() {
+                out.push(TaskConversationRef {
+                    conversation_id: conv_id,
+                    title: rec.title,
+                    tool_name,
+                    requested_at,
+                });
+            }
+        }
+    }
+    Ok(out)
+}
+
+/// TaskDrawer「关联会话」查询 IPC:按 task_id 反查发起过 create/update/advance_task 的主会话,
+/// 按 最近命中 倒序去重返回。只读,无副作用。
+#[tauri::command]
+pub async fn list_task_conversations(
+    state: State<'_, AppState>,
+    task_id: String,
+) -> Result<Vec<TaskConversationRef>, String> {
+    find_task_conversations(&state.db, &task_id).await
+}
+
 // ============================================================
 // AC-5 按工具失败率统计(诊断 IPC,替代 ad-hoc 查库)
 // ============================================================
@@ -421,6 +544,110 @@ mod tests {
         assert_eq!(out.stats[0].tool_name, "read_file");
     }
 
+    /// S4 关联会话反查:插入一条指定 conv/tool/result/requested_at 的审计行(测试构造数据用)。
+    async fn insert_conv_audit(
+        repo: &AiToolExecutionRepo,
+        conv: &str,
+        tool: &str,
+        result: Option<String>,
+        t: i64,
+    ) {
+        repo.insert(AiToolExecutionRecord {
+            id: new_id(),
+            conversation_id: Some(conv.to_string()),
+            message_id: None,
+            tool_call_id: new_id().into(),
+            tool_name: tool.to_string(),
+            arguments: "{}".to_string(),
+            result,
+            status: "completed".to_string(),
+            risk_level: "low".to_string(),
+            requested_at: t.to_string(),
+            executed_at: Some(t.to_string()),
+            decided_by: Some("auto".to_string()),
+        })
+        .await
+        .expect("测试数据插入应成功");
+    }
+
+    /// S4 关联会话反查:插入一条指定 origin 的会话记录(V44 origin 列,测试过滤子会话用)。
+    async fn insert_conversation(db: &Database, id: &str, title: Option<&str>, origin: Option<&str>) {
+        use df_storage::models::AiConversationRecord;
+        AiConversationRepo::new(db)
+            .insert(AiConversationRecord {
+                id: id.to_string(),
+                title: title.map(|s| s.to_string()),
+                messages: "[]".to_string(),
+                provider_id: None,
+                model: None,
+                models: None,
+                archived: false,
+                pinned: false,
+                prompt_tokens: None,
+                completion_tokens: None,
+                pinned_goals: None,
+                pending_approvals: None,
+                created_at: "1".to_string(),
+                updated_at: "1".to_string(),
+                origin: origin.map(|s| s.to_string()),
+            })
+            .await
+            .expect("测试会话插入应成功");
+    }
+
+    /// S4 关联会话反查三件事:① result JSON 顶层 id 精确判等(禁 LIKE,前缀 id 不误命中);
+    /// ② 同会话多行命中按最近一次去重 + 行按 requested_at 倒序;③ origin='sub_agent' 子会话
+    /// 与已不在库的会话 id 剔除,只列主会话。SQL 三个工具名与 TASK_CONV_TOOLS 同步覆盖。
+    #[tokio::test]
+    async fn find_task_conversations_exact_id_dedup_origin_filter() {
+        let db = Database::open_in_memory().await.expect("in-memory db 初始化失败");
+        let repo = AiToolExecutionRepo::new(&db);
+        let t = 2_000_000_000_000i64;
+
+        insert_conversation(&db, "conv-main", Some("主会话"), None).await;
+        insert_conversation(&db, "conv-sub", Some("子代理"), Some("sub_agent")).await;
+
+        // conv-main:create(t-30) + update(t-10) 两次命中 → 去重后取 update(最近)
+        insert_conv_audit(&repo, "conv-main", "create_task",
+            Some(r#"{"id":"t1","title":"A","status":"todo"}"#.into()), t - 30).await;
+        insert_conv_audit(&repo, "conv-main", "update_task",
+            Some(r#"{"id":"t1","field":"title","updated":true}"#.into()), t - 10).await;
+        // conv-sub:advance 命中但 origin='sub_agent' → 过滤
+        insert_conv_audit(&repo, "conv-sub", "advance_task",
+            Some(r#"{"id":"t1","status":"in_progress"}"#.into()), t - 20).await;
+        // 干扰项:非名单工具 / id 前缀(t11 ≠ t1)/ 缺 id / 非法 JSON / 其他任务
+        insert_conv_audit(&repo, "conv-main", "list_tasks",
+            Some(r#"{"id":"t1"}"#.into()), t - 5).await;
+        insert_conv_audit(&repo, "conv-main", "update_task",
+            Some(r#"{"id":"t12","field":"title","updated":true}"#.into()), t - 6).await;
+        insert_conv_audit(&repo, "conv-main", "advance_task",
+            Some(r#"{"error":"boom"}"#.into()), t - 7).await;
+        insert_conv_audit(&repo, "conv-main", "create_task",
+            Some("审批超时,自动取消".into()), t - 8).await;
+        insert_conv_audit(&repo, "conv-main", "create_task",
+            Some(r#"{"id":"t2","title":"B"}"#.into()), t - 9).await;
+        insert_conv_audit(&repo, "conv-ghost", "create_task",
+            Some(r#"{"id":"t1","title":"C"}"#.into()), t - 4).await; // 会话已不在库
+
+        let out = find_task_conversations(&db, "t1").await.unwrap();
+        assert_eq!(out.len(), 1, "只应命中主会话 conv-main(去重后单条)");
+        assert_eq!(out[0].conversation_id, "conv-main");
+        assert_eq!(out[0].title.as_deref(), Some("主会话"));
+        assert_eq!(out[0].tool_name, "update_task", "同会话两行命中取最近一次(update)");
+        assert_eq!(out[0].requested_at, (t - 10).to_string());
+    }
+
+    /// S4 纯函数判定表:名单外工具 / 非 JSON / 缺 id / id 不等 均不命中,精确判等才命中。
+    #[test]
+    fn match_task_audit_row_truth_table() {
+        assert!(match_task_audit_row("create_task", Some(r#"{"id":"t1"}"#), "t1"));
+        assert!(!match_task_audit_row("create_task", Some(r#"{"id":"t12"}"#), "t1"), "禁 LIKE:前缀不误命中");
+        assert!(!match_task_audit_row("list_tasks", Some(r#"{"id":"t1"}"#), "t1"), "名单外工具不命中");
+        assert!(!match_task_audit_row("create_task", Some("人话文本"), "t1"), "非 JSON 不命中");
+        assert!(!match_task_audit_row("create_task", Some(r#"{"error":"x"}"#), "t1"), "缺 id 不命中");
+        assert!(!match_task_audit_row("create_task", None, "t1"), "result 空不命中");
+    }
+
     /// from 时间过滤:仅统计 requested_at >= from 的记录。
     ///
     /// patch_file 追加一条更早(early)的 completed → from=late 时其不计入,

src-tauri/src/commands/ai/commands/conversation.rs

@@ -604,6 +604,42 @@ pub async fn ai_conversation_load_more(
     }))
 }
 
+/// 只读窥视子代理轨迹(S3 子代理轨迹查看)
+///
+/// 复用 ai_conversation_switch 的读路径(ai_messages.list_by_conversation → record_to_message,
+/// 含空表回退旧 messages JSON 列的兼容分支),但**只读不切状态**:不写 activeConversationId/
+/// per_conv/pending_approvals/标题重生成。子会话(ai_conversations.origin='sub_agent')不进侧栏、
+/// 不能走 switch(切走会把父会话 active 顶掉并触发一整套会话态副作用),前端完成卡浮层
+/// 查看轨迹的唯一合理通道就是本只读命令。
+///
+/// 返回全量消息不分页:子代理单任务轨迹体量有限,分页属过度设计(设计宁少勿预留)。
+#[tauri::command]
+pub async fn ai_conversation_peek(
+    state: State<'_, AppState>,
+    conversation_id: String,
+) -> Result<serde_json::Value, String> {
+    let record = state.ai_conversations.get_by_id(&conversation_id).await
+        .map_err(err_str)?
+        .ok_or_else(|| format!("对话不存在: {}", conversation_id))?;
+    // 读路径与 switch 同构:优先 ai_messages 表,空表回退旧 messages JSON 列(双向兼容)
+    let all_records = state.ai_messages.list_by_conversation(&conversation_id).await
+        .map_err(err_str)?;
+    let messages: Vec<ChatMessage> = if !all_records.is_empty() {
+        all_records.iter().map(record_to_message).collect()
+    } else if record.messages != "[]" && !record.messages.is_empty() {
+        serde_json::from_str(&record.messages)
+            .map_err(|e| format!("解析消息失败: {}", e))?
+    } else {
+        Vec::new()
+    };
+    let messages_json = serde_json::to_string(&messages)
+        .map_err(|e| format!("序列化消息失败: {}", e))?;
+    Ok(serde_json::json!({
+        "title": record.title,
+        "messages": messages_json,
+    }))
+}
+
 /// 删除对话
 #[tauri::command]
 pub async fn ai_conversation_delete(

src-tauri/src/lib.rs

@@ -458,6 +458,8 @@ pub fn run() {
             commands::ai::ai_conversation_create,
             commands::ai::ai_conversation_list,
             commands::ai::ai_conversation_switch,
+            // S3 子代理轨迹只读窥视(复用 switch 读路径,不切任何会话状态)
+            commands::ai::ai_conversation_peek,
             commands::ai::ai_conversation_load_more,
             commands::ai::ai_conversation_delete,
             commands::ai::ai_conversation_rename,
@@ -480,6 +482,8 @@ pub fn run() {
             commands::ai::ai_authorize_dir,
             // 审批历史面板查 ai_tool_executions 表,敏感字段截断)
             commands::ai::audit::record::list_tool_executions,
+            // S4 TaskDrawer 关联会话:按 task_id 反查发起过任务工具的主会话(审计表推导,只读)
+            commands::ai::audit::record::list_task_conversations,
             // AC-5 按工具失败率统计(诊断 IPC,聚合 ai_tool_executions 覆盖全量/时间范围)
             commands::ai::audit::record::tool_failure_stats,
             // 知识库
@@ -561,9 +565,11 @@ pub fn run() {
                     let _ = window.hide(); // 无运行会话:隐藏到托盘(原行为)
                     return;
                 }
-                // 有运行会话:每进程一次给主窗发提示事件(payload 带运行数),前端画气泡
+                // 有运行会话:每进程一次给主窗发提示事件(payload 带运行数),前端画气泡。
+                // 键名必须为 running:App.vue 监听端读 e.payload.running,曾发 "active" 致
+                // 端上恒取 undefined→按 0 处理,气泡整条链路静默失效(UX-P3 批次修正)。
                 if !CLOSE_NOTICE_SENT.swap(true, std::sync::atomic::Ordering::SeqCst) {
-                    let _ = app.emit_to("main", "tray-hide-notice", serde_json::json!({ "active": active }));
+                    let _ = app.emit_to("main", "tray-hide-notice", serde_json::json!({ "running": active }));
                 }
                 // 异步等 2 秒再 hide:气泡可见窗口,禁用 std::thread::sleep(同步阻塞卡渲染)
                 let win = window.clone();

src/App.vue

@@ -29,6 +29,9 @@
     </template>
   </AppLayout>
 
+  <!-- 走查 P3-16:全局 toast 统一收进同一纵向堆叠容器。原错误 toast 与托盘提示气泡
+       各自 fixed 同位(bottom:24px 居中),两者同现时互相压盖、后弹者盖住先弹者。 -->
+  <div class="df-toast-stack">
     <!-- 全局错误 toast(根级,主窗口与分离窗口共用)-->
     <transition name="toast">
       <div v-if="errorMsg" class="df-toast">⚠ {{ errorMsg }}</div>
@@ -39,6 +42,7 @@
     <transition name="toast">
       <div v-if="trayNoticeMsg" class="df-toast df-toast--info">{{ trayNoticeMsg }}</div>
     </transition>
+  </div>
 
