chore(deps): upgrade sqlx 0.7→0.8 + libsqlite3-sys 0.27→0.30

Unifies dual sqlx versions caused by pgvector 0.4 pulling sqlx 0.8.x as indirect dependency. Zero source code changes required, 719/719 tests pass. Batch 3 of audit plan (plans/stateless-petting-rossum.md).
fix(audit): Batch 7-9 dead_code 标注 + TODO 清理 + 文档同步
2026-04-19 11:15:05 +08:00 · 2026-04-19 08:54:57 +08:00 · 2026-04-19 08:46:12 +08:00 · 2026-04-19 08:38:09 +08:00 · 2026-04-19 08:30:33 +08:00 · 2026-04-19 00:43:02 +08:00
90 changed files with 3958 additions and 1093 deletions
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -227,21 +227,22 @@ Client  → 负责网络通信和协议转换

 ## 6. 自主能力系统 (Hands)

-ZCLAW 提供 11 个自主能力包（9 启用 + 2 禁用）：
+ZCLAW 提供 12 个自主能力包（7 已注册 + 3 开发中 + 2 禁用）：

 | Hand | 功能 | 状态 |
 |------|------|------|
 | Browser | 浏览器自动化 | ✅ 可用 |
 | Collector | 数据收集聚合 | ✅ 可用 |
 | Researcher | 深度研究 | ✅ 可用 |
-| Predictor | 预测分析 | ❌ 已禁用 (enabled=false)，无 Rust 实现 |
-| Lead | 销售线索发现 | ❌ 已禁用 (enabled=false)，无 Rust 实现 |
 | Clip | 视频处理 | ⚠️ 需 FFmpeg |
 | Twitter | Twitter 自动化 | ✅ 可用（12 个 API v2 真实调用，写操作需 OAuth 1.0a） |
-| Whiteboard | 白板演示 | ✅ 可用（导出功能开发中，标注 demo） |
-| Slideshow | 幻灯片生成 | ✅ 可用 |
-| Speech | 语音合成 | ✅ 可用（Browser TTS 前端集成完成） |
 | Quiz | 测验生成 | ✅ 可用 |
+| _reminder | 系统内部提醒 | ✅ 可用（kernel 编程注册，无 HAND.toml） |
+| Whiteboard | 白板演示 | 🚧 开发中（HAND.toml 未合并到主分支） |
+| Slideshow | 幻灯片生成 | 🚧 开发中（HAND.toml 未合并到主分支） |
+| Speech | 语音合成 | 🚧 开发中（HAND.toml 未合并到主分支） |
+| Predictor | 预测分析 | ❌ 已禁用 (enabled=false)，无 Rust 实现 |
+| Lead | 销售线索发现 | ❌ 已禁用 (enabled=false)，无 Rust 实现 |

 **触发 Hand 时：**
 1. 检查依赖是否满足
@@ -541,10 +542,10 @@ refactor(store): 统一 Store 数据获取方式
 | Hermes 管线 | ✅ 活跃 | 04-12 触发信号持久化 + 经验行业维度 + 注入格式优化 |
 | Intelligence Heartbeat | ✅ 活跃 | 04-15 统一健康快照 (health_snapshot.rs) + HeartbeatManager 重构 + HealthPanel 前端 |
 | 聊天流 (ChatStream) | ✅ 稳定 | 04-02 ChatStore 拆分为 4 Store (stream/conversation/message/chat) |
-| 记忆管道 (Memory) | ✅ 稳定 | 04-02 闭环修复: 对话→提取→FTS5+TF-IDF→检索→注入 |
+| 记忆管道 (Memory) | ✅ 稳定 | 04-17 E2E 验证: 存储+FTS5+TF-IDF+注入闭环，去重+跨会话注入已修复 |
 | SaaS 认证 (Auth) | ✅ 稳定 | Token池 RPM/TPM 轮换 + JWT password_version 失效机制 |
 | Pipeline DSL | ✅ 稳定 | 04-01 17 个 YAML 模板 + DAG 执行器 |
-| Hands 系统 | ✅ 稳定 | 9 启用 (Browser/Collector/Researcher/Twitter/Whiteboard/Slideshow/Speech/Quiz/Clip) |
+| Hands 系统 | ✅ 稳定 | 7 注册 (6 HAND.toml + _reminder)，Whiteboard/Slideshow/Speech 开发中 |
 | 技能系统 (Skills) | ✅ 稳定 | 75 个 SKILL.md + 语义路由 |
 | 中间件链 | ✅ 稳定 | 14 层 (ButlerRouter@80, DataMasking@90, Compaction@100, Memory@150, Title@180, SkillIndex@200, DanglingTool@300, ToolError@350, ToolOutputGuard@360, Guardrail@400, LoopGuard@500, SubagentLimit@550, TrajectoryRecorder@650, TokenCalibration@700) |

@@ -555,17 +556,17 @@ refactor(store): 统一 Store 数据获取方式
 - **聊天流**: 3种实现 → GatewayClient(WebSocket) / KernelClient(Tauri Event) / SaaSRelay(SSE) + 5min超时守护。详见 [ARCHITECTURE_BRIEF.md](docs/ARCHITECTURE_BRIEF.md)
 - **客户端路由**: `getClient()` 4分支决策树 → Admin路由 / SaaS Relay(可降级到本地) / Local Kernel / External Gateway
 - **SaaS 认证**: JWT→OS keyring 存储 + HttpOnly cookie + Token池 RPM/TPM 限流轮换 + SaaS unreachable 自动降级
- **记忆闭环**: 对话→extraction_adapter→FTS5全文+TF-IDF权重→检索→注入系统提示
+- **记忆闭环**: 对话→extraction_adapter→FTS5全文+TF-IDF权重→检索→注入系统提示（E2E 04-17 验证通过，去重+跨会话注入已修复）
 - **LLM 驱动**: 4 Rust Driver (Anthropic/OpenAI/Gemini/Local) + 国内兼容 (DeepSeek/Qwen/Moonshot 通过 base_url)

 ### 最近变更

-1. [04-15] Heartbeat 统一健康系统: health_snapshot.rs 统一收集器(LLM连接/记忆/会话/系统资源) + heartbeat.rs HeartbeatManager 重构 + HealthPanel.tsx 前端面板 + Tauri 命令 182→183 + intelligence 模块 15→16 文件 + 删除 intelligence-client/ 9 废弃文件
-2. [04-12] 行业配置+管家主动性 全栈 5 Phase: 行业数据模型+4内置配置+ButlerRouter动态关键词+触发信号+Tauri加载+Admin管理页面+跨会话连续性+XML fencing注入格式
-2. [04-09] Hermes Intelligence Pipeline 4 Chunk: ExperienceStore+Extractor, UserProfileStore+Profiler, NlScheduleParser, TrajectoryRecorder+Compressor (684 tests, 0 failed)
-3. [04-09] 管家模式6交付物完成: ButlerRouter + 冷启动 + 简洁模式UI + 桥测试 + 发布文档
-3. [04-07] @reserved 标注 5 个 butler Tauri 命令 + 痛点持久化 SQLite
-4. [04-06] 4 个发布前 bug 修复 (身份覆盖/模型配置/agent同步/自动身份)
+1. [04-17] 全系统 E2E 测试 129 链路: 82 PASS / 20 PARTIAL / 1 FAIL / 26 SKIP，有效通过率 79.1%。7 项 Bug 修复 (Dashboard 404/记忆去重/记忆注入/invoice_id/Prompt版本/agent隔离/行业字段)
+2. [04-16] 3 项 P0 修复 + 5 项 E2E Bug 修复 + Agent 面板刷新 + TRUTH.md 数字校准
+3. [04-15] Heartbeat 统一健康系统: health_snapshot.rs 统一收集器(LLM连接/记忆/会话/系统资源) + heartbeat.rs HeartbeatManager 重构 + HealthPanel.tsx 前端面板 + Tauri 命令 182→183 + intelligence 模块 15→16 文件 + 删除 intelligence-client/ 9 废弃文件
+4. [04-12] 行业配置+管家主动性 全栈 5 Phase: 行业数据模型+4内置配置+ButlerRouter动态关键词+触发信号+Tauri加载+Admin管理页面+跨会话连续性+XML fencing注入格式
+5. [04-09] Hermes Intelligence Pipeline 4 Chunk: ExperienceStore+Extractor, UserProfileStore+Profiler, NlScheduleParser, TrajectoryRecorder+Compressor (684 tests, 0 failed)
+6. [04-09] 管家模式6交付物完成: ButlerRouter + 冷启动 + 简洁模式UI + 桥测试 + 发布文档

 <!-- ARCH-SNAPSHOT-END -->

--- a/Cargo.lock
+++ b/Cargo.lock
@@ -61,19 +61,6 @@ dependencies = [
 "subtle",
 ]

-[[package]]
-name = "ahash"
-version = "0.8.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a15f179cd60c4584b8a8c596927aadc462e27f2ca70c04e0071964a73ba7a75"
-dependencies = [
- "cfg-if",
- "getrandom 0.3.4",
- "once_cell",
- "version_check",
- "zerocopy",
-]
-
 [[package]]
 name = "aho-corasick"
 version = "1.1.4"
@@ -179,7 +166,7 @@ version = "0.7.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "435a87a52755b8f27fcf321ac4f04b2802e337c8c4872923137471ec39c37532"
 dependencies = [
- "event-listener 5.4.1",
+ "event-listener",
 "event-listener-strategy",
 "futures-core",
 "pin-project-lite",
@@ -235,7 +222,7 @@ version = "3.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "290f7f2596bd5b78a9fec8088ccd89180d7f9f55b94b0576823bbbdc72ee8311"
 dependencies = [
- "event-listener 5.4.1",
+ "event-listener",
 "event-listener-strategy",
 "pin-project-lite",
 ]
@@ -253,7 +240,7 @@ dependencies = [
 "async-task",
 "blocking",
 "cfg-if",
- "event-listener 5.4.1",
+ "event-listener",
 "futures-lite",
 "rustix 1.1.4",
 ]
@@ -1606,9 +1593,9 @@ dependencies = [
 "secrecy",
 "serde",
 "serde_json",
- "serde_yaml",
+ "serde_yaml_bw",
 "sha2",
- "sqlx 0.7.4",
+ "sqlx",
 "tauri",
 "tauri-build",
 "tauri-plugin-mcp",
@@ -1974,12 +1961,6 @@ dependencies = [
 "num-traits",
 ]

-[[package]]
-name = "event-listener"
-version = "2.5.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0206175f82b8d6bf6652ff7d71a1e27fd2e4efde587fd368662814d6ec1d9ce0"
-
 [[package]]
 name = "event-listener"
 version = "5.4.1"
@@ -1997,7 +1978,7 @@ version = "0.5.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8be9f3dfaaffdae2972880079a491a1a8bb7cbed0b8dd7a347f668b4150a3b93"
 dependencies = [
- "event-listener 5.4.1",
+ "event-listener",
 "pin-project-lite",
 ]

@@ -2699,10 +2680,6 @@ name = "hashbrown"
 version = "0.14.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e5274423e17b7c9fc20b6e7e208532f9b19825d82dfd615708b70edd83df41f1"
-dependencies = [
- "ahash",
- "allocator-api2",
-]

 [[package]]
 name = "hashbrown"
@@ -2726,15 +2703,6 @@ dependencies = [
 "serde_core",
 ]

-[[package]]
-name = "hashlink"
-version = "0.8.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8094feaf31ff591f651a2664fb9cfd92bba7a60ce3197265e9482ebe753c8f7"
-dependencies = [
- "hashbrown 0.14.5",
-]
-
 [[package]]
 name = "hashlink"
 version = "0.10.0"
@@ -2773,9 +2741,6 @@ name = "heck"
 version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
-dependencies = [
- "unicode-segmentation",
-]

 [[package]]
 name = "heck"
@@ -3555,9 +3520,9 @@ dependencies = [

 [[package]]
 name = "libsqlite3-sys"
-version = "0.27.0"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cf4e226dcd58b4be396f7bd3c20da8fdee2911400705297ba7d2d7cc2c30f716"
+checksum = "2e99fb7a497b1e3339bc746195567ed8d3e24945ecd636e3619d20b9de9e9149"
 dependencies = [
 "cc",
 "pkg-config",
@@ -4358,12 +4323,6 @@ dependencies = [
 "subtle",
 ]

-[[package]]
-name = "paste"
-version = "1.0.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "57c0d7b74b563b49d38dae00a0c37d4d6de9b432382b2892f0574ddcae73fd0a"
-
 [[package]]
 name = "pathdiff"
 version = "0.2.3"
@@ -4426,7 +4385,7 @@ version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "fc58e2d255979a31caa7cabfa7aac654af0354220719ab7a68520ae7a91e8c0b"
 dependencies = [
- "sqlx 0.8.6",
+ "sqlx",
 ]

 [[package]]
@@ -5492,6 +5451,7 @@ version = "0.23.37"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "758025cb5fccfd3bc2fd74708fd4682be41d99e5dff73c377c0646c6012c73a4"
 dependencies = [
+ "log",
 "once_cell",
 "ring",
 "rustls-pki-types",
@@ -5912,19 +5872,6 @@ dependencies = [
 "syn 2.0.117",
 ]

