docs: 更新 README feature 表 + 升级指南 + 示例注释 + roadmap 同步

- README 添加 feature 组合表 + 模块级 features 清单 + 升级指南
- 18 个 example 顶部添加 Required features 注释
- roadmap.md 和 roadmap-v0.3.2.md 同步 Phase 26-27 完成状态
- cargo fmt 全量格式化（修复预存格式问题，CI format job 可通过）

README.md

@@ -38,7 +38,7 @@ use std::sync::Arc;
 use agcore::agent::{Agent, AgentBuilder, AgentSession};
 use agcore::llm::hooks::HookExecutor;
 use agcore::llm::mock::MockProvider;
-use agcore::llm::provider::LlmProvider;
+use agcore::llm::LlmProvider;
 use agcore::llm::types::message::{ContentBlock, Message};
 use agcore::llm::types::response_v2::{MessageResponse, StopReason};
 use agcore::llm::types::Usage;
@@ -133,6 +133,89 @@ let provider = create_provider(
 | `bridge_keys_demo` | 桥接键：Agent 间上下文键值透传 |
 | `agent_switch_demo` | Agent 热切换：会话中动态切换 Agent 角色 |
 
+## Feature 组合
+
+AG Core 通过 Cargo features 让下游按需选择模块，跳过不需要的编译单元和重型依赖。`default = ["full"]` 保持向后兼容——不指定 features 时行为与 v0.3.0 一致。
+
+### 快捷组合
+
+| 组合 | 场景 | 包含的 features |
+|------|------|----------------|
+| `full`（default） | 全栈使用，兼容 v0.3.0 | 全部 16 个 feature |
+| `light` | 生产常用，跳过 Anthropic/DeepSeek/Qwen/Ollama | llm + provider-openai + tools + tools-mcp + memory + agent + engine + prompt + document |
+| `chat` | 纯对话（跳过 SQLite 和 MCP） | agent + provider-openai |
+| `multi` | 多 Agent 复合（chat + subagent + switch + checkpointer） | engine + provider-openai |
+
+### Cargo.toml 配置示例
+
+```toml
+# 默认全栈（兼容 v0.3.0）
+[dependencies]
+agcore = "0.3"
+
+# 纯对话场景：跳过 SQLite 和 MCP，编译更快
+[dependencies]
+agcore = { version = "0.3", default-features = false, features = ["chat", "provider-openai"] }
+
+# 生产常用：OpenAI + 工具 + 记忆 + Agent
+[dependencies]
+agcore = { version = "0.3", default-features = false, features = ["light"] }
+
+# 多 Agent 复合 + MCP 工具
+[dependencies]
+agcore = { version = "0.3", default-features = false, features = ["multi", "provider-openai", "tools-mcp"] }
+```
+
+### 模块级 features
+
+如需更细粒度控制，可单独启用模块级 features：
+
+| Feature | 覆盖内容 | imply |
+|---------|---------|-------|
+| `document` | Document + RecursiveCharacterSplitter | — |
+| `llm-types` | Message / ToolDef / Usage 等 IR 类型 | — |
+| `prompt` | PromptTemplate + PromptComposer | `llm-types` |
+| `llm` | Provider trait + LlmCycle + hooks + compact + embedding + mock | `llm-types` |
+| `tools` | BaseTool + ToolRegistry | `llm-types` |
+| `tools-mcp` | McpClient（Stdio/StreamableHttp） | `tools` |
+| `memory` | MemoryStore + Conversation + VectorStore + KnowledgeGraph + Retriever | `document` + `llm` |
+| `memory-sqlite` | SqliteStore | `memory` |
+| `agent` | Agent + Builder + Session + ContextSlot + Summary | `llm` + `tools` + `memory` |
+| `engine` | SessionManager + Checkpointer + SubAgent + Switch | `agent` |
+| `provider-openai` | OpenAI Provider 实现 | `llm` |
+| `provider-anthropic` | Anthropic Provider 实现 | `llm` |
+| `provider-deepseek` | DeepSeek Provider 实现 | `llm` |
+| `provider-qwen` | Qwen Provider 实现 | `llm` |
+| `provider-ollama` | Ollama Provider 实现 | `llm` |
+| `tracing-init` | `init_tracing()` 函数 | — |
+
+## 升级指南（v0.3.0 → v0.3.2）
+
+### LlmProvider trait 路径变更
+
+v0.3.2 起，`LlmProvider` trait 及其关联类型 `ProviderCapabilities` / `ProviderFeatures` 从 `provider` 模块移至 `llm` 模块根级别，归属 `#[cfg(feature = "llm")]` 而非 `any(provider-*)`。纯 Mock 场景不再需要引入任何 provider feature。
+
+| 旧路径（v0.3.0） | 新路径（v0.3.2） |
+|------------------|------------------|
+| `agcore::llm::provider::LlmProvider` | `agcore::llm::LlmProvider` |
+| `agcore::llm::provider::ProviderCapabilities` | `agcore::llm::ProviderCapabilities` |
+| `agcore::llm::provider::ProviderFeatures` | `agcore::llm::ProviderFeatures` |
+
+**向后兼容**：`provider` 模块中保留了 `pub use` 重导出，老路径仍可编译。但推荐迁移至新路径，未来版本可能移除重导出。
+
+`ProviderConfig` / `ProviderType` / `create_provider()` 等 provider 创建逻辑仍在 `agcore::llm::provider` 下，无需迁移。
+
+### 迁移步骤
+
+```bash
+# 1. 全局替换 use 路径
+sed -i 's/agcore::llm::provider::LlmProvider/agcore::llm::LlmProvider/g' src/**/*.rs
+sed -i 's/agcore::llm::provider::{LlmProvider/agcore::llm::{LlmProvider/g' src/**/*.rs
+
+# 2. 验证编译
+cargo build --features "full"
+```
+
 ## 核心模块
 
 | 模块 | 一句话说明 |

docs/roadmap-v0.3.2.md

@@ -1,11 +1,11 @@
 # AG Core Roadmap — v0.3.2
 
-**状态**：✅ Phase 20-25 已交付（Step 1 合并实施）；⏳ Phase 26-27 待实施
+**状态**：✅ Phase 20-27 全部交付（v0.3.2 交付完毕）
 
 > 本文件聚焦 **v0.3.2 版本** 的规划与交付（Phase 20–27）。
 > 返回总入口：[`roadmap.md`](./roadmap.md)
 
-> **进度更新（2026-07-19）**：Step 1 合并完成 Phase 20（Cargo.toml features 定义 + 依赖 optional 化）+ Phase 21-25 的全部 cfg 门控（因编译阻塞提前）。验证矩阵 14 条编译命令全部通过，`cargo test -F full` 427 passed。实施方案见 [`docs/26-step1-phase20-cargo-features-implementation.md`](./26-step1-phase20-cargo-features-implementation.md)。下一步：Phase 26（CI 测试矩阵）+ Phase 27（文档更新）。
+> **进度更新（2026-07-19）**：v0.3.2 全部交付。Step 1 完成 Phase 20-25（Cargo features 定义 + 依赖 optional 化 + 全模块 cfg 门控）；Step 3 完成 Phase 26-27（CI 测试矩阵固化 + LlmProvider trait 归属修正 + examples required-features + 文档更新）。验证矩阵 6 种组合全部通过，clippy 0 警告。实施方案见 [`docs/26-step1-phase20-cargo-features-implementation.md`](./26-step1-phase20-cargo-features-implementation.md) 和 [`docs/27-step3-phase26-ci-verification.md`](./27-step3-phase26-ci-verification.md)。
 
 ## v0.3.2 愿景
 
@@ -218,7 +218,7 @@
 **依赖**：Phase 20-25（所有门控就绪）
 **优先级**：P0
 **预估规模**：约 10 行（CI 配置）
-**状态**：⏳ 待实施
+**状态**：✅ 已交付（2026-07-19）— 6 种 feature 组合测试矩阵 + clippy + format + examples 验证 job 全部通过；`RUSTFLAGS=-D warnings` 强制零警告；LlmProvider trait 归属修正 + bundle() 门控 + llm feature 补充 imply futures-util
 
 ---
 
@@ -235,7 +235,7 @@
 **依赖**：Phase 20-26
 **优先级**：P0
 **预估规模**：约 100 行
-**状态**：⏳ 待实施
+**状态**：✅ 已交付（2026-07-19）— README 添加 feature 表格 + 快捷组合 + 模块级 features 清单 + 升级指南（LlmProvider 路径迁移）；18 个 example 顶部添加 Required features 注释；roadmap 总入口同步
 
 ---
 
@@ -305,8 +305,8 @@ graph TD
 | **M19** | Phase 23 | tools 不含 mcp 编译通过；加 tools-mcp 引入 McpClient | ✅ 2026-07-19（Step 1 合并） |
 | **M20** | Phase 24 | memory 不含 llm/sqlite 编译通过；加 memory-sqlite 引入 SqliteStore | ✅ 2026-07-19（Step 1 合并） |
 | **M21** | Phase 25 | agent + engine 全链路门控编译通过 | ✅ 2026-07-19（Step 1 合并） |
-| **M22** | Phase 26 | 7 种 CI 组合全部编译 + 测试通过；clippy --all-features 0 警告 | ⏳ |
-| **M23** | Phase 27 | 文档 review 通过 | ⏳ |
+| **M22** | Phase 26 | 7 种 CI 组合全部编译 + 测试通过；clippy --all-features 0 警告 | ✅ 2026-07-19 |
+| **M23** | Phase 27 | 文档 review 通过 | ✅ 2026-07-19 |
 
 ## Cargo.toml [features] 草案
 

docs/roadmap.md

@@ -1,7 +1,7 @@
 # AG Core Roadmap
 
 > 拆分式 roadmap：按版本归档 + 未归类内容
-> 最后更新：2026-07-19（v0.3.2 Step 1 完成 — Phase 20-25 cfg 门控全部交付，427 测试通过）
+> 最后更新：2026-07-19（v0.3.2 Step 3 完成 — Phase 26-27 CI 固化 + 文档更新交付，427 测试通过）
 
 ## 文件索引
 
@@ -10,7 +10,7 @@
 | [`roadmap-v0.1.0.md`](./roadmap-v0.1.0.md) | v0.1.0 计划与交付 — Phase 0–4c + v0.1.0 Release | ✅ 已发布 2026-07-04 |
 | [`roadmap-v0.2.0.md`](./roadmap-v0.2.0.md) | v0.2.0 计划与交付 — Phase 5–12 + v0.2.0-rc.1 | 🟡 Phase 5-11 已完成；Phase 12 P2 锦上添花可选 |
 | [`roadmap-v0.3.0.md`](./roadmap-v0.3.0.md) | v0.3.0 计划与交付 - Phase 13–19 | ✅ Phase 13-19 全部完成，v0.3.0 交付完毕 |
-| [`roadmap-v0.3.2.md`](./roadmap-v0.3.2.md) | v0.3.2 计划与交付 — Phase 20–27（Cargo features 拆分） | 🟡 Step 1 完成（Phase 20-25）；Phase 26-27 待实施 |
+| [`roadmap-v0.3.2.md`](./roadmap-v0.3.2.md) | v0.3.2 计划与交付 — Phase 20–27（Cargo features 拆分） | ✅ Phase 20-27 全部完成，v0.3.2 交付完毕 |
 | [`roadmap-unsorted.md`](./roadmap-unsorted.md) | 未归到任何版本的内容 — 全局愿景、当前状态、模块完整性、v0.4+ 展望、风险与建议、下一步行动、阶段总回顾 | — |
 
 ## 阅读建议

examples/agent_session_demo.rs

@@ -1,4 +1,5 @@
 //! agent_session_demo —— Agent 装配 + 会话链路 + SessionMemory 桥接。
+//! Required features: cargo run --example agent_session_demo --features "agent"
 //!
 //! 演示：
 //! 1. 实现 `Agent` trait（定义角色 + system prompt）

examples/agent_switch_demo.rs

@@ -1,4 +1,5 @@
 //! agent_switch_demo —— Agent 角色热切换示例。
+//! Required features: cargo run --example agent_switch_demo --features "engine"
 //!
 //! 演示：
 //! 1. 创建 session（绑定 Analyst agent）
@@ -110,6 +111,9 @@ async fn main() {
     };
     println!("\n[verify] turn_index = {turn_index}, agent = {agent_name_owned}");
     assert_eq!(turn_index, 2, "turn_index should be 2 after 2 turns");
-    assert_eq!(agent_name_owned, "reporter", "current agent should be reporter");
+    assert_eq!(
+        agent_name_owned, "reporter",
+        "current agent should be reporter"
+    );
     println!("✓ context preserved across agent switch");
 }

examples/bridge_keys_demo.rs

@@ -1,4 +1,5 @@
 //! bridge_keys_demo —— bridge_keys 过滤 + 子↔子共享 namespace 示例。
+//! Required features: cargo run --example bridge_keys_demo --features "engine"
 //!
 //! 演示：
 //! 1. 父 session 设置 SessionMemory（key: "project_goal", "constraints", "noise"）
@@ -103,13 +104,24 @@ async fn main() {
         .dispatch(&parent_id, worker.clone(), "do work", config)
         .await
         .expect("dispatch");
-    println!("[3] dispatched sub-agent (child_id={})\n", &result.child_id[..20]);
+    println!(
+        "[3] dispatched sub-agent (child_id={})\n",
+        &result.child_id[..20]
+    );
 
     // 验证过滤效果
     let child_session = sm.get(&result.child_id).await.unwrap();
     let child_guard = child_session.lock().await;
-    let inherited_goal = child_guard.session_memory().get("project_goal").await.unwrap();
-    let inherited_constraint = child_guard.session_memory().get("constraints").await.unwrap();
+    let inherited_goal = child_guard
+        .session_memory()
+        .get("project_goal")
+        .await
+        .unwrap();
+    let inherited_constraint = child_guard
+        .session_memory()
+        .get("constraints")
+        .await
+        .unwrap();
     let filtered_noise = child_guard.session_memory().get("noise").await.unwrap();
     drop(child_guard);
 
@@ -172,12 +184,7 @@ async fn main() {
 
     // dispatch 第二个子 agent
     let _writer_result = sm
-        .dispatch(
-            &parent_id,
-            worker.clone(),
-            "writing task",
-            shared_ns_config,
-        )
+        .dispatch(&parent_id, worker.clone(), "writing task", shared_ns_config)
         .await
         .expect("dispatch writer");
 

examples/context_slot_demo.rs

@@ -1,4 +1,5 @@
 //! context_slot_demo —— 多上下文槽位管理示例。
+//! Required features: cargo run --example context_slot_demo --features "agent"
 //!
 //! 场景：法律咨询入口 → 派生两个独立探索方向 → 切换 → 隔离验证 → 删除。
 //!
@@ -13,12 +14,12 @@
 use std::sync::Arc;
 
 use agcore::agent::{Agent, AgentBuilder, AgentSession};
+use agcore::llm::LlmProvider;
 use agcore::llm::hooks::HookExecutor;
 use agcore::llm::mock::MockProvider;
-use agcore::llm::LlmProvider;
+use agcore::llm::types::Usage;
 use agcore::llm::types::message::{ContentBlock, Message};
 use agcore::llm::types::response_v2::{MessageResponse, StopReason};
-use agcore::llm::types::Usage;
 use agcore::tools::ToolRegistry;
 
 struct LegalAdvisor;
@@ -78,11 +79,19 @@ async fn main() {
 
     println!("\n=== 3. 派生两个独立探索方向的 slot ===");
     session
-        .derive_slot("option_jurisdiction", "default", agcore::agent::DeriveStrategy::Full)
+        .derive_slot(
+            "option_jurisdiction",
+            "default",
+            agcore::agent::DeriveStrategy::Full,
+        )
         .await
         .unwrap();
     session
-        .derive_slot("option_amendment", "default", agcore::agent::DeriveStrategy::Full)
+        .derive_slot(
+            "option_amendment",
+            "default",
+            agcore::agent::DeriveStrategy::Full,
+        )
         .await
         .unwrap();
     let slots: Vec<_> = session.list_slots().cloned().collect();
@@ -158,4 +167,4 @@ fn message_contains(msg: &Message, needle: &str) -> bool {
         }
     }
     false
-}
+}

examples/conversation_memory_demo.rs

@@ -1,4 +1,5 @@
 //! conversation_memory_demo —— 对话记忆滑动窗口与隔离。
+//! Required features: cargo run --example conversation_memory_demo --features "memory"
 //!
 //! 演示：
 //! 1. `ConversationMemoryConfig` 构造（SlidingWindow / Full 策略）

examples/custom_tool.rs

@@ -1,4 +1,5 @@
 //! custom_tool —— 自定义工具注册、单次 / 并行调用、权限检查。
+//! Required features: cargo run --example custom_tool --features "tools,llm"
 //!
 //! 演示：
 //! 1. 实现 `BaseTool` trait（WeatherTool + DeleteFileTool）

examples/dispatch_stream_demo.rs

@@ -1,4 +1,5 @@
 //! dispatch_stream_demo —— 流式子代理调度示例。
+//! Required features: cargo run --example dispatch_stream_demo --features "engine"
 //!
 //! 演示：
 //! 1. 创建父 session
@@ -92,7 +93,11 @@ async fn main() {
                 saw_stream_count += 1;
             }
             SubTaskStreamEvent::Completed(r) => {
-                println!("  → Completed(child_id={}, {} tokens)", &r.child_id[..20], r.usage.total().total_tokens);
+                println!(
+                    "  → Completed(child_id={}, {} tokens)",
+                    &r.child_id[..20],
+                    r.usage.total().total_tokens
+                );
                 completed = Some(r);
                 break;
             }
@@ -114,7 +119,10 @@ async fn main() {
     let child_guard = child_session.lock().await;
     let child_turn_index = child_guard.turn_index();
     drop(child_guard);
-    assert_eq!(child_turn_index, 1, "turn_index should increment after finalize");
+    assert_eq!(
+        child_turn_index, 1,
+        "turn_index should increment after finalize"
+    );
     println!("[4] child session turn_index = {child_turn_index} (finalize works)");
 
     println!("\n✓ dispatch_stream completed successfully");

examples/document_demo.rs

@@ -1,4 +1,5 @@
 //! document_demo —— Document + RecursiveCharacterSplitter + MockEmbedding + RagPipeline 完整衔接示例。
+//! Required features: cargo run --example document_demo --features "memory,tracing-init"
 //!
 //! 演示 RAG 管线：
 //! 1. 创建多段落 Document
@@ -37,11 +38,7 @@ async fn main() {
     let embedder: Arc<dyn Embedding> = Arc::new(MockEmbedding::new(4));
     let store: Arc<dyn VectorStore> = Arc::new(InMemoryVectorStore::new());
     let splitter = RecursiveCharacterSplitter::new(200, 30);
-    let pipeline = RagPipeline::new(
-        Arc::clone(&embedder),
-        Arc::clone(&store),
-        Some(splitter),
-    );
+    let pipeline = RagPipeline::new(Arc::clone(&embedder), Arc::clone(&store), Some(splitter));
 