   <!-- 审批待办浮层(应用内右上角,窗口聚焦/可见时作快速提示)。
        仅在主窗口与 ai-detached 窗口挂载(file-explorer / approval-popup 自身不挂)。
@@ -63,6 +67,7 @@ import { useKnowledgeStore } from './stores/knowledge'
 import { useProjectStore } from './stores/project'
 import { useAppSettingsStore } from './stores/appSettings'
 import i18n from './i18n'
+import { t as ti18n } from './i18n/i18n-helpers'
 import AiChat from './components/AiChat.vue'
 import ApprovalOverlay from './components/ai/ApprovalOverlay.vue'
 import ApprovalDialog from './components/project/ApprovalDialog.vue'
@@ -227,7 +232,8 @@ onMounted(async () => {
     _unlistenTrayHideNotice = await listen<{ running: number }>('tray-hide-notice', (e) => {
       const running = e.payload?.running ?? 0
       if (running <= 0) return
-      trayNoticeMsg.value = `${running} 个会话将继续在后台运行，托盘图标可随时唤起`
+      // 走查 P3-15:气泡文案硬编码中文 → 入 i18n(运行数经命名参数注入)
+      trayNoticeMsg.value = ti18n('common.trayNoticeRunning', { n: running })
       if (trayNoticeTimer) clearTimeout(trayNoticeTimer)
       trayNoticeTimer = setTimeout(() => { trayNoticeMsg.value = '' }, 3500)
     })
@@ -313,16 +319,22 @@ onUnmounted(() => {
 .page-enter-from { opacity: 0; transform: translateY(6px); }
 .page-leave-to { opacity: 0; transform: translateY(-4px); }
 
-/* ═══ 全局错误 toast ═══ */
-.df-toast {
+/* ═══ 全局 toast 纵向堆叠容器(走查 P3-16)═══
+   容器独占 fixed 定位,内部 flex 纵排 + 底对齐,多条 toast 依次向上堆叠不再同位互盖。
+   子项保持 .df-toast 原形态(定宽自适应),定位字段从子项上移到容器。 */
+.df-toast-stack {
   position: fixed; bottom: 24px; left: 50%; transform: translateX(-50%);
+  display: flex; flex-direction: column; align-items: center; gap: 8px;
+  z-index: 9999; max-width: 90vw; pointer-events: none;
+}
+.df-toast {
   background: var(--df-danger); color: #fff;
   padding: 10px 18px; border-radius: var(--df-radius-sm);
-  font-size: 13px; z-index: 9999; max-width: 90vw;
+  font-size: 13px; max-width: 90vw;
   box-shadow: 0 4px 16px rgba(0,0,0,0.3);
 }
 .toast-enter-active, .toast-leave-active { transition: all 0.25s var(--df-ease); }
-.toast-enter-from, .toast-leave-to { opacity: 0; transform: translate(-50%, 10px); }
+.toast-enter-from, .toast-leave-to { opacity: 0; transform: translateY(10px); }
 /* 后台运行提示:同 df-toast 形态,信息语义用 accent 底色与错误红区分 */
 .df-toast--info { background: var(--df-accent); }
 </style>

src/api/ai.ts

@@ -368,4 +368,22 @@ export const aiApi = {
    updateConversationGoals(conversationId: string, goals: string[]): Promise<void> {
      return invoke('ai_update_conversation_goals', { conversationId, goals })
    },
+
+  /**
+   * 只读窥视子代理轨迹(S3):后端复用 switch 读路径但**不切任何会话状态**。
+   * 子会话(origin='sub_agent')不进侧栏、不可 switch,完成卡浮层经此拉轨迹;
+   * 返回全量消息 JSON 字符串(不分页)。
+   */
+  peekConversation(conversationId: string): Promise<{ title: string | null; messages: string }> {
+    return invoke('ai_conversation_peek', { conversationId })
+  },
+
+  /**
+   * S4 TaskDrawer「关联会话」:按任务 id 反查发起过 create/update/advance_task 的会话。
+   * 后端从 ai_tool_executions 审计表推导(result JSON 顶层 id 精确判等,非 LIKE),
+   * 已过滤子代理会话(origin='sub_agent')并按最近命中倒序去重,v1 只列主会话。
+   */
+  listTaskConversations(taskId: string): Promise<{ conversation_id: string; title: string | null; tool_name: string; requested_at: string }[]> {
+    return invoke('list_task_conversations', { taskId })
+  },
  }

src/components/AiChat.vue

@@ -289,6 +289,10 @@ const projectsStore = useProjectStore()
 // 滚动边沿/跟随意图重置随 MessageList 迁移,由子组件自管。
 watch(() => store.state.activeConversationId, () => {
   if (editingMsgId.value) cancelEdit()
+  // UX-P3: 队列编辑态(editingQueueIdx/editingQueueText)绑定旧会话队列在 activeQueue
+  // 视图内的下标,切会话后 activeQueue 已换成新会话的集合,残留 idx/text 会错挂到
+  // 新会话同下标队列项上,随消息编辑态一并重置。
+  if (editingQueueIdx.value !== null) cancelEditQueued()
 })
 
 
@@ -331,7 +335,14 @@ function onGlobalKeydown(e: KeyboardEvent) {
   // Ctrl+R: 重新生成末条回复(替代浏览器刷新,Tauri 下 preventDefault 可靠)
   if (ctrl && (e.key === 'r' || e.key === 'R')) {
     e.preventDefault()
-    if (!store.state.streaming) store.regenerate()
+    // UX-P3: 末条非 assistant 等场景 regenerate() 会 reject(regenerateUnavailable),原样裸调
+    // 成 unhandled rejection 用户无感;对齐 MessageList.handleRetry 的 catch + toast。
+    if (!store.state.streaming) {
+      store.regenerate().catch((e: unknown) => {
+        const msg = e instanceof Error ? e.message : String(e)
+        showToast(t('aiChat.regenerateFailed', { msg }), 'error')
+      })
+    }
     return
   }
   // Ctrl+L: 清空当前对话(带 confirm;生成中跳过避免删流式中消息)
@@ -510,9 +521,14 @@ async function handleSendQueuedNow(idx: number) {
  *  cancelQueued 内部经 activeQueueRealIndex 做过滤视图下标映射,越界 no-op;
  *  项携带技能时按名找回 SkillInfo 回挂 chip,找不到(技能已卸载)只回填文本,
  *  与队列原地编辑只保 text 同一取舍。 */
-function backfillQueued(idx: number) {
+async function backfillQueued(idx: number) {
   const item = activeQueue.value[idx]
   if (!item) return
+  // UX-P3: 输入框已有未发送草稿(文本/图片/技能 chip 任一非空)时回填会静默覆盖,
+  // 先 confirm 确认;取消则直接返回,队列项原样保留(尚未 cancelQueued)。
+  if (chatInputRef.value?.hasUnsentInput()) {
+    if (!await confirmDialog(t('aiChat.backfillOverwriteConfirm'))) return
+  }
   const skill = item.skill
   ? store.state.skills.find(s => s.name === item.skill) ?? null
   : null

src/components/ai/ChatInput.vue

@@ -25,7 +25,8 @@
           : (pendingSkill ? (pendingSkill.argument_hint || $t('aiChat.placeholderWithSkill')) : $t('aiChat.placeholderDefault'))"
         rows="1"
         @keydown="handleKeydown"
-        @input="autoResize"
+        @beforeinput="onBeforeInputText"
+        @input="onInputText"
         @paste="onPaste"
         @dragover="onDragOver"
         @drop="onDrop"
@@ -401,6 +402,38 @@ const mentionStart = ref(-1)
 // 后端据 mentionSpans resolve 投影成 Augmentation 注入;MessageList 据此精确切区间渲染 chip。
 const pendingMentionSpans = ref<MentionSpan[]>([])
 
+// ── UX-P3: mention span 编辑平移 ──
+// span.start 是 selectMention 时刻的绝对偏移,用户在 chip 文本之前增删字符后全部 span 错位
+// (MessageList 渲染守卫对不上文本即降级纯文本,chip 全灭)。借 beforeinput(仅用户编辑触发,
+// 程序赋值不派发事件)快照旧文本,input 时做公共前后缀差分求出编辑区间,按区间相对各 span
+// 的位置处理:编辑点在 span 前→整体平移;在 span 后→不动;落在 span 文本内→丢弃该 span
+// (文本已被改写区间失真,宁丢一枚 chip 也不让后端 resolve 到错位区间)。
+let mentionTextBeforeEdit = ''
+function onBeforeInputText(): void {
+  mentionTextBeforeEdit = inputEl.value?.value ?? inputText.value
+}
+function onInputText(): void {
+  autoResize()
+  shiftMentionSpans(mentionTextBeforeEdit, inputEl.value?.value ?? inputText.value)
+}
+function shiftMentionSpans(oldText: string, newText: string): void {
+  if (pendingMentionSpans.value.length === 0 || oldText === newText) return
+  // 公共前后缀夹出旧文本中被编辑的区间(粘贴/剪切的多字符增删同为一段区间,该差分通用)
+  const minLen = Math.min(oldText.length, newText.length)
+  let p = 0
+  while (p < minLen && oldText[p] === newText[p]) p++
+  let s = 0
+  while (s < minLen - p && oldText[oldText.length - 1 - s] === newText[newText.length - 1 - s]) s++
+  const editStart = p
+  const editEnd = oldText.length - s
+  const delta = newText.length - oldText.length
+  pendingMentionSpans.value = pendingMentionSpans.value.flatMap((sp) => {
+    if (editEnd <= sp.start) return [{ ...sp, start: sp.start + delta }]
+    if (editStart >= sp.start + sp.length) return [sp]
+    return []
+  })
+}
+
 // ── ⑥.4 Phase 4: @项目展开摘要(输入区 enrichment 预览) ──
 // 当用户 @[项目:xxx] 后,从 projectsStore 反查项目+关联任务/灵感,
 // 在输入区下方显式 badge + 可展开的上下文参考面板。
@@ -685,6 +718,14 @@ function handleKeydown(e: KeyboardEvent) {
   // P1-1 输入法组词守卫:拼音/双拼组词未上屏时 Enter=选词、Esc=取消组词,均非用户发送/取消意图,
   // 直接放行给 IME。isComposing 覆盖标准浏览器;keyCode===229 兜部分环境 keydown 阶段不上报 isComposing。
   if (e.isComposing || e.keyCode === 229) return
+  // UX-P3: AI 输入框聚焦时 Ctrl/Cmd+R 会触发 WebView 刷新(浏览器直开/开发态整页刷新丢会话态),
+  // 此处 preventDefault 掐掉刷新默认行为。只拦输入框聚焦这一场景(本 handler 仅 textarea 触发,
+  // 全局其他区域不拦);不 stopPropagation,冒泡到 AiChat onGlobalKeydown 的 Ctrl+R 分支,
+  // 重新生成语义照常生效。
+  if ((e.ctrlKey || e.metaKey) && (e.key === 'r' || e.key === 'R')) {
+    e.preventDefault()
+    return
+  }
   // 各 Esc 消费分支 stopPropagation:P1-2 双层防,分离窗口 AiChat 的 window 级 Esc 会关整个窗口,
   // 内层已消费的 Esc 不再外抛(主防线是 AiChat 侧的 e.target 输入控件守卫,此处为辅)
   // 编辑态 Escape 取消编辑(最高优先级)
@@ -732,6 +773,14 @@ function handleKeydown(e: KeyboardEvent) {
       return
     }
   }
+  // UX-P3: 技能联想浮层开着但无匹配时,Enter 原会落到下方发送分支把 "/xxx" 原文当普通
+  // 消息发出。改为不发送 + toast 提示无此技能;确要发送斜杠原文可先 Esc 关浮层再 Enter。
+  if (skillOpen.value && e.key === 'Enter' && !e.shiftKey) {
+    e.preventDefault()
+    e.stopPropagation()
+    emit('error', { msg: t('aiChat.skillNoMatchQuery', { q: skillQuery.value.trim() }), type: 'warning' })
+    return
+  }
   // @ 实体引用浮层导航(仅有匹配项时拦截方向/回车;与技能联想互斥)
   if (mentionOpen.value && mentionItems.value.length) {
     if (e.key === 'ArrowDown') {
@@ -775,7 +824,12 @@ async function handleSend() {
   const text = inputText.value.trim()
   // UX-09 编辑态:走 editMessage(不 push 新 user,编辑末条 user 后重生成)
   if (props.editingMsgId) {
-    if (!text) return
+    // UX-P3 空编辑静默:清空文本后 Enter 原来静默 return,用户无反馈卡在编辑态。
+    // 仍阻止空提交(editMessage 空文本语义不明),但补轻提示告知原因。
+    if (!text) {
+      emit('error', { msg: t('aiChat.editEmptyBlocked'), type: 'warning' })
+      return
+    }
     inputText.value = ''
     if (inputEl.value) inputEl.value.style.height = 'auto'
     try {