-[[package]]
-name = "serde_yaml"
-version = "0.9.34+deprecated"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a8b1a1a2ebf674015cc02edccce75287f1a0130d394307b36743c2f5d504b47"
-dependencies = [
- "indexmap 2.13.0",
- "itoa 1.0.18",
- "ryu",
- "serde",
- "unsafe-libyaml",
-]
-
 [[package]]
 name = "serde_yaml_bw"
 version = "2.5.3"
@@ -6187,78 +6134,17 @@ dependencies = [
 "der",
 ]

-[[package]]
-name = "sqlformat"
-version = "0.2.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7bba3a93db0cc4f7bdece8bb09e77e2e785c20bfebf79eb8340ed80708048790"
-dependencies = [
- "nom",
- "unicode_categories",
-]
-
-[[package]]
-name = "sqlx"
-version = "0.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c9a2ccff1a000a5a59cd33da541d9f2fdcd9e6e8229cc200565942bff36d0aaa"
-dependencies = [
- "sqlx-core 0.7.4",
- "sqlx-macros 0.7.4",
- "sqlx-mysql",
- "sqlx-postgres 0.7.4",
- "sqlx-sqlite",
-]
-
 [[package]]
 name = "sqlx"
 version = "0.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1fefb893899429669dcdd979aff487bd78f4064e5e7907e4269081e0ef7d97dc"
 dependencies = [
- "sqlx-core 0.8.6",
- "sqlx-macros 0.8.6",
- "sqlx-postgres 0.8.6",
-]
-
-[[package]]
-name = "sqlx-core"
-version = "0.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "24ba59a9342a3d9bab6c56c118be528b27c9b60e490080e9711a04dccac83ef6"
-dependencies = [
- "ahash",
- "atoi",
- "byteorder",
- "bytes",
- "chrono",
- "crc",
- "crossbeam-queue",
- "either",
- "event-listener 2.5.3",
- "futures-channel",
- "futures-core",
- "futures-intrusive",
- "futures-io",
- "futures-util",
- "hashlink 0.8.4",
- "hex",
- "indexmap 2.13.0",
- "log",
- "memchr",
- "once_cell",
- "paste",
- "percent-encoding",
- "serde",
- "serde_json",
- "sha2",
- "smallvec",
- "sqlformat",
- "thiserror 1.0.69",
- "tokio",
- "tokio-stream",
- "tracing",
- "url",
+ "sqlx-core",
+ "sqlx-macros",
+ "sqlx-mysql",
+ "sqlx-postgres",
+ "sqlx-sqlite",
 ]

 [[package]]
@@ -6269,16 +6155,17 @@ checksum = "ee6798b1838b6a0f69c007c133b8df5866302197e404e8b6ee8ed3e3a5e68dc6"
 dependencies = [
 "base64 0.22.1",
 "bytes",
+ "chrono",
 "crc",
 "crossbeam-queue",
 "either",
- "event-listener 5.4.1",
+ "event-listener",
 "futures-core",
 "futures-intrusive",
 "futures-io",
 "futures-util",
 "hashbrown 0.15.5",
- "hashlink 0.10.0",
+ "hashlink",
 "indexmap 2.13.0",
 "log",
 "memchr",
@@ -6289,23 +6176,12 @@ dependencies = [
 "sha2",
 "smallvec",
 "thiserror 2.0.18",
+ "tokio",
+ "tokio-stream",
 "tracing",
 "url",
 ]

-[[package]]
-name = "sqlx-macros"
-version = "0.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4ea40e2345eb2faa9e1e5e326db8c34711317d2b5e08d0d5741619048a803127"
-dependencies = [
- "proc-macro2",
- "quote",
- "sqlx-core 0.7.4",
- "sqlx-macros-core 0.7.4",
- "syn 1.0.109",
-]
-
 [[package]]
 name = "sqlx-macros"
 version = "0.8.6"
@@ -6314,37 +6190,11 @@ checksum = "a2d452988ccaacfbf5e0bdbc348fb91d7c8af5bee192173ac3636b5fb6e6715d"
 dependencies = [
 "proc-macro2",
 "quote",
- "sqlx-core 0.8.6",
- "sqlx-macros-core 0.8.6",
+ "sqlx-core",
+ "sqlx-macros-core",
 "syn 2.0.117",
 ]

-[[package]]
-name = "sqlx-macros-core"
-version = "0.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5833ef53aaa16d860e92123292f1f6a3d53c34ba8b1969f152ef1a7bb803f3c8"
-dependencies = [
- "dotenvy",
- "either",
- "heck 0.4.1",
- "hex",
- "once_cell",
- "proc-macro2",
- "quote",
- "serde",
- "serde_json",
- "sha2",
- "sqlx-core 0.7.4",
- "sqlx-mysql",
- "sqlx-postgres 0.7.4",
- "sqlx-sqlite",
- "syn 1.0.109",
- "tempfile",
- "tokio",
- "url",
-]
-
 [[package]]
 name = "sqlx-macros-core"
 version = "0.8.6"
@@ -6361,20 +6211,23 @@ dependencies = [
 "serde",
 "serde_json",
 "sha2",
- "sqlx-core 0.8.6",
- "sqlx-postgres 0.8.6",
+ "sqlx-core",
+ "sqlx-mysql",
+ "sqlx-postgres",
+ "sqlx-sqlite",
 "syn 2.0.117",
+ "tokio",
 "url",
 ]

 [[package]]
 name = "sqlx-mysql"
-version = "0.7.4"
+version = "0.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1ed31390216d20e538e447a7a9b959e06ed9fc51c37b514b46eb758016ecd418"
+checksum = "aa003f0038df784eb8fecbbac13affe3da23b45194bd57dba231c8f48199c526"
 dependencies = [
 "atoi",
- "base64 0.21.7",
+ "base64 0.22.1",
 "bitflags 2.11.0",
 "byteorder",
 "bytes",
@@ -6403,48 +6256,9 @@ dependencies = [
 "sha1 0.10.6",
 "sha2",
 "smallvec",
- "sqlx-core 0.7.4",
+ "sqlx-core",
 "stringprep",
- "thiserror 1.0.69",
- "tracing",
- "whoami",
-]
-
-[[package]]
-name = "sqlx-postgres"
-version = "0.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c824eb80b894f926f89a0b9da0c7f435d27cdd35b8c655b114e58223918577e"
-dependencies = [
- "atoi",
- "base64 0.21.7",
- "bitflags 2.11.0",
- "byteorder",
- "chrono",
- "crc",
- "dotenvy",
- "etcetera",
- "futures-channel",
- "futures-core",
- "futures-io",
- "futures-util",
- "hex",
- "hkdf",
- "hmac",
- "home",
- "itoa 1.0.18",
- "log",
- "md-5",
- "memchr",
- "once_cell",
- "rand 0.8.5",
- "serde",
- "serde_json",
- "sha2",
- "smallvec",
- "sqlx-core 0.7.4",
- "stringprep",
- "thiserror 1.0.69",
+ "thiserror 2.0.18",
 "tracing",
 "whoami",
 ]
@@ -6459,6 +6273,7 @@ dependencies = [
 "base64 0.22.1",
 "bitflags 2.11.0",
 "byteorder",
+ "chrono",
 "crc",
 "dotenvy",
 "etcetera",
@@ -6479,7 +6294,7 @@ dependencies = [
 "serde_json",
 "sha2",
 "smallvec",
- "sqlx-core 0.8.6",
+ "sqlx-core",
 "stringprep",
 "thiserror 2.0.18",
 "tracing",
@@ -6488,9 +6303,9 @@ dependencies = [

 [[package]]
 name = "sqlx-sqlite"
-version = "0.7.4"
+version = "0.8.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b244ef0a8414da0bed4bb1910426e890b19e5e9bccc27ada6b797d05c55ae0aa"
+checksum = "c2d12fe70b2c1b4401038055f90f151b78208de1f9f89a7dbfd41587a10c3eea"
 dependencies = [
 "atoi",
 "chrono",
@@ -6504,10 +6319,11 @@ dependencies = [
 "log",
 "percent-encoding",
 "serde",
- "sqlx-core 0.7.4",
+ "serde_urlencoded",
+ "sqlx-core",
+ "thiserror 2.0.18",
 "tracing",
 "url",
- "urlencoding",
 ]

 [[package]]
@@ -7824,12 +7640,6 @@ version = "0.2.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"

-[[package]]
-name = "unicode_categories"
-version = "0.1.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "39ec24b3121d976906ece63c9daad25b85969647682eee313cb5779fdd69e14e"
-
 [[package]]
 name = "universal-hash"
 version = "0.5.1"
@@ -7840,12 +7650,6 @@ dependencies = [
 "subtle",
 ]

-[[package]]
-name = "unsafe-libyaml"
-version = "0.2.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "673aac59facbab8a9007c7f6108d11f63b603f7cabff99fabf650fea5c32b861"
-
 [[package]]
 name = "unsafe-libyaml-norway"
 version = "0.2.15"
@@ -7858,6 +7662,35 @@ version = "0.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"

+[[package]]
+name = "ureq"
+version = "3.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dea7109cdcd5864d4eeb1b58a1648dc9bf520360d7af16ec26d0a9354bafcfc0"
+dependencies = [
+ "base64 0.22.1",
+ "flate2",
+ "log",
+ "percent-encoding",
+ "rustls",
+ "rustls-pki-types",
+ "ureq-proto",
+ "utf8-zero",
+ "webpki-roots",
+]
+
+[[package]]
+name = "ureq-proto"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e994ba84b0bd1b1b0cf92878b7ef898a5c1760108fe7b6010327e274917a808c"
+dependencies = [
+ "base64 0.22.1",
+ "http 1.4.0",
+ "httparse",
+ "log",
+]
+
 [[package]]
 name = "url"
 version = "2.5.8"
@@ -7895,6 +7728,12 @@ version = "0.7.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"

+[[package]]
+name = "utf8-zero"
+version = "0.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8c0a043c9540bae7c578c88f91dda8bd82e59ae27c21baca69c8b191aaf5a6e"
+
 [[package]]
 name = "utf8_iter"
 version = "1.0.4"
@@ -9573,7 +9412,7 @@ dependencies = [
 "async-trait",
 "blocking",
 "enumflags2",
- "event-listener 5.4.1",
+ "event-listener",
 "futures-core",
 "futures-lite",
 "hex",
@@ -9630,7 +9469,7 @@ dependencies = [
 "libsqlite3-sys",
 "serde",
 "serde_json",
- "sqlx 0.7.4",
+ "sqlx",
 "thiserror 2.0.18",
 "tokio",
 "tokio-test",
@@ -9696,7 +9535,7 @@ dependencies = [
 "libsqlite3-sys",
 "serde",
 "serde_json",
- "sqlx 0.7.4",
+ "sqlx",
 "thiserror 2.0.18",
 "tokio",
 "tracing",
@@ -9723,7 +9562,6 @@ dependencies = [
 "tracing",
 "uuid",
 "zclaw-hands",
- "zclaw-kernel",
 "zclaw-runtime",
 "zclaw-skills",
 "zclaw-types",
@@ -9809,7 +9647,7 @@ dependencies = [
 "serde_json",
 "sha2",
 "socket2 0.5.10",
- "sqlx 0.7.4",
+ "sqlx",
 "tempfile",
 "thiserror 2.0.18",
 "tokio",
@@ -9840,6 +9678,8 @@ dependencies = [
 "thiserror 2.0.18",
 "tokio",
 "tracing",
+ "ureq",
+ "url",
 "uuid",
 "wasmtime",
 "wasmtime-wasi",
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -57,12 +57,15 @@ chrono = { version = "0.4", features = ["serde"] }
 uuid = { version = "1", features = ["v4", "v5", "serde"] }

 # Database
-sqlx = { version = "0.7", features = ["runtime-tokio", "sqlite", "postgres", "chrono"] }
-libsqlite3-sys = { version = "0.27", features = ["bundled"] }
+sqlx = { version = "0.8", features = ["runtime-tokio", "sqlite", "postgres", "chrono"] }
+libsqlite3-sys = { version = "0.30", features = ["bundled"] }

 # HTTP client (for LLM drivers)
 reqwest = { version = "0.12", default-features = false, features = ["json", "stream", "rustls-tls"] }

+# Synchronous HTTP (for WASM host functions in blocking threads)
+ureq = { version = "3", features = ["rustls"] }
+
 # URL parsing
 url = "2"