     // 3. 一次性 ingest：自动 split → embed → add
     pipeline.ingest(std::slice::from_ref(&doc)).await.unwrap();
@@ -67,4 +64,4 @@ async fn main() {
     );
 
     println!("\n✓ document_demo 完成");
-}
+}

examples/end_to_end.rs

@@ -1,4 +1,5 @@
 //! end_to_end —— 3 工具 + 3 轮对话 + SqliteStore 持久化跨连接验证。
+//! Required features: cargo run --example end_to_end --features "agent,memory-sqlite,provider-openai"
 //!
 //! 运行：`cargo run --example end_to_end`（离线，零配置）
 //!
@@ -16,11 +17,15 @@ use std::env;
 use std::sync::Arc;
 
 use agcore::agent::{Agent, AgentBuilder, AgentSession};
+use agcore::llm::LlmProvider;
 use agcore::llm::hooks::HookExecutor;
 use agcore::llm::mock::MockProvider;
-use agcore::llm::LlmProvider;
-use agcore::llm::provider::{create_provider, ProviderConfig, ProviderType};
-use agcore::llm::types::{Usage, message::{ContentBlock, Message}, response_v2::{MessageResponse, StopReason}};
+use agcore::llm::provider::{ProviderConfig, ProviderType, create_provider};
+use agcore::llm::types::{
+    Usage,
+    message::{ContentBlock, Message},
+    response_v2::{MessageResponse, StopReason},
+};
 use agcore::memory::store::{MemoryStore, SqliteStore};
 use agcore::memory::types::{MemoryFilter, MemoryItem};
 use agcore::tools::{BaseTool, ToolContext, ToolError, ToolRegistry};
@@ -33,8 +38,12 @@ use time::OffsetDateTime;
 
 struct AssistantAgent;
 impl Agent for AssistantAgent {
-    fn name(&self) -> &str { "end-to-end assistant" }
-    fn system_prompt(&self) -> Option<&str> { Some("简洁助手，必要时调用工具完成任务。") }
+    fn name(&self) -> &str {
+        "end-to-end assistant"
+    }
+    fn system_prompt(&self) -> Option<&str> {
+        Some("简洁助手，必要时调用工具完成任务。")
+    }
 }
 
 // === Tools ===
@@ -42,14 +51,19 @@ impl Agent for AssistantAgent {
 struct EchoTool;
 #[async_trait]
 impl BaseTool for EchoTool {
-    fn name(&self) -> &str { "echo" }
-    fn description(&self) -> &str { "回显输入文本" }
+    fn name(&self) -> &str {
+        "echo"
+    }
+    fn description(&self) -> &str {
+        "回显输入文本"
+    }
     fn parameters(&self) -> Value {
         json!({"type":"object","properties":{"text":{"type":"string"}},"required":["text"]})
     }
     async fn execute(&self, args: Value, _: &ToolContext<'_>) -> Result<Value, ToolError> {
-        let text = args.get("text").and_then(|v| v.as_str())
-            .ok_or_else(|| ToolError::InvalidArguments("text".into(), "需要 string 类型的 text 参数".into()))?;
+        let text = args.get("text").and_then(|v| v.as_str()).ok_or_else(|| {
+            ToolError::InvalidArguments("text".into(), "需要 string 类型的 text 参数".into())
+        })?;
         Ok(json!({"echoed": format!("收到: {text}")}))
     }
 }
@@ -58,8 +72,12 @@ impl BaseTool for EchoTool {
 struct CalcTool;
 #[async_trait]
 impl BaseTool for CalcTool {
-    fn name(&self) -> &str { "calc" }
-    fn description(&self) -> &str { "四则运算：'a op b' 格式，op ∈ {+, -, *, /}" }
+    fn name(&self) -> &str {
+        "calc"
+    }
+    fn description(&self) -> &str {
+        "四则运算：'a op b' 格式，op ∈ {+, -, *, /}"
+    }
     fn parameters(&self) -> Value {
         json!({"type":"object","properties":{"expr":{"type":"string"}},"required":["expr"]})
     }
@@ -67,18 +85,30 @@ impl BaseTool for CalcTool {
         let expr = args["expr"].as_str().unwrap_or("");
         let parts: Vec<&str> = expr.split_whitespace().collect();
         if parts.len() != 3 {
-            return Err(ToolError::InvalidArguments("expr".into(), "需要 'a op b' 三段式".into()));
+            return Err(ToolError::InvalidArguments(
+                "expr".into(),
+                "需要 'a op b' 三段式".into(),
+            ));
         }
-        let a: i64 = parts[0].parse().map_err(|_| ToolError::InvalidArguments("expr".into(), format!("无法解析 '{}'", parts[0])))?;
-        let b: i64 = parts[2].parse().map_err(|_| ToolError::InvalidArguments("expr".into(), format!("无法解析 '{}'", parts[2])))?;
+        let a: i64 = parts[0].parse().map_err(|_| {
+            ToolError::InvalidArguments("expr".into(), format!("无法解析 '{}'", parts[0]))
+        })?;
+        let b: i64 = parts[2].parse().map_err(|_| {
+            ToolError::InvalidArguments("expr".into(), format!("无法解析 '{}'", parts[2]))
+        })?;
         let result = match parts[1] {
             "+" => a + b,
             "-" => a - b,
             "*" => a * b,
-            "/" => a.checked_div(b).ok_or_else(|| {
-                ToolError::InvalidArguments("expr".into(), "除数不能为 0".into())
-            })?,
-            op => return Err(ToolError::InvalidArguments("expr".into(), format!("不支持的运算符: {op}"))),
+            "/" => a
+                .checked_div(b)
+                .ok_or_else(|| ToolError::InvalidArguments("expr".into(), "除数不能为 0".into()))?,
+            op => {
+                return Err(ToolError::InvalidArguments(
+                    "expr".into(),
+                    format!("不支持的运算符: {op}"),
+                ));
+            }
         };
         Ok(json!({"result": result}))
     }
@@ -87,13 +117,21 @@ impl BaseTool for CalcTool {
 /// 通过 MemoryStore trait 读写笔记：直接持有 Arc<dyn MemoryStore>，
 /// 绕开 AgentSession 封装（NoteTool 在 tool.execute 中直接操作 store）。
 /// 关键前缀 "note:" 用于 list 过滤。
-struct NoteTool { store: Arc<dyn MemoryStore> }
-impl NoteTool { const PREFIX: &'static str = "note:"; }
+struct NoteTool {
+    store: Arc<dyn MemoryStore>,
+}
+impl NoteTool {
+    const PREFIX: &'static str = "note:";
+}
 
 #[async_trait]
 impl BaseTool for NoteTool {
-    fn name(&self) -> &str { "note" }
-    fn description(&self) -> &str { "笔记 save/query: save(key, content) / query()" }
+    fn name(&self) -> &str {
+        "note"
+    }
+    fn description(&self) -> &str {
+        "笔记 save/query: save(key, content) / query()"
+    }
     fn parameters(&self) -> Value {
         json!({
             "type":"object",
@@ -117,18 +155,29 @@ impl BaseTool for NoteTool {
                     metadata: json!({}),
                     created_at: OffsetDateTime::now_utc(),
                 };
-                self.store.save(item).await
+                self.store
+                    .save(item)
+                    .await
                     .map_err(|e| ToolError::ExecutionFailed("note".into(), e.to_string()))?;
                 Ok(json!({"saved": key}))
             }
             "query" => {
-                let filter = MemoryFilter { prefix: Some(Self::PREFIX.into()), ..Default::default() };
-                let items = self.store.list(&filter).await
+                let filter = MemoryFilter {
+                    prefix: Some(Self::PREFIX.into()),
+                    ..Default::default()
+                };
+                let items = self
+                    .store
+                    .list(&filter)
+                    .await
                     .map_err(|e| ToolError::ExecutionFailed("note".into(), e.to_string()))?;
                 let notes: Vec<String> = items.into_iter().map(|i| i.content).collect();
                 Ok(json!({"notes": notes}))
             }
-            _ => Err(ToolError::InvalidArguments("action".into(), format!("未知 action: {action}"))),
+            _ => Err(ToolError::InvalidArguments(
+                "action".into(),
+                format!("未知 action: {action}"),
+            )),
         }
     }
 }
@@ -136,31 +185,91 @@ impl BaseTool for NoteTool {
 // === Mock response helper ===
 
 fn resp(content: Vec<ContentBlock>, stop: StopReason, u: (u32, u32)) -> MessageResponse {
-    MessageResponse { id: String::new(), model: "mock".into(),
+    MessageResponse {
+        id: String::new(),
+        model: "mock".into(),
         message: Message::Assistant { content },
         usage: Usage::from_input_output(u.0, u.1),
-        stop_reason: stop, extra: Default::default() }
+        stop_reason: stop,
+        extra: Default::default(),
+    }
 }
 
 fn mock_responses() -> Vec<MessageResponse> {
     vec![
         // 第 1 轮：calc(25 * 4) → tool_result(100) → 文本回答
-        resp(vec![ContentBlock::ToolUse { id: "t1".into(), name: "calc".into(),
-                input: json!({"expr": "25 * 4"}) }], StopReason::ToolUse, (5, 8)),
-        resp(vec![ContentBlock::Text { text: "25 * 4 = 100".into() }], StopReason::Stop, (8, 12)),
+        resp(
+            vec![ContentBlock::ToolUse {
+                id: "t1".into(),
+                name: "calc".into(),
+                input: json!({"expr": "25 * 4"}),
+            }],
+            StopReason::ToolUse,
+            (5, 8),
+        ),
+        resp(
+            vec![ContentBlock::Text {
+                text: "25 * 4 = 100".into(),
+            }],
+            StopReason::Stop,
+            (8, 12),
+        ),
         // 第 2 轮：note(save, last_calc, "100") → tool_result(saved) → 文本回答
-        resp(vec![ContentBlock::ToolUse { id: "t2".into(), name: "note".into(),
-                input: json!({"action": "save", "key": "last_calc", "content": "100"}) }],
-             StopReason::ToolUse, (10, 14)),
-        resp(vec![ContentBlock::Text { text: "已记录：last_calc = 100".into() }], StopReason::Stop, (12, 16)),
+        resp(
+            vec![ContentBlock::ToolUse {
+                id: "t2".into(),
+                name: "note".into(),
+                input: json!({"action": "save", "key": "last_calc", "content": "100"}),
+            }],
+            StopReason::ToolUse,
+            (10, 14),
+        ),
+        resp(
+            vec![ContentBlock::Text {
+                text: "已记录：last_calc = 100".into(),
+            }],
+            StopReason::Stop,
+            (12, 16),
+        ),
         // 第 3 轮：note(query) → tool_result([100]) → 文本回答
-        resp(vec![ContentBlock::ToolUse { id: "t3".into(), name: "note".into(),
-                input: json!({"action": "query"}) }], StopReason::ToolUse, (8, 8)),
-        resp(vec![ContentBlock::Text { text: "您刚才的计算结果是 100".into() }], StopReason::Stop, (10, 14)),
+        resp(
+            vec![ContentBlock::ToolUse {
+                id: "t3".into(),
+                name: "note".into(),
+                input: json!({"action": "query"}),
+            }],
+            StopReason::ToolUse,
+            (8, 8),
+        ),
+        resp(
+            vec![ContentBlock::Text {
+                text: "您刚才的计算结果是 100".into(),
+            }],
+            StopReason::Stop,
+            (10, 14),
+        ),
         // 后续冗余响应（防止队列耗尽报错）
-        resp(vec![ContentBlock::Text { text: "done".into() }], StopReason::Stop, (1, 1)),
-        resp(vec![ContentBlock::Text { text: "done".into() }], StopReason::Stop, (1, 1)),
-        resp(vec![ContentBlock::Text { text: "done".into() }], StopReason::Stop, (1, 1)),
+        resp(
+            vec![ContentBlock::Text {
+                text: "done".into(),
+            }],
+            StopReason::Stop,
+            (1, 1),
+        ),
+        resp(
+            vec![ContentBlock::Text {
+                text: "done".into(),
+            }],
+            StopReason::Stop,
+            (1, 1),
+        ),
+        resp(
+            vec![ContentBlock::Text {
+                text: "done".into(),
+            }],
+            StopReason::Stop,
+            (1, 1),
+        ),
     ]
 }
 
@@ -169,14 +278,21 @@ fn mock_responses() -> Vec<MessageResponse> {
 fn select_provider() -> Arc<dyn LlmProvider> {
     if env::var("AG_LLM_BASE_URL").is_ok() && env::var("AG_LLM_API_KEY").is_ok() {
         let cfg = ProviderConfig::from_env("AG_LLM").expect("AG_LLM_* 环境变量解析失败");
-        let provider_type = env::var("AG_LLM_PROVIDER").ok()
+        let provider_type = env::var("AG_LLM_PROVIDER")
+            .ok()
             .and_then(|s| s.parse::<ProviderType>().ok())
             .unwrap_or(ProviderType::OpenaiChat);
         Arc::from(create_provider(provider_type, cfg).expect("Provider 创建失败"))
     } else {
         let found: Vec<&str> = ["AG_LLM_BASE_URL", "AG_LLM_API_KEY", "AG_LLM_MODEL"]
-            .iter().filter(|k| env::var(k).is_ok()).copied().collect();
-        eprintln!("AG_LLM_* 环境变量不完整（检测到: {:?}），回退到 MockProvider", found);
+            .iter()
+            .filter(|k| env::var(k).is_ok())
+            .copied()
+            .collect();
+        eprintln!(
+            "AG_LLM_* 环境变量不完整（检测到: {:?}），回退到 MockProvider",
+            found
+        );
         Arc::new(MockProvider::new(mock_responses()))
     }
 }
@@ -191,52 +307,74 @@ async fn main() {
     let backend: Arc<dyn MemoryStore> =
         Arc::new(SqliteStore::open(&db_path).expect("SqliteStore 打开失败"));
     println!("💾 SqliteStore: {}", db_path.display());
-    let provider_label = if env::var("AG_LLM_BASE_URL").is_ok() && env::var("AG_LLM_API_KEY").is_ok() {
-        "真实 LLM Provider"
-    } else {
-        "MockProvider (离线回退模式)"
-    };
+    let provider_label =
+        if env::var("AG_LLM_BASE_URL").is_ok() && env::var("AG_LLM_API_KEY").is_ok() {
+            "真实 LLM Provider"
+        } else {
+            "MockProvider (离线回退模式)"
+        };
     println!("🔄 Provider: {provider_label}");
 
     let mut registry = ToolRegistry::new();
     registry.register(Arc::new(EchoTool)).unwrap();
     registry.register(Arc::new(CalcTool)).unwrap();
-    registry.register(Arc::new(NoteTool { store: backend.clone() })).unwrap();
+    registry
+        .register(Arc::new(NoteTool {
+            store: backend.clone(),
+        }))
+        .unwrap();
     println!("🔧 注册工具: {:?}", registry.list_tools());
 
-    let bundle = Arc::new(AgentBuilder::new()
-        .provider(select_provider())
-        .tool_registry(Arc::new(registry))
-        .hook_executor(Arc::new(HookExecutor::new()))
-        .build().expect("RuntimeBundle 装配失败"));
+    let bundle = Arc::new(
+        AgentBuilder::new()
+            .provider(select_provider())
+            .tool_registry(Arc::new(registry))
+            .hook_executor(Arc::new(HookExecutor::new()))
+            .build()
+            .expect("RuntimeBundle 装配失败"),
+    );
 
     let mut session = AgentSession::new(Arc::new(AssistantAgent), "e2e-1", bundle.clone());
 
     println!("\n第 1 轮 用户: 帮我算 25 * 4");
-    let r1 = session.submit_turn("帮我算 25 * 4").await.expect("turn 1 失败");
+    let r1 = session
+        .submit_turn("帮我算 25 * 4")
+        .await
+        .expect("turn 1 失败");
     println!("  → 回答: {}", r1.text());
 
     println!("\n第 2 轮 用户: 记下来：结果是 100");
-    let r2 = session.submit_turn("记下来：结果是 100").await.expect("turn 2 失败");
+    let r2 = session
+        .submit_turn("记下来：结果是 100")
+        .await
+        .expect("turn 2 失败");
     println!("  → 回答: {}", r2.text());
 
     println!("\n第 3 轮 用户: 我刚才算了什么？");
-    let r3 = session.submit_turn("我刚才算了什么？").await.expect("turn 3 失败");
+    let r3 = session
+        .submit_turn("我刚才算了什么？")
+        .await
+        .expect("turn 3 失败");
     println!("  → 回答: {}", r3.text());
 
     let total = session.usage().total();
-    println!("\n📊 用量: prompt={}, completion={}, total={}",
-        total.prompt_tokens, total.completion_tokens, total.total_tokens);
+    println!(
+        "\n📊 用量: prompt={}, completion={}, total={}",
+        total.prompt_tokens, total.completion_tokens, total.total_tokens
+    );
 
     println!("\n=== 持久化验证 ===");
     // 显式释放所有对 backend 的 Arc 引用，确保 SqliteStore Connection 真正关闭。
     // 释放顺序：session → bundle（间接持有 NoteTool → backend clone）→ backend 局部变量。
-    drop(session);  // session.bundle Arc 计数 -1
-    drop(bundle);   // bundle Arc 计数归零 → registry → NoteTool → backend clone Arc 计数 2→1
-    drop(backend);  // backend 局部变量 Arc 计数 1→0 → SqliteStore::drop → Connection 自动 close
+    drop(session); // session.bundle Arc 计数 -1
+    drop(bundle); // bundle Arc 计数归零 → registry → NoteTool → backend clone Arc 计数 2→1
+    drop(backend); // backend 局部变量 Arc 计数 1→0 → SqliteStore::drop → Connection 自动 close
     let backend2: Arc<dyn MemoryStore> =
         Arc::new(SqliteStore::open(&db_path).expect("重开 SqliteStore 失败"));
-    let filter = MemoryFilter { prefix: Some("note:".into()), ..Default::default() };
+    let filter = MemoryFilter {
+        prefix: Some("note:".into()),
+        ..Default::default()
+    };
     let items = backend2.list(&filter).await.expect("list 失败");
     println!("✓ 跨连接数据存活: 找到 {} 条 note", items.len());
     assert!(!items.is_empty(), "持久化验证失败：重开后无数据");
@@ -245,4 +383,4 @@ async fn main() {
     }
 
     println!("\n✓ 端到端演示完成");
-}
+}

examples/engine_demo.rs

@@ -1,4 +1,5 @@
 //! engine_demo —— SessionManager + Checkpointer 端到端示例。
+//! Required features: cargo run --example engine_demo --features "engine"
 //!
 //! 演示：
 //! 1. SessionManager::create 创建 session
@@ -190,8 +191,8 @@ async fn main() {
         snapshot_turn, snapshot_data_count
     );
 