src/components/ai/MessageList.vue

@@ -23,6 +23,7 @@ import { useStreamRenderer } from '../../composables/ai/useStreamRenderer'
 import { loadMoreHistory, getLoadMoreState } from '../../composables/ai/useAiConversations'
 import { collectTurnStats, turnSummaryText, sumTurnTokens, formatTokensK, type TurnCallStats } from '../../composables/ai/useTurnTrace'
 import { getConvAgentFinished, type AgentFinishedCard } from '../../composables/ai/aiShared'
+import { aiApi } from '../../api'
 import ToolCardList from '../ToolCardList.vue'
 import MessageItem from './MessageItem.vue'
 import EmptyState from './EmptyState.vue'
@@ -273,9 +274,32 @@ watch(() => store.state.streaming, (s) => {
     saveSelection()
     clearStreamingState()
     nextTick(() => restoreSelection())
+    anchorCollapseScroll()
   }
 })
 
+// ── UX-P3: 长回复完成瞬间折叠的滚动锚 ──
+// 末条 AI 回复超阈值(LONG_MSG_THRESHOLD)时,完成瞬间 shouldCollapseLong 从 false 翻 true
+// (流式中不折叠的约束解除),气泡收缩到 400px;跟随意图下的 scrollToBottom 把视口钉在收缩后
+// 的底部,内容整体上跳,刚流完的答案开头/结尾都看不全。改锚消息顶:本回调在流式结束当拍注册
+// nextTick(此时折叠 class 尚未渲染),渲染拍后把该消息顶对齐视口顶,视口落点确定(答案开头,
+// 回底按钮照常出现),不再随内容收缩被浏览器 clamp 到任意位置。
+// 仅跟随意图成立时干预(用户上滑读历史不拉扯);手动展开过的消息不会折叠,无收跳不处理。
+// 注册时序:currentText watch(先注册)的 scrollToBottom nextTick 先入队,本 nextTick 后入队
+// 后执行,锚定是滚动链的最后一手,不被钉底覆盖。
+const COLLAPSE_ANCHOR_PAD_PX = 8
+function anchorCollapseScroll(): void {
+  if (!isFollowingBottom.value) return
+  const last = store.state.messages[store.state.messages.length - 1]
+  if (!last || !shouldCollapseLong(last)) return
+  nextTick(() => {
+    const c = messagesContainer.value
+    const el = c?.querySelector(`[data-msg-id="${last.id}"]`)
+    if (!c || !el) return
+    c.scrollTop += el.getBoundingClientRect().top - c.getBoundingClientRect().top - COLLAPSE_ANCHOR_PAD_PX
+  })
+}
+
 // ── 工具卡片自动收起:新内容追加(新消息/toolCall 状态变化)时,
 // 收起旧的已完成/已拒绝卡片,保留 running/pending/write_file 活跃卡片。
 // 折叠态由 ToolCardList 自治,父级仅经 expose 调 collapseInactive ──
@@ -1083,6 +1107,68 @@ function agentFinSummary(summary: string): string {
   return s.length > 160 ? s.slice(0, 160) + '…' : s
 }
 
+// ── S3 子代理轨迹浮层(完成卡点击展开,只读) ──
+// 状态全部组件局部:浮层关即弃、不缓存轨迹(每次点开重新 peek,宁少勿预留),
+// 不进 aiShared/stores——子会话轨迹纯只读,与主会话状态机(消息流/生成态)完全无关。
+interface AgentPeekMsg {
+  role: 'user' | 'assistant' | 'tool' | 'system'
+  text: string
+  toolNames: string[]
+}
+const agentPeekVisible = ref(false)
+const agentPeekLoading = ref(false)
+const agentPeekMsgs = ref<AgentPeekMsg[]>([])
+// 单条内容截断上限:tool 结果(文件全文/长输出)可能巨大,简版渲染按 4000 字截断兜底,
+// 完整数据仍在库,不做全文渲染(不要求富文本)。
+const AGENT_PEEK_TEXT_MAX = 4000
+
+/**
+ * 完成卡点击:peek 子会话轨迹并开浮层。
+ * agentId 即子会话 conv id(形态 {parent}::agent-{n},后端 spawn_agent 落库 origin='sub_agent')。
+ * 加载中/已开时忽略重复点击(防连点重复 IPC);失败走既有 toast 通道给可见反馈,不开空浮层。
+ */
+async function openAgentTrace(card: AgentFinishedCard): Promise<void> {
+  if (agentPeekVisible.value || agentPeekLoading.value) return
+  agentPeekLoading.value = true
+  try {
+    const res = await aiApi.peekConversation(card.agentId)
+    const raw: any[] = typeof res.messages === 'string' ? JSON.parse(res.messages) : (res.messages || [])
+    agentPeekMsgs.value = raw.map((m) => {
+      const text = String(m.content || '')
+      const role = m.role === 'user' || m.role === 'assistant' || m.role === 'tool' || m.role === 'system'
+        ? m.role
+        : 'assistant'
+      return {
+        role,
+        text: text.length > AGENT_PEEK_TEXT_MAX ? text.slice(0, AGENT_PEEK_TEXT_MAX) + '…' : text,
+        toolNames: Array.isArray(m.tool_calls)
+          ? m.tool_calls.map((tc: any) => tc?.function?.name || '').filter(Boolean)
+          : [],
+      }
+    })
+    agentPeekVisible.value = true
+  } catch (e) {
+    console.error('[AI] 拉取子代理轨迹失败:', e)
+    emit('toast', { msg: t('ai.agentTrace.loadFail'), type: 'error' })
+  } finally {
+    agentPeekLoading.value = false
+  }
+}
+
+/** 关浮层即弃:清空局部消息态,不缓存轨迹 */
+function closeAgentTrace(): void {
+  agentPeekVisible.value = false
+  agentPeekMsgs.value = []
+}
+
+/** role → 本地化角色名(简版渲染,无头像气泡) */
+function agentPeekRoleText(role: AgentPeekMsg['role']): string {
+  if (role === 'user') return t('ai.agentTrace.roleUser')
+  if (role === 'tool') return t('ai.agentTrace.roleTool')
+  if (role === 'system') return t('ai.agentTrace.roleSystem')
+  return t('ai.agentTrace.roleAssistant')
+}
+
 // UX 密度改造:轮次导航条(仅多轮长会话 ≥3 张轨迹卡出现,sticky 置顶,只做跳转不做统计)
 const traceNav = computed(() =>
   renderItems.value.filter((it): it is Extract<RenderItem, { kind: 'trace' }> => it.kind === 'trace')
@@ -1230,9 +1316,16 @@ defineExpose({
       </div>
       <!-- S1 子代理完成卡(AiAgentFinished):状态点/状态词三态区分(completed 中性/truncated
            警示/failed 错误),摘要单行截断 + 轮数 + tokens;骨架对齐 ai-step-line/ai-retry-line
-           轻量细条范式,非消息数据不进气泡 -->
+           轻量细条范式,非消息数据不进气泡。
+           S3:整卡可点,点击 peek 子会话轨迹开只读浮层(数据关即弃,见 openAgentTrace) -->
       <div v-else-if="item.kind === 'agentFin'" class="ai-msg-slot">
-        <div class="ai-agent-fin" :class="'ai-agent-fin--' + agentFinTone(item.card.status)">
+        <div
+          class="ai-agent-fin"
+          :class="'ai-agent-fin--' + agentFinTone(item.card.status)"
+          role="button"
+          :title="$t('ai.agentFinished.viewHint')"
+          @click="openAgentTrace(item.card)"
+        >
           <span class="ai-agent-fin-dot"></span>
           <span class="ai-agent-fin-label">{{ $t('ai.agentFinished.title') }}</span>
           <span class="ai-agent-fin-status">{{ agentFinStatusText(item.card.status) }}</span>
@@ -1485,6 +1578,35 @@ defineExpose({
       <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><line x1="12" y1="5" x2="12" y2="19"/><polyline points="19 12 12 19 5 12"/></svg>
     </button>
   </Transition>
+
+  <!-- S3 子代理轨迹浮层(只读,关即弃):与回底部按钮同为模板根级兄弟,position:fixed
+       不受 .ai-messages 滚动容器 overflow 裁剪;backdrop 点击关闭,数据不缓存 -->
+  <div v-if="agentPeekVisible" class="ai-agent-peek-mask" @click.self="closeAgentTrace">
+    <div class="ai-agent-peek" role="dialog" :aria-label="$t('ai.agentTrace.title')">
+      <div class="ai-agent-peek-head">
+        <span class="ai-agent-peek-title">{{ $t('ai.agentTrace.title') }}</span>
+        <button class="ai-agent-peek-close" :title="$t('common.close')" @click="closeAgentTrace">
+          <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg>
+        </button>
+      </div>
+      <div class="ai-agent-peek-body">
+        <div v-if="agentPeekMsgs.length === 0" class="ai-agent-peek-empty">{{ $t('ai.agentTrace.empty') }}</div>
+        <!-- 简版卡片流:role 名 + 工具调用名 chips + 纯文本内容(不做完整 markdown 富渲染) -->
+        <div
+          v-for="(m, i) in agentPeekMsgs"
+          :key="i"
+          class="ai-agent-peek-msg"
+          :class="'ai-agent-peek-msg--' + m.role"
+        >
+          <div class="ai-agent-peek-role">
+            <span>{{ agentPeekRoleText(m.role) }}</span>
+            <span v-for="name in m.toolNames" :key="name" class="ai-agent-peek-tool">{{ name }}</span>
+          </div>
+          <div v-if="m.text" class="ai-agent-peek-text">{{ m.text }}</div>
+        </div>
+      </div>
+    </div>
+  </div>
 </template>
 