--- a/admin-v2/src/layouts/AdminLayout.tsx
+++ b/admin-v2/src/layouts/AdminLayout.tsx
@@ -117,7 +117,7 @@ function Sidebar({
              const isActive =
                item.path === '/'
                  ? activePath === '/'
-                  : activePath.startsWith(item.path)
+                  : activePath === item.path || activePath.startsWith(item.path + '/')

              const btn = (
                <button
--- a/admin-v2/src/services/stats.ts
+++ b/admin-v2/src/services/stats.ts
@@ -3,5 +3,5 @@ import type { DashboardStats } from '@/types'

 export const statsService = {
  dashboard: (signal?: AbortSignal) =>
-    request.get<DashboardStats>('/stats/dashboard', withSignal({}, signal)).then((r) => r.data),
+    request.get<DashboardStats>('/admin/dashboard', withSignal({}, signal)).then((r) => r.data),
 }
--- a/crates/zclaw-growth/src/evolution_engine.rs
+++ b/crates/zclaw-growth/src/evolution_engine.rs
@@ -0,0 +1,305 @@
+//! 进化引擎中枢
+//! 协调 L1/L2/L3 三层进化的触发和执行
+//! L1 (记忆进化) 在 GrowthIntegration 中处理
+//! L2 (技能进化) 通过 PatternAggregator + SkillGenerator + QualityGate 协调
+//! L3 (工作流进化) 通过 WorkflowComposer 协调
+//! 反馈闭环通过 FeedbackCollector 管理
+
+use std::sync::Arc;
+
+use crate::experience_store::ExperienceStore;
+use crate::feedback_collector::{
+    FeedbackCollector, FeedbackEntry, TrustUpdate,
+};
+use crate::pattern_aggregator::{AggregatedPattern, PatternAggregator};
+use crate::quality_gate::{QualityGate, QualityReport};
+use crate::skill_generator::{SkillCandidate, SkillGenerator};
+use crate::workflow_composer::{ToolChainPattern, WorkflowComposer};
+use crate::VikingAdapter;
+use zclaw_types::Result;
+
+/// 进化引擎配置
+#[derive(Debug, Clone)]
+pub struct EvolutionConfig {
+    /// 经验复用次数达到此阈值触发 L2
+    pub min_reuse_for_skill: u32,
+    /// 置信度阈值
+    pub quality_confidence_threshold: f32,
+    /// 是否启用进化引擎
+    pub enabled: bool,
+}
+
+impl Default for EvolutionConfig {
+    fn default() -> Self {
+        Self {
+            min_reuse_for_skill: 3,
+            quality_confidence_threshold: 0.7,
+            enabled: true,
+        }
+    }
+}
+
+/// 进化引擎中枢
+pub struct EvolutionEngine {
+    viking: Arc<VikingAdapter>,
+    feedback: Arc<tokio::sync::Mutex<FeedbackCollector>>,
+    config: EvolutionConfig,
+}
+
+impl EvolutionEngine {
+    pub fn new(viking: Arc<VikingAdapter>) -> Self {
+        Self {
+            viking: viking.clone(),
+            feedback: Arc::new(tokio::sync::Mutex::new(
+                FeedbackCollector::with_viking(viking),
+            )),
+            config: EvolutionConfig::default(),
+        }
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// Backward-compatible constructor
+    /// 从 ExperienceStore 中提取共享的 VikingAdapter 实例
+    pub fn from_experience_store(experience_store: Arc<ExperienceStore>) -> Self {
+        let viking = experience_store.viking().clone();
+        Self {
+            viking: viking.clone(),
+            feedback: Arc::new(tokio::sync::Mutex::new(
+                FeedbackCollector::with_viking(viking),
+            )),
+            config: EvolutionConfig::default(),
+        }
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    pub fn with_config(mut self, config: EvolutionConfig) -> Self {
+        self.config = config;
+        self
+    }
+
+    pub fn set_enabled(&mut self, enabled: bool) {
+        self.config.enabled = enabled;
+    }
+
+    /// L2 检查：是否有可进化的模式
+    pub async fn check_evolvable_patterns(
+        &self,
+        agent_id: &str,
+    ) -> Result<Vec<AggregatedPattern>> {
+        if !self.config.enabled {
+            return Ok(Vec::new());
+        }
+        let store = ExperienceStore::new(self.viking.clone());
+        let aggregator = PatternAggregator::new(store);
+        aggregator
+            .find_evolvable_patterns(agent_id, self.config.min_reuse_for_skill)
+            .await
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// L2 执行：为给定模式构建技能生成 prompt
+    /// 返回 (prompt_string, pattern) 供上层通过 LLM 调用后 parse
+    pub fn build_skill_prompt(&self, pattern: &AggregatedPattern) -> String {
+        SkillGenerator::build_prompt(pattern)
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// L2 执行：解析 LLM 返回的技能 JSON 并进行质量门控
+    pub fn validate_skill_candidate(
+        &self,
+        json_str: &str,
+        pattern: &AggregatedPattern,
+        existing_triggers: Vec<String>,
+    ) -> Result<(SkillCandidate, QualityReport)> {
+        let candidate = SkillGenerator::parse_response(json_str, pattern)?;
+        let gate = QualityGate::new(self.config.quality_confidence_threshold, existing_triggers);
+        let report = gate.validate_skill(&candidate);
+        Ok((candidate, report))
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// 获取当前配置
+    pub fn config(&self) -> &EvolutionConfig {
+        &self.config
+    }
+
+    // -----------------------------------------------------------------------
+    // L3: 工作流进化
+    // -----------------------------------------------------------------------
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// L3: 从轨迹数据中提取重复的工具链模式
+    pub fn analyze_trajectory_patterns(
+        &self,
+        trajectories: &[(String, Vec<String>)], // (session_id, tools_used)
+    ) -> Vec<(ToolChainPattern, Vec<String>)> {
+        if !self.config.enabled {
+            return Vec::new();
+        }
+        WorkflowComposer::extract_patterns(trajectories)
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// L3: 为给定工具链模式构建工作流生成 prompt
+    pub fn build_workflow_prompt(
+        &self,
+        pattern: &ToolChainPattern,
+        frequency: usize,
+        industry: Option<&str>,
+    ) -> String {
+        WorkflowComposer::build_prompt(pattern, frequency, industry)
+    }
+
+    // -----------------------------------------------------------------------
+    // 反馈闭环
+    // -----------------------------------------------------------------------
+
+    /// 提交反馈并获取信任度更新，自动持久化
+    pub async fn submit_feedback(&self, entry: FeedbackEntry) -> TrustUpdate {
+        let mut feedback = self.feedback.lock().await;
+        let update = feedback.submit_feedback(entry);
+        // 非阻塞持久化：失败仅打日志，不影响返回值
+        if let Err(e) = feedback.save().await {
+            tracing::warn!("[EvolutionEngine] Failed to persist trust records: {}", e);
+        }
+        update
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// 获取需要优化的进化产物
+    pub async fn get_artifacts_needing_optimization(&self) -> Vec<String> {
+        self.feedback
+            .lock()
+            .await
+            .get_artifacts_needing_optimization()
+            .iter()
+            .map(|r| r.artifact_id.clone())
+            .collect()
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// 获取建议归档的进化产物
+    pub async fn get_artifacts_to_archive(&self) -> Vec<String> {
+        self.feedback
+            .lock()
+            .await
+            .get_artifacts_to_archive()
+            .iter()
+            .map(|r| r.artifact_id.clone())
+            .collect()
+    }
+
+    /// @reserved: EvolutionEngine L2/L3 feature, post-release integration
+    /// 获取推荐产物
+    pub async fn get_recommended_artifacts(&self) -> Vec<String> {
+        self.feedback
+            .lock()
+            .await
+            .get_recommended_artifacts()
+            .iter()
+            .map(|r| r.artifact_id.clone())
+            .collect()
+    }
+
+    /// 启动时加载已持久化的信任度记录
+    pub async fn load_feedback(&self) -> Result<usize> {
+        self.feedback
+            .lock()
+            .await
+            .load()
+            .await
+            .map_err(|e| zclaw_types::ZclawError::Internal(e))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::experience_store::Experience;
+
+    #[tokio::test]
+    async fn test_disabled_returns_empty() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let mut engine = EvolutionEngine::new(viking);
+        engine.set_enabled(false);
+
+        let patterns = engine.check_evolvable_patterns("agent-1").await.unwrap();
+        assert!(patterns.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_no_evolvable_patterns() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let engine = EvolutionEngine::new(viking);
+
+        let patterns = engine.check_evolvable_patterns("unknown-agent").await.unwrap();
+        assert!(patterns.is_empty());
+    }
+
+    #[tokio::test]
+    async fn test_finds_evolvable_pattern() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let store_inner = ExperienceStore::new(viking.clone());
+
+        let mut exp = Experience::new(
+            "agent-1",
+            "report generation",
+            "researcher",
+            vec!["query db".into(), "format".into()],
+            "success",
+        );
+        exp.reuse_count = 5;
+        store_inner.store_experience(&exp).await.unwrap();
+
+        let engine = EvolutionEngine::new(viking);
+
+        let patterns = engine.check_evolvable_patterns("agent-1").await.unwrap();
+        assert_eq!(patterns.len(), 1);
+        assert_eq!(patterns[0].pain_pattern, "report generation");
+    }
+
+    #[test]
+    fn test_build_skill_prompt() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let engine = EvolutionEngine::new(viking);
+
+        let exp = Experience::new(
+            "a", "report", "researcher", vec!["step1".into()], "ok",
+        );
+        let pattern = AggregatedPattern {
+            pain_pattern: "report".to_string(),
+            experiences: vec![exp],
+            common_steps: vec!["step1".into()],
+            total_reuse: 5,
+            tools_used: vec!["researcher".into()],
+            industry_context: None,
+        };
+        let prompt = engine.build_skill_prompt(&pattern);
+        assert!(prompt.contains("report"));
+    }
+
+    #[test]
+    fn test_validate_skill_candidate() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let engine = EvolutionEngine::new(viking);
+
+        let exp = Experience::new(
+            "a", "report", "researcher", vec!["step1".into()], "ok",
+        );
+        let pattern = AggregatedPattern {
+            pain_pattern: "report".to_string(),
+            experiences: vec![exp],
+            common_steps: vec!["step1".into()],
+            total_reuse: 5,
+            tools_used: vec!["researcher".into()],
+            industry_context: None,
+        };
+
+        let json = r##"{"name":"报表技能","description":"生成报表","triggers":["报表","日报"],"tools":["researcher"],"body_markdown":"# 报表\n步骤","confidence":0.9}"##;
+        let (candidate, report) = engine
+            .validate_skill_candidate(json, &pattern, vec!["搜索".to_string()])
+            .unwrap();
+        assert_eq!(candidate.name, "报表技能");
+        assert!(report.passed);
+    }
+}
--- a/crates/zclaw-growth/src/experience_extractor.rs
+++ b/crates/zclaw-growth/src/experience_extractor.rs
@@ -0,0 +1,119 @@
+//! 结构化经验提取器
+//! 从对话中提取 ExperienceCandidate（pain_pattern → solution_steps → outcome）
+//! 持久化到 ExperienceStore
+
+use std::sync::Arc;
+
+use crate::experience_store::ExperienceStore;
+use crate::types::{CombinedExtraction, Outcome};
+
+/// 结构化经验提取器
+/// LLM 调用已由上层 MemoryExtractor 完成，这里只做解析和持久化
+pub struct ExperienceExtractor {
+    store: Option<Arc<ExperienceStore>>,
+}
+
+impl ExperienceExtractor {
+    pub fn new() -> Self {
+        Self { store: None }
+    }
+
+    pub fn with_store(mut self, store: Arc<ExperienceStore>) -> Self {
+        self.store = Some(store);
+        self
+    }
+
+    /// 从 CombinedExtraction 中提取经验并持久化
+    /// LLM 调用已由上层完成，这里只做解析和存储
+    pub async fn persist_experiences(
+        &self,
+        agent_id: &str,
+        extraction: &CombinedExtraction,
+    ) -> zclaw_types::Result<usize> {
+        let store = match &self.store {
+            Some(s) => s,
+            None => return Ok(0),
+        };
+
+        let mut count = 0;
+        for candidate in &extraction.experiences {
+            if candidate.confidence < 0.6 {
+                continue;
+            }
+            let outcome_str = match candidate.outcome {
+                Outcome::Success => "success",
+                Outcome::Partial => "partial",
+                Outcome::Failed => "failed",
+            };
+            let mut exp = crate::experience_store::Experience::new(
+                agent_id,
+                &candidate.pain_pattern,
+                &candidate.context,
+                candidate.solution_steps.clone(),
+                outcome_str,
+            );
+            // 填充 tool_used：取 tools_used 中的第一个作为主要工具
+            exp.tool_used = candidate.tools_used.first().cloned();
+            exp.industry_context = candidate.industry_context.clone();
+            store.store_experience(&exp).await?;
+            count += 1;
+        }
+        Ok(count)
+    }
+}
+
+impl Default for ExperienceExtractor {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::types::{ExperienceCandidate, Outcome};
+
+    #[test]
+    fn test_extractor_new_without_store() {
+        let ext = ExperienceExtractor::new();
+        assert!(ext.store.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_persist_no_store_returns_zero() {
+        let ext = ExperienceExtractor::new();
+        let extraction = CombinedExtraction::default();
+        let count = ext.persist_experiences("agent1", &extraction).await.unwrap();
+        assert_eq!(count, 0);
+    }
+
+    #[tokio::test]
+    async fn test_persist_filters_low_confidence() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let store = Arc::new(ExperienceStore::new(viking));
+        let ext = ExperienceExtractor::new().with_store(store);
+
+        let mut extraction = CombinedExtraction::default();
+        extraction.experiences.push(ExperienceCandidate {
+            pain_pattern: "low confidence task".to_string(),
+            context: "should be filtered".to_string(),
+            solution_steps: vec!["step1".to_string()],
+            outcome: Outcome::Success,
+            confidence: 0.3, // 低于 0.6 阈值
+            tools_used: vec![],
+            industry_context: None,
+        });
+        extraction.experiences.push(ExperienceCandidate {
+            pain_pattern: "high confidence task".to_string(),
+            context: "should be stored".to_string(),
+            solution_steps: vec!["step1".to_string(), "step2".to_string()],
+            outcome: Outcome::Success,
+            confidence: 0.9,
+            tools_used: vec!["researcher".to_string()],
+            industry_context: Some("healthcare".to_string()),
+        });
+
+        let count = ext.persist_experiences("agent-1", &extraction).await.unwrap();
+        assert_eq!(count, 1); // 只有 1 个通过置信度过滤
+    }
+}
--- a/crates/zclaw-growth/src/experience_store.rs
+++ b/crates/zclaw-growth/src/experience_store.rs
@@ -48,6 +48,9 @@ pub struct Experience {
    /// Which trigger signal produced this experience.
    #[serde(default)]
    pub source_trigger: Option<String>,
+    /// Primary tool/skill used to resolve this pain point.
+    #[serde(default)]
+    pub tool_used: Option<String>,
 }

 impl Experience {
@@ -72,6 +75,7 @@ impl Experience {
            updated_at: now,
            industry_context: None,
            source_trigger: None,
+            tool_used: None,
        }
    }