-    let mut rolled_back =
-        AgentSession::from_snapshot(snapshot, agent.clone(), bundle.clone()).expect("from_snapshot");
+    let mut rolled_back = AgentSession::from_snapshot(snapshot, agent.clone(), bundle.clone())
+        .expect("from_snapshot");
     rolled_back
         .restore_memory()
         .await
@@ -216,8 +217,14 @@ async fn main() {
         after_turn,
         before_turn
     );
-    assert_eq!(after_turn, snapshot_turn, "rollback 后 turn_index 应等于 checkpoint 时刻值");
-    assert!(after_cost <= before_cost, "rollback 后 cost 应 ≤ rollback 前");
+    assert_eq!(
+        after_turn, snapshot_turn,
+        "rollback 后 turn_index 应等于 checkpoint 时刻值"
+    );
+    assert!(
+        after_cost <= before_cost,
+        "rollback 后 cost 应 ≤ rollback 前"
+    );
     println!("✓ rollback + replace 一致性验证通过");
 
     // 9. destroy 父子 session
@@ -249,4 +256,4 @@ async fn main() {
     sm.destroy(&c_id).await.unwrap();
 
     println!("\n✓ engine_demo 完成");
-}
+}

examples/knowledge_graph_demo.rs

@@ -1,4 +1,5 @@
 //! knowledge_graph_demo -- 知识图谱 + 双通道检索演示。
+//! Required features: cargo run --example knowledge_graph_demo --features "memory"
 //!
 //! 演示：
 //! 1. 构建 KnowledgeGraph（实体 + 关系）
@@ -52,15 +53,30 @@ async fn main() {
         graph.add_entity(e.clone()).await.unwrap();
     }
     graph
-        .add_relation(GraphRelation::new("langchain", "langgraph", "includes", 0.9))
+        .add_relation(GraphRelation::new(
+            "langchain",
+            "langgraph",
+            "includes",
+            0.9,
+        ))
         .await
         .unwrap();
     graph
-        .add_relation(GraphRelation::new("langchain", "langsmith", "includes", 0.7))
+        .add_relation(GraphRelation::new(
+            "langchain",
+            "langsmith",
+            "includes",
+            0.7,
+        ))
         .await
         .unwrap();
     graph
-        .add_relation(GraphRelation::new("langchain", "python", "built_with", 0.95))
+        .add_relation(GraphRelation::new(
+            "langchain",
+            "python",
+            "built_with",
+            0.95,
+        ))
         .await
         .unwrap();
     graph
@@ -86,7 +102,10 @@ async fn main() {
     // ── 3. 标签管理 ──
     println!("\n=== 3. 标签管理 ===");
     graph
-        .set_entity_tags("langchain", vec!["ai".into(), "framework".into(), "llm".into()])
+        .set_entity_tags(
+            "langchain",
+            vec!["ai".into(), "framework".into(), "llm".into()],
+        )
         .await
         .unwrap();
     graph
@@ -118,8 +137,8 @@ async fn main() {
     .unwrap();
 
     // Hybrid 策略（默认）
-    let retriever = MemoryRetriever::new(ks, RetrieverConfig::default())
-        .with_knowledge_graph(graph.clone());
+    let retriever =
+        MemoryRetriever::new(ks, RetrieverConfig::default()).with_knowledge_graph(graph.clone());
     println!("\n--- Hybrid 检索: 'langchain' ---");
     let result = retriever.retrieve("langchain").await.unwrap();
     println!("策略: {:?}", result.strategy);
@@ -129,7 +148,10 @@ async fn main() {
                 println!("  [Store] {} (score={:.3})", page.title, score);
             }
             RetrievalItem::GraphEntity {
-                entity, score, path, ..
+                entity,
+                score,
+                path,
+                ..
             } => {
                 println!(
                     "  [Graph] {} (score={:.3}, path={:?})",
@@ -172,7 +194,10 @@ async fn main() {
         }
     }
     assert!(
-        result.items.iter().all(|i| matches!(i, RetrievalItem::GraphEntity { .. })),
+        result
+            .items
+            .iter()
+            .all(|i| matches!(i, RetrievalItem::GraphEntity { .. })),
         "GraphOnly 应只返回 Graph 结果"
     );
 

examples/knowledge_search_demo.rs

@@ -1,4 +1,5 @@
 //! knowledge_search_demo —— 知识页面存储与关键词检索。
+//! Required features: cargo run --example knowledge_search_demo --features "memory"
 //!
 //! 演示：
 //! 1. `KnowledgeStore` 存储多个 `KnowledgePage`

examples/prompt_composer.rs

@@ -1,4 +1,5 @@
 //! prompt_composer —— 提示词模板与组合器离线示例。
+//! Required features: cargo run --example prompt_composer --features "prompt,llm"
 //!
 //! 演示：
 //! 1. `PromptTemplate::compile` + `render` 变量插值（`{{var}}` 语法）

examples/quick_start.rs

@@ -1,41 +1,61 @@
 //! quick_start —— 30 行最小可运行示例，展示 Agent / BaseTool / Builder / Session 四层抽象。
+//! Required features: cargo run --example quick_start --features "agent"
 //!
 //! 运行：`cargo run --example quick_start`（离线，零配置）
 
-use std::sync::Arc;
 use agcore::agent::{Agent, AgentBuilder, AgentSession};
+use agcore::llm::LlmProvider;
 use agcore::llm::hooks::HookExecutor;
 use agcore::llm::mock::MockProvider;
-use agcore::llm::LlmProvider;
-use agcore::llm::types::{Usage, message::{ContentBlock, Message}, response_v2::{MessageResponse, StopReason}};
+use agcore::llm::types::{
+    Usage,
+    message::{ContentBlock, Message},
+    response_v2::{MessageResponse, StopReason},
+};
 use agcore::tools::{BaseTool, ToolContext, ToolError, ToolRegistry};
 use async_trait::async_trait;
 use serde_json::{Value, json};
+use std::sync::Arc;
 
 struct Greeter;
 impl Agent for Greeter {
-    fn name(&self) -> &str { "greeter" }
-    fn system_prompt(&self) -> Option<&str> { Some("中文助手，先调用 echo 工具，再总结。") }
+    fn name(&self) -> &str {
+        "greeter"
+    }
+    fn system_prompt(&self) -> Option<&str> {
+        Some("中文助手，先调用 echo 工具，再总结。")
+    }
 }
 
 struct EchoTool;
 #[async_trait]
 impl BaseTool for EchoTool {
-    fn name(&self) -> &str { "echo" }
-    fn description(&self) -> &str { "回显文本" }
+    fn name(&self) -> &str {
+        "echo"
+    }
+    fn description(&self) -> &str {
+        "回显文本"
+    }
     fn parameters(&self) -> Value {
         json!({"type":"object","properties":{"text":{"type":"string"}},"required":["text"]})
     }
     async fn execute(&self, args: Value, _: &ToolContext<'_>) -> Result<Value, ToolError> {
-        let text = args.get("text").and_then(|v| v.as_str())
-            .ok_or_else(|| ToolError::InvalidArguments("text".into(), "需要 string 类型的 text 参数".into()))?;
+        let text = args.get("text").and_then(|v| v.as_str()).ok_or_else(|| {
+            ToolError::InvalidArguments("text".into(), "需要 string 类型的 text 参数".into())
+        })?;
         Ok(json!({"echoed": format!("收到: {text}")}))
     }
 }
 
 fn resp(content: Vec<ContentBlock>, stop: StopReason, u: (u32, u32)) -> MessageResponse {
-    MessageResponse { id: String::new(), model: "mock".into(), message: Message::Assistant { content },
-        usage: Usage::from_input_output(u.0, u.1), stop_reason: stop, extra: Default::default() }
+    MessageResponse {
+        id: String::new(),
+        model: "mock".into(),
+        message: Message::Assistant { content },
+        usage: Usage::from_input_output(u.0, u.1),
+        stop_reason: stop,
+        extra: Default::default(),
+    }
 }
 
 #[tokio::main]
@@ -43,14 +63,31 @@ async fn main() {
     let mut registry = ToolRegistry::new();
     registry.register(Arc::new(EchoTool)).unwrap();
     let provider: Arc<dyn LlmProvider> = Arc::new(MockProvider::new(vec![
-        resp(vec![ContentBlock::ToolUse { id: "c1".into(), name: "echo".into(),
-                input: json!({"text": "你好"}) }], StopReason::ToolUse, (5, 8)),
-        resp(vec![ContentBlock::Text { text: "EchoTool 已收到您的消息并完成回传。".into() }],
-             StopReason::Stop, (8, 16)),
+        resp(
+            vec![ContentBlock::ToolUse {
+                id: "c1".into(),
+                name: "echo".into(),
+                input: json!({"text": "你好"}),
+            }],
+            StopReason::ToolUse,
+            (5, 8),
+        ),
+        resp(
+            vec![ContentBlock::Text {
+                text: "EchoTool 已收到您的消息并完成回传。".into(),
+            }],
+            StopReason::Stop,
+            (8, 16),
+        ),
     ]));
-    let bundle = Arc::new(AgentBuilder::new()
-        .provider(provider).tool_registry(Arc::new(registry))
-        .hook_executor(Arc::new(HookExecutor::new())).build().unwrap());
+    let bundle = Arc::new(
+        AgentBuilder::new()
+            .provider(provider)
+            .tool_registry(Arc::new(registry))
+            .hook_executor(Arc::new(HookExecutor::new()))
+            .build()
+            .unwrap(),
+    );
     let mut session = AgentSession::new(Arc::new(Greeter), "qs", bundle);
     let resp = session.submit_turn("你好").await.unwrap();
     let text = resp.text();

examples/simple_visit.rs

@@ -1,3 +1,5 @@
+//! Required features: cargo run --example simple_visit --features "llm,provider-openai,tracing-init"
+
 use std::env;
 
 use agcore::init_tracing;

examples/streaming_events_demo.rs

@@ -1,4 +1,5 @@
 //! streaming_events_demo —— LLM 流式响应事件流消费（含错误路径）。
+//! Required features: cargo run --example streaming_events_demo --features "llm,provider-openai"
 //!
 //! 演示：
 //! 1. `MockProvider::chat_stream` 输出标准 `StreamEvent` 流（离线可跑）
@@ -14,9 +15,9 @@
 
 use std::sync::Arc;
 
+use agcore::llm::LlmProvider;
 use agcore::llm::cycle::{CycleConfig, LlmCycle};
 use agcore::llm::mock::MockProvider;
-use agcore::llm::LlmProvider;
 use agcore::llm::types::Usage;
 use agcore::llm::types::message::{ContentBlock, Message};
 use agcore::llm::types::response_v2::{MessageResponse, StopReason, StreamEvent};

examples/sub_agent_dispatch_demo.rs

@@ -1,4 +1,5 @@
 //! sub_agent_dispatch_demo —— SubAgent 并行派发示例。
+//! Required features: cargo run --example sub_agent_dispatch_demo --features "engine"
 //!
 //! 演示：
 //! 1. 创建父 session（"主编" agent）

examples/task_agent_demo.rs

@@ -1,4 +1,5 @@
 //! task_agent_demo —— Plan 解析、Step 状态机、错误路径。
+//! Required features: cargo run --example task_agent_demo --features "agent"
 //!
 //! 演示：
 //! 1. `JsonPlanParser::parse` 解析合法 JSON 输入
@@ -12,9 +13,9 @@
 use std::collections::HashMap;
 
 use agcore::agent::{AgentError, JsonPlanParser, PlanParser, Step, StepStatus};
+use agcore::llm::types::Usage;
 use agcore::llm::types::message::Message;
 use agcore::llm::types::response_v2::{MessageResponse, StopReason};
-use agcore::llm::types::Usage;
 
 #[tokio::main]
 async fn main() {
@@ -111,4 +112,4 @@ async fn main() {
     assert!(matches!(err, AgentError::PlanParse(_)));
 
     println!("\n✓ task_agent_demo 完成");
-}
+}

src/agent.rs

@@ -23,8 +23,8 @@ pub mod task;
 pub use agent::Agent;
 pub use builder::AgentBuilder;
 pub use context::{
-    ContextBudget, ContextSlot, DeriveStrategy, FocusedConfig, MergeStrategy, SlotConfig,
-    SlotMeta, SlotMode, SlotSource,
+    ContextBudget, ContextSlot, DeriveStrategy, FocusedConfig, MergeStrategy, SlotConfig, SlotMeta,
+    SlotMode, SlotSource,
 };
 pub use error::AgentError;
 pub use runtime::{AgentConfig, RuntimeBundle};

src/agent/builder.rs

@@ -12,8 +12,8 @@ use std::sync::Arc;
 use crate::agent::error::AgentError;
 use crate::agent::runtime::{AgentConfig, RuntimeBundle};
 use crate::agent::summary::SummaryConfig;
-use crate::llm::hooks::HookExecutor;
 use crate::llm::LlmProvider;
+use crate::llm::hooks::HookExecutor;
 use crate::memory::retriever::MemoryRetriever;
 use crate::memory::store::MemoryStore;
 use crate::tools::ToolRegistry;

src/agent/context.rs

@@ -250,12 +250,12 @@ impl ContextSlot {
 
     /// 保存 slot 数据到存储后端（全量写入，含 config）。
     pub async fn save(&self, store: &dyn MemoryStore) -> Result<(), AgentError> {
-        let data = serde_json::to_string(&self.messages)
-            .map_err(|e| AgentError::Other(e.to_string()))?;
-        let meta = serde_json::to_string(&self.meta)
-            .map_err(|e| AgentError::Other(e.to_string()))?;
-        let config = serde_json::to_string(&self.config)
-            .map_err(|e| AgentError::Other(e.to_string()))?;
+        let data =
+            serde_json::to_string(&self.messages).map_err(|e| AgentError::Other(e.to_string()))?;
+        let meta =
+            serde_json::to_string(&self.meta).map_err(|e| AgentError::Other(e.to_string()))?;
+        let config =
+            serde_json::to_string(&self.config).map_err(|e| AgentError::Other(e.to_string()))?;
 
         store
             .save(Self::make_item(
@@ -551,8 +551,10 @@ mod tests {
     async fn slot_save_load_roundtrip() {
         let store = make_store();
         let mut slot = make_slot("default", "s1");
-        slot.append_messages(vec![Message::user_text("hi")]).unwrap();
-        slot.append_messages(vec![Message::assistant("hello")]).unwrap();
+        slot.append_messages(vec![Message::user_text("hi")])
+            .unwrap();
+        slot.append_messages(vec![Message::assistant("hello")])
+            .unwrap();
 
         slot.save(&*store).await.unwrap();
         let loaded = ContextSlot::load("default", "s1", &*store).await.unwrap();
@@ -576,8 +578,14 @@ mod tests {
             .unwrap();
         b.save(&*store).await.unwrap();
 
-        let loaded_a = ContextSlot::load("main", "sA", &*store).await.unwrap().unwrap();
-        let loaded_b = ContextSlot::load("main", "sB", &*store).await.unwrap().unwrap();
+        let loaded_a = ContextSlot::load("main", "sA", &*store)
+            .await
+            .unwrap()
+            .unwrap();
+        let loaded_b = ContextSlot::load("main", "sB", &*store)
+            .await
+            .unwrap()
+            .unwrap();
         assert_eq!(extract_text(&loaded_a.messages[0]), "only in A");
         assert_eq!(extract_text(&loaded_b.messages[0]), "only in B");
     }
@@ -622,10 +630,13 @@ mod tests {
     async fn slot_delete_then_load_none() {
         let store = make_store();
         let mut slot = make_slot("to_delete", "s1");
-        slot.append_messages(vec![Message::user_text("hi")]).unwrap();
+        slot.append_messages(vec![Message::user_text("hi")])
+            .unwrap();
         slot.save(&*store).await.unwrap();
 
-        ContextSlot::delete("to_delete", "s1", &*store).await.unwrap();
+        ContextSlot::delete("to_delete", "s1", &*store)
+            .await
+            .unwrap();
         let loaded = ContextSlot::load("to_delete", "s1", &*store).await.unwrap();
         assert!(loaded.is_none());
     }
@@ -723,7 +734,10 @@ mod tests {
         let store = make_store();
         let slot = make_slot("empty", "s1");
         slot.save(&*store).await.unwrap();
-        let loaded = ContextSlot::load("empty", "s1", &*store).await.unwrap().unwrap();
+        let loaded = ContextSlot::load("empty", "s1", &*store)
+            .await
+            .unwrap()
+            .unwrap();
         assert!(loaded.messages.is_empty());
         assert_eq!(loaded.meta.message_count, 0);
     }
@@ -745,7 +759,8 @@ mod tests {
     async fn derive_full_copies_parent_messages() {
         let mut parent = make_slot("p", "s1");
         for i in 0..3 {
-            parent.append_messages(vec![Message::user_text(format!("u{i}"))])
+            parent
+                .append_messages(vec![Message::user_text(format!("u{i}"))])
                 .unwrap();
         }
 
@@ -769,7 +784,10 @@ mod tests {
         let store = make_store();
         child.save(&*store).await.unwrap();
 