 <style scoped>
@@ -2268,7 +2390,8 @@ defineExpose({
 
 /* S1 子代理完成卡(AiAgentFinished):三态色分层——completed 中性灰(常态结果不抢视线)
    / truncated 琥珀(结果不完整,与 retry 细条同警示层)/ failed 红(需关注,对齐错误卡层)。
-   骨架对齐 .ai-retry-line 轻量细条;摘要单行截断,全文经 title hover 查看 */
+   骨架对齐 .ai-retry-line 轻量细条;摘要单行截断,全文经 title hover 查看。
+   S3:整卡可点开轨迹浮层,cursor pointer + hover 提示可交互 */
 .ai-agent-fin {
   display: flex;
   align-items: center;
@@ -2280,6 +2403,10 @@ defineExpose({
   font-size: 11px;
   color: var(--df-text-secondary);
   user-select: none;
+  cursor: pointer;
+}
+.ai-agent-fin:hover {
+  background: color-mix(in srgb, var(--df-text-secondary) 12%, transparent);
 }
 .ai-agent-fin-dot {
   width: 5px;
@@ -2323,4 +2450,106 @@ defineExpose({
   font-family: var(--df-font-mono);
   font-size: 10.5px;
 }
+
+/* ═══ S3 子代理轨迹浮层(只读,关即弃) ═══ */
+.ai-agent-peek-mask {
+  position: fixed;
+  inset: 0;
+  z-index: 300;
+  background: rgba(0, 0, 0, 0.45);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+}
+.ai-agent-peek {
+  width: min(680px, 86vw);
+  max-height: 76vh;
+  display: flex;
+  flex-direction: column;
+  background: var(--df-bg-raised);
+  border: 0.5px solid var(--df-border-strong);
+  border-radius: var(--df-radius-lg);
+  box-shadow: 0 12px 40px rgba(0, 0, 0, 0.4);
+  overflow: hidden;
+}
+.ai-agent-peek-head {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 10px 14px;
+  border-bottom: 0.5px solid var(--df-border);
+  flex-shrink: 0;
+}
+.ai-agent-peek-title {
+  font-size: 12.5px;
+  font-weight: 600;
+  color: var(--df-text);
+}
+.ai-agent-peek-close {
+  border: none;
+  background: transparent;
+  color: var(--df-text-dim);
+  cursor: pointer;
+  padding: 4px;
+  border-radius: var(--df-radius-sm);
+  display: flex;
+}
+.ai-agent-peek-close:hover {
+  color: var(--df-text);
+  background: color-mix(in srgb, var(--df-text-secondary) 10%, transparent);
+}
+.ai-agent-peek-body {
+  overflow-y: auto;
+  padding: 12px 14px;
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+}
+.ai-agent-peek-empty {
+  color: var(--df-text-dim);
+  font-size: 12px;
+  text-align: center;
+  padding: 24px 0;
+}
+.ai-agent-peek-msg {
+  border: 0.5px solid var(--df-border);
+  border-radius: var(--df-radius);
+  padding: 8px 10px;
+}
+/* role 左侧色条:user 中性/accent、assistant 默认、tool 结果弱化、system 弱化斜体 */
+.ai-agent-peek-msg--user { border-left: 2px solid var(--df-accent); }
+.ai-agent-peek-msg--assistant { border-left: 2px solid var(--df-success); }
+.ai-agent-peek-msg--tool,
+.ai-agent-peek-msg--system { border-left: 2px solid var(--df-border-strong); }
+.ai-agent-peek-role {
+  display: flex;
+  align-items: center;
+  flex-wrap: wrap;
+  gap: 6px;
+  font-size: 10.5px;
+  font-weight: 600;
+  color: var(--df-text-secondary);
+  margin-bottom: 4px;
+}
+.ai-agent-peek-tool {
+  font-family: var(--df-font-mono);
+  font-size: 10px;
+  font-weight: 500;
+  color: var(--df-info);
+  background: var(--df-info-bg);
+  border-radius: var(--df-radius-sm);
+  padding: 1px 6px;
+}
+.ai-agent-peek-text {
+  font-size: 12px;
+  line-height: 1.55;
+  color: var(--df-text);
+  white-space: pre-wrap;
+  word-break: break-word;
+}
+.ai-agent-peek-msg--tool .ai-agent-peek-text,
+.ai-agent-peek-msg--system .ai-agent-peek-text {
+  color: var(--df-text-secondary);
+  font-size: 11.5px;
+}
 </style>

src/components/project/GitChanges.vue

@@ -185,8 +185,6 @@ const loading = ref(false)
 const gitStatus = ref<GitStatusResult | null>(null)
 const showHistory = ref(false)
 const selectedFile = ref('')
-const diffContent = ref('')
-const diffLoading = ref(false)
 // 行内错误提示:loadError 主区展示(loadStatus 失败),actionError 分支条下方展示(loadMore/selectFile/切分支/提交详情 失败)
 const loadError = ref('')
 const actionError = ref('')
@@ -194,6 +192,12 @@ const actionError = ref('')
 // (对齐 ProjectDetail taskReqSeq 模式)
 let statusReqSeq = 0
 
+// UX走查 P3-4: 点文件 diff 单槽去重缓存 — 重复点击同一文件不再重复发 git diff IPC,
+// 直接复用上次结果回填。切模块/refreshKey 重拉状态均经 loadStatus,缓存在彼处失效,
+// 防 AI 改文件后旧 diff 被复用(单槽不设 Map,免过期条目堆积)
+let lastDiffKey = ''
+let lastDiffContent = ''
+
 /** 单条提交记录形状(对齐 getModuleCommits / GitRecentCommit)。 */
 interface CommitItem {
   hash: string
@@ -353,6 +357,10 @@ async function loadStatus() {
   const mid = props.moduleId
   if (!mid) return
   const seq = ++statusReqSeq
+  // UX走查 P3-4: 状态重拉(切模块/外部 refreshKey,如 AI 改文件)即失效 diff 去重缓存,
+  // 防改文件后旧 diff 被复用
+  lastDiffKey = ''
+  lastDiffContent = ''
   loading.value = true
   loadError.value = ''
   actionError.value = ''
@@ -448,23 +456,35 @@ async function switchToHistory() {
 
 async function selectFile(path: string) {
   selectedFile.value = path
-  // 先切预览到该文件(内容视图),diff 拉到后回填,点文件即展示变化
+  // 先切预览到该文件(内容视图兜底),diff 拉到后回填切 diff 视图,点文件即展示变化
   emit('select-file', path)
-  diffContent.value = ''
   actionError.value = ''
-  diffLoading.value = true
+  // UX走查 P3-4 同参数去重:重复点击同一文件直接复用缓存 diff,不再发 IPC
+  const key = `${props.moduleId}::${path}`
+  if (key === lastDiffKey) {
+    emit('select-file', path, lastDiffContent)
+    return
+  }
+  // 捕获请求时的状态序号:期间 refreshKey 重拉(缓存已失效)则本次结果不入缓存,防改文件前旧 diff 滞留
+  const seqAtReq = statusReqSeq
+  let diff = ''
   try {
     const res = await moduleApi.getModuleFileDiff(props.moduleId, path)
-    diffContent.value = res.diff || ''
+    diff = res.diff || ''
+    if (seqAtReq === statusReqSeq) {
+      lastDiffKey = key
+      lastDiffContent = diff
+    }
   } catch (e) {
-    diffContent.value = ''
+    // 失败清缓存使下次点击可重试;不回填 diff,预览停在内容视图(错误经 actionError 条就近提示)
+    lastDiffKey = ''
+    lastDiffContent = ''
     actionError.value = String(e)
-  } finally {
-    diffLoading.value = false
+    return
   }
   // 仅当仍是当前选中文件时才回填 diff,防快速切换时旧响应覆盖新文件
   if (selectedFile.value === path) {
-    emit('select-file', path, diffContent.value)
+    emit('select-file', path, diff)
   }
 }
 

src/components/project/ProjectCard.vue

@@ -24,7 +24,10 @@
     <div class="card-footer">
       <div class="footer-stat">
         <span class="stat-icon">📁</span>
-        <span>{{ moduleCount }}</span>
+        <!-- UX走查 P3-6: 工程数异步入场,加载中显骨架条占位(对齐 GitChanges shimmer 范式),
+             防未加载数先显 0 再跳变——真 0 与"还没加载"不可混淆 -->
+        <span v-if="moduleCountLoading" class="stat-skeleton" aria-hidden="true"></span>
+        <span v-else>{{ moduleCount }}</span>
       </div>
       <div class="footer-stat">
         <span class="stat-icon">📅</span>
@@ -48,7 +51,7 @@
 // 依赖均为无副作用纯函数 + 共享常量;删除走 emit 交回父组件处理(confirm 弹层 + store 调用)。
 // view prop:'list'(紧凑,默认)/'card'(宽松);描述 max-height 截断 + 展开/收起。
 import { useRouter } from 'vue-router'
-import { ref, onMounted, watch, nextTick } from 'vue'
+import { ref, onMounted, onBeforeUnmount, watch, nextTick } from 'vue'
 import { formatDate } from '@/utils/time'
 import { parseStack } from '@/utils/project'
 import { stripMd } from '@/utils/markdown'
@@ -71,11 +74,14 @@ const router = useRouter()
 
 // 工程数(异步拉取,展示在卡片底部)
 const moduleCount = ref(0)
+// UX走查 P3-6: 加载态标记,模板据此显骨架占位而非 0(成功/失败都收 loading,失败仍显 0 属既有约定)
+const moduleCountLoading = ref(true)
 onMounted(async () => {
   try {
     const list = await moduleApi.listProjectModules(props.project.id)
     moduleCount.value = list.length
   } catch { /* 老项目无工程记录不报错 */ }
+  moduleCountLoading.value = false
   // DOM 渲染完成后测描述是否溢出(决定是否显示「展开」)
   await nextTick()
   measureDesc()
@@ -91,6 +97,8 @@ const descExpanded = ref(false)
 function measureDesc() {
   const el = descRef.value
   if (!el) return
+  // 展开态无行截断,测出的一律是"未溢出",会误收起按钮致无法「收起」;展开期间维持原判定
+  if (descExpanded.value) return
   // scrollHeight > clientHeight 说明被 max-height 截断了
   descOverflow.value = el.scrollHeight - el.clientHeight > 1
 }
@@ -98,6 +106,13 @@ function measureDesc() {
 // 描述/视图模式变化后重新测量(描述异步入库或切 view 行数阈值变,按钮显隐需重算)
 watch(() => props.project.description, () => nextTick(measureDesc))
 watch(() => props.view, () => nextTick(measureDesc))
+// 收起后重测:展开期间判定被冻结,若窗口已放宽到不溢出,收起时自校正防残留多余「展开」按钮
+watch(descExpanded, (v) => { if (!v) nextTick(measureDesc) })
+
+// UX走查 P3-6: 窗口尺寸变化会改描述可用宽度,截断与否需重测,否则「展开」按钮显隐失准
+function onWinResize() { measureDesc() }
+onMounted(() => window.addEventListener('resize', onWinResize))
+onBeforeUnmount(() => window.removeEventListener('resize', onWinResize))
 </script>
 
 <style scoped>
@@ -196,6 +211,22 @@ watch(() => props.view, () => nextTick(measureDesc))
 }
 .stat-icon { font-size: 13px; }
 