@@ -109,6 +113,11 @@ impl ExperienceStore {
        Self { viking }
    }

+    /// Get a reference to the underlying VikingAdapter.
+    pub fn viking(&self) -> &Arc<VikingAdapter> {
+        &self.viking
+    }
+
    /// Store (or overwrite) an experience. The URI is derived from
    /// `agent_id + pain_pattern`, ensuring one experience per pattern.
    pub async fn store_experience(&self, exp: &Experience) -> zclaw_types::Result<()> {
@@ -119,6 +128,9 @@ impl ExperienceStore {
        if let Some(ref industry) = exp.industry_context {
            keywords.push(industry.clone());
        }
+        if let Some(ref tool) = exp.tool_used {
+            keywords.push(tool.clone());
+        }

        let entry = MemoryEntry {
            uri,
--- a/crates/zclaw-growth/src/extractor.rs
+++ b/crates/zclaw-growth/src/extractor.rs
@@ -19,6 +19,34 @@ pub trait LlmDriverForExtraction: Send + Sync {
        messages: &[Message],
        extraction_type: MemoryType,
    ) -> Result<Vec<ExtractedMemory>>;
+
+    /// 单次 LLM 调用提取全部类型（记忆 + 经验 + 画像信号）
+    /// 默认实现：退化到 3 次独立调用（experiences 和 profile_signals 为空）
+    async fn extract_combined_all(
+        &self,
+        messages: &[Message],
+    ) -> Result<crate::types::CombinedExtraction> {
+        let mut combined = crate::types::CombinedExtraction::default();
+        for mt in [MemoryType::Preference, MemoryType::Knowledge, MemoryType::Experience] {
+            if let Ok(mems) = self.extract_memories(messages, mt).await {
+                combined.memories.extend(mems);
+            }
+        }
+        Ok(combined)
+    }
+
+    /// 使用自定义 prompt 进行单次 LLM 调用，返回原始文本响应
+    /// 用于统一提取场景，默认返回不支持错误
+    async fn extract_with_prompt(
+        &self,
+        _messages: &[Message],
+        _system_prompt: &str,
+        _user_prompt: &str,
+    ) -> Result<String> {
+        Err(zclaw_types::ZclawError::Internal(
+            "extract_with_prompt not implemented".to_string(),
+        ))
+    }
 }

 /// Memory Extractor - extracts memories from conversations
@@ -85,13 +113,10 @@ impl MemoryExtractor {
        session_id: SessionId,
    ) -> Result<Vec<ExtractedMemory>> {
        // Check if LLM driver is available
-        let _llm_driver = match &self.llm_driver {
-            Some(driver) => driver,
-            None => {
-                tracing::debug!("[MemoryExtractor] No LLM driver configured, skipping extraction");
-                return Ok(Vec::new());
-            }
-        };
+        if self.llm_driver.is_none() {
+            tracing::debug!("[MemoryExtractor] No LLM driver configured, skipping extraction");
+            return Ok(Vec::new());
+        }

        let mut results = Vec::new();

@@ -227,6 +252,299 @@ impl MemoryExtractor {
        tracing::info!("[MemoryExtractor] Stored {} memories to OpenViking", stored);
        Ok(stored)
    }
+
+    /// 统一提取：单次 LLM 调用同时产出 memories + experiences + profile_signals
+    ///
+    /// 优先使用 `extract_with_prompt()` 进行单次调用；若 driver 不支持则
+    /// 退化为 `extract()` + 从记忆推断经验/画像。
+    pub async fn extract_combined(
+        &self,
+        messages: &[Message],
+        session_id: SessionId,
+    ) -> Result<crate::types::CombinedExtraction> {
+        let llm_driver = match &self.llm_driver {
+            Some(driver) => driver,
+            None => {
+                tracing::debug!(
+                    "[MemoryExtractor] No LLM driver configured, skipping combined extraction"
+                );
+                return Ok(crate::types::CombinedExtraction::default());
+            }
+        };
+
+        // 尝试单次 LLM 调用路径
+        let system_prompt = "You are a memory extraction assistant. Analyze conversations and extract \
+            structured memories, experiences, and profile signals in valid JSON format. \
+            Always respond with valid JSON only, no additional text or markdown formatting.";
+        let user_prompt = format!(
+            "{}{}",
+            crate::extractor::prompts::COMBINED_EXTRACTION_PROMPT,
+            format_conversation_text(messages)
+        );
+
+        match llm_driver
+            .extract_with_prompt(messages, system_prompt, &user_prompt)
+            .await
+        {
+            Ok(raw_text) if !raw_text.trim().is_empty() => {
+                match parse_combined_response(&raw_text, session_id.clone()) {
+                    Ok(combined) => {
+                        tracing::info!(
+                            "[MemoryExtractor] Combined extraction: {} memories, {} experiences, {} profile signals",
+                            combined.memories.len(),
+                            combined.experiences.len(),
+                            combined.profile_signals.signal_count(),
+                        );
+                        return Ok(combined);
+                    }
+                    Err(e) => {
+                        tracing::warn!(
+                            "[MemoryExtractor] Combined response parse failed, falling back: {}",
+                            e
+                        );
+                    }
+                }
+            }
+            Ok(_) => {
+                tracing::debug!("[MemoryExtractor] extract_with_prompt returned empty, falling back");
+            }
+            Err(e) => {
+                tracing::debug!(
+                    "[MemoryExtractor] extract_with_prompt not supported ({}), falling back",
+                    e
+                );
+            }
+        }
+
+        // 退化路径：使用已有的 extract() 然后推断 experiences 和 profile_signals
+        let memories = self.extract(messages, session_id).await?;
+        let experiences = infer_experiences_from_memories(&memories);
+        let profile_signals = infer_profile_signals_from_memories(&memories);
+
+        Ok(crate::types::CombinedExtraction {
+            memories,
+            experiences,
+            profile_signals,
+        })
+    }
+}
+
+/// 格式化对话消息为文本
+fn format_conversation_text(messages: &[Message]) -> String {
+    messages
+        .iter()
+        .filter_map(|msg| match msg {
+            Message::User { content } => Some(format!("[User]: {}", content)),
+            Message::Assistant { content, .. } => Some(format!("[Assistant]: {}", content)),
+            Message::System { content } => Some(format!("[System]: {}", content)),
+            Message::ToolUse { .. } | Message::ToolResult { .. } => None,
+        })
+        .collect::<Vec<_>>()
+        .join("\n\n")
+}
+
+/// 从 LLM 原始响应解析 CombinedExtraction
+pub fn parse_combined_response(
+    raw: &str,
+    session_id: SessionId,
+) -> Result<crate::types::CombinedExtraction> {
+    use crate::types::CombinedExtraction;
+
+    let json_str = crate::json_utils::extract_json_block(raw);
+    let parsed: serde_json::Value = serde_json::from_str(json_str).map_err(|e| {
+        zclaw_types::ZclawError::Internal(format!("Failed to parse combined JSON: {}", e))
+    })?;
+
+    // 解析 memories
+    let memories = parsed
+        .get("memories")
+        .and_then(|v| v.as_array())
+        .map(|arr| {
+            arr.iter()
+                .filter_map(|item| parse_memory_item(item, &session_id))
+                .collect::<Vec<_>>()
+        })
+        .unwrap_or_default();
+
+    // 解析 experiences
+    let experiences = parsed
+        .get("experiences")
+        .and_then(|v| v.as_array())
+        .map(|arr| {
+            arr.iter()
+                .filter_map(parse_experience_item)
+                .collect::<Vec<_>>()
+        })
+        .unwrap_or_default();
+
+    // 解析 profile_signals
+    let profile_signals = parse_profile_signals(&parsed);
+
+    Ok(CombinedExtraction {
+        memories,
+        experiences,
+        profile_signals,
+    })
+}
+
+/// 解析单个 memory 项
+fn parse_memory_item(
+    value: &serde_json::Value,
+    session_id: &SessionId,
+) -> Option<ExtractedMemory> {
+    let content = value.get("content")?.as_str()?.to_string();
+    let category = value
+        .get("category")
+        .and_then(|v| v.as_str())
+        .unwrap_or("unknown")
+        .to_string();
+    let memory_type_str = value
+        .get("memory_type")
+        .and_then(|v| v.as_str())
+        .unwrap_or("knowledge");
+    let memory_type = crate::types::MemoryType::parse(memory_type_str);
+    let confidence = value
+        .get("confidence")
+        .and_then(|v| v.as_f64())
+        .unwrap_or(0.7) as f32;
+    let keywords = crate::json_utils::extract_string_array(value, "keywords");
+
+    Some(
+        ExtractedMemory::new(memory_type, category, content, session_id.clone())
+            .with_confidence(confidence)
+            .with_keywords(keywords),
+    )
+}
+
+/// 解析单个 experience 项
+fn parse_experience_item(value: &serde_json::Value) -> Option<crate::types::ExperienceCandidate> {
+    use crate::types::Outcome;
+
+    let pain_pattern = value.get("pain_pattern")?.as_str()?.to_string();
+    let context = value
+        .get("context")
+        .and_then(|v| v.as_str())
+        .unwrap_or("")
+        .to_string();
+    let solution_steps = crate::json_utils::extract_string_array(value, "solution_steps");
+    let outcome_str = value
+        .get("outcome")
+        .and_then(|v| v.as_str())
+        .unwrap_or("partial");
+    let outcome = match outcome_str {
+        "success" => Outcome::Success,
+        "failed" => Outcome::Failed,
+        _ => Outcome::Partial,
+    };
+    let confidence = value
+        .get("confidence")
+        .and_then(|v| v.as_f64())
+        .unwrap_or(0.6) as f32;
+    let tools_used = crate::json_utils::extract_string_array(value, "tools_used");
+    let industry_context = value
+        .get("industry_context")
+        .and_then(|v| v.as_str())
+        .map(String::from);
+
+    Some(crate::types::ExperienceCandidate {
+        pain_pattern,
+        context,
+        solution_steps,
+        outcome,
+        confidence,
+        tools_used,
+        industry_context,
+    })
+}
+
+/// 解析 profile_signals
+fn parse_profile_signals(obj: &serde_json::Value) -> crate::types::ProfileSignals {
+    let signals = obj.get("profile_signals");
+    crate::types::ProfileSignals {
+        industry: signals
+            .and_then(|s| s.get("industry"))
+            .and_then(|v| v.as_str())
+            .map(String::from),
+        recent_topic: signals
+            .and_then(|s| s.get("recent_topic"))
+            .and_then(|v| v.as_str())
+            .map(String::from),
+        pain_point: signals
+            .and_then(|s| s.get("pain_point"))
+            .and_then(|v| v.as_str())
+            .map(String::from),
+        preferred_tool: signals
+            .and_then(|s| s.get("preferred_tool"))
+            .and_then(|v| v.as_str())
+            .map(String::from),
+        communication_style: signals
+            .and_then(|s| s.get("communication_style"))
+            .and_then(|v| v.as_str())
+            .map(String::from),
+    }
+}
+
+/// 从已有记忆推断结构化经验（退化路径）
+fn infer_experiences_from_memories(
+    memories: &[ExtractedMemory],
+) -> Vec<crate::types::ExperienceCandidate> {
+    memories
+        .iter()
+        .filter(|m| m.memory_type == crate::types::MemoryType::Experience)
+        .filter_map(|m| {
+            // 经验类记忆 → ExperienceCandidate
+            let content = &m.content;
+            if content.len() < 10 {
+                return None;
+            }
+            Some(crate::types::ExperienceCandidate {
+                pain_pattern: m.category.clone(),
+                context: content.clone(),
+                solution_steps: Vec::new(),
+                outcome: crate::types::Outcome::Partial,
+                confidence: m.confidence * 0.7, // 降低推断置信度
+                tools_used: m.keywords.clone(),
+                industry_context: None,
+            })
+        })
+        .collect()
+}
+
+/// 从已有记忆推断画像信号（退化路径）
+fn infer_profile_signals_from_memories(
+    memories: &[ExtractedMemory],
+) -> crate::types::ProfileSignals {
+    use crate::types::ProfileSignals;
+
+    let mut signals = ProfileSignals::default();
+    for m in memories {
+        match m.memory_type {
+            crate::types::MemoryType::Preference => {
+                if m.category.contains("style") || m.category.contains("风格") {
+                    if signals.communication_style.is_none() {
+                        signals.communication_style = Some(m.content.clone());
+                    }
+                }
+            }
+            crate::types::MemoryType::Knowledge => {
+                if signals.recent_topic.is_none() && !m.keywords.is_empty() {
+                    signals.recent_topic = Some(m.keywords.first().cloned().unwrap_or_default());
+                }
+            }
+            crate::types::MemoryType::Experience => {
+                for kw in &m.keywords {
+                    if signals.preferred_tool.is_none()
+                        && m.content.contains(kw.as_str())
+                    {
+                        signals.preferred_tool = Some(kw.clone());
+                        break;
+                    }
+                }
+            }
+            _ => {}
+        }
+    }
+    signals
 }

