优化: aichat效率批(save无变化捷径/断路器尾窗/system缓存指纹/会话列表刷新收敛) + 新增 LLM prompt caching(cache_control开关默认关/system稳定段易变段分离治缓存命中率) + 销账

src-tauri/src/commands/ai/agentic/helpers.rs

@@ -4,8 +4,8 @@
 use tauri::{AppHandle, Emitter, Manager};
 use crate::state::AppState;
 use super::approval_timeout;
-use crate::commands::ai::prompt::{build_system_prompt, get_active_provider};
-use crate::commands::ai::knowledge_inject::inject_knowledge_into_prompt;
+use crate::commands::ai::prompt::{build_system_prompt_parts, cache_enabled, get_active_provider};
+use crate::commands::ai::knowledge_inject::{build_knowledge_context_for_conv, inject_knowledge_into_prompt};
 use crate::commands::ai::{AiChatEvent, ErrorType, SessionState, GoalEntry};
 use super::conv_state::ConvState;
 
@@ -94,7 +94,12 @@ pub async fn try_continue_agent_loop(
     };
     let lang = snap.agent_language.clone().unwrap_or_else(|| "zh-CN".to_string());
     let conv_id_owned = conv_id.to_string();
-    let system_prompt = build_system_prompt(state, &lang).await;
+    // Anthropic prompt caching 模式组装(开关关 → 与旧路径逐字等价;仅 anthropic provider 拆分,
+    // OpenAI 零影响)。续跑路径无新 @ mention,不注入 augmentation(与旧路径一致)。
+    let cache_on = cache_enabled(state).await
+        && matches!(provider_config.provider_type.as_str(), "anthropic" | "anthropic_compat");
+    let (system_prompt, base_volatile) =
+        build_system_prompt_parts(state, &lang, &[], &[], cache_on).await;
 
     let session_arc = state.ai_session.clone();
     let tools_arc = state.ai_tools.clone();
@@ -103,7 +108,23 @@ pub async fn try_continue_agent_loop(
     let knowledge_config = state.knowledge_config.lock().await.clone();
     let llm_concurrency = state.llm_concurrency.clone();
 
-    let system_prompt = inject_knowledge_into_prompt(state, conv_id, system_prompt, &knowledge_config).await;
+    // 知识注入:缓存模式拼到易变段末尾(追加消息流,不失效 system 缓存),否则拼 system
+    // (与旧路径逐字等价)。
+    let volatile_tail = if cache_on {
+        let knowledge = build_knowledge_context_for_conv(state, conv_id, &knowledge_config).await;
+        let mut tail = base_volatile;
+        if !knowledge.is_empty() {
+            tail = if tail.is_empty() { knowledge } else { format!("{}\n\n---\n{}", tail, knowledge) };
+        }
+        if tail.is_empty() { None } else { Some(tail) }
+    } else {
+        None
+    };
+    let system_prompt = if cache_on {
+        system_prompt
+    } else {
+        inject_knowledge_into_prompt(state, conv_id, system_prompt, &knowledge_config).await
+    };
 
     let max_iterations = state.agent_max_iterations.load(std::sync::atomic::Ordering::SeqCst);
     let max_retries = state.agent_max_retries.load(std::sync::atomic::Ordering::SeqCst);
@@ -141,7 +162,7 @@ pub async fn try_continue_agent_loop(
     };
 
     tauri::async_runtime::spawn(async move {
-        super::run_agentic_loop(session_arc, tools_arc, db, app_handle, provider_config, system_prompt, conv_id_owned, knowledge_config, llm_concurrency, max_iterations, max_retries, start_iteration, model_override, loop_epoch).await;
+        super::run_agentic_loop(session_arc, tools_arc, db, app_handle, provider_config, system_prompt, volatile_tail, conv_id_owned, knowledge_config, llm_concurrency, max_iterations, max_retries, start_iteration, model_override, loop_epoch).await;
     });
 }