+/* UX走查 P3-6: 工程数加载占位骨架条(shimmer 配方对齐 GitChanges .gc-skeleton-row),
+   宽度约两位数字占位,高度对齐 12px 字号,防 0 起显跳变 */
+.stat-skeleton {
+  display: inline-block;
+  width: 18px;
+  height: 12px;
+  border-radius: 3px;
+  background: linear-gradient(90deg, rgba(255,255,255,0.04) 0%, rgba(255,255,255,0.08) 50%, rgba(255,255,255,0.04) 100%);
+  background-size: 200% 100%;
+  animation: pc-shimmer 1.5s infinite linear;
+}
+@keyframes pc-shimmer {
+  0% { background-position: 200% 0; }
+  100% { background-position: -200% 0; }
+}
+
 .card-top-actions { display: flex; align-items: center; gap: 8px; }
 .card-del-btn {
   background: transparent; border: none; cursor: pointer;

src/components/settings/AppearanceSection.vue

@@ -6,20 +6,21 @@
     </div>
     <div class="general-settings">
       <SettingRow ref="rowTheme" :label="$t('settings.labelTheme')" :desc="$t('settings.descTheme')">
-        <select v-model="settings.theme" class="setting-select" @change="markSaved('theme')">
+        <!-- 走查 P3-11:不再 @change 即显「已保存」,反馈移到 persist 完成后(防假成功) -->
+        <select v-model="settings.theme" class="setting-select">
           <option value="dark">{{ $t('settings.themeDark') }}</option>
           <option value="light">{{ $t('settings.themeLight') }}</option>
           <option value="system">{{ $t('settings.themeSystem') }}</option>
         </select>
       </SettingRow>
       <SettingRow ref="rowLanguage" :label="$t('settings.labelLanguage')" :desc="$t('settings.descLanguage')">
-        <select v-model="settings.language" class="setting-select" @change="markSaved('language')">
+        <select v-model="settings.language" class="setting-select">
           <option value="zh-CN">{{ $t('settings.aiLanguageZh') }}</option>
           <option value="en">{{ $t('settings.aiLanguageEn') }}</option>
         </select>
       </SettingRow>
       <SettingRow ref="rowAiLanguage" :label="$t('settings.labelAiLanguage')" :desc="$t('settings.descAiLanguage')">
-        <select v-model="settings.aiLanguage" class="setting-select" @change="markSaved('aiLanguage')">
+        <select v-model="settings.aiLanguage" class="setting-select">
           <option value="auto">{{ $t('settings.aiLanguageAuto') }}</option>
           <option value="zh-CN">{{ $t('settings.aiLanguageZh') }}</option>
           <option value="en">{{ $t('settings.aiLanguageEn') }}</option>
@@ -27,13 +28,13 @@
       </SettingRow>
       <SettingRow ref="rowToken" :label="$t('settings.labelShowTokenUsage')" :desc="$t('settings.descShowTokenUsage')">
         <label class="toggle">
-          <input type="checkbox" v-model="settings.showTokenUsage" @change="markSaved('token')" />
+          <input type="checkbox" v-model="settings.showTokenUsage" />
           <span class="toggle-slider"></span>
         </label>
       </SettingRow>
       <SettingRow ref="rowStreamingMd" :label="$t('settings.labelStreamingMd')" :desc="$t('settings.descStreamingMd')">
         <label class="toggle">
-          <input type="checkbox" v-model="settings.streamingMd" @change="markSaved('streamingMd')" />
+          <input type="checkbox" v-model="settings.streamingMd" />
           <span class="toggle-slider"></span>
         </label>
       </SettingRow>
@@ -91,32 +92,42 @@ const rowLanguage = useTemplateRef<InstanceType<typeof SettingRow>>('rowLanguage
 const rowAiLanguage = useTemplateRef<InstanceType<typeof SettingRow>>('rowAiLanguage')
 const rowToken = useTemplateRef<InstanceType<typeof SettingRow>>('rowToken')
 const rowStreamingMd = useTemplateRef<InstanceType<typeof SettingRow>>('rowStreamingMd')
-function markSaved(key: 'theme' | 'language' | 'aiLanguage' | 'token' | 'streamingMd') {
-  const map = { theme: rowTheme.value, language: rowLanguage.value, aiLanguage: rowAiLanguage.value, token: rowToken.value, streamingMd: rowStreamingMd.value }
-  map[key]?.markSaved()
+
+// 走查 P3-11 假成功族:原 @change 立即 markSaved,先于持久化完成,失败无人知晓。
+// 改为 set 完成后才点亮「已保存」,失败点亮行内红字。注:appSettings.set 当前对落库
+// 失败仅 console 内吞(stores/appSettings.ts 降级内存态设计),catch 实际不会触发;
+// 待 store 层把失败上抛后本处失败通道即自动生效(登记项,见批次报告)。
+async function persist(key: string, value: unknown, row: { value: { markSaved: () => void; markFailed: () => void } | null }) {
+  try {
+    await appSettings.set(key, value)
+    row.value?.markSaved()
+  } catch (e) {
+    console.error(`保存设置 ${key} 失败:`, e)
+    row.value?.markFailed()
+  }
 }
 
 watch(() => settings.theme, (val) => {
   applyTheme(val)
-  void appSettings.set('df-theme', val)
+  void persist('df-theme', val, rowTheme)
 })
 
 watch(() => settings.aiLanguage, (val) => {
-  void appSettings.set('df-ai-language', val)
+  void persist('df-ai-language', val, rowAiLanguage)
 })
 
 watch(() => settings.showTokenUsage, (val) => {
-  void appSettings.set('df-show-token-usage', val)
+  void persist('df-show-token-usage', val, rowToken)
 })
 
 watch(() => settings.streamingMd, (val) => {
-  void appSettings.set('df-ai-streaming-md', val)
+  void persist('df-ai-streaming-md', val, rowStreamingMd)
 })
 
 watch(() => settings.language, (val) => {
-  void appSettings.set('df-language', val)
   // 同步 i18n 实例，让界面立即跟随语言切换
   locale.value = val as 'zh-CN' | 'en'
+  void persist('df-language', val, rowLanguage)
 })
 
 onMounted(() => {

src/components/settings/PerformanceSection.vue

@@ -10,7 +10,8 @@
         next-round
         :label="$t('settings.labelGlobalConcurrency')"
         :desc="$t('settings.descGlobalConcurrency')">
-        <input type="number" class="setting-number" min="1" style="width:80px"
+        <!-- 走查 P3-14:max=64 前端 sanity 上界(原生 min/max 只约束步进按钮,越界拦截在 JS clamp) -->
+        <input type="number" class="setting-number" min="1" max="64" style="width:80px"
                v-model.number="settings.llmGlobalConcurrency"
                @change="onConcurrencyChange" />
       </SettingRow>
@@ -19,7 +20,7 @@
         next-round
         :label="$t('settings.labelPerConvConcurrency')"
         :desc="$t('settings.descPerConvConcurrency')">
-        <input type="number" class="setting-number" min="1" style="width:80px"
+        <input type="number" class="setting-number" min="1" max="64" style="width:80px"
                v-model.number="settings.llmPerConvConcurrency"
                @change="onConcurrencyChange" />
       </SettingRow>
@@ -29,7 +30,8 @@
         :label="$t('settings.labelAgentMaxIterations')"
         :desc="$t('settings.descAgentMaxIterations')">
         <label class="unlimited-toggle">
-          <input type="checkbox" v-model="unlimited" @change="markSaved('iter')" />
+          <!-- 走查 P3-11:勾选反馈不再即时显,由 syncAgentMaxIterations 的 IPC 成功回调点亮 -->
+          <input type="checkbox" v-model="unlimited" />
           <span>{{ $t('settings.labelUnlimited') }}</span>
         </label>
         <input type="number" class="setting-number" min="1" max="50" style="width:80px"
@@ -52,6 +54,7 @@
 
 <script setup lang="ts">
 import { reactive, watch, computed, onMounted, onUnmounted, useTemplateRef } from 'vue'
+import { useI18n } from 'vue-i18n'
 import { aiApi } from '@/api'
 import { useAppSettingsStore } from '@/stores/appSettings'
 import SettingRow from './SettingRow.vue'
@@ -59,10 +62,16 @@ import SettingRow from './SettingRow.vue'
 // ============================================================
 // 性能域 — 本地设置(localStorage):LLM 并发 / 单对话并发 / Agent 轮次 / 流式重试
 // 拆自 GeneralPanel,本 Section 仅持有性能相关 key + 各自 debounce IPC 同步
-// 阶段5:SettingRow 封装行布局 + nextRound badge(后端快照锁定边界,下次对话生效);
-//        4 项变更后调 markSaved 显"已保存"
+// 阶段5:SettingRow 封装行布局 + nextRound badge(后端快照锁定边界,下次对话生效)
+// 走查 P3-11:「已保存」改为 IPC 成功回调后点亮,失败点亮行内红字(防假成功)
 // ============================================================
 const appSettings = useAppSettingsStore()
+const { t } = useI18n()
+
+// 走查 P3-14:并发越界钳制需要上抛 warning toast,由壳(Settings.vue)的 showToast 消费
+const emit = defineEmits<{
+  (e: 'toast', msg: string, type: 'error' | 'warning' | 'info' | 'success'): void
+}>()
 
 const settings = reactive({
   llmGlobalConcurrency: appSettings.get<number>('df-llm-global-concurrency', 3),
@@ -73,7 +82,7 @@ const settings = reactive({
   agentMaxRetries: appSettings.get<number>('df-ai-agent-max-retries', 3),
 })
 
-// 即时生效反馈:SettingRow refs
+// 即时生效反馈:SettingRow refs(成功/失败两态,走查 P3-11)
 const rowGlobalConcurrency = useTemplateRef<InstanceType<typeof SettingRow>>('rowGlobalConcurrency')
 const rowPerConvConcurrency = useTemplateRef<InstanceType<typeof SettingRow>>('rowPerConvConcurrency')
 const rowMaxIterations = useTemplateRef<InstanceType<typeof SettingRow>>('rowMaxIterations')
@@ -82,50 +91,73 @@ function markSaved(key: 'global' | 'perConv' | 'iter' | 'retry') {
   const map = { global: rowGlobalConcurrency.value, perConv: rowPerConvConcurrency.value, iter: rowMaxIterations.value, retry: rowMaxRetries.value }
   map[key]?.markSaved()
 }
+function markFailed(key: 'global' | 'perConv' | 'iter' | 'retry') {
+  const map = { global: rowGlobalConcurrency.value, perConv: rowPerConvConcurrency.value, iter: rowMaxIterations.value, retry: rowMaxRetries.value }
+  map[key]?.markFailed()
+}
 
+// SQLite 持久化(watch 触发):落库失败点亮行内红字。注:appSettings.set 当前对落库失败
+// 仅 console 内吞(降级内存态设计,stores/appSettings.ts),catch 是 store 上抛错误后的
+// 失败通道,今日实际不触发(登记项见批次报告);IPC 同步失败通道在下方 timer 内,是真实信号。
 watch(() => settings.llmGlobalConcurrency, (v) => {
-  void appSettings.set('df-llm-global-concurrency', v)
+  appSettings.set('df-llm-global-concurrency', v).catch(() => markFailed('global'))
 })
 
 watch(() => settings.llmPerConvConcurrency, (v) => {
-  void appSettings.set('df-llm-per-conv-concurrency', v)
+  appSettings.set('df-llm-per-conv-concurrency', v).catch(() => markFailed('perConv'))
 })
 
 watch(() => settings.agentMaxIterations, (v) => {
-  void appSettings.set('df-ai-agent-max-iterations', v)
+  appSettings.set('df-ai-agent-max-iterations', v).catch(() => markFailed('iter'))
 })
 
 watch(() => settings.agentMaxRetries, (v) => {
-  void appSettings.set('df-ai-agent-max-retries', v)
+  appSettings.set('df-ai-agent-max-retries', v).catch(() => markFailed('retry'))
 })
 