 /// Default extraction prompts for LLM
@@ -243,6 +561,55 @@ pub mod prompts {
        }
    }

+    /// 统一提取 prompt — 单次 LLM 调用同时提取记忆、结构化经验、画像信号
+    pub const COMBINED_EXTRACTION_PROMPT: &str = r#"
+分析以下对话，一次性提取三类信息。严格按 JSON 格式返回。
+
+## 输出格式
+
+```json
+{
+  "memories": [
+    {
+      "memory_type": "preference|knowledge|experience",
+      "category": "分类标签",
+      "content": "记忆内容",
+      "confidence": 0.0-1.0,
+      "keywords": ["关键词"]
+    }
+  ],
+  "experiences": [
+    {
+      "pain_pattern": "痛点模式简述",
+      "context": "问题发生的上下文",
+      "solution_steps": ["步骤1", "步骤2"],
+      "outcome": "success|partial|failed",
+      "confidence": 0.0-1.0,
+      "tools_used": ["使用的工具/技能"],
+      "industry_context": "行业标识(可选)"
+    }
+  ],
+  "profile_signals": {
+    "industry": "用户所在行业(可选)",
+    "recent_topic": "最近讨论的主要话题(可选)",
+    "pain_point": "用户当前痛点(可选)",
+    "preferred_tool": "用户偏好的工具/技能(可选)",
+    "communication_style": "沟通风格: concise|detailed|formal|casual(可选)"
+  }
+}
+```
+
+## 提取规则
+
+1. **memories**: 提取用户偏好(沟通风格/格式/语言)、知识(事实/领域知识/经验教训)、使用经验(技能/工具使用模式和结果)
+2. **experiences**: 仅提取明确的"问题→解决"模式，要求有清晰的痛点和步骤，confidence >= 0.6
+3. **profile_signals**: 从对话中推断用户画像信息，只在有明确信号时填写，留空则不填
+4. 每个字段都要有实际内容，不确定的宁可省略
+5. 只返回 JSON，不要附加其他文本
+
+对话内容：
+"#;
+
    const PREFERENCE_EXTRACTION_PROMPT: &str = r#"
 分析以下对话，提取用户的偏好设置。关注：
 - 沟通风格偏好（简洁/详细、正式/随意）
@@ -362,11 +729,103 @@ mod tests {
        assert!(!result.is_empty());
    }