-        let loaded = ContextSlot::load("c", "s1", &*store).await.unwrap().unwrap();
+        let loaded = ContextSlot::load("c", "s1", &*store)
+            .await
+            .unwrap()
+            .unwrap();
         assert_eq!(loaded.messages.len(), 3);
         assert!(matches!(loaded.config.source, SlotSource::Derived { .. }));
     }
@@ -777,9 +795,12 @@ mod tests {
     #[tokio::test]
     async fn derive_focused_filters_parent_messages() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::system("sys")]).unwrap();
+        parent
+            .append_messages(vec![Message::system("sys")])
+            .unwrap();
         for i in 0..5 {
-            parent.append_messages(vec![Message::user_text(format!("u{i}"))])
+            parent
+                .append_messages(vec![Message::user_text(format!("u{i}"))])
                 .unwrap();
         }
 
@@ -818,7 +839,9 @@ mod tests {
     async fn derived_slot_loadable_independently() {
         let store = make_store();
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("u")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("u")])
+            .unwrap();
         parent.save(&*store).await.unwrap();
 
         // 派生 child
@@ -835,12 +858,16 @@ mod tests {
                 compact: true,
             },
         );
-        child.append_messages(vec![Message::user_text("derived msg")])
+        child
+            .append_messages(vec![Message::user_text("derived msg")])
             .unwrap();
         child.save(&*store).await.unwrap();
 
         // child 可独立加载
-        let loaded = ContextSlot::load("c", "s1", &*store).await.unwrap().unwrap();
+        let loaded = ContextSlot::load("c", "s1", &*store)
+            .await
+            .unwrap()
+            .unwrap();
         assert_eq!(loaded.messages.len(), 1);
         assert_eq!(extract_text(&loaded.messages[0]), "derived msg");
     }
@@ -867,17 +894,59 @@ mod tests {
         slot.save(&*store).await.unwrap();
 
         // 确认所有记录存在
-        assert!(store.get(&ContextSlot::data_key("s1", "x")).await.unwrap().is_some());
-        assert!(store.get(&ContextSlot::meta_key("s1", "x")).await.unwrap().is_some());
-        assert!(store.get(&ContextSlot::config_key("s1", "x")).await.unwrap().is_some());
-        assert!(store.get(&ContextSlot::rel_key("s1", "x")).await.unwrap().is_some());
+        assert!(
+            store
+                .get(&ContextSlot::data_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_some()
+        );
+        assert!(
+            store
+                .get(&ContextSlot::meta_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_some()
+        );
+        assert!(
+            store
+                .get(&ContextSlot::config_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_some()
+        );
+        assert!(
+            store
+                .get(&ContextSlot::rel_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_some()
+        );
 
         ContextSlot::delete("x", "s1", &*store).await.unwrap();
 
         // data/meta/config 已删
-        assert!(store.get(&ContextSlot::data_key("s1", "x")).await.unwrap().is_none());
-        assert!(store.get(&ContextSlot::meta_key("s1", "x")).await.unwrap().is_none());
-        assert!(store.get(&ContextSlot::config_key("s1", "x")).await.unwrap().is_none());
+        assert!(
+            store
+                .get(&ContextSlot::data_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_none()
+        );
+        assert!(
+            store
+                .get(&ContextSlot::meta_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_none()
+        );
+        assert!(
+            store
+                .get(&ContextSlot::config_key("s1", "x"))
+                .await
+                .unwrap()
+                .is_none()
+        );
     }
 
     // ===== 基础类型测试 =====
@@ -893,7 +962,10 @@ mod tests {
     #[test]
     fn context_budget_default_sum_128k() {
         let b = ContextBudget::default();
-        assert_eq!(b.system + b.history + b.tools + b.tool_results + b.reserve, 128_000);
+        assert_eq!(
+            b.system + b.history + b.tools + b.tool_results + b.reserve,
+            128_000
+        );
     }
 
     #[test]
@@ -939,10 +1011,7 @@ mod tests {
 
     #[test]
     fn filter_focused_injects_summary() {
-        let messages = vec![
-            Message::user_text("u"),
-            Message::assistant("a"),
-        ];
+        let messages = vec![Message::user_text("u"), Message::assistant("a")];
         let cfg = FocusedConfig {
             keep_system: false,
             recent_messages: 100,
@@ -964,8 +1033,12 @@ mod tests {
     #[test]
     fn fork_full_copies_messages() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("a")]).unwrap();
-        parent.append_messages(vec![Message::assistant("b")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("a")])
+            .unwrap();
+        parent
+            .append_messages(vec![Message::assistant("b")])
+            .unwrap();
         let child = parent.fork("c".into(), DeriveStrategy::Full);
         assert_eq!(child.messages.len(), 2);
         assert!(matches!(child.config.mode, SlotMode::Full));
@@ -974,7 +1047,9 @@ mod tests {
     #[test]
     fn fork_focused_filters_messages() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::system("sys")]).unwrap();
+        parent
+            .append_messages(vec![Message::system("sys")])
+            .unwrap();
         for i in 0..5 {
             parent
                 .append_messages(vec![Message::user_text(format!("u{i}"))])
@@ -994,7 +1069,9 @@ mod tests {
     #[test]
     fn fork_preserves_independence() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("a")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("a")])
+            .unwrap();
         let mut child = parent.fork("c".into(), DeriveStrategy::Full);
         let child_count_at_fork = child.messages.len();
 
@@ -1003,7 +1080,9 @@ mod tests {
             .append_messages(vec![Message::user_text("b")])
             .unwrap();
         // 子 slot 追加
-        child.append_messages(vec![Message::user_text("c")]).unwrap();
+        child
+            .append_messages(vec![Message::user_text("c")])
+            .unwrap();
 
         assert_eq!(parent.messages.len(), 2);
         assert_eq!(child.messages.len(), child_count_at_fork + 1);
@@ -1013,10 +1092,15 @@ mod tests {
     #[test]
     fn fork_sets_derived_source() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("a")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("a")])
+            .unwrap();
         let child = parent.fork("c".into(), DeriveStrategy::Full);
         match &child.config.source {
-            SlotSource::Derived { parent_id, strategy } => {
+            SlotSource::Derived {
+                parent_id,
+                strategy,
+            } => {
                 assert_eq!(parent_id, "p");
                 assert!(matches!(strategy, DeriveStrategy::Full));
             }
@@ -1030,7 +1114,9 @@ mod tests {
     #[test]
     fn merge_append_appends_messages() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("p1")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("p1")])
+            .unwrap();
         let child = {
             let mut c = parent.fork("c".into(), DeriveStrategy::Full);
             // fork 时 child 继承父的 "p1"；再追加一条 c1
@@ -1049,8 +1135,12 @@ mod tests {
     #[test]
     fn merge_replace_replaces_messages() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("p1")]).unwrap();
-        parent.append_messages(vec![Message::user_text("p2")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("p1")])
+            .unwrap();
+        parent
+            .append_messages(vec![Message::user_text("p2")])
+            .unwrap();
         let child = {
             let mut c = parent.fork("c".into(), DeriveStrategy::Full);
             // 清空 child 再追加
@@ -1088,7 +1178,9 @@ mod tests {
     #[test]
     fn merge_cross_session_rejected() {
         let mut parent = make_slot("p", "s1");
-        parent.append_messages(vec![Message::user_text("a")]).unwrap();
+        parent
+            .append_messages(vec![Message::user_text("a")])
+            .unwrap();
         let child = ContextSlot::new("OTHER_SESSION", "c", SlotConfig::default());
         let err = parent.merge(child, MergeStrategy::Append).unwrap_err();
         assert!(matches!(err, AgentError::Config(_)));
@@ -1121,4 +1213,4 @@ mod tests {
             .unwrap()
             .block_on(f)
     }
-}
+}

src/agent/runtime.rs

@@ -16,9 +16,9 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use crate::agent::summary::SummaryConfig;
+use crate::llm::LlmProvider;
 use crate::llm::compact::CompactConfig;
 use crate::llm::hooks::HookExecutor;
-use crate::llm::LlmProvider;
 use crate::memory::retriever::MemoryRetriever;
 use crate::memory::store::MemoryStore;
 use crate::tools::ToolRegistry;

src/agent/session.rs

@@ -15,24 +15,22 @@ use std::sync::Arc;
 use futures_core::Stream;
 
 use crate::agent::agent::Agent;
-use crate::agent::context::{
-    ContextSlot, DeriveStrategy, SlotConfig, SlotMode,
-};
+use crate::agent::context::{ContextSlot, DeriveStrategy, SlotConfig, SlotMode};
 // SlotSource 仅在 `mod tests` 中使用（通过 `use super::*;` 引入），lib 主体保留以避免测试 import 变更。
 #[allow(unused_imports)]
 use crate::agent::context::SlotSource;
 use crate::agent::error::AgentError;
 use crate::agent::runtime::RuntimeBundle;
 use crate::agent::session_memory::SessionMemory;
-use crate::agent::summary::{format_messages_as_text, SummaryConfig};
-#[cfg(feature = "engine")]
-use crate::engine::snapshot::{SessionMemoryEntry, SessionSnapshot};
+use crate::agent::summary::{SummaryConfig, format_messages_as_text};
 #[cfg(feature = "engine")]
 use crate::engine::EngineError;
+#[cfg(feature = "engine")]
+use crate::engine::snapshot::{SessionMemoryEntry, SessionSnapshot};
+use crate::llm::LlmProvider;
 use crate::llm::cycle::{CostTracker, CycleConfig, LlmCycle};
 use crate::llm::error::LlmError;
 use crate::llm::hooks::{HookContext, HookEvent};
-use crate::llm::LlmProvider;
 use crate::llm::stream::StreamEvent;
 use crate::llm::types::message::Message;
 use crate::llm::types::response_v2::MessageResponse;
@@ -112,11 +110,7 @@ impl AgentSession {
         let session_memory = SessionMemory::new(backend, &session_id_str);
 
         // 自动创建 "default" slot
-        let default_slot = ContextSlot::new(
-            &session_id_str,
-            "default",
-            SlotConfig::default(),
-        );
+        let default_slot = ContextSlot::new(&session_id_str, "default", SlotConfig::default());
         let mut slots = HashMap::new();
         slots.insert("default".to_string(), default_slot);
 
@@ -162,9 +156,7 @@ impl AgentSession {
         key: impl Into<String>,
         value: impl Into<String>,
     ) -> Result<(), AgentError> {
-        self.session_memory
-            .set(&key.into(), &value.into())
-            .await
+        self.session_memory.set(&key.into(), &value.into()).await
     }
 
     /// 读取一条会话级数据。
@@ -204,11 +196,7 @@ impl AgentSession {
         if self.slots.contains_key(&id) {
             return Err(AgentError::SlotAlreadyExists(id));
         }
-        let slot = ContextSlot::new(
-            &self.session_id,
-            &id,
-            config.unwrap_or_default(),
-        );
+        let slot = ContextSlot::new(&self.session_id, &id, config.unwrap_or_default());
         slot.save(&*self.resolve_store()).await?;
         self.slots.insert(id, slot);
         Ok(())
@@ -264,7 +252,9 @@ impl AgentSession {
     /// - 已删除后再 load 返回 None
     pub async fn delete_slot(&mut self, id: &str) -> Result<(), AgentError> {
         if id == "default" {
-            return Err(AgentError::Config("Cannot delete the 'default' slot".into()));
+            return Err(AgentError::Config(
+                "Cannot delete the 'default' slot".into(),
+            ));
         }
         if self.slots.len() <= 1 {
             return Err(AgentError::Config("Cannot delete the last slot".into()));
@@ -354,12 +344,7 @@ impl AgentSession {
         // 6. 只将本轮新增消息追加到当前 slot（保留全量历史，确保 Focused 模式的"读时过滤"语义不丢失数据）
         //    cycle.messages() 包含 [system_prompt?, history..., user_input, tool_calls..., final_response]
         //    新增消息 = cycle.messages()[input_len..]（跳过 initial_messages，即跳过已被持久化的内容）
-        let new_messages: Vec<Message> = cycle
-            .messages()
-            .iter()
-            .skip(input_len)
-            .cloned()
-            .collect();
+        let new_messages: Vec<Message> = cycle.messages().iter().skip(input_len).cloned().collect();
         let store = self.resolve_store();
         if let Some(slot) = self.slots.get_mut(&self.current_slot_id) {
             slot.append_messages(new_messages)?;
@@ -437,10 +422,7 @@ impl AgentSession {
 
         // 5. 调用流式工具循环
         let stream = cycle
-            .submit_with_tools_stream(
-                user_input.into(),
-                Arc::clone(&self.bundle.tool_registry),
-            )
+            .submit_with_tools_stream(user_input.into(), Arc::clone(&self.bundle.tool_registry))
             .await?;
 
         // 6. turn_index 递增 —— 配合 finalize_turn 用 (turn_index - 1) 传递正确的 OnTurnEnd 序号
@@ -492,7 +474,8 @@ impl AgentSession {
             .await;
 
         // Phase 16: 摘要检查点（流式路径 turn_index 已被 submit_turn_stream 提前 ++1）
-        self.maybe_summarize(self.turn_index.saturating_sub(1)).await;
+        self.maybe_summarize(self.turn_index.saturating_sub(1))
+            .await;
 
         Ok(())
     }
@@ -569,11 +552,7 @@ impl AgentSession {
         if slots.is_empty() {
             slots.insert(
                 "default".to_string(),
-                ContextSlot::new(
-                    &snapshot.session_id,
-                    "default",
-                    SlotConfig::default(),
-                ),
+                ContextSlot::new(&snapshot.session_id, "default", SlotConfig::default()),
             );
         }
 
@@ -619,12 +598,7 @@ impl AgentSession {
 
         for (key, entry) in entries {
             self.session_memory
-                .set_with_meta(
-                    &key,
-                    &entry.value,
-                    entry.metadata.clone(),
-                    entry.created_at,
-                )
+                .set_with_meta(&key, &entry.value, entry.metadata.clone(), entry.created_at)
                 .await
                 .map_err(EngineError::Agent)?;
         }
@@ -686,13 +660,22 @@ impl AgentSession {
         let model = cfg.summary_model.clone();
         let prompt = cfg.summary_prompt.clone();
 
-        let result =
-            Self::generate_summary(&provider, &messages, &prompt, model.as_deref(), max_tool_result_chars)
-                .await;
+        let result = Self::generate_summary(
+            &provider,
+            &messages,
+            &prompt,
+            model.as_deref(),
+            max_tool_result_chars,
+        )
+        .await;
 
         match result {
             Ok(text) => {
-                tracing::info!(turn = current_turn, summary_len = text.len(), "摘要自动生成成功");
+                tracing::info!(
+                    turn = current_turn,
+                    summary_len = text.len(),
+                    "摘要自动生成成功"
+                );
                 // Resolve store first (immutable borrow on self) before mutable borrow on slots.
                 let store = self.resolve_store();
                 if let Some(slot) = self.slots.get_mut(&self.current_slot_id)
@@ -755,8 +738,8 @@ impl AgentSession {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::agent::builder::AgentBuilder;
     use crate::agent::FocusedConfig;
+    use crate::agent::builder::AgentBuilder;
     use crate::llm::hooks::{Hook, HookContext, HookExecutor, HookResult};
     use crate::llm::mock::MockProvider;
     use crate::llm::stream::StreamEvent;
@@ -815,7 +798,9 @@ mod tests {
         }
     }
 
-    fn build_session(provider_responses: Vec<MessageResponse>) -> (AgentSession, Arc<CountHook>, Arc<CountHook>) {
+    fn build_session(
+        provider_responses: Vec<MessageResponse>,
+    ) -> (AgentSession, Arc<CountHook>, Arc<CountHook>) {
         let mut hook_executor = HookExecutor::new();
         let start_count = Arc::new(CountHook(AtomicU32::new(0)));
         let end_count = Arc::new(CountHook(AtomicU32::new(0)));
@@ -849,7 +834,8 @@ mod tests {
     /// 烟雾测试 1：AgentSession::submit_turn 跑通 mock provider（向后兼容）。
     #[tokio::test]
     async fn submit_turn_runs_with_mock_provider() {
-        let (mut session, start_count, end_count) = build_session(vec![assistant_text("hello back")]);
+        let (mut session, start_count, end_count) =
+            build_session(vec![assistant_text("hello back")]);
         assert_eq!(session.turn_index(), 0);
 
         let response = session.submit_turn("hi").await.unwrap();
@@ -884,10 +870,8 @@ mod tests {
     /// 烟雾测试 3：submit_turn 触发 OnTurnStart / OnTurnEnd hook。
     #[tokio::test]
     async fn submit_turn_triggers_turn_hooks() {
-        let (mut session, start_count, end_count) = build_session(vec![
-            assistant_text("ok"),
-            assistant_text("ok 2"),
-        ]);
+        let (mut session, start_count, end_count) =
+            build_session(vec![assistant_text("ok"), assistant_text("ok 2")]);
 
         session.submit_turn("hi").await.unwrap();
         assert_eq!(start_count.0.load(Ordering::SeqCst), 1);
@@ -940,9 +924,15 @@ mod tests {
         // 即 [user_input, tool_results?, final_response]（不含 system_prompt，system 由 agent 提供）
         assert!(slot.messages.len() >= 2, "应至少包含 user 和 assistant");
         // 验证 user 输入和 assistant 响应都已写入
-        let has_user = slot.messages.iter().any(|m| extract_text(m) == "user input");
+        let has_user = slot
+            .messages
+            .iter()
+            .any(|m| extract_text(m) == "user input");
         let has_resp = slot.messages.iter().any(|m| extract_text(m) == "resp");
-        assert!(has_user && has_resp, "slot 应包含 user input 和 assistant response");
+        assert!(
+            has_user && has_resp,
+            "slot 应包含 user input 和 assistant response"
+        );
     }
 
     /// Phase 10: create_slot 创建新 slot。
@@ -986,7 +976,11 @@ mod tests {
         // 3. 检查 slot_a 的消息数
         let slot_a = session.slots.get("slot_a").unwrap();
         let slot_a_count = slot_a.messages.len();
-        assert!(slot_a_count >= 2, "slot_a 至少 2 条消息，实际 {}", slot_a_count);
+        assert!(
+            slot_a_count >= 2,
+            "slot_a 至少 2 条消息，实际 {}",
+            slot_a_count
+        );
 