 // onMounted 期间调 sync 不应触发"已保存"反馈(那是启动恢复,非用户改动)
 let _userInteracting = false
 
+// 走查 P3-14:并发上界 64 为前端 sanity 值(后端 ai_set_concurrency_config 仅 max(1) 无上界,
+// 超大并发会让限流器形同虚设、放大上游限流/超时风险);越界时静默修正 + toast 告知修正结果
+const MAX_CONCURRENCY = 64
+
+/** 非法输入(NaN/清空)回退默认值,合法值 clamp [1,64] */
+function clampConcurrency(v: number | undefined, fallback: number): number {
+  const n = typeof v === 'number' && Number.isFinite(v) ? Math.trunc(v) : fallback
+  return Math.min(MAX_CONCURRENCY, Math.max(1, n))
+}
+
 // 同步 LLM 并发上限到后端(debounce 300ms,防快速连续调整时频繁 IPC)
 // 持久化已由上方 watch 写 appSettings(SQLite) 完成;此处只负责同步内存限流器
 let _concurrencyTimer: ReturnType<typeof setTimeout> | null = null
 function syncConcurrencyConfig() {
   // clamp 到合理区间,防越界输入(缓存被篡改 / 手动输超限)致限流失效;
   // 修正后值同步回 v-model(input 显示) + watch 持久化
-  // 下限 1,无上限;约束单对话并发不超过全局并发
-  settings.llmGlobalConcurrency = Math.max(1, settings.llmGlobalConcurrency || 3)
-  settings.llmPerConvConcurrency = Math.max(1, settings.llmPerConvConcurrency || 2)
-  if (settings.llmPerConvConcurrency > settings.llmGlobalConcurrency) {
-    settings.llmPerConvConcurrency = settings.llmGlobalConcurrency
+  // 下限 1,上界 64(前端 sanity);约束单对话并发不超过全局并发
+  let clamped = false
+  const g = clampConcurrency(settings.llmGlobalConcurrency, 3)
+  if (g !== settings.llmGlobalConcurrency) clamped = true
+  settings.llmGlobalConcurrency = g
+  let p = clampConcurrency(settings.llmPerConvConcurrency, 2)
+  if (p > g) { p = g; clamped = true }
+  settings.llmPerConvConcurrency = p
+  if (clamped && _userInteracting) {
+    emit('toast', t('settings.toastConcurrencyClamped', { max: MAX_CONCURRENCY }), 'warning')
   }
   if (_concurrencyTimer) clearTimeout(_concurrencyTimer)
+  // 走查 P3-11:interactive 在排程时捕获(debounce 回调执行时 _userInteracting 已复位 false)
+  const interactive = _userInteracting
   _concurrencyTimer = setTimeout(async () => {
     try {
-      await aiApi.setConcurrencyConfig(settings.llmGlobalConcurrency, settings.llmPerConvConcurrency)
+      await aiApi.setConcurrencyConfig(g, p)
+      // IPC 成功才显「已保存」,失败显行内红字(原实现先显成功、失败仅 console = 假成功)
+      if (interactive) { markSaved('global'); markSaved('perConv') }
     } catch (e) {
       console.error('同步 LLM 并发配置失败:', e)
+      if (interactive) { markFailed('global'); markFailed('perConv') }
     }
   }, 300)
-  if (_userInteracting) {
-    markSaved('global')
-    markSaved('perConv')
-  }
 }
 
 // wrapper:并发变更(@change 用户触发)同时反馈两项(global/perConv 任一变动都同步两项)
@@ -158,28 +190,32 @@ function syncAgentMaxIterations() {
     settings.agentMaxIterations = Math.min(50, Math.max(1, settings.agentMaxIterations || 10))
   }
   if (_agentIterTimer) clearTimeout(_agentIterTimer)
+  const interactive = _userInteracting
   _agentIterTimer = setTimeout(async () => {
     try {
       await aiApi.setAgentMaxIterations(settings.agentMaxIterations)
+      if (interactive) markSaved('iter')
     } catch (e) {
       console.error('同步 Agent 最大轮次失败:', e)
+      if (interactive) markFailed('iter')
     }
   }, 300)
-  if (_userInteracting) {
-    markSaved('iter')
-  }
 }
 
 // 「不限」勾选:agentMaxIterations===0 即不限(后端 0→usize::MAX,loop 靠 stop/收敛/审批退出)
 const unlimited = computed({
   get: () => settings.agentMaxIterations === 0,
   set(v: boolean) {
+    // 走查 P3-11:勾选属用户交互,标记 interactive 让「已保存」等 IPC 成功后再显
+    // (原实现在模板 @change 即时 markSaved,先于持久化 = 假成功)
+    _userInteracting = true
     if (v) {
       settings.agentMaxIterations = 0
     } else if (settings.agentMaxIterations === 0) {
       settings.agentMaxIterations = 10
     }
     syncAgentMaxIterations()
+    _userInteracting = false
   },
 })
 
@@ -191,16 +227,16 @@ function syncAgentMaxRetries() {
   // clamp 0-10，防越界
   settings.agentMaxRetries = Math.min(10, Math.max(0, settings.agentMaxRetries ?? 3))
   if (_agentRetryTimer) clearTimeout(_agentRetryTimer)
+  const interactive = _userInteracting
   _agentRetryTimer = setTimeout(async () => {
     try {
       await aiApi.setAgentMaxRetries(settings.agentMaxRetries)
+      if (interactive) markSaved('retry')
     } catch (e) {
       console.error('同步流式重试次数失败:', e)
+      if (interactive) markFailed('retry')
     }
   }, 300)
-  if (_userInteracting) {
-    markSaved('retry')
-  }
 }
 
 onMounted(() => {

src/components/settings/ProviderPanel.vue

@@ -67,7 +67,10 @@
     <div class="form-grid">
       <div class="form-field">
         <label class="form-label">{{ $t('settings.labelName') }}</label>
-        <input v-model="providerForm.name" class="setting-input" :placeholder="$t('settings.phProviderName')" />
+        <!-- 走查 P3-10:校验失败定位字段——红框 + @input 清错(示范级,仅本表单) -->
+        <input ref="inputName" v-model="providerForm.name" class="setting-input"
+               :class="{ 'setting-input--invalid': fieldErrors.name }"
+               :placeholder="$t('settings.phProviderName')" @input="fieldErrors.name = false" />
       </div>
       <!-- 项 D:厂商预设下拉 — 选中自动填充 providerType/baseUrl/defaultModel(用户可改覆盖)。
            新建态独立"快捷选择";编辑态回填后切预设仍可覆盖(等价快捷填充),不阻塞手改。
@@ -88,12 +91,16 @@
       </div>
       <div class="form-field">
         <label class="form-label">{{ $t('settings.labelBaseUrl') }}</label>
-        <input v-model="providerForm.baseUrl" class="setting-input" :placeholder="$t('settings.phBaseUrl')" />
+        <input ref="inputBaseUrl" v-model="providerForm.baseUrl" class="setting-input"
+               :class="{ 'setting-input--invalid': fieldErrors.baseUrl }"
+               :placeholder="$t('settings.phBaseUrl')" @input="fieldErrors.baseUrl = false" />
       </div>
       <div class="form-field">
         <label class="form-label">{{ $t('settings.labelApiKey') }}</label>
         <!-- 走查 P3-9:type 恒 password 无法核对已粘贴密钥,加显隐切换 -->
-        <input v-model="providerForm.apiKey" class="setting-input" :type="showApiKey ? 'text' : 'password'" :placeholder="$t('settings.phApiKey')" />
+        <input ref="inputApiKey" v-model="providerForm.apiKey" class="setting-input" :type="showApiKey ? 'text' : 'password'"
+               :class="{ 'setting-input--invalid': fieldErrors.apiKey }"
+               :placeholder="$t('settings.phApiKey')" @input="fieldErrors.apiKey = false" />
         <button type="button" class="btn-link" :title="$t(showApiKey ? 'settings.apiKeyHide' : 'settings.apiKeyShow')"
                 @click="showApiKey = !showApiKey">
           {{ $t(showApiKey ? 'settings.apiKeyHide' : 'settings.apiKeyShow') }}
@@ -104,7 +111,9 @@
       </div>
       <div class="form-field">
         <label class="form-label">{{ $t('settings.labelDefaultModel') }}</label>
-        <input v-model="providerForm.defaultModel" class="setting-input" :placeholder="$t('settings.phDefaultModel')" />
+        <input ref="inputDefaultModel" v-model="providerForm.defaultModel" class="setting-input"
+               :class="{ 'setting-input--invalid': fieldErrors.defaultModel }"
+               :placeholder="$t('settings.phDefaultModel')" @input="fieldErrors.defaultModel = false" />
         <!-- 方向 A(轻量):defaultModel 仅作"建议起点"——预设 model id 易过时(DeepSeek chat 停用 /
              Gemini 2.5→3.5 等),提示用户点「测试连接」拉真实模型列表覆盖,降低踩过时默认值的困惑。 -->
         <div class="form-field-hint">{{ $t('settings.defaultModelHint') }}</div>
@@ -112,6 +121,7 @@
       <!-- 供应商级采样温度覆盖:部分受限模型(如 Kimi 系)仅接受 temperature=1,留空走默认 0.7 -->
       <div class="form-field">
         <label class="form-label">{{ $t('settings.labelSamplingTemp') }}</label>
+        <!-- 走查 P3-14:原生 min/max 只约束步进按钮,手输越界由 saveProvider clamp + toast 提示 -->
         <input type="number" v-model.number="providerForm.temperature" class="setting-input"
                :placeholder="$t('settings.phSamplingTemp')" step="0.1" min="0" max="2" />
       </div>
@@ -192,7 +202,7 @@
 </template>
 
 <script setup lang="ts">
-import { reactive, ref, computed, onMounted } from 'vue'
+import { reactive, ref, computed, onMounted, useTemplateRef } from 'vue'
 import { useI18n } from 'vue-i18n'
 import { aiApi } from '@/api'
 import type { AiProviderConfig, ModelConfig, Modality, Capability, ProbeSource } from '@/api/types'
@@ -228,6 +238,25 @@ const aiProviders = ref<AiProviderConfig[]>([])
 // 密钥显隐切换(走查 P3-9):默认隐藏;重开表单时复位,防上一家 provider 的显钥状态延续
 const showApiKey = ref(false)
 
+// 走查 P3-10:校验失败只 toast 不定位字段——补字段级红框 + 滚动聚焦首个缺失字段。
+// 仅保存必填四项(name/baseUrl/apiKey/defaultModel)做示范级实现;红框在对应输入框
+// @input 时即清(用户开始修正就不该继续红着)。
+const fieldErrors = reactive({ name: false, baseUrl: false, apiKey: false, defaultModel: false })
+const inputName = useTemplateRef<HTMLInputElement>('inputName')
+const inputBaseUrl = useTemplateRef<HTMLInputElement>('inputBaseUrl')
+const inputApiKey = useTemplateRef<HTMLInputElement>('inputApiKey')
+const inputDefaultModel = useTemplateRef<HTMLInputElement>('inputDefaultModel')
+
+/** 聚焦首个红框字段(focus 默认行为即滚动到该元素,无需额外 scrollIntoView) */
+function focusFirstError() {
+  const el = fieldErrors.name ? inputName.value
+    : fieldErrors.baseUrl ? inputBaseUrl.value
+    : fieldErrors.apiKey ? inputApiKey.value
+    : fieldErrors.defaultModel ? inputDefaultModel.value
+    : null
+  el?.focus()
+}
+
 function errMsg(e: unknown): string { return e instanceof Error ? e.message : String(e) }
 