+    #[tokio::test]
+    async fn test_extract_combined_all_default_impl() {
+        let driver = MockLlmDriver;
+        let messages = vec![Message::user("Hello")];
+        let result = driver.extract_combined_all(&messages).await.unwrap();
+        assert_eq!(result.memories.len(), 3); // 3 types
+    }
+
    #[test]
    fn test_prompts_available() {
        assert!(!prompts::get_extraction_prompt(MemoryType::Preference).is_empty());
        assert!(!prompts::get_extraction_prompt(MemoryType::Knowledge).is_empty());
        assert!(!prompts::get_extraction_prompt(MemoryType::Experience).is_empty());
        assert!(!prompts::get_extraction_prompt(MemoryType::Session).is_empty());
+        assert!(!prompts::COMBINED_EXTRACTION_PROMPT.is_empty());
+    }
+
+    #[test]
+    fn test_parse_combined_response_full() {
+        let raw = r#"```json
+{
+  "memories": [
+    {
+      "memory_type": "preference",
+      "category": "communication-style",
+      "content": "用户偏好简洁回复",
+      "confidence": 0.9,
+      "keywords": ["简洁", "风格"]
+    },
+    {
+      "memory_type": "knowledge",
+      "category": "user-facts",
+      "content": "用户是医院行政人员",
+      "confidence": 0.85,
+      "keywords": ["医院", "行政"]
+    }
+  ],
+  "experiences": [
+    {
+      "pain_pattern": "报表生成耗时",
+      "context": "月度报表需要手动汇总多个Excel",
+      "solution_steps": ["使用researcher工具自动抓取", "格式化输出为Excel"],
+      "outcome": "success",
+      "confidence": 0.85,
+      "tools_used": ["researcher"],
+      "industry_context": "healthcare"
+    }
+  ],
+  "profile_signals": {
+    "industry": "healthcare",
+    "recent_topic": "报表自动化",
+    "pain_point": "手动汇总Excel太慢",
+    "preferred_tool": "researcher",
+    "communication_style": "concise"
+  }
+}
+```"#;
+
+        let result = super::parse_combined_response(raw, SessionId::new()).unwrap();
+        assert_eq!(result.memories.len(), 2);
+        assert_eq!(result.experiences.len(), 1);
+        assert_eq!(result.experiences[0].pain_pattern, "报表生成耗时");
+        assert_eq!(result.experiences[0].outcome, crate::types::Outcome::Success);
+        assert_eq!(result.profile_signals.industry.as_deref(), Some("healthcare"));
+        assert_eq!(result.profile_signals.pain_point.as_deref(), Some("手动汇总Excel太慢"));
+        assert!(result.profile_signals.has_any_signal());
+    }
+
+    #[test]
+    fn test_parse_combined_response_minimal() {
+        let raw = r#"{"memories": [], "experiences": [], "profile_signals": {}}"#;
+        let result = super::parse_combined_response(raw, SessionId::new()).unwrap();
+        assert!(result.memories.is_empty());
+        assert!(result.experiences.is_empty());
+        assert!(!result.profile_signals.has_any_signal());
+    }
+
+    #[test]
+    fn test_parse_combined_response_invalid() {
+        let raw = "not json at all";
+        let result = super::parse_combined_response(raw, SessionId::new());
+        assert!(result.is_err());
+    }
+
+    #[tokio::test]
+    async fn test_extract_combined_fallback() {
+        // MockLlmDriver doesn't implement extract_with_prompt, so it falls back
+        let driver = Arc::new(MockLlmDriver);
+        let extractor = MemoryExtractor::new(driver);
+        let messages = vec![Message::user("Hello"), Message::assistant("Hi there!")];
+
+        let result = extractor
+            .extract_combined(&messages, SessionId::new())
+            .await
+            .unwrap();
+
+        // Fallback: extract() produces 3 memories, infer produces experiences from them
+        assert!(!result.memories.is_empty());
    }
 }
--- a/crates/zclaw-growth/src/feedback_collector.rs
+++ b/crates/zclaw-growth/src/feedback_collector.rs
@@ -0,0 +1,448 @@
+//! 反馈信号收集与信任度管理（Phase 5 反馈闭环）
+//! 收集用户对进化产物（技能/Pipeline）的显式/隐式反馈
+//! 管理信任度衰减和优化循环
+//! 信任度记录通过 VikingAdapter 持久化
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use chrono::{DateTime, Utc};
+use serde::{Deserialize, Serialize};
+
+use crate::types::MemoryType;
+use crate::viking_adapter::VikingAdapter;
+
+/// 反馈信号类型
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum FeedbackSignal {
+    /// 用户直接表达的意见
+    Explicit,
+    /// 从使用行为推断
+    ImplicitUsage,
+    /// 使用频率
+    UsageCount,
+    /// 任务完成率
+    CompletionRate,
+}
+
+/// 情感倾向
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum Sentiment {
+    Positive,
+    Negative,
+    Neutral,
+}
+
+/// 进化产物类型
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum EvolutionArtifact {
+    Skill,
+    Pipeline,
+}
+
+/// 单条反馈记录
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct FeedbackEntry {
+    pub artifact_id: String,
+    pub artifact_type: EvolutionArtifact,
+    pub signal: FeedbackSignal,
+    pub sentiment: Sentiment,
+    pub details: Option<String>,
+    pub timestamp: DateTime<Utc>,
+}
+
+/// 信任度记录
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct TrustRecord {
+    pub artifact_id: String,
+    pub artifact_type: EvolutionArtifact,
+    pub trust_score: f32,
+    pub total_feedback: u32,
+    pub positive_count: u32,
+    pub negative_count: u32,
+    pub last_updated: DateTime<Utc>,
+}
+
+/// 反馈收集器
+/// 管理反馈记录和信任度评分
+/// 通过 VikingAdapter 持久化信任度记录（可选）
+pub struct FeedbackCollector {
+    trust_records: HashMap<String, TrustRecord>,
+    viking: Option<Arc<VikingAdapter>>,
+    /// 是否已从持久化存储加载信任度记录
+    loaded: bool,
+}
+
+impl FeedbackCollector {
+    pub fn new() -> Self {
+        Self {
+            trust_records: HashMap::new(),
+            viking: None,
+            loaded: false,
+        }
+    }
+
+    /// 创建带 VikingAdapter 的 FeedbackCollector
+    pub fn with_viking(viking: Arc<VikingAdapter>) -> Self {
+        Self {
+            trust_records: HashMap::new(),
+            viking: Some(viking),
+            loaded: false,
+        }
+    }
+
+    /// 从 VikingAdapter 加载已持久化的信任度记录
+    pub async fn load(&mut self) -> Result<usize, String> {
+        let viking = match &self.viking {
+            Some(v) => v,
+            None => return Ok(0),
+        };
+
+        // MemoryEntry::new("feedback", Session, artifact_id) 生成
+        // URI: agent://feedback/sessions/{artifact_id}
+        let entries = viking
+            .find_by_prefix("agent://feedback/sessions/")
+            .await
+            .map_err(|e| format!("Failed to load trust records: {}", e))?;
+
+        let mut count = 0;
+        for entry in entries {
+            match serde_json::from_str::<TrustRecord>(&entry.content) {
+                Ok(record) => {
+                    // 只合并不覆盖：保留内存中的较新记录
+                    self.trust_records
+                        .entry(record.artifact_id.clone())
+                        .or_insert(record);
+                    count += 1;
+                }
+                Err(e) => {
+                    tracing::warn!(
+                        "[FeedbackCollector] Failed to deserialize trust record at {}: {}",
+                        entry.uri,
+                        e
+                    );
+                }
+            }
+        }
+
+        tracing::debug!(
+            "[FeedbackCollector] Loaded {} trust records from storage",
+            count
+        );
+        Ok(count)
+    }
+
+    /// 将信任度记录持久化到 VikingAdapter
+    /// 首次调用时自动从存储加载已有记录，避免覆盖
+    pub async fn save(&mut self) -> Result<usize, String> {
+        // 首次保存前自动加载已有记录，防止丢失历史数据
+        if !self.loaded {
+            match self.load().await {
+                Ok(_) => {
+                    self.loaded = true;
+                }
+                Err(e) => {
+                    // 加载失败时保留 loaded=false，下次 save 会重试
+                    tracing::warn!(
+                        "[FeedbackCollector] Auto-load before save failed, will retry next save: {}",
+                        e
+                    );
+                }
+            }
+        }
+
+        let viking = match &self.viking {
+            Some(v) => v,
+            None => return Ok(0),
+        };
+
+        let mut saved = 0;
+        for record in self.trust_records.values() {
+            let content = match serde_json::to_string(record) {
+                Ok(c) => c,
+                Err(e) => {
+                    tracing::warn!(
+                        "[FeedbackCollector] Failed to serialize trust record {}: {}",
+                        record.artifact_id,
+                        e
+                    );
+                    continue;
+                }
+            };
+            let entry = crate::types::MemoryEntry::new(
+                "feedback",
+                MemoryType::Session,
+                &record.artifact_id,
+                content,
+            )
+            .with_importance((record.trust_score * 10.0) as u8);
+
+            match viking.store(&entry).await {
+                Ok(_) => saved += 1,
+                Err(e) => {
+                    tracing::warn!(
+                        "[FeedbackCollector] Failed to save trust record {}: {}",
+                        record.artifact_id,
+                        e
+                    );
+                }
+            }
+        }
+
+        tracing::debug!(
+            "[FeedbackCollector] Saved {} trust records to storage",
+            saved
+        );
+        Ok(saved)
+    }
+
+    /// 提交一条反馈
+    pub fn submit_feedback(&mut self, entry: FeedbackEntry) -> TrustUpdate {
+        let record = self
+            .trust_records
+            .entry(entry.artifact_id.clone())
+            .or_insert_with(|| TrustRecord {
+                artifact_id: entry.artifact_id.clone(),
+                artifact_type: entry.artifact_type.clone(),
+                trust_score: 0.5,
+                total_feedback: 0,
+                positive_count: 0,
+                negative_count: 0,
+                last_updated: Utc::now(),
+            });
+
+        // 更新计数
+        record.total_feedback += 1;
+        match entry.sentiment {
+            Sentiment::Positive => record.positive_count += 1,
+            Sentiment::Negative => record.negative_count += 1,
+            Sentiment::Neutral => {}
+        }
+
+        // 重新计算信任度
+        let old_score = record.trust_score;
+        record.trust_score = Self::calculate_trust_internal(
+            record.positive_count,
+            record.negative_count,
+            record.total_feedback,
+            record.last_updated,
+        );
+        record.last_updated = Utc::now();
+
+        let new_score = record.trust_score;
+        let total = record.total_feedback;
+        let action = Self::recommend_action_internal(new_score, total);
+
+        TrustUpdate {
+            artifact_id: entry.artifact_id.clone(),
+            old_score,
+            new_score,
+            action,
+        }
+    }
+
+    /// 获取信任度记录
+    pub fn get_trust(&self, artifact_id: &str) -> Option<&TrustRecord> {
+        self.trust_records.get(artifact_id)
+    }
+
+    /// 获取所有需要优化的产物（信任度 < 0.4）
+    pub fn get_artifacts_needing_optimization(&self) -> Vec<&TrustRecord> {
+        self.trust_records
+            .values()
+            .filter(|r| r.trust_score < 0.4 && r.total_feedback >= 2)
+            .collect()
+    }
+
+    /// 获取所有应该归档的产物（信任度 < 0.2 且反馈 >= 5）
+    pub fn get_artifacts_to_archive(&self) -> Vec<&TrustRecord> {
+        self.trust_records
+            .values()
+            .filter(|r| r.trust_score < 0.2 && r.total_feedback >= 5)
+            .collect()
+    }
+
+    /// 获取所有高信任产物（信任度 >= 0.8）
+    pub fn get_recommended_artifacts(&self) -> Vec<&TrustRecord> {
+        self.trust_records
+            .values()
+            .filter(|r| r.trust_score >= 0.8)
+            .collect()
+    }
+
+    fn calculate_trust_internal(
+        positive: u32,
+        negative: u32,
+        total: u32,
+        last_updated: DateTime<Utc>,
+    ) -> f32 {
+        if total == 0 {
+            return 0.5;
+        }
+        let positive_ratio = positive as f32 / total as f32;
+        let negative_penalty = negative as f32 * 0.1;
+        let days_since = (Utc::now() - last_updated).num_days().max(0) as f32;
+        let time_decay = 1.0 - (days_since * 0.005).min(0.5);
+        (positive_ratio * time_decay - negative_penalty).clamp(0.0, 1.0)
+    }
+
+    fn recommend_action_internal(trust_score: f32, total_feedback: u32) -> RecommendedAction {
+        if trust_score >= 0.8 {
+            RecommendedAction::Promote
+        } else if trust_score < 0.2 && total_feedback >= 5 {
+            RecommendedAction::Archive
+        } else if trust_score < 0.4 && total_feedback >= 2 {
+            RecommendedAction::Optimize
+        } else {
+            RecommendedAction::Monitor
+        }
+    }
+}
+
+impl Default for FeedbackCollector {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// 信任度更新结果
+#[derive(Debug, Clone)]
+pub struct TrustUpdate {
+    pub artifact_id: String,
+    pub old_score: f32,
+    pub new_score: f32,
+    pub action: RecommendedAction,
+}
+
+/// 建议动作
+#[derive(Debug, Clone, PartialEq)]
+pub enum RecommendedAction {
+    /// 继续观察
+    Monitor,
+    /// 需要优化
+    Optimize,
+    /// 建议归档（降级为记忆）
+    Archive,
+    /// 建议提升为推荐技能
+    Promote,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_feedback(artifact_id: &str, sentiment: Sentiment) -> FeedbackEntry {
+        FeedbackEntry {
+            artifact_id: artifact_id.to_string(),
+            artifact_type: EvolutionArtifact::Skill,
+            signal: FeedbackSignal::Explicit,
+            sentiment,
+            details: None,
+            timestamp: Utc::now(),
+        }
+    }
+
+    #[test]
+    fn test_initial_trust() {
+        let collector = FeedbackCollector::new();
+        assert!(collector.get_trust("skill-1").is_none());
+    }
+
+    #[test]
+    fn test_positive_feedback_increases_trust() {
+        let mut collector = FeedbackCollector::new();
+        collector.submit_feedback(make_feedback("skill-1", Sentiment::Positive));
+        let record = collector.get_trust("skill-1").unwrap();
+        assert!(record.trust_score > 0.5);
+        assert_eq!(record.positive_count, 1);
+    }
+
+    #[test]
+    fn test_negative_feedback_decreases_trust() {
+        let mut collector = FeedbackCollector::new();
+        collector.submit_feedback(make_feedback("skill-1", Sentiment::Negative));
+        let record = collector.get_trust("skill-1").unwrap();
+        assert!(record.trust_score < 0.5);
+    }
+
+    #[test]
+    fn test_mixed_feedback() {
+        let mut collector = FeedbackCollector::new();
+        collector.submit_feedback(make_feedback("skill-1", Sentiment::Positive));
+        collector.submit_feedback(make_feedback("skill-1", Sentiment::Positive));
+        collector.submit_feedback(make_feedback("skill-1", Sentiment::Negative));
+        let record = collector.get_trust("skill-1").unwrap();
+        assert_eq!(record.total_feedback, 3);
+        assert!(record.trust_score > 0.3); // 2/3 positive
+    }
+
+    #[test]
+    fn test_recommend_optimize() {
+        let mut collector = FeedbackCollector::new();
+        collector.submit_feedback(make_feedback("skill-1", Sentiment::Negative));
+        let update = collector.submit_feedback(make_feedback("skill-1", Sentiment::Negative));
+        assert_eq!(update.action, RecommendedAction::Optimize);
+    }
+
+    #[test]
+    fn test_needs_optimization_filter() {
+        let mut collector = FeedbackCollector::new();
+        collector.submit_feedback(make_feedback("bad-skill", Sentiment::Negative));
+        collector.submit_feedback(make_feedback("bad-skill", Sentiment::Negative));
+        collector.submit_feedback(make_feedback("good-skill", Sentiment::Positive));
+
+        let needs = collector.get_artifacts_needing_optimization();
+        assert_eq!(needs.len(), 1);
+        assert_eq!(needs[0].artifact_id, "bad-skill");
+    }
+
+    #[test]
+    fn test_promote_recommendation() {
+        let mut collector = FeedbackCollector::new();
+        for _ in 0..5 {
+            collector.submit_feedback(make_feedback("great-skill", Sentiment::Positive));
+        }
+        let recommended = collector.get_recommended_artifacts();
+        assert_eq!(recommended.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn test_save_and_load_roundtrip() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+
+        // 写入阶段
+        let mut collector = FeedbackCollector::with_viking(viking.clone());
+        collector.submit_feedback(make_feedback("skill-a", Sentiment::Positive));
+        collector.submit_feedback(make_feedback("skill-a", Sentiment::Positive));
+        collector.submit_feedback(make_feedback("skill-b", Sentiment::Negative));
+
+        let saved = collector.save().await.unwrap();
+        assert_eq!(saved, 2); // 2 个 artifact
+
+        // 读取阶段：新 collector 从存储加载
+        let mut collector2 = FeedbackCollector::with_viking(viking);
+        let loaded = collector2.load().await.unwrap();
+        assert_eq!(loaded, 2);
+
+        let record_a = collector2.get_trust("skill-a").unwrap();
+        assert_eq!(record_a.positive_count, 2);
+        assert_eq!(record_a.total_feedback, 2);
+
+        let record_b = collector2.get_trust("skill-b").unwrap();
+        assert_eq!(record_b.negative_count, 1);
+    }
+
+    #[tokio::test]
+    async fn test_load_without_viking_returns_zero() {
+        let mut collector = FeedbackCollector::new();
+        let loaded = collector.load().await.unwrap();
+        assert_eq!(loaded, 0);
+    }
+
+    #[tokio::test]
+    async fn test_save_without_viking_returns_zero() {
+        let mut collector = FeedbackCollector::new();
+        let saved = collector.save().await.unwrap();
+        assert_eq!(saved, 0);
+    }
+}
--- a/crates/zclaw-growth/src/json_utils.rs
+++ b/crates/zclaw-growth/src/json_utils.rs
@@ -0,0 +1,148 @@
+//! 共享 JSON 工具函数
+//! 从 LLM 返回的文本中提取 JSON 块
+
+/// 从 LLM 返回文本中提取 JSON 块
+/// 支持三种格式：```json...``` 围栏、```...``` 围栏、裸 {...}
+/// 使用括号平衡算法找到第一个完整 JSON 块，避免误匹配
+pub fn extract_json_block(text: &str) -> &str {
+    // 尝试匹配 ```json ... ```
+    if let Some(start) = text.find("```json") {
+        let json_start = start + 7;
+        if let Some(end) = text[json_start..].find("```") {
+            return text[json_start..json_start + end].trim();
+        }
+    }
+    // 尝试匹配 ``` ... ```
+    if let Some(start) = text.find("```") {
+        let json_start = start + 3;
+        if let Some(end) = text[json_start..].find("```") {
+            return text[json_start..json_start + end].trim();
+        }
+    }
+    // 用括号平衡算法找第一个完整 {...} 块
+    if let Some(slice) = find_balanced_json(text) {
+        return slice;
+    }
+    text.trim()
+}
+
+/// 使用括号平衡计数找到第一个完整的 {...} JSON 块
+/// 正确处理字符串字面量中的花括号
+fn find_balanced_json(text: &str) -> Option<&str> {
+    let start = text.find('{')?;
+    let mut depth = 0i32;
+    let mut in_string = false;
+    let mut escape_next = false;
+
+    for (i, c) in text[start..].char_indices() {
+        if escape_next {
+            escape_next = false;
+            continue;
+        }
+        match c {
+            '\\' if in_string => escape_next = true,
+            '"' => in_string = !in_string,
+            '{' if !in_string => {
+                depth += 1;
+            }
+            '}' if !in_string => {
+                depth -= 1;
+                if depth == 0 {
+                    return Some(&text[start..=start + i]);
+                }
+            }
+            _ => {}
+        }
+    }
+    None
+}
+
+/// 从 serde_json::Value 中提取字符串数组
+/// 用于解析 LLM 返回 JSON 中的 triggers/tools 等字段
+pub fn extract_string_array(raw: &serde_json::Value, key: &str) -> Vec<String> {
+    raw.get(key)
+        .and_then(|v| v.as_array())
+        .map(|a| {
+            a.iter()
+                .filter_map(|v| v.as_str().map(String::from))
+                .collect()
+        })
+        .unwrap_or_default()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_json_block_with_markdown() {
+        let text = "Here is the result:\n```json\n{\"key\": \"value\"}\n```\nDone.";
+        assert_eq!(extract_json_block(text), "{\"key\": \"value\"}");
+    }
+
+    #[test]
+    fn test_json_block_bare() {
+        let text = "{\"key\": \"value\"}";
+        assert_eq!(extract_json_block(text), "{\"key\": \"value\"}");
+    }
+
+    #[test]
+    fn test_json_block_plain_fences() {
+        let text = "Result:\n```\n{\"a\": 1}\n```";
+        assert_eq!(extract_json_block(text), "{\"a\": 1}");
+    }
+
+    #[test]
+    fn test_json_block_nested_braces() {
+        let text = r#"{"outer": {"inner": "val"}}"#;
+        assert_eq!(extract_json_block(text), r#"{"outer": {"inner": "val"}}"#);
+    }
+
+    #[test]
+    fn test_json_block_no_json() {
+        let text = "no json here";
+        assert_eq!(extract_json_block(text), "no json here");
+    }
+
+    #[test]
+    fn test_balanced_json_skips_outer_text() {
+        // 第一个 { 到最后一个 } 会包含多余文本，但平衡算法只取第一个完整块
+        let text = "prefix {\"a\": 1} suffix {\"b\": 2}";
+        assert_eq!(extract_json_block(text), "{\"a\": 1}");
+    }
+
+    #[test]
+    fn test_balanced_json_handles_braces_in_strings() {
+        let text = r#"{"body": "function() { return x; }", "name": "test"}"#;
+        assert_eq!(
+            extract_json_block(text),
+            r#"{"body": "function() { return x; }", "name": "test"}"#
+        );
+    }
+
+    #[test]
+    fn test_balanced_json_handles_escaped_quotes() {
+        let text = r#"{"msg": "He said \"hello {world}\""}"#;
+        assert_eq!(
+            extract_json_block(text),
+            r#"{"msg": "He said \"hello {world}\""}"#
+        );
+    }
+
+    #[test]
+    fn test_extract_string_array() {
+        let raw: serde_json::Value = serde_json::from_str(
+            r#"{"triggers": ["报表", "日报"], "name": "test"}"#,
+        )
+        .unwrap();
+        let arr = extract_string_array(&raw, "triggers");
+        assert_eq!(arr, vec!["报表", "日报"]);
+    }
+
+    #[test]
+    fn test_extract_string_array_missing_key() {
+        let raw: serde_json::Value = serde_json::from_str(r#"{"name": "test"}"#).unwrap();
+        let arr = extract_string_array(&raw, "triggers");
+        assert!(arr.is_empty());
+    }
+}
--- a/crates/zclaw-growth/src/lib.rs
+++ b/crates/zclaw-growth/src/lib.rs
@@ -5,10 +5,13 @@
 //!
 //! # Architecture
 //!
-//! The growth system consists of four main components:
+//! The growth system consists of several subsystems:
+//!
+//! ## Memory Pipeline (L0-L2)
 //!
 //! 1. **MemoryExtractor** (`extractor`) - Analyzes conversations and extracts
-//!    preferences, knowledge, and experience using LLM.
+//!    preferences, knowledge, and experience using LLM. Supports combined extraction
+//!    (single LLM call for memories + experiences + profile signals).
 //!
 //! 2. **MemoryRetriever** (`retriever`) - Performs semantic search over
 //!    stored memories to find contextually relevant information.
@@ -19,6 +22,28 @@
 //! 4. **GrowthTracker** (`tracker`) - Tracks growth metrics and evolution
 //!    over time.
 //!
+//! ## Evolution Engine (L1-L3)
+//!
+//! 5. **ExperienceStore** (`experience_store`) - FTS5-backed structured experience storage.
+//!
+//! 6. **PatternAggregator** (`pattern_aggregator`) - Collects high-frequency patterns for L2.
+//!
+//! 7. **SkillGenerator** (`skill_generator`) - LLM-driven SKILL.md content generation.
+//!
+//! 8. **QualityGate** (`quality_gate`) - Validates candidate skills (confidence, conflicts).
+//!
+//! 9. **EvolutionEngine** (`evolution_engine`) - Orchestrates L1/L2/L3 evolution phases.
+//!
+//! 10. **WorkflowComposer** (`workflow_composer`) - Extracts tool chain patterns for Pipeline YAML.
+//!
+//! 11. **FeedbackCollector** (`feedback_collector`) - Trust score management with decay.
+//!
+//! ## Support Modules
+//!
+//! 12. **VikingAdapter** (`viking_adapter`) - Storage abstraction (in-memory + SQLite backends).
+//! 13. **Summarizer** (`summarizer`) - L0/L1 summary generation.
+//! 14. **JsonUtils** (`json_utils`) - Shared JSON parsing utilities.
+//!
 //! # Storage
 //!
 //! All memories are stored in OpenViking with a URI structure:
@@ -65,6 +90,15 @@ pub mod storage;
 pub mod retrieval;
 pub mod summarizer;
 pub mod experience_store;
+pub mod json_utils;
+pub mod experience_extractor;
+pub mod profile_updater;
+pub mod pattern_aggregator;
+pub mod skill_generator;
+pub mod quality_gate;
+pub mod evolution_engine;
+pub mod workflow_composer;
+pub mod feedback_collector;