         // 4. 切回 default，验证 default 不包含 slot_a 的消息
         session.switch_slot("default").await.unwrap();
@@ -1281,7 +1275,10 @@ mod tests {
             .iter()
             .any(|m| matches!(m, Message::User { .. }));
         let has_resp = slot.messages.iter().any(|m| extract_text(m) == "hi back");
-        assert!(has_user && has_resp, "default slot 应包含 user 和 assistant 消息");
+        assert!(
+            has_user && has_resp,
+            "default slot 应包含 user 和 assistant 消息"
+        );
     }
 
     /// Phase 9 Step 5.2 — `submit_turn_stream` 触发 OnTurnStart / OnTurnEnd hook。
@@ -1380,7 +1377,11 @@ mod tests {
     #[tokio::test]
     async fn summary_not_generated_below_threshold() {
         let mut session = build_session_with_summary(
-            vec![assistant_text("a"), assistant_text("b"), assistant_text("c")],
+            vec![
+                assistant_text("a"),
+                assistant_text("b"),
+                assistant_text("c"),
+            ],
             vec![assistant_text("should_not_appear")],
             SummaryConfig::default(),
         );
@@ -1400,16 +1401,20 @@ mod tests {
     #[tokio::test]
     async fn summary_generated_above_threshold() {
         let mut session = build_session_with_summary(
-            vec![assistant_text("a"), assistant_text("b"), assistant_text("c")],
+            vec![
+                assistant_text("a"),
+                assistant_text("b"),
+                assistant_text("c"),
+            ],
             vec![
                 assistant_text("summary-1"),
                 assistant_text("summary-2"),
                 assistant_text("summary-3"),
             ],
             SummaryConfig {
-                max_context_tokens: 20,   // 阈值 20 * 0.5 = 10
+                max_context_tokens: 20, // 阈值 20 * 0.5 = 10
                 trigger_token_ratio: 0.5,
-                debounce_turns: 0,        // 关闭防抖便于测试
+                debounce_turns: 0, // 关闭防抖便于测试
                 ..SummaryConfig::default()
             },
         );
@@ -1459,7 +1464,11 @@ mod tests {
     #[tokio::test]
     async fn summary_written_to_session_memory_but_full_mode_does_not_inject() {
         let mut session = build_session_with_summary(
-            vec![assistant_text("a"), assistant_text("b"), assistant_text("c")],
+            vec![
+                assistant_text("a"),
+                assistant_text("b"),
+                assistant_text("c"),
+            ],
             vec![assistant_text("captured-summary")],
             SummaryConfig {
                 max_context_tokens: 20,
@@ -1485,7 +1494,7 @@ mod tests {
         // 故意只提供 1 个对话响应；摘要调用时队列耗尽，MockProvider 返回 LlmError::Other
         let mut session = build_session_with_summary(
             vec![assistant_text("only-one")], // 后续摘要会失败
-            vec![],                          // 无摘要响应
+            vec![],                           // 无摘要响应
             SummaryConfig {
                 max_context_tokens: 20,
                 trigger_token_ratio: 0.5,
@@ -1639,4 +1648,4 @@ mod tests {
         let summary = session.get_conversation_summary().await.unwrap();
         assert!(summary.is_none(), "巨型 max_context_tokens 应永不触发");
     }
-}
+}

src/agent/session_memory.rs

@@ -48,7 +48,8 @@ impl SessionMemory {
     /// **不保留 metadata 和 created_at** —— 写入时 metadata 为空 JSON `{}`，created_at 为 `now_utc()`。
     /// 若需保留这两个字段（如 checkpoint rollback），使用 [`Self::set_with_meta`]。
     pub async fn set(&self, key: &str, value: &str) -> Result<(), AgentError> {
-        self.set_with_meta(key, value, serde_json::json!({}), None).await
+        self.set_with_meta(key, value, serde_json::json!({}), None)
+            .await
     }
 
     /// 写入一条 key-value 条目（含完整 metadata + created_at）。

src/agent/summary.rs

@@ -98,7 +98,11 @@ pub fn format_messages_as_text(messages: &[Message], max_tool_result_chars: usiz
                 content,
                 is_error,
             } => {
-                let label = if *is_error { "Tool Error" } else { "Tool Result" };
+                let label = if *is_error {
+                    "Tool Error"
+                } else {
+                    "Tool Result"
+                };
                 if let Some(text) = first_text(content) {
                     let truncated = truncate_chars(text, max_tool_result_chars);
                     lines.push(format!("{} [{}]: {}", label, tool_call_id, truncated));

src/document.rs

@@ -339,7 +339,14 @@ impl RecursiveCharacterSplitter {
             let take_n = overlap.min(prev_chars_count);
 
             // 字符级安全地取 prev 末尾 take_n 个字符
-            let tail: String = prev.chars().rev().take(take_n).collect::<Vec<_>>().into_iter().rev().collect();
+            let tail: String = prev
+                .chars()
+                .rev()
+                .take(take_n)
+                .collect::<Vec<_>>()
+                .into_iter()
+                .rev()
+                .collect();
             chunks[i] = format!("{}{}", tail, chunks[i]);
         }
 
@@ -408,8 +415,14 @@ mod tests {
         let chunks = splitter.split(&[doc]);
         assert_eq!(chunks.len(), 1);
         assert_eq!(chunks[0].content, "hello");
-        assert_eq!(chunks[0].metadata.get("chunk_index").map(|s| s.as_str()), Some("0"));
-        assert_eq!(chunks[0].metadata.get("chunk_count").map(|s| s.as_str()), Some("1"));
+        assert_eq!(
+            chunks[0].metadata.get("chunk_index").map(|s| s.as_str()),
+            Some("0")
+        );
+        assert_eq!(
+            chunks[0].metadata.get("chunk_count").map(|s| s.as_str()),
+            Some("1")
+        );
     }
 
     #[test]
@@ -464,7 +477,11 @@ mod tests {
         // para1 (5 chars) > chunk_size=4 → 递归降级到 char 级拆分
         // para2 同理
         // 总共应该产生多个 chunk
-        assert!(chunks.len() >= 2, "expected >= 2 chunks, got {}", chunks.len());
+        assert!(
+            chunks.len() >= 2,
+            "expected >= 2 chunks, got {}",
+            chunks.len()
+        );
     }
 
     #[test]
@@ -474,10 +491,18 @@ mod tests {
         let text: String = "a".repeat(200);
         let doc = Document::from_raw("long", &text);
         let chunks = splitter.split(&[doc]);
-        assert!(chunks.len() >= 3, "expected >= 3 chunks, got {}", chunks.len());
+        assert!(
+            chunks.len() >= 3,
+            "expected >= 3 chunks, got {}",
+            chunks.len()
+        );
         for chunk in &chunks {
             // chunk 内容 = overlap_tail(≤5) + new_content(≤50)，故 ≤ 55
-            assert!(chars_len(&chunk.content) <= 55, "chunk too long: {} chars", chars_len(&chunk.content));
+            assert!(
+                chars_len(&chunk.content) <= 55,
+                "chunk too long: {} chars",
+                chars_len(&chunk.content)
+            );
         }
     }
 
@@ -488,7 +513,11 @@ mod tests {
         let doc = Document::from_raw("g", "aa\n\nbb\n\ncc\n\ndd");
         let chunks = splitter.split(&[doc]);
         // 短段应被合并：总共应该少于 4 个 chunk
-        assert!(chunks.len() <= 3, "expected <= 3 chunks after merge, got {}", chunks.len());
+        assert!(
+            chunks.len() <= 3,
+            "expected <= 3 chunks after merge, got {}",
+            chunks.len()
+        );
     }
 
     #[test]
@@ -538,11 +567,19 @@ mod tests {
         let doc = Document::from_raw("cjk", &text);
         let chunks = splitter.split(&[doc]);
         // 30 字符 / 10 chunk_size = 3 个 chunk
-        assert!(chunks.len() >= 3, "expected >= 3 chunks for 30 chars / chunk_size=10, got {}", chunks.len());
+        assert!(
+            chunks.len() >= 3,
+            "expected >= 3 chunks for 30 chars / chunk_size=10, got {}",
+            chunks.len()
+        );
         for chunk in &chunks {
             let char_count = chars_len(&chunk.content);
             // chunk = overlap_tail(≤2) + new_content(≤10)，故 ≤ 12
-            assert!(char_count <= 12, "chunk char count {} exceeds 10+overlap", char_count);
+            assert!(
+                char_count <= 12,
+                "chunk char count {} exceeds 10+overlap",
+                char_count
+            );
         }
     }
 
@@ -569,12 +606,25 @@ mod tests {
     fn split_metadata_inheritance() {
         let splitter = RecursiveCharacterSplitter::new(100, 10);
         let mut doc = Document::new("m", "short content", "text/plain");
-        doc.metadata.insert("author".to_string(), "alice".to_string());
+        doc.metadata
+            .insert("author".to_string(), "alice".to_string());
         let chunks = splitter.split(&[doc]);
         assert_eq!(chunks.len(), 1);
-        assert_eq!(chunks[0].metadata.get("author").map(|s| s.as_str()), Some("alice"));
-        assert_eq!(chunks[0].metadata.get("source_id").map(|s| s.as_str()), Some("m"));
-        assert_eq!(chunks[0].metadata.get("chunk_index").map(|s| s.as_str()), Some("0"));
-        assert_eq!(chunks[0].metadata.get("chunk_count").map(|s| s.as_str()), Some("1"));
+        assert_eq!(
+            chunks[0].metadata.get("author").map(|s| s.as_str()),
+            Some("alice")
+        );
+        assert_eq!(
+            chunks[0].metadata.get("source_id").map(|s| s.as_str()),
+            Some("m")
+        );
+        assert_eq!(
+            chunks[0].metadata.get("chunk_index").map(|s| s.as_str()),
+            Some("0")
+        );
+        assert_eq!(
+            chunks[0].metadata.get("chunk_count").map(|s| s.as_str()),
+            Some("1")
+        );
     }
 }

src/engine/checkpointer.rs

@@ -16,8 +16,8 @@ use serde::{Deserialize, Serialize};
 use time::OffsetDateTime;
 
 use crate::agent::session::AgentSession;
-use crate::engine::snapshot::SessionSnapshot;
 use crate::engine::EngineError;
+use crate::engine::snapshot::SessionSnapshot;
 use crate::memory::store::MemoryStore;
 use crate::memory::types::{MemoryFilter, MemoryItem};
 
@@ -80,9 +80,8 @@ impl Checkpointer {
         let ckpt_id = generate_ckpt_id();
         let key = ckpt_key(&session.session_id, &ckpt_id);
 
-        let json = serde_json::to_string(&snapshot).map_err(|e| {
-            EngineError::Serialization(format!("snapshot serialize failed: {e}"))
-        })?;
+        let json = serde_json::to_string(&snapshot)
+            .map_err(|e| EngineError::Serialization(format!("snapshot serialize failed: {e}")))?;
 
         let item = MemoryItem {
             id: key,
@@ -141,10 +140,7 @@ impl Checkpointer {
     /// prefix 查询 `ckpt:{session_id}:` → 反序列化 `SessionSnapshot` → 提取元数据。
     /// 不需要 `CkptMeta` 单独存储——`SessionSnapshot` 已含 `turn_index` 字段，
     /// `created_at` 用 `MemoryItem.created_at` 转换。
-    pub async fn list_checkpoints(
-        &self,
-        session_id: &str,
-    ) -> Result<Vec<CkptMeta>, EngineError> {
+    pub async fn list_checkpoints(&self, session_id: &str) -> Result<Vec<CkptMeta>, EngineError> {
         let prefix = format!("ckpt:{}:", session_id);
         let filter = MemoryFilter {
             prefix: Some(prefix),
@@ -346,10 +342,7 @@ mod tests {
         let mut session = new_session_for_test("latest-session");
         cp.checkpoint(&session).await.unwrap();
         tokio::time::sleep(std::time::Duration::from_millis(2)).await;
-        session
-            .set_session_data("v", "2")
-            .await
-            .unwrap();
+        session.set_session_data("v", "2").await.unwrap();
         cp.checkpoint(&session).await.unwrap();
 
         let latest = cp.latest_snapshot("latest-session").await.unwrap().unwrap();
@@ -374,4 +367,4 @@ mod tests {
         assert_eq!(metas_a[0].session_id, "iso-a");
         assert_eq!(metas_b[0].session_id, "iso-b");
     }
-}
+}

src/engine/error.rs

@@ -48,4 +48,4 @@ pub enum EngineError {
     /// 调用方收到此错误时，子 session 已通过 `destroy()` 清理（SessionMeta + checkpoint 全部清空）。
     #[error("Dispatch failed: {0}")]
     DispatchFailed(String),
-}
+}

src/engine/mod.rs

@@ -20,4 +20,4 @@ pub use checkpointer::{Checkpointer, CkptMeta};
 pub use error::EngineError;
 pub use session_manager::{SessionManager, SessionManagerConfig};
 pub use snapshot::{SessionMemoryEntry, SessionSnapshot};
-pub use sub_agent::{DispatchConfig, SubTaskResult, SubTaskStreamEvent};
+pub use sub_agent::{DispatchConfig, SubTaskResult, SubTaskStreamEvent};

src/engine/session_manager.rs

@@ -141,11 +141,11 @@ impl SessionManager {
         Ok(())
     }
 
-    pub(crate) async fn load_session_meta(&self, session_id: &str) -> Result<Option<SessionMeta>, EngineError> {
-        let item = self
-            .store
-            .get(&SessionMeta::meta_key(session_id))
-            .await?;
+    pub(crate) async fn load_session_meta(
+        &self,
+        session_id: &str,
+    ) -> Result<Option<SessionMeta>, EngineError> {
+        let item = self.store.get(&SessionMeta::meta_key(session_id)).await?;
         match item {
             Some(item) => {
                 let meta: SessionMeta = serde_json::from_str(&item.content).map_err(|e| {
@@ -241,10 +241,7 @@ impl SessionManager {
     /// 按 ID 获取 session（**仅查内存**，不自动从存储恢复）。
     ///
     /// 冷启动时 `get()` 未命中返回 `SessionNotFound`。如需从存储恢复，使用 `recover()` 方法。
-    pub async fn get(
-        &self,
-        session_id: &str,
-    ) -> Result<Arc<Mutex<AgentSession>>, EngineError> {
+    pub async fn get(&self, session_id: &str) -> Result<Arc<Mutex<AgentSession>>, EngineError> {
         let sessions = self.sessions.read().await;
         let result = sessions.get(session_id).cloned();
         tracing::debug!(
@@ -392,12 +389,7 @@ impl SessionManager {
         session_id: &str,
         user_input: impl Into<String>,
     ) -> Result<
-        std::pin::Pin<
-            Box<
-                dyn futures_core::Stream<Item = crate::llm::stream::StreamEvent>
-                    + Send,
-            >,
-        >,
+        std::pin::Pin<Box<dyn futures_core::Stream<Item = crate::llm::stream::StreamEvent> + Send>>,
         EngineError,
     > {
         let session = self.get(session_id).await?;
@@ -593,8 +585,14 @@ mod tests {
             .await
             .unwrap();
         let session = sm.get(&id).await.unwrap();
-        sm.checkpointer().checkpoint(&*session.lock().await).await.unwrap();
-        assert_eq!(sm.checkpointer().list_checkpoints(&id).await.unwrap().len(), 1);
+        sm.checkpointer()
+            .checkpoint(&*session.lock().await)
+            .await
+            .unwrap();
+        assert_eq!(
+            sm.checkpointer().list_checkpoints(&id).await.unwrap().len(),
+            1
+        );
 
         sm.destroy(&id).await.unwrap();
 
@@ -603,7 +601,10 @@ mod tests {
         // SessionMeta 已删除
         assert!(sm.load_session_meta(&id).await.unwrap().is_none());
         // Checkpoint 已删除
-        assert_eq!(sm.checkpointer().list_checkpoints(&id).await.unwrap().len(), 0);
+        assert_eq!(
+            sm.checkpointer().list_checkpoints(&id).await.unwrap().len(),
+            0
+        );
 
         // destroy 不存在的 session 不报错
         sm.destroy(&id).await.unwrap();
@@ -682,7 +683,10 @@ mod tests {
         {
             let s = sm.get(&id).await.unwrap();
             s.lock().await.set_session_data("k", "v1").await.unwrap();
-            sm.checkpointer().checkpoint(&*s.lock().await).await.unwrap();
+            sm.checkpointer()
+                .checkpoint(&*s.lock().await)
+                .await
+                .unwrap();
         }
 
         // 2. 模拟"进程重启"——清空内存但保留 store
@@ -710,7 +714,10 @@ mod tests {
             .unwrap();
         {
             let s = sm.get(&id).await.unwrap();
-            sm.checkpointer().checkpoint(&*s.lock().await).await.unwrap();
+            sm.checkpointer()
+                .checkpoint(&*s.lock().await)
+                .await
+                .unwrap();
         }
 
         let err = sm
@@ -750,7 +757,10 @@ mod tests {
         let err = sm.submit_turn(&id, "hello").await.unwrap_err();
         match err {
             EngineError::Agent(crate::agent::error::AgentError::Llm(_)) => {}
-            other => panic!("expected EngineError::Agent(AgentError::Llm), got {:?}", other),
+            other => panic!(
+                "expected EngineError::Agent(AgentError::Llm), got {:?}",
+                other
+            ),
         }
     }
 
@@ -772,8 +782,14 @@ mod tests {
         let _ = sm.submit_turn(&id, "x").await;
 
         // 手动 checkpoint 仍可工作
-        sm.checkpointer().checkpoint(&*s.lock().await).await.unwrap();
-        assert_eq!(sm.checkpointer().list_checkpoints(&id).await.unwrap().len(), 1);
+        sm.checkpointer()
+            .checkpoint(&*s.lock().await)
+            .await
+            .unwrap();
+        assert_eq!(
+            sm.checkpointer().list_checkpoints(&id).await.unwrap().len(),
+            1
+        );
     }
 
     #[tokio::test]
@@ -787,7 +803,8 @@ mod tests {
             .unwrap();
 