 const providerForm = reactive({
@@ -433,6 +462,7 @@ function openProviderForm(p?: AiProviderConfig) {
   }
   providerForm.fetching = false
   providerForm.saveError = null  // 重开表单清旧错误横幅
+  fieldErrors.name = fieldErrors.baseUrl = fieldErrors.apiKey = fieldErrors.defaultModel = false  // 重开表单清字段红框
   showApiKey.value = false  // 密钥默认隐藏,不延续上次表单的显钥状态(防窥)
   providerForm.visible = true
   // L19/M32: 记初始快照(在 visible=true、字段回填完成后),isFormDirty 据此比对。
@@ -547,10 +577,28 @@ function classifySaveError(raw: string): { title: string; detail: string } {
 async function saveProvider() {
   // apiKey:新建必填;编辑可留空(不改,FR-S1 后端保留原 DB 值)
   const needKey = !providerForm.editId
-  if (!providerForm.name || !providerForm.baseUrl || (needKey && !providerForm.apiKey) || !providerForm.defaultModel) {
+  // 走查 P3-10:逐字段判定缺失,红框标出并滚动聚焦首个缺失字段(toast 保留,双通道反馈)
+  fieldErrors.name = !providerForm.name
+  fieldErrors.baseUrl = !providerForm.baseUrl
+  fieldErrors.apiKey = needKey && !providerForm.apiKey
+  fieldErrors.defaultModel = !providerForm.defaultModel
+  if (fieldErrors.name || fieldErrors.baseUrl || fieldErrors.apiKey || fieldErrors.defaultModel) {
     emit('toast', t('settings.toastSaveIncomplete'), 'warning')
+    focusFirstError()
     return
   }
+  // 走查 P3-14:采样温度钳制 [0,2](原生 min/max 只约束步进按钮,手输 99 照存)。
+  // 越界有限值 → clamp + toast 告知修正值;NaN/清空残留 → 归一 null(走默认 0.7,同步清显示)
+  let temperatureOverride: number | undefined
+  if (typeof providerForm.temperature === 'number' && Number.isFinite(providerForm.temperature)) {
+    if (providerForm.temperature < 0 || providerForm.temperature > 2) {
+      providerForm.temperature = Math.min(2, Math.max(0, providerForm.temperature))
+      emit('toast', t('settings.toastTempClamped', { v: providerForm.temperature }), 'warning')
+    }
+    temperatureOverride = providerForm.temperature
+  } else {
+    providerForm.temperature = null
+  }
   providerForm.saving = true
   try {
     await aiApi.saveProvider({
@@ -561,11 +609,7 @@ async function saveProvider() {
       apiKey: providerForm.apiKey,
       defaultModel: providerForm.defaultModel,
       modelConfigs: providerForm.models,
-      // 采样温度非空且为有效数字才带(清空输入 v-model.number 可能残留 ''/NaN,不过滤会污染后端 config)
-      temperatureOverride:
-        typeof providerForm.temperature === 'number' && Number.isFinite(providerForm.temperature)
-          ? providerForm.temperature
-          : undefined,
+      temperatureOverride,
     })
     // 成功明确反馈——清错误横幅 + 关表单 + 绿色「已保存」toast
     providerForm.saveError = null
@@ -788,6 +832,9 @@ onMounted(loadProviders)
 /* 输入/下拉在 form-field 内 flex 填满(基础样式走全局 settings.css .setting-input/.setting-select) */
 .form-field .setting-input,
 .form-field .setting-select { flex: 1; min-width: 0; }
+/* 走查 P3-10:校验缺失字段红框(@input 即清,用户开始修正就不再红) */
+.setting-input--invalid { border-color: var(--df-danger) !important; }
+.setting-input--invalid:focus { border-color: var(--df-danger) !important; box-shadow: 0 0 0 1px var(--df-danger); }
 .setting-input--sm { flex: 0 0 auto; min-width: 0; width: 70px; padding: 4px 8px; font-size: 12px; }
 
 /* Toggle 小号变体(本面板负载均衡池用;.toggle 基础走全局) */

src/components/settings/SettingRow.vue

@@ -11,7 +11,11 @@
     <div class="setting-control">
       <slot />
       <Transition name="saved-fade">
-        <span v-if="savedVisible" class="saved-hint">{{ $t('settings.savedHint') }}</span>
+        <!-- 走查 P3-11 假成功族:持久化/IPC 有成功/失败两态,反馈位同理补失败红字,
+             防止「已保存」承诺落空后用户无从察觉 -->
+        <span v-if="hintState !== 'none'" class="saved-hint" :class="{ 'saved-hint--fail': hintState === 'failed' }">
+          {{ hintState === 'saved' ? $t('settings.savedHint') : $t('settings.saveFailedHint') }}
+        </span>
       </Transition>
     </div>
   </div>
@@ -21,9 +25,10 @@
 import { shallowRef } from 'vue'
 
 // ============================================================
-// 统一设置行:label/desc 入 props,控件走默认 slot,变更后父级调 markSaved() 显反馈。
+// 统一设置行:label/desc 入 props,控件走默认 slot,变更后父级调 markSaved() 显反馈,
+// 持久化失败时调 markFailed() 显红字(走查 P3-11:只报成功不报失败 = 半个反馈位)。
 // nextRound=true 显"下次对话生效" badge(性能域并发/轮次/重试白名单项)。
-// savedVisible 用 shallowRef(布尔翻转无深响应需求);timer ref 持有供连续触发时清旧。
+// hintState 用 shallowRef(字符串翻转无深响应需求);timer ref 持有供连续触发时清旧。
 // ============================================================
 defineProps<{
   /** 设置项标题 */
@@ -34,30 +39,41 @@ defineProps<{
   nextRound?: boolean
 }>()
 
-const savedVisible = shallowRef(false)
+const hintState = shallowRef<'none' | 'saved' | 'failed'>('none')
 const timer = shallowRef<ReturnType<typeof setTimeout> | null>(null)
 
-/** 显"✓已保存"行内提示,2s 后消失。连续触发清旧 timer 防提前消失。 */
-function markSaved() {
-  savedVisible.value = true
+/** 点亮行内反馈提示,2s 后消失。连续触发清旧 timer 防提前消失。 */
+function flashHint(state: 'saved' | 'failed') {
+  hintState.value = state
   if (timer.value) clearTimeout(timer.value)
   timer.value = setTimeout(() => {
-    savedVisible.value = false
+    hintState.value = 'none'
     timer.value = null
   }, 2000)
 }
 
-defineExpose({ markSaved })
+/** 显"✓已保存"行内提示(持久化/IPC 成功后才调,防假成功) */
+function markSaved() {
+  flashHint('saved')
+}
+
+/** 显"✗ 保存失败"行内提示(持久化/IPC 失败路径调用) */
+function markFailed() {
+  flashHint('failed')
+}
+
+defineExpose({ markSaved, markFailed })
 </script>
 
 <style scoped>
-/* 反馈位"✓已保存" — 行内紧贴控件右侧,2s 淡出 */
+/* 反馈位"✓已保存" — 行内紧贴控件右侧,2s 淡出;失败态红字区分 */
 .saved-hint {
   margin-left: 10px;
   font-size: 12px;
   color: var(--df-accent);
   white-space: nowrap;
 }
+.saved-hint--fail { color: var(--df-danger); }
 /* badge"下次对话生效" — label 右侧小标 */
 .next-round-badge {
   display: inline-block;

src/components/settings/SettingsNav.vue

@@ -1,8 +1,9 @@
 <template>
   <!-- ═══ 设置左导航(master-detail 左栏)═══
        阶段3 UX 重构:6 类导航(外观/AI模型/知识库/性能/安全/高级)+ emoji 图标。
-       阶段4 UX 重构:顶部搜索框 → 匹配 label/desc → 高亮匹配类 + 选中首匹配类
+       阶段4 UX 重构:顶部搜索框 → 匹配 label/desc → 即时高亮匹配类 + 防抖后跳首匹配类
        + emit scroll-target(父 Settings.vue 滚动到对应 setting-row)。
+       走查 P3-17:跳转防抖 350ms + 仅命中才跳,避免每键入强制切分类打断输入。
        activeCategory 受控(双向),v-model 由 Settings.vue 绑定 + appSettings 持久化。
        底部留导入导出占位(阶段6 接入)。 -->
   <nav class="settings-nav" :class="{ 'is-horizontal': horizontal }">
@@ -52,7 +53,7 @@
 //   advanced    高级(autoExecuteMode/logLevel + 连接)
 // ============================================================
 
-import { computed, ref, watch } from 'vue'
+import { computed, ref, watch, onUnmounted } from 'vue'
 import { useI18n } from 'vue-i18n'
 import { SETTINGS_INDEX, type SettingsCategory, type SettingIndexEntry } from './searchIndex'
 
@@ -126,12 +127,28 @@ function isMatchedCategory(cat: SettingsCategory): boolean {
   return matchedCategories.value.includes(cat)
 }
 
-/** 关键词变化 → 选中首匹配类 + 上抛首匹配项 key(触发右侧滚动定位)。
- *  空关键词时不改 activeCategory(用户清空搜索应留在当前类),也不上抛。 */
+/** 关键词变化 → (防抖后)选中首匹配类 + 上抛首匹配项 key(触发右侧滚动定位)。
+ *  空关键词时不改 activeCategory(用户清空搜索应留在当前类),也不上抛。
+ *  走查 P3-17:原实现每敲一键立即切分类+滚动,浏览 A 类打字被反复拽到 B 类打断输入。
+ *  改为 350ms 防抖——连续键入只更新左侧命中高亮(即时,computed),停顿后才跳转定位;
+ *  且仅当有命中时才跳(无命中保持原地,不打断)。 */
+const JUMP_DEBOUNCE_MS = 350
+let _jumpTimer: ReturnType<typeof setTimeout> | null = null
+function cancelPendingJump() {
+  if (_jumpTimer) {
+    clearTimeout(_jumpTimer)
+    _jumpTimer = null
+  }
+}
 watch(
   () => searchKeyword.value,
-  (kw) => {
-    const trimmed = kw.trim()
+  () => {
+    cancelPendingJump()
+    if (!searchKeyword.value.trim()) return
+    _jumpTimer = setTimeout(() => {
+      _jumpTimer = null
+      // 防抖窗口后关键词可能又变了,重读当前值;无命中不跳(仅命中时跳)
+      const trimmed = searchKeyword.value.trim()
       if (!trimmed) return
       const first = matchedEntries.value[0]
       if (!first) return
@@ -140,12 +157,17 @@ watch(
       }
       // 分类切换后需等右侧 Section 挂载,父组件 watch scroll-target 时自带 nextTick
       emit('scroll-target', first.key)
+    }, JUMP_DEBOUNCE_MS)
   },
 )
 
+// 卸载清防抖 timer,防 unmount 后旧 timer fire 时 emit 到已销毁组件
+onUnmounted(cancelPendingJump)
+
 defineExpose({
   /** 清空搜索框(父组件切分类时可调用,但当前保留搜索状态便于用户继续搜) */
   clearSearch() {
+    cancelPendingJump()
     searchKeyword.value = ''
   },
 })

src/components/task/TaskDrawer.vue

@@ -65,6 +65,19 @@
               <router-link :to="`/ideas/${task.idea_id}`" class="project-link" @click="$emit('close')">{{ ideaTitle }}</router-link>
             </span>
           </div>
+          <!-- S4 关联会话:后端审计表反查的主会话,无关联时入口隐藏;点击关抽屉 + 跳会话 -->
+          <div v-if="taskConversations.length > 0" class="info-item">
+            <span class="label">{{ $t('taskDetail.relatedConversations') }}</span>
+            <span class="value conv-links">
+              <a
+                v-for="c in taskConversations"
+                :key="c.conversation_id"
+                class="conv-link"
+                :title="`${c.tool_name} · ${formatDate(c.requested_at)}`"
+                @click.prevent="handleOpenConversation(c.conversation_id)"
+              >{{ c.title || c.conversation_id }}</a>
+            </span>
+          </div>
           <div v-if="task.branch_name" class="info-item">
             <span class="label">{{ $t('taskDetail.branch') }}</span>
             <span class="value">
@@ -169,10 +182,11 @@ import { ref, computed, watch } from 'vue'
 import { useRouter } from 'vue-router'
 import { useI18n } from 'vue-i18n'
 import { Drawer } from '@arco-design/web-vue'
-import { taskApi, projectApi, ideaApi } from '@/api'
+import { taskApi, projectApi, ideaApi, aiApi } from '@/api'
 import { useProjectStore } from '@/stores/project'
 import { formatDate } from '@/utils/time'
 import { useRendered } from '@/composables/useMarkdown'
+import { useAiConversations } from '@/composables/ai/useAiConversations'
 import { taskStatusLabel, taskStatusClass, priorityLabel, priorityClass, ADVANCE_MAP } from '@/constants/project'
 import type { AdvanceAction } from '@/constants/project'
 import type { TaskRecord, ProjectRecord, IdeaRecord } from '@/api/types'
@@ -192,6 +206,11 @@ const { t } = useI18n()
 const router = useRouter()
 const store = useProjectStore()
 const { showToast } = useToast()
+const { openConversation } = useAiConversations()
+
+// 关联会话引用(S4:审计表反查,无关联=空数组 → 入口隐藏)
+type TaskConvRef = { conversation_id: string; title: string | null; tool_name: string; requested_at: string }
+const taskConversations = ref<TaskConvRef[]>([])
 
 const loading = ref(true)
 const errorMsg = ref('')
@@ -248,6 +267,9 @@ async function load() {
   if (!tid) return
   loading.value = true
   errorMsg.value = ''
+  taskConversations.value = [] // 新一轮加载先清旧引用,防切换任务后残留上一任务的会话
+  // 关联会话为辅助入口,旁路加载不阻塞主数据;失败静默(入口不显示即降级)
+  void loadTaskConversations(tid)
   try {
     const [t, ps, ideasList] = await Promise.all([
       taskApi.get(tid),
@@ -268,6 +290,21 @@ async function load() {
   }
 }
 