 // Re-export main types for convenience
 pub use types::{
@@ -78,6 +112,14 @@ pub use types::{
    RetrievalResult,
    UriBuilder,
    effective_importance,
+    ArtifactType,
+    CombinedExtraction,
+    EvolutionEvent,
+    EvolutionEventType,
+    EvolutionStatus,
+    ExperienceCandidate,
+    Outcome,
+    ProfileSignals,
 };

 pub use extractor::{LlmDriverForExtraction, MemoryExtractor};
@@ -89,6 +131,18 @@ pub use storage::SqliteStorage;
 pub use experience_store::{Experience, ExperienceStore};
 pub use retrieval::{EmbeddingClient, MemoryCache, QueryAnalyzer, SemanticScorer};
 pub use summarizer::SummaryLlmDriver;
+pub use experience_extractor::ExperienceExtractor;
+pub use json_utils::{extract_json_block, extract_string_array};
+pub use profile_updater::{ProfileFieldUpdate, ProfileUpdateKind, UserProfileUpdater};
+pub use pattern_aggregator::{AggregatedPattern, PatternAggregator};
+pub use skill_generator::{SkillCandidate, SkillGenerator};
+pub use quality_gate::{QualityGate, QualityReport};
+pub use evolution_engine::{EvolutionConfig, EvolutionEngine};
+pub use workflow_composer::{PipelineCandidate, ToolChainPattern, WorkflowComposer};
+pub use feedback_collector::{
+    EvolutionArtifact, FeedbackCollector, FeedbackEntry, FeedbackSignal,
+    RecommendedAction, Sentiment, TrustRecord, TrustUpdate,
+};

 /// Growth system configuration
 #[derive(Debug, Clone)]
--- a/crates/zclaw-growth/src/pattern_aggregator.rs
+++ b/crates/zclaw-growth/src/pattern_aggregator.rs
@@ -0,0 +1,245 @@
+//! 经验模式聚合器
+//! 收集同一 pain_pattern 下的所有 Experience，找出共同步骤
+//! 用于 L2 技能进化触发判断
+
+use std::collections::HashMap;
+
+use crate::experience_store::{Experience, ExperienceStore};
+use zclaw_types::Result;
+
+/// 聚合后的经验模式
+#[derive(Debug, Clone)]
+pub struct AggregatedPattern {
+    pub pain_pattern: String,
+    pub experiences: Vec<Experience>,
+    pub common_steps: Vec<String>,
+    pub total_reuse: u32,
+    pub tools_used: Vec<String>,
+    pub industry_context: Option<String>,
+}
+
+/// 经验模式聚合器
+/// 从 ExperienceStore 中收集高频复用的模式，作为 L2 技能生成的输入
+pub struct PatternAggregator {
+    store: ExperienceStore,
+}
+
+impl PatternAggregator {
+    pub fn new(store: ExperienceStore) -> Self {
+        Self { store }
+    }
+
+    /// 查找可固化的模式：reuse_count >= threshold 的经验
+    pub async fn find_evolvable_patterns(
+        &self,
+        agent_id: &str,
+        min_reuse: u32,
+    ) -> Result<Vec<AggregatedPattern>> {
+        let all = self.store.find_by_agent(agent_id).await?;
+        let mut grouped: HashMap<String, Vec<Experience>> = HashMap::new();
+
+        for exp in all {
+            if exp.reuse_count >= min_reuse {
+                grouped
+                    .entry(exp.pain_pattern.clone())
+                    .or_default()
+                    .push(exp);
+            }
+        }
+
+        let mut patterns = Vec::new();
+        for (pattern, experiences) in grouped {
+            let total_reuse: u32 = experiences.iter().map(|e| e.reuse_count).sum();
+            let common_steps = Self::find_common_steps(&experiences);
+
+            // 从 tool_used 字段提取工具名
+            let tools: Vec<String> = experiences
+                .iter()
+                .filter_map(|e| e.tool_used.clone())
+                .filter(|s| !s.is_empty())
+                .collect::<std::collections::HashSet<_>>()
+                .into_iter()
+                .collect();
+
+            let industry = experiences
+                .iter()
+                .filter_map(|e| e.industry_context.clone())
+                .next();
+
+            patterns.push(AggregatedPattern {
+                pain_pattern: pattern,
+                experiences,
+                common_steps,
+                total_reuse,
+                tools_used: tools,
+                industry_context: industry,
+            });
+        }
+
+        // 按 reuse 排序
+        patterns.sort_by(|a, b| b.total_reuse.cmp(&a.total_reuse));
+        Ok(patterns)
+    }
+
+    /// 找出多条经验中共同的解决步骤
+    fn find_common_steps(experiences: &[Experience]) -> Vec<String> {
+        if experiences.is_empty() {
+            return Vec::new();
+        }
+        if experiences.len() == 1 {
+            return experiences[0].solution_steps.clone();
+        }
+
+        // 取所有经验的交集步骤
+        let mut step_counts: HashMap<String, u32> = HashMap::new();
+        for exp in experiences {
+            for step in &exp.solution_steps {
+                *step_counts.entry(step.clone()).or_insert(0) += 1;
+            }
+        }
+
+        let threshold = experiences.len() as f32 * 0.5; // 出现在 50%+ 的经验中
+        let mut common: Vec<_> = step_counts
+            .into_iter()
+            .filter(|(_, count)| (*count as f32) >= threshold)
+            .map(|(step, _)| step)
+            .collect();
+        common.dedup();
+        common
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::sync::Arc;
+
+    #[test]
+    fn test_find_common_steps_empty() {
+        let steps = PatternAggregator::find_common_steps(&[]);
+        assert!(steps.is_empty());
+    }
+
+    #[test]
+    fn test_find_common_steps_single() {
+        let exp = Experience::new(
+            "a",
+            "packaging",
+            "ctx",
+            vec!["step1".into(), "step2".into()],
+            "ok",
+        );
+        let steps = PatternAggregator::find_common_steps(&[exp]);
+        assert_eq!(steps.len(), 2);
+    }
+
+    #[test]
+    fn test_find_common_steps_multiple() {
+        let exp1 = Experience::new(
+            "a",
+            "packaging",
+            "ctx",
+            vec!["step1".into(), "step2".into(), "step3".into()],
+            "ok",
+        );
+        let exp2 = Experience::new(
+            "a",
+            "packaging",
+            "ctx",
+            vec!["step1".into(), "step2".into(), "step4".into()],
+            "ok",
+        );
+        // step1 and step2 appear in both (100% >= 50%)
+        let steps = PatternAggregator::find_common_steps(&[exp1, exp2]);
+        assert!(steps.contains(&"step1".to_string()));
+        assert!(steps.contains(&"step2".to_string()));
+    }
+
+    #[tokio::test]
+    async fn test_find_evolvable_patterns_filters_low_reuse() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let store = ExperienceStore::new(viking);
+
+        // 经验 1: reuse_count = 0 (低于阈值)
+        let mut exp_low = Experience::new(
+            "agent-1",
+            "low reuse task",
+            "ctx",
+            vec!["step".into()],
+            "ok",
+        );
+        exp_low.reuse_count = 0;
+        store.store_experience(&exp_low).await.unwrap();
+
+        // 经验 2: reuse_count = 5 (高于阈值)
+        let mut exp_high = Experience::new(
+            "agent-1",
+            "high reuse task",
+            "ctx",
+            vec!["step1".into()],
+            "ok",
+        );
+        exp_high.reuse_count = 5;
+        store.store_experience(&exp_high).await.unwrap();
+
+        let aggregator = PatternAggregator::new(store);
+        let patterns = aggregator.find_evolvable_patterns("agent-1", 3).await.unwrap();
+
+        assert_eq!(patterns.len(), 1);
+        assert_eq!(patterns[0].pain_pattern, "high reuse task");
+        assert_eq!(patterns[0].total_reuse, 5);
+    }
+
+    #[tokio::test]
+    async fn test_find_evolvable_patterns_groups_by_pain() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let store = ExperienceStore::new(viking);
+
+        let mut exp1 = Experience::new(
+            "agent-1",
+            "report generation",
+            "ctx1",
+            vec!["query db".into(), "format".into()],
+            "ok",
+        );
+        exp1.reuse_count = 3;
+        store.store_experience(&exp1).await.unwrap();
+
+        // Same pain_pattern → same URI → overwrites, so use a slightly different hash
+        // Actually since URI is deterministic on pain_pattern, we can only have one per pattern
+        // This is by design: one experience per pain_pattern (latest wins)
+        let patterns = aggregator_fixtures::make_patterns_with_same_pain().await;
+        assert_eq!(patterns.len(), 1);
+    }
+
+    mod aggregator_fixtures {
+        use super::*;
+
+        pub async fn make_patterns_with_same_pain() -> Vec<AggregatedPattern> {
+            let viking = Arc::new(crate::VikingAdapter::in_memory());
+            let store = ExperienceStore::new(viking);
+
+            let mut exp = Experience::new(
+                "agent-1",
+                "report generation",
+                "ctx1",
+                vec!["query db".into(), "format".into()],
+                "ok",
+            );
+            exp.reuse_count = 3;
+            store.store_experience(&exp).await.unwrap();
+
+            let aggregator = PatternAggregator::new(store);
+            aggregator.find_evolvable_patterns("agent-1", 2).await.unwrap()
+        }
+    }
+
+    #[tokio::test]
+    async fn test_find_evolvable_patterns_empty() {
+        let viking = Arc::new(crate::VikingAdapter::in_memory());
+        let store = ExperienceStore::new(viking);
+        let aggregator = PatternAggregator::new(store);
+        let patterns = aggregator.find_evolvable_patterns("unknown-agent", 3).await.unwrap();
+        assert!(patterns.is_empty());
+    }
+}
--- a/crates/zclaw-growth/src/profile_updater.rs
+++ b/crates/zclaw-growth/src/profile_updater.rs
@@ -0,0 +1,157 @@
+//! 用户画像增量更新器
+//! 从 CombinedExtraction 的 profile_signals 提取需要更新的字段
+//! 不额外调用 LLM，纯规则驱动
+
+use crate::types::CombinedExtraction;
+
+/// 更新类型：字段覆盖 vs 数组追加
+#[derive(Debug, Clone, PartialEq)]
+pub enum ProfileUpdateKind {
+    /// 直接覆盖字段值（industry, communication_style）
+    SetField,
+    /// 追加到 JSON 数组字段（recent_topic, pain_point, preferred_tool）
+    AppendArray,
+}
+
+/// 待更新的画像字段
+#[derive(Debug, Clone, PartialEq)]
+pub struct ProfileFieldUpdate {
+    pub field: String,
+    pub value: String,
+    pub kind: ProfileUpdateKind,
+}
+
+/// 用户画像更新器
+/// 从 CombinedExtraction 的 profile_signals 中提取需更新的字段列表
+/// 调用方（zclaw-runtime）负责实际写入 UserProfileStore
+pub struct UserProfileUpdater;
+
+impl UserProfileUpdater {
+    pub fn new() -> Self {
+        Self
+    }
+
+    /// 从提取结果中收集需要更新的画像字段
+    /// 返回 (field, value, kind) 列表，由调用方根据 kind 选择写入方式
+    pub fn collect_updates(
+        &self,
+        extraction: &CombinedExtraction,
+    ) -> Vec<ProfileFieldUpdate> {
+        let signals = &extraction.profile_signals;
+        let mut updates = Vec::new();
+
+        if let Some(ref industry) = signals.industry {
+            updates.push(ProfileFieldUpdate {
+                field: "industry".to_string(),
+                value: industry.clone(),
+                kind: ProfileUpdateKind::SetField,
+            });
+        }
+
+        if let Some(ref style) = signals.communication_style {
+            updates.push(ProfileFieldUpdate {
+                field: "communication_style".to_string(),
+                value: style.clone(),
+                kind: ProfileUpdateKind::SetField,
+            });
+        }
+
+        if let Some(ref topic) = signals.recent_topic {
+            updates.push(ProfileFieldUpdate {
+                field: "recent_topic".to_string(),
+                value: topic.clone(),
+                kind: ProfileUpdateKind::AppendArray,
+            });
+        }
+
+        if let Some(ref pain) = signals.pain_point {
+            updates.push(ProfileFieldUpdate {
+                field: "pain_point".to_string(),
+                value: pain.clone(),
+                kind: ProfileUpdateKind::AppendArray,
+            });
+        }
+
+        if let Some(ref tool) = signals.preferred_tool {
+            updates.push(ProfileFieldUpdate {
+                field: "preferred_tool".to_string(),
+                value: tool.clone(),
+                kind: ProfileUpdateKind::AppendArray,
+            });
+        }
+
+        updates
+    }
+}
+
+impl Default for UserProfileUpdater {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_collect_updates_industry() {
+        let mut extraction = CombinedExtraction::default();
+        extraction.profile_signals.industry = Some("healthcare".to_string());
+
+        let updater = UserProfileUpdater::new();
+        let updates = updater.collect_updates(&extraction);
+
+        assert_eq!(updates.len(), 1);
+        assert_eq!(updates[0].field, "industry");
+        assert_eq!(updates[0].value, "healthcare");
+        assert_eq!(updates[0].kind, ProfileUpdateKind::SetField);
+    }
+
+    #[test]
+    fn test_collect_updates_no_signals() {
+        let extraction = CombinedExtraction::default();
+        let updater = UserProfileUpdater::new();
+        let updates = updater.collect_updates(&extraction);
+        assert!(updates.is_empty());
+    }
+
+    #[test]
+    fn test_collect_updates_multiple_signals() {
+        let mut extraction = CombinedExtraction::default();
+        extraction.profile_signals.industry = Some("ecommerce".to_string());
+        extraction.profile_signals.communication_style = Some("concise".to_string());
+
+        let updater = UserProfileUpdater::new();
+        let updates = updater.collect_updates(&extraction);
+
+        assert_eq!(updates.len(), 2);
+    }
+
+    #[test]
+    fn test_collect_updates_all_five_dimensions() {
+        let mut extraction = CombinedExtraction::default();
+        extraction.profile_signals.industry = Some("healthcare".to_string());
+        extraction.profile_signals.communication_style = Some("concise".to_string());
+        extraction.profile_signals.recent_topic = Some("报表自动化".to_string());
+        extraction.profile_signals.pain_point = Some("手动汇总太慢".to_string());
+        extraction.profile_signals.preferred_tool = Some("researcher".to_string());
+
+        let updater = UserProfileUpdater::new();
+        let updates = updater.collect_updates(&extraction);
+
+        assert_eq!(updates.len(), 5);
+        let set_fields: Vec<_> = updates
+            .iter()
+            .filter(|u| u.kind == ProfileUpdateKind::SetField)
+            .map(|u| u.field.as_str())
+            .collect();
+        let append_fields: Vec<_> = updates
+            .iter()
+            .filter(|u| u.kind == ProfileUpdateKind::AppendArray)
+            .map(|u| u.field.as_str())
+            .collect();
+        assert_eq!(set_fields, vec!["industry", "communication_style"]);
+        assert_eq!(append_fields, vec!["recent_topic", "pain_point", "preferred_tool"]);
+    }
+}
--- a/crates/zclaw-growth/src/quality_gate.rs
+++ b/crates/zclaw-growth/src/quality_gate.rs
@@ -0,0 +1,160 @@
+//! 质量门控
+//! 验证生成的技能/工作流是否满足质量标准
+//! 包括：置信度阈值、触发词冲突检查、格式校验
+
+use crate::skill_generator::SkillCandidate;
+
+/// 质量验证报告
+#[derive(Debug, Clone)]
+pub struct QualityReport {
+    pub passed: bool,
+    pub issues: Vec<String>,
+    pub confidence: f32,
+}
+
+/// 质量门控验证器
+pub struct QualityGate {
+    min_confidence: f32,
+    existing_triggers: Vec<String>,
+}
+
+impl QualityGate {
+    pub fn new(min_confidence: f32, existing_triggers: Vec<String>) -> Self {
+        Self {
+            min_confidence,
+            existing_triggers,
+        }
+    }
+
+    /// 验证技能候选项
+    pub fn validate_skill(&self, candidate: &SkillCandidate) -> QualityReport {
+        let mut issues = Vec::new();
+
+        // 1. 置信度检查
+        if candidate.confidence < self.min_confidence {
+            issues.push(format!(
+                "置信度 {:.2} 低于阈值 {:.2}",
+                candidate.confidence, self.min_confidence
+            ));
+        }
+
+        // 2. 名称非空
+        if candidate.name.trim().is_empty() {
+            issues.push("技能名称不能为空".to_string());
+        }
+
+        // 3. 至少一个触发词
+        if candidate.triggers.is_empty() {
+            issues.push("至少需要一个触发词".to_string());
+        }
+
+        // 4. 触发词不与现有技能冲突
+        let conflicts: Vec<_> = candidate
+            .triggers
+            .iter()
+            .filter(|t| self.existing_triggers.iter().any(|et| et == *t))
+            .collect();
+        if !conflicts.is_empty() {
+            issues.push(format!("触发词冲突: {:?}", conflicts));
+        }
+
+        // 5. SKILL.md 正文非空
+        if candidate.body_markdown.trim().is_empty() {
+            issues.push("技能正文不能为空".to_string());
+        }
+
+        QualityReport {
+            passed: issues.is_empty(),
+            issues,
+            confidence: candidate.confidence,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn make_valid_candidate() -> SkillCandidate {
+        SkillCandidate {
+            name: "每日报表".to_string(),
+            description: "生成每日报表".to_string(),
+            triggers: vec!["报表".to_string(), "日报".to_string()],
+            tools: vec!["researcher".to_string()],
+            body_markdown: "# 每日报表\n步骤1\n步骤2".to_string(),
+            source_pattern: "报表生成".to_string(),
+            confidence: 0.85,
+            version: 1,
+        }
+    }
+
+    #[test]
+    fn test_validate_valid_skill() {
+        let gate = QualityGate::new(0.7, vec!["搜索".to_string()]);
+        let candidate = make_valid_candidate();
+        let report = gate.validate_skill(&candidate);
+        assert!(report.passed);
+        assert!(report.issues.is_empty());
+    }
+
+    #[test]
+    fn test_validate_low_confidence() {
+        let gate = QualityGate::new(0.7, vec![]);
+        let mut candidate = make_valid_candidate();
+        candidate.confidence = 0.5;
+        let report = gate.validate_skill(&candidate);
+        assert!(!report.passed);
+        assert!(report.issues.iter().any(|i| i.contains("置信度")));
+    }
+
+    #[test]
+    fn test_validate_empty_name() {
+        let gate = QualityGate::new(0.5, vec![]);
+        let mut candidate = make_valid_candidate();
+        candidate.name = "".to_string();
+        let report = gate.validate_skill(&candidate);
+        assert!(!report.passed);
+        assert!(report.issues.iter().any(|i| i.contains("名称")));
+    }
+
+    #[test]
+    fn test_validate_empty_triggers() {
+        let gate = QualityGate::new(0.5, vec![]);
+        let mut candidate = make_valid_candidate();
+        candidate.triggers = vec![];
+        let report = gate.validate_skill(&candidate);
+        assert!(!report.passed);
+        assert!(report.issues.iter().any(|i| i.contains("触发词")));
+    }
+
+    #[test]
+    fn test_validate_trigger_conflict() {
+        let gate = QualityGate::new(0.5, vec!["报表".to_string()]);
+        let candidate = make_valid_candidate();
+        let report = gate.validate_skill(&candidate);
+        assert!(!report.passed);
+        assert!(report.issues.iter().any(|i| i.contains("冲突")));
+    }
+
+    #[test]
+    fn test_validate_empty_body() {
+        let gate = QualityGate::new(0.5, vec![]);
+        let mut candidate = make_valid_candidate();
+        candidate.body_markdown = "".to_string();
+        let report = gate.validate_skill(&candidate);
+        assert!(!report.passed);
+        assert!(report.issues.iter().any(|i| i.contains("正文")));
+    }
+
+    #[test]
+    fn test_validate_multiple_issues() {
+        let gate = QualityGate::new(0.9, vec![]);
+        let mut candidate = make_valid_candidate();
+        candidate.confidence = 0.3;
+        candidate.triggers = vec![];
+        candidate.body_markdown = "".to_string();
+        let report = gate.validate_skill(&candidate);
+        assert!(!report.passed);
+        assert!(report.issues.len() >= 3);
+    }
+}
--- a/crates/zclaw-growth/src/retrieval/cache.rs
+++ b/crates/zclaw-growth/src/retrieval/cache.rs
@@ -19,7 +19,7 @@ struct CacheEntry {
 }