         // 构造一个新 session（同 session_id）然后 replace
-        let mut new_session = AgentSession::new(Arc::new(StubAgent("a".into())), &id, make_bundle());
+        let mut new_session =
+            AgentSession::new(Arc::new(StubAgent("a".into())), &id, make_bundle());
         new_session
             .set_session_data("replaced", "yes")
             .await
@@ -838,8 +855,7 @@ mod tests {
     /// restore_memory 幂等性：第二次调用应立即返回 Ok(())（pending 已被清空）。
     #[tokio::test]
     async fn restore_memory_is_idempotent() {
-        let store: Arc<dyn MemoryStore> =
-            Arc::new(crate::memory::store::InMemoryStore::new());
+        let store: Arc<dyn MemoryStore> = Arc::new(crate::memory::store::InMemoryStore::new());
         let sm = SessionManager::new(store.clone());
 
         let id = sm
@@ -848,12 +864,11 @@ mod tests {
             .unwrap();
         {
             let s = sm.get(&id).await.unwrap();
-            s.lock()
-                .await
-                .set_session_data("k", "v")
+            s.lock().await.set_session_data("k", "v").await.unwrap();
+            sm.checkpointer()
+                .checkpoint(&*s.lock().await)
                 .await
                 .unwrap();
-            sm.checkpointer().checkpoint(&*s.lock().await).await.unwrap();
         }
 
         // 模拟"进程重启"——新建 SessionManager，复用 store
@@ -877,8 +892,7 @@ mod tests {
     /// 10 并发 session 创建：验证 RwLock 写锁争用下不冲突，所有 ID 唯一。
     #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
     async fn concurrent_create_ten_sessions() {
-        let store: Arc<dyn MemoryStore> =
-            Arc::new(crate::memory::store::InMemoryStore::new());
+        let store: Arc<dyn MemoryStore> = Arc::new(crate::memory::store::InMemoryStore::new());
         let sm = Arc::new(SessionManager::new(store));
 
         let mut handles = Vec::with_capacity(10);
@@ -905,4 +919,4 @@ mod tests {
             assert!(sm.get(id).await.is_ok());
         }
     }
-}
+}

src/engine/snapshot.rs

@@ -36,4 +36,4 @@ pub struct SessionSnapshot {
     pub last_summary_turn: Option<u32>,
     #[serde(default)]
     pub session_memory_data: std::collections::HashMap<String, SessionMemoryEntry>,
-}
+}

src/engine/sub_agent.rs

@@ -77,10 +77,7 @@ pub enum SubTaskStreamEvent {
     /// 执行完成，携带完整结果。
     Completed(SubTaskResult),
     /// 流式调度中的错误。
-    Error {
-        child_id: String,
-        error: String,
-    },
+    Error { child_id: String, error: String },
 }
 
 impl std::fmt::Display for SubTaskStreamEvent {
@@ -150,7 +147,7 @@ impl SessionManager {
 
         // 3. 过滤（按 bridge_keys）
         let filtered: Vec<_> = match &config.bridge_keys {
-            None => Vec::new(), // None = 不继承任何
+            None => Vec::new(),                       // None = 不继承任何
             Some(keys) if keys.is_empty() => entries, // 空列表 = 全部继承
             Some(keys) => entries
                 .into_iter()
@@ -343,9 +340,7 @@ impl SessionManager {
         task: impl Into<String>,
         config: DispatchConfig,
     ) -> Result<
-        std::pin::Pin<
-            Box<dyn futures_core::Stream<Item = SubTaskStreamEvent> + Send>,
-        >,
+        std::pin::Pin<Box<dyn futures_core::Stream<Item = SubTaskStreamEvent> + Send>>,
         EngineError,
     > {
         use futures_util::StreamExt;
@@ -473,9 +468,7 @@ impl SessionManager {
                         }
                     };
                     let mut guard = session.lock().await;
-                    guard
-                        .finalize_turn(response, new_messages)
-                        .await
+                    guard.finalize_turn(response, new_messages).await
                 };
 
                 if let Err(e) = lock_result {
@@ -575,8 +568,7 @@ mod tests {
     }
 
     async fn make_manager_and_bundle() -> (Arc<SessionManager>, Arc<RuntimeBundle>) {
-        let store: Arc<dyn crate::memory::store::MemoryStore> =
-            Arc::new(InMemoryStore::new());
+        let store: Arc<dyn crate::memory::store::MemoryStore> = Arc::new(InMemoryStore::new());
         let provider = Arc::new(MockProvider::new(vec![
             assistant_text("child response 1"),
             assistant_text("child response 2"),
@@ -702,7 +694,12 @@ mod tests {
 
         // dispatch 到不存在的 parent_id
         let result = sm
-            .dispatch("nonexistent_parent", child, "task", DispatchConfig::default())
+            .dispatch(
+                "nonexistent_parent",
+                child,
+                "task",
+                DispatchConfig::default(),
+            )
             .await;
         assert!(result.is_err());
 
@@ -716,7 +713,12 @@ mod tests {
         let (sm, _bundle) = make_manager_and_bundle().await;
         let child: Arc<dyn Agent> = Arc::new(MockAgent::new("child"));
         let result = sm
-            .dispatch("nonexistent_parent", child, "task", DispatchConfig::default())
+            .dispatch(
+                "nonexistent_parent",
+                child,
+                "task",
+                DispatchConfig::default(),
+            )
             .await;
         assert!(matches!(result, Err(EngineError::SessionNotFound(_))));
     }
@@ -724,9 +726,10 @@ mod tests {
     // ====== dispatch_all 测试 ======
 
     /// 提供充足的 mock response（>= 3）
-    async fn make_manager_and_bundle_for_all(n: usize) -> (Arc<SessionManager>, Arc<RuntimeBundle>) {
-        let store: Arc<dyn crate::memory::store::MemoryStore> =
-            Arc::new(InMemoryStore::new());
+    async fn make_manager_and_bundle_for_all(
+        n: usize,
+    ) -> (Arc<SessionManager>, Arc<RuntimeBundle>) {
+        let store: Arc<dyn crate::memory::store::MemoryStore> = Arc::new(InMemoryStore::new());
         let responses: Vec<_> = (0..n)
             .map(|i| assistant_text(&format!("response {i}")))
             .collect();
@@ -747,8 +750,7 @@ mod tests {
     /// 创建空 mock responses 的 manager 和 bundle —— 后续 dispatch 会触发
     /// "MockProvider: 预设响应已用完" 错误，可用于测试错误传播。
     async fn make_manager_and_bundle_empty_mock() -> (Arc<SessionManager>, Arc<RuntimeBundle>) {
-        let store: Arc<dyn crate::memory::store::MemoryStore> =
-            Arc::new(InMemoryStore::new());
+        let store: Arc<dyn crate::memory::store::MemoryStore> = Arc::new(InMemoryStore::new());
         let provider = Arc::new(MockProvider::empty());
         let bundle = Arc::new(
             AgentBuilder::new()
@@ -1048,8 +1050,7 @@ mod tests {
         }
 
         // 验证：收到 Error 事件
-        let (err_child_id, err_message) =
-            error_received.expect("Error event should be received");
+        let (err_child_id, err_message) = error_received.expect("Error event should be received");
         assert_eq!(
             Some(err_child_id.as_str()),
             child_id_from_event.as_deref(),

src/engine/switch.rs

@@ -129,8 +129,7 @@ mod tests {
     }
 
     async fn make_manager_and_bundle() -> (Arc<SessionManager>, Arc<RuntimeBundle>) {
-        let store: Arc<dyn crate::memory::store::MemoryStore> =
-            Arc::new(InMemoryStore::new());
+        let store: Arc<dyn crate::memory::store::MemoryStore> = Arc::new(InMemoryStore::new());
         let provider = Arc::new(MockProvider::new(vec![
             assistant_text("response1"),
             assistant_text("response2"),
@@ -175,10 +174,7 @@ mod tests {
         {
             let session = sm.get(&sid).await.unwrap();
             let mut guard = session.lock().await;
-            guard
-                .set_session_data("key1", "value1")
-                .await
-                .unwrap();
+            guard.set_session_data("key1", "value1").await.unwrap();
         }
         sm.submit_turn(&sid, "hello").await.unwrap();
 

src/lib.rs

@@ -1,19 +1,19 @@
 //! agcore —— 智能体（Agent）核心工具箱。
 
-#[cfg(any(feature = "llm-types", feature = "llm"))]
-pub mod llm;
+#[cfg(feature = "agent")]
+pub mod agent;
 #[cfg(feature = "document")]
 pub mod document;
+#[cfg(feature = "engine")]
+pub mod engine;
+#[cfg(any(feature = "llm-types", feature = "llm"))]
+pub mod llm;
+#[cfg(feature = "memory")]
+pub mod memory;
 #[cfg(feature = "prompt")]
 pub mod prompt;
 #[cfg(feature = "tools")]
 pub mod tools;
-#[cfg(feature = "memory")]
-pub mod memory;
-#[cfg(feature = "agent")]
-pub mod agent;
-#[cfg(feature = "engine")]
-pub mod engine;
 
 #[cfg(feature = "document")]
 pub use document::Document;

src/llm.rs

@@ -1,7 +1,5 @@
 //! LLM 调用周期 —— 大模型基础调用周期控制。
 
-#[cfg(feature = "llm-types")]
-pub mod types;
 #[cfg(feature = "llm")]
 pub mod compact;
 #[cfg(feature = "llm")]
@@ -16,6 +14,8 @@ pub mod error;
 pub mod hooks;
 #[cfg(feature = "llm")]
 pub mod mock;
+#[cfg(feature = "llm-types")]
+pub mod types;
 // provider 模块依赖 reqwest（通过 reqwest::Client），仅在任一 provider feature 启用时编译
 #[cfg(any(
     feature = "provider-openai",

src/llm/cycle.rs

@@ -15,17 +15,17 @@ use serde_json::Value;
 use tokio::sync::mpsc;
 use tokio_stream::wrappers::UnboundedReceiverStream;
 
+use crate::llm::LlmProvider;
 use crate::llm::compact::{CompactConfig, CompactState, microcompact, should_compact};
 use crate::llm::cycle::retry::should_retry;
 use crate::llm::error::LlmError;
 use crate::llm::hooks::{HookContext, HookEvent, HookExecutor};
-use crate::llm::LlmProvider;
 use crate::llm::stream::StreamEvent;
+use crate::llm::types::ToolChoice;
 use crate::llm::types::message::{ContentBlock, Message};
 use crate::llm::types::request_v2::MessageRequest;
 use crate::llm::types::response_v2::{MessageResponse, PartialMessageResponse, StopReason};
 use crate::llm::types::tool::ToolDef;
-use crate::llm::types::ToolChoice;
 #[cfg(feature = "tools")]
 use crate::tools::ToolRegistry;
 
@@ -452,10 +452,7 @@ impl LlmCycle {
     /// 内部请求方法（与 `submit` 共享重试逻辑，但不 push user message 和 Assistant 响应）。
     ///
     /// 用于 `submit_with_tools()` 的多轮 tool 循环。
-    async fn submit_request(
-        &mut self,
-        tools: &[ToolDef],
-    ) -> Result<MessageResponse, LlmError> {
+    async fn submit_request(&mut self, tools: &[ToolDef]) -> Result<MessageResponse, LlmError> {
         let mut attempts = 0;
 
         loop {
@@ -913,11 +910,10 @@ async fn run_tool_loop(
                 Ok(v) => {
                     // ponytail: 与 submit_with_tools 行为对齐 —— 用 truncate_tool_result
                     // 截断结果以防止超大工具输出在 tool 循环中膨胀 messages 上下文窗口
-                    let serialized =
-                        serde_json::to_string(v).unwrap_or_else(|e| {
-                            tracing::warn!("工具结果序列化失败: {}", e);
-                            "{}".to_string()
-                        });
+                    let serialized = serde_json::to_string(v).unwrap_or_else(|e| {
+                        tracing::warn!("工具结果序列化失败: {}", e);
+                        "{}".to_string()
+                    });
                     truncate_tool_result(&serialized, max_bytes)
                 }
                 Err(e) if e.is_recoverable() => format!("错误: {}", e),
@@ -1204,8 +1200,7 @@ mod tests {
     #[tokio::test(flavor = "multi_thread")]
     async fn test_submit_with_tools_stream_pure_text() {
         let provider = Mock::new(vec![assistant_text_response("你好")]);
-        let mut cycle =
-            LlmCycle::new(Box::new(provider), CycleConfig::default());
+        let mut cycle = LlmCycle::new(Box::new(provider), CycleConfig::default());
         let mut registry = ToolRegistry::new();
         registry.register(std::sync::Arc::new(AddTool)).unwrap();
 
@@ -1216,23 +1211,36 @@ mod tests {
         let events = drain(stream).await;
 
         // 期望序列：MessageStart → ContentBlockStart(Text) → TextDelta → ContentBlockEnd → CostUpdate → MessageComplete
-        assert!(matches!(events.first(), Some(StreamEvent::MessageStart { .. })));
-        assert!(events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::TextDelta { text } if text == "你好")));
-        assert!(events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::MessageComplete { .. })));
+        assert!(matches!(
+            events.first(),
+            Some(StreamEvent::MessageStart { .. })
+        ));
+        assert!(
+            events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::TextDelta { text } if text == "你好"))
+        );
+        assert!(
+            events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::MessageComplete { .. }))
+        );
         // 纯文本流不应有 ToolExecutionStarted/Completed 事件
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::ToolExecutionStarted { .. })));
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::ToolExecutionCompleted { .. })));
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::Error { .. })));
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::ToolExecutionStarted { .. }))
+        );
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::ToolExecutionCompleted { .. }))
+        );
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::Error { .. }))
+        );
     }
 
     /// Phase 9 测试 3.2 — 单轮工具调用
@@ -1242,8 +1250,7 @@ mod tests {
             assistant_tool_call_response(vec![("call_1", "add", r#"{"a":1,"b":2}"#)]),
             assistant_text_response("答案是 3"),
         ]);
-        let mut cycle =
-            LlmCycle::new(Box::new(provider), CycleConfig::default());
+        let mut cycle = LlmCycle::new(Box::new(provider), CycleConfig::default());
         let mut registry = ToolRegistry::new();
         registry.register(std::sync::Arc::new(AddTool)).unwrap();
 
@@ -1320,8 +1327,7 @@ mod tests {
             assistant_tool_call_response(vec![("call_3", "add", r#"{"a":5,"b":6}"#)]),
             assistant_text_response("完成"),
         ]);
-        let mut cycle =
-            LlmCycle::new(Box::new(provider), CycleConfig::default());
+        let mut cycle = LlmCycle::new(Box::new(provider), CycleConfig::default());
         let mut registry = ToolRegistry::new();
         registry.register(std::sync::Arc::new(AddTool)).unwrap();
 
@@ -1381,8 +1387,11 @@ mod tests {
             })
             .collect();
         assert!(
-            error_events.iter().any(|m| m.contains("达到最大工具循环轮次")),
-            "应包含最大轮次超限 Error，实际: {:?}", error_events
+            error_events
+                .iter()
+                .any(|m| m.contains("达到最大工具循环轮次")),
+            "应包含最大轮次超限 Error，实际: {:?}",
+            error_events
         );
 
         // 工具调用次数应 ≤ 2
@@ -1409,10 +1418,8 @@ mod tests {
             async fn chat_stream(
                 &self,
                 _r: MessageRequest,
-            ) -> Result<
-                Pin<Box<dyn Stream<Item = Result<StreamEvent, LlmError>> + Send>>,
-                LlmError,
-            > {
+            ) -> Result<Pin<Box<dyn Stream<Item = Result<StreamEvent, LlmError>> + Send>>, LlmError>
+            {
                 Err(LlmError::Other("网络错误".into()))
             }
             fn capabilities(&self) -> ProviderCapabilities {
@@ -1436,25 +1443,29 @@ mod tests {
 
         // 第一个事件应是 Error（chat_stream Err 立即事件化）
         assert!(
-            events.first().map(|e| matches!(e, StreamEvent::Error { .. }))
+            events
+                .first()
+                .map(|e| matches!(e, StreamEvent::Error { .. }))
                 == Some(true),
-            "流中首个事件应是 Error，实际: {:?}", events.first()
+            "流中首个事件应是 Error，实际: {:?}",
+            events.first()
+        );
+        assert!(
+            events
+                .iter()
+                .filter_map(|e| match e {
+                    StreamEvent::Error { message } => Some(message.as_str()),
+                    _ => None,
+                })
+                .any(|m| m.contains("网络错误"))
         );
-        assert!(events
-            .iter()
-            .filter_map(|e| match e {
-                StreamEvent::Error { message } => Some(message.as_str()),
-                _ => None,
-            })
-            .any(|m| m.contains("网络错误")));
     }
 
     /// Phase 9 测试 3.6 — 空 tool_registry
     #[tokio::test(flavor = "multi_thread")]
     async fn test_submit_with_tools_stream_empty_registry() {
         let provider = Mock::new(vec![assistant_text_response("纯文本回答")]);
-        let mut cycle =
-            LlmCycle::new(Box::new(provider), CycleConfig::default());
+        let mut cycle = LlmCycle::new(Box::new(provider), CycleConfig::default());
         let registry = ToolRegistry::new();
 
         let stream = cycle
@@ -1464,18 +1475,26 @@ mod tests {
         let events = drain(stream).await;
 
         // 流退化为纯文本流 —— 无 ToolExecution 事件，无 Error
-        assert!(events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::TextDelta { text } if text == "纯文本回答")));
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::ToolExecutionStarted { .. })));
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::ToolExecutionCompleted { .. })));
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::Error { .. })));
+        assert!(
+            events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::TextDelta { text } if text == "纯文本回答"))
+        );
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::ToolExecutionStarted { .. }))
+        );
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::ToolExecutionCompleted { .. }))
+        );
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::Error { .. }))
+        );
     }
 
     /// Phase 9 测试 3.7 — 不可恢复工具错误
@@ -1507,8 +1526,7 @@ mod tests {
             assistant_tool_call_response(vec![("call_x", "fail_unrecoverable", "{}")]),
             assistant_text_response("忽略"),
         ]);
-        let mut cycle =
-            LlmCycle::new(Box::new(provider), CycleConfig::default());
+        let mut cycle = LlmCycle::new(Box::new(provider), CycleConfig::default());
         let mut registry = ToolRegistry::new();
         registry
             .register(std::sync::Arc::new(UnrecoverableTool))
@@ -1566,8 +1584,7 @@ mod tests {
             assistant_tool_call_response(vec![("call_y", "fail_recoverable", "{}")]),
             assistant_text_response("已恢复"),
         ]);
-        let mut cycle =
-            LlmCycle::new(Box::new(provider), CycleConfig::default());
+        let mut cycle = LlmCycle::new(Box::new(provider), CycleConfig::default());
         let mut registry = ToolRegistry::new();
         registry
             .register(std::sync::Arc::new(RecoverableTool))
@@ -1580,9 +1597,11 @@ mod tests {
         let events = drain(stream).await;
 
