优化: token分项显示(in/cache/out/reasoning)+ 详情面板 + base前置

token分项(各计费不同,不显 total):df-ai 解析 provider cache/reasoning(openai_compat prompt_cache_hit/miss/reasoning_tokens + anthropic cache_read/creation)+ TokenUsage 加字段(全构造点)+ AiMessage/AiCompleted/DB V39(ai_messages 加 cache_hit/miss/reasoning 列)+ message_repo 映射(持久化)+ 前端 MessageList 显 in·cache·out·reason(in=cache_miss 全价,reasoning 有才显)+ 点击 token 弹详情面板(完整 usage+缓存命中率+model)+ df-miniapp 同步

base前置(提升 prompt cache 命中率):chat.rs aug 拼 base 后(4处)+ knowledge_inject 知识拼 base 后(固定 base 前缀,cache 命中)

附修:replace_conversation 原 13 列 INSERT 丢消息级 token → 改 18 列

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

@@ -25,10 +25,16 @@ const SAVE_CONVERSATION_TIMEOUT_SECS: u64 = 5;
 ///
 /// 纯结构 + 方法:抽自 run_agentic_loop 的 `total_prompt`/`total_completion` 双计数器,
 /// 保证多轮累加、None 起始、跨 loop 实例叠加语义一致且可单测。
+///
+/// token 分项显示(2026-08-02):新增 cache_hit/cache_miss/reasoning 三计数器,
+/// 各轮 provider usage 累加,emit AiCompleted 时透传前端做分计费展示。
 #[derive(Debug, Clone, Default)]
 pub(crate) struct TokenAccumulator {
     prompt: u32,
     completion: u32,
+    cache_hit: u32,
+    cache_miss: u32,
+    reasoning: u32,
 }
 
 impl TokenAccumulator {
@@ -40,6 +46,15 @@ impl TokenAccumulator {
         self.completion = self.completion.saturating_add(completion);
     }
 
+    /// 叠加一轮全量用量(含 cache/reasoning 分项)。
+    /// 调用方传 round TokenUsage,本方法拆分 prompt/completion/cache/reasoning 累加。
+    pub(crate) fn add_usage(&mut self, u: &df_ai::provider::TokenUsage) {
+        self.add(u.prompt_tokens, u.completion_tokens);
+        self.cache_hit = self.cache_hit.saturating_add(u.prompt_cache_hit_tokens);
+        self.cache_miss = self.cache_miss.saturating_add(u.prompt_cache_miss_tokens);
+        self.reasoning = self.reasoning.saturating_add(u.reasoning_tokens);
+    }
+
     pub(crate) fn prompt(&self) -> u32 {
         self.prompt
     }
@@ -51,6 +66,18 @@ impl TokenAccumulator {
     pub(crate) fn total(&self) -> u32 {
         self.prompt.saturating_add(self.completion)
     }
+
+    pub(crate) fn cache_hit(&self) -> u32 {
+        self.cache_hit
+    }
+
+    pub(crate) fn cache_miss(&self) -> u32 {
+        self.cache_miss
+    }
+
+    pub(crate) fn reasoning(&self) -> u32 {
+        self.reasoning
+    }
 }
 
 /// 把单轮增量叠加到 DB 的 Option<i64> 字段(读旧值+增量,跨 loop 实例防覆盖)
@@ -611,6 +638,11 @@ mod tests {
             model: None,
             status: None,
             reasoning_content: None,
+            prompt_tokens: None,
+            completion_tokens: None,
+            prompt_cache_hit_tokens: None,
+            prompt_cache_miss_tokens: None,
+            reasoning_tokens: None,
             timestamp: None,
         }
     }