+/** S4:拉当前任务的关联会话(独立于主 try,查失败不拖垮抽屉);响应回来校验 taskId 未变 */
+async function loadTaskConversations(tid: string) {
+  try {
+    const refs = await aiApi.listTaskConversations(tid)
+    if (tid !== props.taskId) return // 任务已切换,旧任务响应丢弃
+    taskConversations.value = refs
+  } catch { taskConversations.value = [] }
+}
+
+/** 点击关联会话:先关抽屉(面板在抽屉下层,不关会被遮挡),再走 openConversation 跳转 */
+async function handleOpenConversation(convId: string) {
+  emit('close')
+  await openConversation(convId)
+}
+
 async function loadChildren() {
   const cur = task.value
   if (!cur || cur.parent_id) { children.value = []; return }
@@ -356,4 +393,13 @@ async function confirmSubtask() {
 .dr-badges { display: flex; align-items: center; gap: 6px; flex-wrap: wrap; }
 .dr-header-actions { display: flex; align-items: center; gap: 6px; flex-shrink: 0; }
 .loading-hint { text-align: center; padding: 40px 0; }
+/* 关联会话:多会话纵向排列,链接样式对齐 .project-link 语义(TaskDrawer 内无其 scoped 定义) */
+.conv-links { display: flex; flex-direction: column; gap: 4px; }
+.conv-link {
+  color: var(--df-accent);
+  font-weight: 500;
+  cursor: pointer;
+  word-break: break-all;
+}
+.conv-link:hover { text-decoration: underline; }
 </style>

src/composables/ai/useAiConversations.ts

@@ -656,6 +656,17 @@ function toggleSidebar() {
   persistUiState()
 }
 
+/** 跨视图跳转会话(S4 TaskDrawer「关联会话」入口):AI 无路由,跳转 = 面板展开 + 切会话。
+ * 主窗口面板未展开时先展开再切换;detached(分离窗口)下主窗口无面板,只切会话不动面板。 */
+async function openConversation(id: string) {
+  // 已展开不重复写 persistUiState(省一次 debounced SQLite 写)
+  if (!state.detached && !state.panelOpen) {
+    state.panelOpen = true
+    persistUiState()
+  }
+  await switchConversation(id)
+}
+
 export function useAiConversations() {
   return {
     loadConversations,
@@ -670,6 +681,7 @@ export function useAiConversations() {
     toggleArchivedFold,
     toggleGroupFold,
     toggleSidebar,
+    openConversation,
   }
 }
 

src/i18n/en/ai.ts

@@ -89,6 +89,19 @@ export default {
       truncated: 'truncated',
       failed: 'failed',
       meta: '{rounds} rounds · {tokens} tokens',
+      // S3: finished card is clickable to open the read-only trace (hover hint)
+      viewHint: 'Click to view execution trace',
+    },
+    // S3 sub-agent trace overlay (opened from the finished card, read-only simple render,
+    // discarded on close)
+    agentTrace: {
+      title: 'Sub-agent Trace',
+      empty: 'No messages recorded',
+      loadFail: 'Failed to load sub-agent trace',
+      roleUser: 'User',
+      roleAssistant: 'Assistant',
+      roleTool: 'Tool result',
+      roleSystem: 'System',
     },
   },
 }

src/i18n/en/aiChat.ts

@@ -292,5 +292,10 @@ export default {
     sendErrRateLimit: 'Rate limit reached, please retry later (error: {msg})',
     sendErrQuota: 'AI quota or balance insufficient, check your provider account (error: {msg})',
     sendErrNetwork: 'Network error or timeout, check your connection and retry (error: {msg})',
+
+    // ── UX-P3 chat polish additions ──
+    editEmptyBlocked: 'Edit content is empty, submission cancelled',
+    skillNoMatchQuery: 'No skill matches "{q}", send blocked',
+    backfillOverwriteConfirm: 'The input box has an unsent draft; backfilling will overwrite it. Continue?',
   },
 }

src/i18n/en/common.ts

@@ -24,5 +24,7 @@ export default {
       perPage: 'Per page',
       total: '{n} total',
     },
+    // Close-to-tray notice bubble (App.vue consumes the tray-hide-notice event; moved from hardcoded Chinese)
+    trayNoticeRunning: '{n} conversation(s) keep running in the background — the tray icon brings the window back anytime',
   },
 }

src/i18n/en/settings.ts

@@ -17,7 +17,15 @@ export default {
 
     // ===== Instant feedback (phase 5 SettingRow) =====
     savedHint: '✓ Saved',
+    // False-success family: red inline feedback for persist/IPC failure
+    saveFailedHint: '✗ Save failed',
     nextRoundBadge: 'Next message',
+    // 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}',
+    // Prompt caching toggle save feedback (previously silent)
+    toastAnthropicCacheSaved: 'Prompt caching setting saved',
+    toastAnthropicCacheFail: 'Failed to save the prompt caching setting',
 
     // ===== Connection mgmt (phase 7: experimental badge, not wired to AI yet) =====
     experimentalBadge: 'Experimental · not wired to AI tools',

src/i18n/en/taskDetail.ts

@@ -91,5 +91,7 @@ export default {
     descSaveFailed: 'Failed to save description: {msg}',
     saveDesc: 'Save',
     cancelEdit: 'Cancel',
+    // S4 related conversations (reverse lookup of main sessions that ran task tools, click to jump)
+    relatedConversations: 'Conversations',
   },
 }

src/i18n/zh-CN/ai.ts

@@ -88,6 +88,18 @@ export default {
       truncated: '已截断',
       failed: '失败',
       meta: '{rounds} 轮 · {tokens} tokens',
+      // S3:完成卡可点击查看只读轨迹(hover 提示)
+      viewHint: '点击查看执行轨迹',
+    },
+    // S3 子代理轨迹浮层(完成卡点击展开,只读简版渲染,关即弃)
+    agentTrace: {
+      title: '子代理执行轨迹',
+      empty: '暂无消息记录',
+      loadFail: '拉取子代理轨迹失败',
+      roleUser: '用户',
+      roleAssistant: '助手',
+      roleTool: '工具结果',
+      roleSystem: '系统',
     },
   },
 }

src/i18n/zh-CN/aiChat.ts

@@ -293,5 +293,10 @@ export default {
     sendErrRateLimit: '发送频率超限,请稍后重试(原始错误:{msg})',
     sendErrQuota: 'AI 服务额度或余额不足,请检查 Provider 账户(原始错误:{msg})',
     sendErrNetwork: '网络异常或请求超时,请检查网络后重试(原始错误:{msg})',
+
+    // ── UX-P3 对话域打磨追加 ──
+    editEmptyBlocked: '编辑内容为空,已取消提交',
+    skillNoMatchQuery: '未找到技能「{q}」,已阻止发送',
+    backfillOverwriteConfirm: '输入框已有未发送草稿,回填将覆盖,确定继续?',
   },
 }

src/i18n/zh-CN/common.ts

@@ -24,5 +24,7 @@ export default {
       perPage: '每页',
       total: '共 {n} 条',
     },
+    // 关窗后台运行气泡(App.vue 消费 tray-hide-notice 事件;走查 P3-15 由硬编码中文迁入)
+    trayNoticeRunning: '{n} 个会话将继续在后台运行，托盘图标可随时唤起',
   },
 }

src/i18n/zh-CN/settings.ts

@@ -17,7 +17,15 @@ export default {
 
     // ===== 即时生效反馈(阶段5 SettingRow) =====
     savedHint: '✓ 已保存',
+    // 走查 P3-11 假成功族:持久化/IPC 失败的红字反馈位
+    saveFailedHint: '✗ 保存失败',
     nextRoundBadge: '下次对话生效',
+    // 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}',
+    // 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',
 
     // ===== 连接管理(阶段7:实验功能 badge 明示暂未接入 AI) =====
     experimentalBadge: '实验功能·暂未接入 AI 工具',

src/i18n/zh-CN/taskDetail.ts

@@ -91,5 +91,7 @@ export default {
     descSaveFailed: '保存描述失败: {msg}',
     saveDesc: '保存',
     cancelEdit: '取消',
+    // S4 关联会话(审计表反查发起过任务工具的主会话,点击跳转)
+    relatedConversations: '关联会话',
   },
 }

src/views/Settings.vue

@@ -44,7 +44,7 @@
               <input
                 type="checkbox"
                 :checked="anthropicCache"
-                @change="anthropicCache = ($event.target as HTMLInputElement).checked"
+                @change="onAnthropicCacheChange(($event.target as HTMLInputElement).checked)"
               />
               <span class="toggle-slider"></span>
             </label>
@@ -57,8 +57,11 @@
           :ai-providers="aiProviders"
         />
 
-        <!-- 性能:global/per-conv concurrency/max-iterations/max-retries -->
-        <PerformanceSection v-else-if="activeCategory === 'performance'" />
+        <!-- 性能:global/per-conv concurrency/max-iterations/max-retries(@toast 接走查 P3-14 越界修正提示) -->
+        <PerformanceSection
+          v-else-if="activeCategory === 'performance'"
+          @toast="showToast"
+        />
 
         <!-- 安全:allowed_dirs -->
         <AllowedDirsPanel
@@ -153,6 +156,20 @@ onBeforeRouteLeave(async () => {
 // 值落库为 JSON 布尔,后端 "true"/"1" 判定兼容)
 const anthropicCache = appSettings.useSetting<boolean>('df-ai-anthropic-cache', false)
 
+// 走查 P3-11:原开关切换零反馈,用户无从得知是否生效。改为 set 完成后 toast 反馈成败。
+// 注:appSettings.set 当前落库失败仅 console 内吞(降级内存态),catch 是 store 上抛错误后
+// 的失败通道,今日实际不触发(登记项见批次报告),成功反馈先补齐。
+async function onAnthropicCacheChange(checked: boolean) {
+  anthropicCache.value = checked
+  try {
+    await appSettings.set('df-ai-anthropic-cache', checked)
+    showToast(ti18n('settings.toastAnthropicCacheSaved'), 'success')
+  } catch (e) {
+    console.error('保存 Prompt 缓存开关失败:', e)
+    showToast(ti18n('settings.toastAnthropicCacheFail'), 'error')
+  }
+}
+
 // 窄屏响应式:监听窗口宽度,nav 折叠为横向 tab(避免依赖仅 CSS 的 :has,
 // 并让 SettingsNav 按横向模式渲染 list)
 const isNarrowScreen = ref(false)

src/views/Tasks.vue

@@ -541,7 +541,11 @@ async function quickDelete(task: TaskRecord, childCount = 0) {
     // deleteTask 返成功布尔,失败 toast(store.error 同时走 error-banner)
     const ok = await store.deleteTask(task.id)
     if (!ok) showToast(t('tasks.err.deleteFailed'), 'error')
-  } catch (e) { console.error('删除失败:', e) }
+  } catch (e) {
+    // UX走查 P3-3: throw 型异常原仅 console.error 静默,用户无感知;对齐 quickAdvance/quickPriority 的 toast 反馈
+    console.error('删除失败:', e)
+    showToast(e?.toString() ?? t('tasks.err.deleteFailed'), 'error')
+  }
 }
 // 点击外部关闭快捷菜单
 function closeQuickMenu() { quickMenuId.value = null }