         // 可恢复错误：tool_result 回传 LLM，最终流正常结束
-        assert!(!events
-            .iter()
-            .any(|e| matches!(e, StreamEvent::Error { .. })));
+        assert!(
+            !events
+                .iter()
+                .any(|e| matches!(e, StreamEvent::Error { .. }))
+        );
         // 最终 MessageComplete 应是 Stop（不是 ToolUse）
         let final_response = events
             .iter()
@@ -1602,7 +1621,10 @@ mod tests {
                 _ => None,
             })
             .unwrap();
-        assert!(completed, "可恢复错误的 ToolExecutionCompleted.is_error 应为 true");
+        assert!(
+            completed,
+            "可恢复错误的 ToolExecutionCompleted.is_error 应为 true"
+        );
     }
 
     /// Phase 9 测试 3.9 — 工具超时
@@ -1651,9 +1673,7 @@ mod tests {
         ]);
         let mut cycle = LlmCycle::new(Box::new(provider), config);
         let mut registry = ToolRegistry::new();
-        registry
-            .register(std::sync::Arc::new(SlowTool))
-            .unwrap();
+        registry.register(std::sync::Arc::new(SlowTool)).unwrap();
 
         let stream = cycle
             .submit_with_tools_stream("test".to_string(), std::sync::Arc::new(registry))
@@ -1673,8 +1693,14 @@ mod tests {
                 _ => None,
             })
             .collect();
-        assert!(!tool_completed.is_empty(), "应有 ToolExecutionCompleted 事件");
-        assert!(tool_completed[0].0, "超时后 ToolExecutionCompleted.is_error 应为 true");
+        assert!(
+            !tool_completed.is_empty(),
+            "应有 ToolExecutionCompleted 事件"
+        );
+        assert!(
+            tool_completed[0].0,
+            "超时后 ToolExecutionCompleted.is_error 应为 true"
+        );
         assert_eq!(tool_completed[0].1, "slow_tool");
 
         // 2. 流中应有不可恢复错误终止事件（tool_timeout → McpTimeout → 不可恢复 → Error）
@@ -1686,10 +1712,9 @@ mod tests {
             })
             .collect();
         assert!(
+            error_events.iter().any(|m| m.contains("不可恢复错误")),
+            "应有不可恢复错误事件终止流，实际事件: {:?}",
             error_events
-                .iter()
-                .any(|m| m.contains("不可恢复错误")),
-            "应有不可恢复错误事件终止流，实际事件: {:?}", error_events
         );
 
         // 3. 第一轮的 MessageComplete { stop_reason: ToolUse } 在 Error 之前已发出

src/llm/embedding.rs

@@ -151,17 +151,18 @@ mod tests {
         let result = embedder.embed(&inputs).await.unwrap();
         for vec in &result {
             let norm = l2_norm(vec);
-            assert!((norm - 1.0).abs() < 1e-5, "vector norm should be ~1.0, got {}", norm);
+            assert!(
+                (norm - 1.0).abs() < 1e-5,
+                "vector norm should be ~1.0, got {}",
+                norm
+            );
         }
     }
 
     #[tokio::test]
     async fn embed_different_inputs_different_vectors() {
         let embedder = MockEmbedding::new(16);
-        let r1 = embedder
-            .embed(&["hello world".to_string()])
-            .await
-            .unwrap();
+        let r1 = embedder.embed(&["hello world".to_string()]).await.unwrap();
         let r2 = embedder
             .embed(&["completely different".to_string()])
             .await
@@ -178,6 +179,10 @@ mod tests {
         assert_eq!(result.len(), 1);
         assert_eq!(result[0].len(), 4);
         let norm = l2_norm(&result[0]);
-        assert!((norm - 1.0).abs() < 1e-5, "empty-string vector norm should be ~1.0, got {}", norm);
+        assert!(
+            (norm - 1.0).abs() < 1e-5,
+            "empty-string vector norm should be ~1.0, got {}",
+            norm
+        );
     }
 }

src/llm/provider/anthropic.rs

@@ -18,7 +18,6 @@ use serde::{Deserialize, Serialize};
 use serde_json::{Value, json};
 use tracing::{debug, error, info, warn};
 
-use crate::llm::{LlmProvider, ProviderCapabilities, ProviderFeatures};
 use crate::llm::error::LlmError;
 use crate::llm::types::message::{ContentBlock, ContentBlockType, Message};
 use crate::llm::types::request_v2::MessageRequest;
@@ -26,6 +25,7 @@ use crate::llm::types::response_v2::{
     MessageResponse, PartialMessageResponse, PartialUsage, StopReason, StreamEvent,
 };
 use crate::llm::types::usage::Usage;
+use crate::llm::{LlmProvider, ProviderCapabilities, ProviderFeatures};
 
 /// Anthropic Provider 默认 `max_tokens` 兜底值。
 ///

src/llm/provider/ollama.rs

@@ -11,10 +11,10 @@ use futures_core::Stream;
 use reqwest::Client;
 
 use super::openai::GenericOpenaiProvider;
-use crate::llm::{LlmProvider, ProviderCapabilities};
 use crate::llm::error::LlmError;
 use crate::llm::types::request_v2::MessageRequest;
 use crate::llm::types::response_v2::{MessageResponse, StreamEvent};
+use crate::llm::{LlmProvider, ProviderCapabilities};
 
 /// Ollama 本地 Provider —— OpenAI-compatible 协议的 newtype 包装。
 ///

src/llm/provider/openai.rs

@@ -21,7 +21,6 @@ use serde::Serialize;
 use serde_json::Value;
 use tracing::{debug, error, info};
 
-use crate::llm::{LlmProvider, ProviderCapabilities, ProviderFeatures};
 use crate::llm::convert::{from_openai, to_openai};
 use crate::llm::error::LlmError;
 use crate::llm::types::message::{ContentBlock, ContentBlockType, Message};
@@ -32,6 +31,7 @@ use crate::llm::types::response_v2::{
 };
 use crate::llm::types::shared::{FinishReason, ResponseFormat, ServiceTier, StopSequence};
 use crate::llm::types::tool::{OpenaiToolCall, OpenaiToolDefinition, ToolChoice};
+use crate::llm::{LlmProvider, ProviderCapabilities, ProviderFeatures};
 use serde::Deserialize;
 
 // =============================================================================

src/llm/provider/registry.rs

@@ -2,9 +2,9 @@
 
 use std::collections::HashMap;
 
+use crate::llm::LlmProvider;
 use crate::llm::error::LlmError;
 use crate::llm::provider::{ProviderConfig, ProviderType, create_provider};
-use crate::llm::LlmProvider;
 
 /// Provider 注册表 —— 管理多个 LLM Provider 实例。
 ///

src/llm/types/request_v2.rs

@@ -232,10 +232,16 @@ mod tests {
         let decoded: MessageRequest = serde_json::from_str(&json).expect("deserialize");
         assert_eq!(decoded.tools.len(), 1);
         assert_eq!(decoded.tools[0].name, "add");
-        assert_eq!(decoded.tools[0].description.as_deref(), Some("add two numbers"));
+        assert_eq!(
+            decoded.tools[0].description.as_deref(),
+            Some("add two numbers")
+        );
         assert_eq!(decoded.tools[0].parameters, params);
 
         // 验证序列化 JSON 不含 ToolDef 没有的字段（如 strict），保持 wire-format 兼容
-        assert!(!json.contains("strict"), "ToolDef 序列化不应包含 strict 字段");
+        assert!(
+            !json.contains("strict"),
+            "ToolDef 序列化不应包含 strict 字段"
+        );
     }
 }

src/llm/types/response_v2.rs

@@ -372,7 +372,8 @@ impl PartialMessageResponse {
             }
             // 元事件：不参与内容块累积，不修改 partial 状态
             //（Phase 9 —— 工具执行透明化，由 run_tool_loop 在工具前后插入）
-            StreamEvent::ToolExecutionStarted { .. } | StreamEvent::ToolExecutionCompleted { .. } => true,
+            StreamEvent::ToolExecutionStarted { .. }
+            | StreamEvent::ToolExecutionCompleted { .. } => true,
         }
     }
 

src/llm/types/usage.rs

@@ -63,9 +63,7 @@ impl CostTracker {
 
 impl From<Usage> for CostTracker {
     fn from(usage: Usage) -> Self {
-        CostTracker {
-            accumulated: usage,
-        }
+        CostTracker { accumulated: usage }
     }
 }
 

src/memory.rs

@@ -13,12 +13,14 @@ pub mod vector_store;
 // 高频类型（大多数下游需要）
 pub use conversation::{ConversationMemory, ConversationMemoryConfig};
 pub use error::MemoryError;
-pub use graph::{GraphEntity, GraphRelation, InMemoryGraph, KnowledgeGraph, RelationDirection, ScoredEntity};
+pub use graph::{
+    GraphEntity, GraphRelation, InMemoryGraph, KnowledgeGraph, RelationDirection, ScoredEntity,
+};
 pub use knowledge::KnowledgeStore;
 pub use retriever::MemoryRetriever;
-pub use store::{InMemoryStore, MemoryStore};
 #[cfg(feature = "memory-sqlite")]
 pub use store::SqliteStore;
+pub use store::{InMemoryStore, MemoryStore};
 #[allow(deprecated)]
 pub use vector::{InMemoryVectorRetriever, VectorRetriever};
 pub use vector_store::{InMemoryVectorStore, PersistentVectorStore, RagPipeline, VectorStore};

src/memory/graph.rs

@@ -35,7 +35,11 @@ pub struct GraphEntity {
 
 impl GraphEntity {
     /// 创建一个最小实体（仅 id + name + type，其余为空）。
-    pub fn new(id: impl Into<String>, name: impl Into<String>, entity_type: impl Into<String>) -> Self {
+    pub fn new(
+        id: impl Into<String>,
+        name: impl Into<String>,
+        entity_type: impl Into<String>,
+    ) -> Self {
         Self {
             id: id.into(),
             name: name.into(),
@@ -81,7 +85,10 @@ impl GraphRelation {
 
     /// 复合键：`source_id:target_id:relation_type`，用于去重和查找。
     pub fn composite_key(&self) -> String {
-        format!("{}:{}:{}", self.source_id, self.target_id, self.relation_type)
+        format!(
+            "{}:{}:{}",
+            self.source_id, self.target_id, self.relation_type
+        )
     }
 }
 
@@ -253,9 +260,10 @@ impl KnowledgeGraph for InMemoryGraph {
         if entity.id.is_empty() {
             return Err(MemoryError::InvalidInput("entity id is empty".into()));
         }
-        let mut inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let mut inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         // upsert：若已存在，先收集旧标签用于清理反向引用（避免同时借用 entities 和 tag_index）
         let old_tags: Vec<String> = inner
             .entities
@@ -283,16 +291,18 @@ impl KnowledgeGraph for InMemoryGraph {
     }
 
     async fn get_entity(&self, id: &str) -> Result<Option<GraphEntity>, MemoryError> {
-        let inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         Ok(inner.entities.get(id).cloned())
     }
 
     async fn remove_entity(&self, id: &str) -> Result<(), MemoryError> {
-        let mut inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let mut inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         // 移除实体并清理其标签反向引用
         if let Some(entity) = inner.entities.remove(id) {
             for tag in &entity.tags {
@@ -312,9 +322,10 @@ impl KnowledgeGraph for InMemoryGraph {
     }
 
     async fn add_relation(&self, relation: GraphRelation) -> Result<(), MemoryError> {
-        let mut inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let mut inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         // 校验两端实体存在
         if !inner.entities.contains_key(&relation.source_id) {
             return Err(MemoryError::InvalidInput(format!(
@@ -348,11 +359,14 @@ impl KnowledgeGraph for InMemoryGraph {
         target_id: &str,
         relation_type: &str,
     ) -> Result<(), MemoryError> {
-        let mut inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let mut inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         inner.relations.retain(|r| {
-            !(r.source_id == source_id && r.target_id == target_id && r.relation_type == relation_type)
+            !(r.source_id == source_id
+                && r.target_id == target_id
+                && r.relation_type == relation_type)
         });
         Ok(())
     }
@@ -364,9 +378,10 @@ impl KnowledgeGraph for InMemoryGraph {
         direction: RelationDirection,
         relation_types: Option<&[&str]>,
     ) -> Result<Vec<ScoredEntity>, MemoryError> {
-        let inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
 
         // 1. 验证起点存在
         if !inner.entities.contains_key(entity_id) {
@@ -470,9 +485,10 @@ impl KnowledgeGraph for InMemoryGraph {
     }
 
     async fn find_by_keywords(&self, keywords: &[String]) -> Result<Vec<GraphEntity>, MemoryError> {
-        let inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         if keywords.is_empty() {
             return Ok(Vec::new());
         }
@@ -495,9 +511,10 @@ impl KnowledgeGraph for InMemoryGraph {
     }
 
     async fn find_tags(&self, prefix: &str) -> Result<Vec<String>, MemoryError> {
-        let inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
         let prefix_l = prefix.to_lowercase();
         let mut tags: Vec<String> = inner
             .tag_index
@@ -514,9 +531,10 @@ impl KnowledgeGraph for InMemoryGraph {
         entity_id: &str,
         tags: Vec<String>,
     ) -> Result<usize, MemoryError> {
-        let mut inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
+        let mut inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
 
         // 先收集旧标签（避免同时借用 entities 和 tag_index）
         let old_tags: Vec<String> = inner
@@ -554,13 +572,11 @@ impl KnowledgeGraph for InMemoryGraph {
     }
 
     async fn entity_count_by_tag(&self, tag: &str) -> Result<usize, MemoryError> {
-        let inner = self.inner.lock().map_err(|e| {
-            MemoryError::RetrievalError(format!("lock poisoned: {e}"))
-        })?;
-        Ok(inner
-            .tag_index
-            .get(tag)
-            .map_or(0, |ids| ids.len()))
+        let inner = self
+            .inner
+            .lock()
+            .map_err(|e| MemoryError::RetrievalError(format!("lock poisoned: {e}")))?;
+        Ok(inner.tag_index.get(tag).map_or(0, |ids| ids.len()))
     }
 
     fn tag_constraints(&self) -> TagConstraints {
@@ -660,10 +676,7 @@ mod tests {
     #[tokio::test]
     async fn add_relation_validates_entities() {
         let graph = InMemoryGraph::new();
-        graph
-            .add_entity(make_entity("a", "A", "x"))
-            .await
-            .unwrap();
+        graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
         // target 不存在
         let result = graph
             .add_relation(GraphRelation::new("a", "ghost", "r", 0.5))
@@ -674,14 +687,8 @@ mod tests {
     #[tokio::test]
     async fn add_relation_upsert_overrides_weight() {
         let graph = InMemoryGraph::new();
-        graph
-            .add_entity(make_entity("a", "A", "x"))
-            .await
-            .unwrap();
-        graph
-            .add_entity(make_entity("b", "B", "x"))
-            .await
-            .unwrap();
+        graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
+        graph.add_entity(make_entity("b", "B", "x")).await.unwrap();
         graph
             .add_relation(GraphRelation::new("a", "b", "r", 0.5))
             .await
@@ -701,14 +708,8 @@ mod tests {
     #[tokio::test]
     async fn remove_relation() {
         let graph = InMemoryGraph::new();
-        graph
-            .add_entity(make_entity("a", "A", "x"))
-            .await
-            .unwrap();
-        graph
-            .add_entity(make_entity("b", "B", "x"))
-            .await
-            .unwrap();
+        graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
+        graph.add_entity(make_entity("b", "B", "x")).await.unwrap();
         graph
             .add_relation(GraphRelation::new("a", "b", "r", 0.5))
             .await
@@ -726,10 +727,7 @@ mod tests {
     #[tokio::test]
     async fn get_related_empty_graph() {
         let graph = InMemoryGraph::new();
-        graph
-            .add_entity(make_entity("a", "A", "x"))
-            .await
-            .unwrap();
+        graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
         let related = graph
             .get_related("a", 3, RelationDirection::Outgoing, None)
             .await
@@ -740,10 +738,7 @@ mod tests {
     #[tokio::test]
     async fn get_related_depth_0() {
         let graph = InMemoryGraph::new();
-        graph
-            .add_entity(make_entity("a", "A", "x"))
-            .await
-            .unwrap();
+        graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
         let related = graph
             .get_related("a", 0, RelationDirection::Outgoing, None)
             .await
@@ -853,10 +848,16 @@ mod tests {
     #[tokio::test]
     async fn get_related_star_topology() {
         let graph = InMemoryGraph::new();
-        graph.add_entity(make_entity("center", "C", "x")).await.unwrap();
+        graph
+            .add_entity(make_entity("center", "C", "x"))
+            .await
+            .unwrap();
         for i in 0..5 {
             let leaf = format!("leaf{i}");
-            graph.add_entity(make_entity(&leaf, &leaf, "x")).await.unwrap();
+            graph
+                .add_entity(make_entity(&leaf, &leaf, "x"))
+                .await
+                .unwrap();
             graph
                 .add_relation(GraphRelation::new("center", &leaf, "r", 0.8))
                 .await
@@ -1016,10 +1017,7 @@ mod tests {
     #[tokio::test]
     async fn set_and_find_tags() {
         let graph = InMemoryGraph::new();
-        graph
-            .add_entity(make_entity("a", "A", "x"))
-            .await
-            .unwrap();
+        graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
         let n = graph
             .set_entity_tags("a", vec!["rust".into(), "ai".into()])
             .await
@@ -1055,7 +1053,10 @@ mod tests {
         let graph = InMemoryGraph::with_constraints(constraints);
         graph.add_entity(make_entity("a", "A", "x")).await.unwrap();
         let n = graph
-            .set_entity_tags("a", vec!["t1".into(), "t2".into(), "t3".into(), "t4".into()])
+            .set_entity_tags(
+                "a",
+                vec!["t1".into(), "t2".into(), "t3".into(), "t4".into()],
+            )
             .await
             .unwrap();
         assert_eq!(n, 2, "should truncate to max_tags_per_entity");
@@ -1075,10 +1076,7 @@ mod tests {
             .set_entity_tags("b", vec!["rust".into(), "ai".into()])
             .await
             .unwrap();
-        graph
-            .set_entity_tags("c", vec!["ai".into()])
-            .await
-            .unwrap();
+        graph.set_entity_tags("c", vec!["ai".into()]).await.unwrap();
         assert_eq!(graph.entity_count_by_tag("rust").await.unwrap(), 2);
         assert_eq!(graph.entity_count_by_tag("ai").await.unwrap(), 2);
     }
@@ -1086,9 +1084,7 @@ mod tests {
     #[tokio::test]
     async fn set_entity_tags_not_found() {
         let graph = InMemoryGraph::new();
-        let result = graph
-            .set_entity_tags("ghost", vec!["x".into()])
-            .await;
+        let result = graph.set_entity_tags("ghost", vec!["x".into()]).await;
         assert!(matches!(result, Err(MemoryError::NotFound(_))));
     }
 

src/memory/retriever.rs

@@ -270,7 +270,12 @@ impl MemoryRetriever {
             }
             // BFS 找相关实体
             let related: Vec<ScoredEntity> = graph
-                .get_related(&start.id, depth, crate::memory::graph::RelationDirection::Both, None)
+                .get_related(
+                    &start.id,
+                    depth,
+                    crate::memory::graph::RelationDirection::Both,
+                    None,
+                )
                 .await?;
             for se in related {
                 if seen.insert(se.entity.id.clone()) {
@@ -361,7 +366,7 @@ fn default_stop_words() -> HashSet<String> {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::memory::graph::{GraphEntity, InMemoryGraph, GraphRelation};
+    use crate::memory::graph::{GraphEntity, GraphRelation, InMemoryGraph};
     use crate::memory::knowledge::KnowledgeStore;
     use crate::memory::{InMemoryStore, MemoryStore};
     use std::sync::Arc;
@@ -502,7 +507,12 @@ mod tests {
         graph.add_entity(langgraph).await.unwrap();
         graph.add_entity(python).await.unwrap();
         graph
-            .add_relation(GraphRelation::new("langchain", "langgraph", "includes", 0.8))
+            .add_relation(GraphRelation::new(
+                "langchain",
+                "langgraph",
+                "includes",
+                0.8,
+            ))
             .await
             .unwrap();
         graph
@@ -523,7 +533,12 @@ mod tests {
         let result = retriever.retrieve("langchain").await.unwrap();
         assert_eq!(result.strategy, RetrievalStrategy::GraphOnly);
         // 应该全部是 GraphEntity 变体
-        assert!(result.items.iter().all(|i| matches!(i, RetrievalItem::GraphEntity { .. })));
+        assert!(
+            result
+                .items
+                .iter()
+                .all(|i| matches!(i, RetrievalItem::GraphEntity { .. }))
+        );
         // langchain 是起始实体（score=1.0），langgraph 和 python 是 BFS 结果
         assert!(!result.items.is_empty(), "should find graph entities");
     }
@@ -544,8 +559,8 @@ mod tests {
         // KnowledgeGraph 也有匹配
         let graph = make_graph_with_data().await;
 
-        let retriever = MemoryRetriever::new(ks, RetrieverConfig::default())
-            .with_knowledge_graph(graph);
+        let retriever =
+            MemoryRetriever::new(ks, RetrieverConfig::default()).with_knowledge_graph(graph);
         let result = retriever.retrieve("langchain").await.unwrap();
         assert_eq!(result.strategy, RetrievalStrategy::Hybrid);
         // 应该同时包含 KnowledgePage 和 GraphEntity
@@ -569,8 +584,8 @@ mod tests {
             .unwrap();
         // 空图
         let graph = Arc::new(InMemoryGraph::new());
-        let retriever = MemoryRetriever::new(ks, RetrieverConfig::default())
-            .with_knowledge_graph(graph);
+        let retriever =
+            MemoryRetriever::new(ks, RetrieverConfig::default()).with_knowledge_graph(graph);
         let result = retriever.retrieve("langchain").await.unwrap();
         assert_eq!(result.strategy, RetrievalStrategy::Hybrid);
         // 图空，只有 Store 结果
@@ -587,15 +602,18 @@ mod tests {
         let ks = KnowledgeStore::new(store);
         // Store 空，图有数据
         let graph = make_graph_with_data().await;
-        let retriever = MemoryRetriever::new(ks, RetrieverConfig::default())
-            .with_knowledge_graph(graph);
+        let retriever =
+            MemoryRetriever::new(ks, RetrieverConfig::default()).with_knowledge_graph(graph);
         let result = retriever.retrieve("langchain").await.unwrap();
         assert_eq!(result.strategy, RetrievalStrategy::Hybrid);
         let has_entity = result
             .items
             .iter()
             .any(|i| matches!(i, RetrievalItem::GraphEntity { .. }));
-        assert!(has_entity, "should have graph results even when store is empty");
+        assert!(
+            has_entity,
+            "should have graph results even when store is empty"
+        );
     }
 
     #[tokio::test]
@@ -615,7 +633,10 @@ mod tests {
             .items
             .iter()
             .any(|i| matches!(i, RetrievalItem::GraphEntity { .. }));
-        assert!(!has_entity, "KnowledgeOnly should not return graph entities");
+        assert!(
+            !has_entity,
+            "KnowledgeOnly should not return graph entities"
+        );
     }
 
     // ── 辅助函数测试 ──

src/memory/store/sqlite_store.rs

@@ -6,9 +6,9 @@ use std::path::Path;
 use std::sync::{Arc, Mutex};
 
 use async_trait::async_trait;
-use rusqlite::{params, params_from_iter, Connection, ErrorCode};
-use time::format_description::well_known::Rfc3339;
+use rusqlite::{Connection, ErrorCode, params, params_from_iter};
 use time::OffsetDateTime;
+use time::format_description::well_known::Rfc3339;
 use tracing::{debug, error, instrument, warn};
 
 use crate::memory::error::MemoryError;
@@ -114,9 +114,7 @@ impl MemoryStore for SqliteStore {
                 .map_err(|e| map_sqlite_error(e, "query get"))?;
             match rows.next() {
                 None => Ok(None),
-                Some(row) => row
-                    .map(Some)
-                    .map_err(|e| map_sqlite_error(e, "decode row")),
+                Some(row) => row.map(Some).map_err(|e| map_sqlite_error(e, "decode row")),
             }
         })
         .await
@@ -130,11 +128,8 @@ impl MemoryStore for SqliteStore {
 
         tokio::task::spawn_blocking(move || -> Result<(), MemoryError> {
             let conn = conn.lock().unwrap_or_else(|e| e.into_inner());
-            conn.execute(
-                "DELETE FROM memory_items WHERE id = ?1",
-                params![id_owned],
-            )
-            .map_err(|e| map_sqlite_error(e, "delete"))?;
+            conn.execute("DELETE FROM memory_items WHERE id = ?1", params![id_owned])
+                .map_err(|e| map_sqlite_error(e, "delete"))?;
             Ok(())
         })
         .await
@@ -143,9 +138,8 @@ impl MemoryStore for SqliteStore {
 
     #[instrument(skip(self, filter))]
     async fn list(&self, filter: &MemoryFilter) -> Result<Vec<MemoryItem>, MemoryError> {
-        let mut sql = String::from(
-            "SELECT id, content, metadata, created_at FROM memory_items WHERE 1=1",
-        );
+        let mut sql =
+            String::from("SELECT id, content, metadata, created_at FROM memory_items WHERE 1=1");
         let mut param_values: Vec<String> = Vec::new();
         let mut ph_idx = 0usize;
 
@@ -309,9 +303,7 @@ fn map_sqlite_error(e: rusqlite::Error, ctx: &str) -> MemoryError {
         rusqlite::Error::InvalidQuery
         | rusqlite::Error::InvalidParameterName(_)
         | rusqlite::Error::InvalidColumnIndex(_)
-        | rusqlite::Error::InvalidColumnName(_) => {
-            MemoryError::InvalidInput(format!("{ctx}: {e}"))
-        }
+        | rusqlite::Error::InvalidColumnName(_) => MemoryError::InvalidInput(format!("{ctx}: {e}")),
         rusqlite::Error::FromSqlConversionFailure(_, _, _)
         | rusqlite::Error::ToSqlConversionFailure(_) => {
             MemoryError::Serialization(format!("{ctx}: {e}"))
@@ -527,8 +519,7 @@ mod tests {
         // ponytail: 回归验证 SqliteStore 可作为 Arc<dyn MemoryStore> 与 InMemoryStore 互换
         // 所有现有消费者（Conversation / Knowledge / Retriever / SessionMemory）均通过 trait object 引用，
         // 此测试确保 trait 接口契约在 SqliteStore 上同样成立。
-        let sqlite: Arc<dyn MemoryStore> =
-            Arc::new(SqliteStore::open(":memory:").unwrap());
+        let sqlite: Arc<dyn MemoryStore> = Arc::new(SqliteStore::open(":memory:").unwrap());
         let in_mem: Arc<dyn MemoryStore> = Arc::new(InMemoryStore::new());
 
         let stores: Vec<Arc<dyn MemoryStore>> = vec![Arc::clone(&sqlite), Arc::clone(&in_mem)];
@@ -556,12 +547,7 @@ mod tests {
             handles.push(tokio::spawn(async move {
                 let id = format!("concurrent_{i}");
                 // 设置每次 save 的 per-call timeout —— busy_timeout=5000ms 应足够
-                match tokio::time::timeout(
-                    Duration::from_secs(10),
-                    s.save(make_item(&id)),
-                )
-                .await
-                {
+                match tokio::time::timeout(Duration::from_secs(10), s.save(make_item(&id))).await {
                     Ok(res) => res.unwrap(),
                     Err(_) => panic!("save({id}) timed out under 100-way concurrency"),
                 }

src/memory/vector.rs

@@ -30,11 +30,7 @@ pub trait VectorRetriever: Send + Sync {
     ///
     /// 返回 `Vec<(id, score)>`，按 score 降序排列，score ∈ [0.0, 1.0]
     /// （余弦相似度）。当 `k == 0`、索引为空或 query 为零向量时返回空 Vec。
-    async fn search(
-        &self,
-        query: Vec<f32>,
-        k: usize,
-    ) -> Result<Vec<(String, f32)>, MemoryError>;
+    async fn search(&self, query: Vec<f32>, k: usize) -> Result<Vec<(String, f32)>, MemoryError>;
 }
 
 /// 进程内向量检索器 —— 基于 HashMap + 全量余弦相似度扫描。
@@ -79,11 +75,7 @@ impl VectorRetriever for InMemoryVectorRetriever {
         Ok(())
     }
 
-    async fn search(
-        &self,
-        query: Vec<f32>,
-        k: usize,
-    ) -> Result<Vec<(String, f32)>, MemoryError> {
+    async fn search(&self, query: Vec<f32>, k: usize) -> Result<Vec<(String, f32)>, MemoryError> {
         let vectors = self
             .vectors
             .lock()
@@ -241,4 +233,4 @@ mod tests {
             h.await.unwrap();
         }
     }
-}
+}

src/memory/vector_store.rs

@@ -11,8 +11,8 @@ use std::sync::{Arc, Mutex};
 
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
-use time::format_description::well_known::Rfc3339;
 use time::OffsetDateTime;
+use time::format_description::well_known::Rfc3339;
 use tracing::{debug, info};
 
 use crate::document::{Document, RecursiveCharacterSplitter};
@@ -43,11 +43,8 @@ pub trait VectorStore: Send + Sync {
     /// - 截取 `min(len)` 对处理（部分写入已发生）
     /// - 返回 `Err(MemoryError::InvalidInput)` 告知截断
     /// - 调用方可以 `let _ = store.add(...)` 忽略错误
-    async fn add(
-        &self,
-        documents: &[Document],
-        embeddings: &[Vec<f32>],
-    ) -> Result<(), MemoryError>;
+    async fn add(&self, documents: &[Document], embeddings: &[Vec<f32>])
+    -> Result<(), MemoryError>;
 
     /// 检索与 `query` 向量最相似的 `k` 条记录。
     ///
@@ -59,11 +56,7 @@ pub trait VectorStore: Send + Sync {
     /// - 空索引 → 返回 `vec![]`
     /// - `k == 0` → 返回 `vec![]`
     /// - 零向量（norm ≈ 0）→ 返回 `vec![]`
-    async fn search(
-        &self,
-        query: &[f32],
-        k: usize,
-    ) -> Result<Vec<(Document, f32)>, MemoryError>;
+    async fn search(&self, query: &[f32], k: usize) -> Result<Vec<(Document, f32)>, MemoryError>;
 
     /// 批量删除文档（幂等）。
     ///
@@ -105,9 +98,7 @@ impl InMemoryVectorStore {
     }
 
     /// 从预填充的 entries 构造（供 `PersistentVectorStore` 使用）。
-    pub(crate) fn with_entries(
-        entries: HashMap<String, (Document, Vec<f32>)>,
-    ) -> Self {
+    pub(crate) fn with_entries(entries: HashMap<String, (Document, Vec<f32>)>) -> Self {
         Self {
             entries: Mutex::new(entries),
         }
@@ -142,7 +133,10 @@ impl VectorStore for InMemoryVectorStore {
         }
 
         for i in 0..n {
-            entries.insert(documents[i].id.clone(), (documents[i].clone(), embeddings[i].clone()));
+            entries.insert(
+                documents[i].id.clone(),
+                (documents[i].clone(), embeddings[i].clone()),
+            );
         }
 
         if documents.len() != embeddings.len() {
@@ -156,11 +150,7 @@ impl VectorStore for InMemoryVectorStore {
         Ok(())
     }
 
-    async fn search(
-        &self,
-        query: &[f32],
-        k: usize,
-    ) -> Result<Vec<(Document, f32)>, MemoryError> {
+    async fn search(&self, query: &[f32], k: usize) -> Result<Vec<(Document, f32)>, MemoryError> {
         if k == 0 {
             return Ok(Vec::new());
         }
@@ -265,10 +255,7 @@ impl PersistentVectorStore {
     ///
     /// `MemoryStore::list()` 由 `SqliteStore` 内部使用 `spawn_blocking` 卸载，
     /// 加载过程本身在 async context 中即可，无需额外 spawn_blocking。
-    pub async fn new(
-        store: Arc<dyn MemoryStore>,
-        namespace: &str,
-    ) -> Result<Self, MemoryError> {
+    pub async fn new(store: Arc<dyn MemoryStore>, namespace: &str) -> Result<Self, MemoryError> {
         let prefix = format!("vec:{namespace}:");
         let filter = MemoryFilter {
             prefix: Some(prefix.clone()),
@@ -292,7 +279,10 @@ impl PersistentVectorStore {
             entries.insert(doc.id.clone(), (doc, entry.embedding));
         }
 
-        info!(entries = entries.len(), "PersistentVectorStore — 内存索引重建完成");
+        info!(
+            entries = entries.len(),
+            "PersistentVectorStore — 内存索引重建完成"
+        );
         Ok(Self {
             inner: InMemoryVectorStore::with_entries(entries),
             store,
@@ -340,11 +330,7 @@ impl VectorStore for PersistentVectorStore {
         self.inner.add(documents, embeddings).await
     }
 
-    async fn search(
-        &self,
-        query: &[f32],
-        k: usize,
-    ) -> Result<Vec<(Document, f32)>, MemoryError> {
+    async fn search(&self, query: &[f32], k: usize) -> Result<Vec<(Document, f32)>, MemoryError> {
         tracing::trace!(k, "PersistentVectorStore::search");
         self.inner.search(query, k).await
     }
@@ -575,10 +561,7 @@ mod tests {
     async fn remove_items() {
         let store = InMemoryVectorStore::new();
         let docs = vec![make_doc("a", "alpha"), make_doc("b", "beta")];
-        let embeddings = vec![
-            make_vec(&[1.0, 0.0, 0.0]),
-            make_vec(&[0.0, 1.0, 0.0]),
-        ];
+        let embeddings = vec![make_vec(&[1.0, 0.0, 0.0]), make_vec(&[0.0, 1.0, 0.0])];
         store.add(&docs, &embeddings).await.unwrap();
 
         store.remove(&["a".to_string()]).await.unwrap();
@@ -599,7 +582,9 @@ mod tests {
         let embeddings = vec![make_vec(&[1.0, 0.0, 0.0])];
         store.add(&docs, &embeddings).await.unwrap();
 
-        let result = store.remove(&["nonexistent".to_string(), "also_nonexistent".to_string()]).await;
+        let result = store
+            .remove(&["nonexistent".to_string(), "also_nonexistent".to_string()])
+            .await;
         assert!(result.is_ok(), "批量删除不存在 id 不应报错");
 
         let results = store.search(&[1.0, 0.0, 0.0], 5).await.unwrap();
@@ -657,7 +642,9 @@ mod tests {
         backend: Arc<dyn MemoryStore>,
         namespace: &str,
     ) -> PersistentVectorStore {
-        PersistentVectorStore::new(backend, namespace).await.unwrap()
+        PersistentVectorStore::new(backend, namespace)
+            .await
+            .unwrap()
     }
 
     #[tokio::test]
@@ -755,10 +742,7 @@ mod tests {
         let store = make_persistent(Arc::clone(&backend), "default").await;
 
         // 正常写入前 2 条
-        let docs_first = vec![
-            make_doc("doc_0", "first"),
-            make_doc("doc_1", "second"),
-        ];
+        let docs_first = vec![make_doc("doc_0", "first"), make_doc("doc_1", "second")];
         let embeddings_first = vec![make_vec(&[1.0, 0.0, 0.0]), make_vec(&[0.0, 1.0, 0.0])];
         store.add(&docs_first, &embeddings_first).await.unwrap();
 
@@ -797,11 +781,7 @@ mod tests {
         let store: Arc<dyn VectorStore> = Arc::new(InMemoryVectorStore::new());
         let splitter = RecursiveCharacterSplitter::new(50, 5);
 
-        let pipeline = RagPipeline::new(
-            Arc::clone(&embedder),
-            Arc::clone(&store),
-            Some(splitter),
-        );
+        let pipeline = RagPipeline::new(Arc::clone(&embedder), Arc::clone(&store), Some(splitter));
 
         // 创建多段落文档
         let doc = Document::new(
@@ -934,4 +914,4 @@ mod tests {
             elapsed.as_millis()
         );
     }
-}
+}