 /// Cache key for efficient lookups (reserved for future cache optimization)
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: post-release cache optimization lookups
 #[derive(Debug, Clone, Hash, Eq, PartialEq)]
 struct CacheKey {
    agent_id: String,
--- a/crates/zclaw-growth/src/skill_generator.rs
+++ b/crates/zclaw-growth/src/skill_generator.rs
@@ -0,0 +1,164 @@
+//! 技能生成器
+//! 将聚合的经验模式通过 LLM 转化为 SKILL.md 内容
+//! 提供 prompt 构建和 JSON 结果解析
+
+use crate::pattern_aggregator::AggregatedPattern;
+use zclaw_types::Result;
+
+/// 技能候选项
+#[derive(Debug, Clone)]
+pub struct SkillCandidate {
+    pub name: String,
+    pub description: String,
+    pub triggers: Vec<String>,
+    pub tools: Vec<String>,
+    pub body_markdown: String,
+    pub source_pattern: String,
+    pub confidence: f32,
+    /// 技能版本号，用于后续迭代追踪
+    pub version: u32,
+}
+
+/// LLM 驱动的技能生成 prompt
+const SKILL_GENERATION_PROMPT: &str = r#"
+你是一个技能设计专家。根据以下用户反复出现的问题和解决步骤，生成一个可复用的技能定义。
+
+问题模式：{pain_pattern}
+解决步骤：{steps}
+使用的工具：{tools}
+行业背景：{industry}
+
+请生成以下 JSON：
+```json
+{
+  "name": "技能名称（简短中文）",
+  "description": "技能描述（一段话）",
+  "triggers": ["触发词1", "触发词2", "触发词3"],
+  "tools": ["tool1", "tool2"],
+  "body_markdown": "技能的 Markdown 正文，包含步骤说明",
+  "confidence": 0.85
+}
+```
+"#;
+
+/// 技能生成器
+/// 负责 prompt 构建和 LLM 返回的 JSON 解析
+pub struct SkillGenerator;
+
+impl SkillGenerator {
+    pub fn new() -> Self {
+        Self
+    }
+
+    /// 从聚合模式构建 LLM prompt
+    pub fn build_prompt(pattern: &AggregatedPattern) -> String {
+        SKILL_GENERATION_PROMPT
+            .replace("{pain_pattern}", &pattern.pain_pattern)
+            .replace("{steps}", &pattern.common_steps.join(" → "))
+            .replace("{tools}", &pattern.tools_used.join(", "))
+            .replace("{industry}", pattern.industry_context.as_deref().unwrap_or("通用"))
+    }
+
+    /// 解析 LLM 返回的 JSON 为 SkillCandidate
+    pub fn parse_response(json_str: &str, pattern: &AggregatedPattern) -> Result<SkillCandidate> {
+        let json_str = crate::json_utils::extract_json_block(json_str);
+
+        let raw: serde_json::Value = serde_json::from_str(&json_str).map_err(|e| {
+            zclaw_types::ZclawError::ConfigError(format!("Invalid skill JSON: {}", e))
+        })?;
+
+        Ok(SkillCandidate {
+            name: raw["name"]
+                .as_str()
+                .unwrap_or("未命名技能")
+                .to_string(),
+            description: raw["description"].as_str().unwrap_or("").to_string(),
+            triggers: crate::json_utils::extract_string_array(&raw, "triggers"),
+            tools: crate::json_utils::extract_string_array(&raw, "tools"),
+            body_markdown: raw["body_markdown"].as_str().unwrap_or("").to_string(),
+            source_pattern: pattern.pain_pattern.clone(),
+            confidence: raw["confidence"].as_f64().unwrap_or(0.5) as f32,
+            version: raw["version"].as_u64().unwrap_or(1) as u32,
+        })
+    }
+}
+
+impl Default for SkillGenerator {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::experience_store::Experience;
+
+    fn make_pattern() -> AggregatedPattern {
+        let exp = Experience::new(
+            "agent-1",
+            "报表生成",
+            "researcher",
+            vec!["查询数据库".into(), "格式化输出".into()],
+            "success",
+        );
+        AggregatedPattern {
+            pain_pattern: "报表生成".to_string(),
+            experiences: vec![exp],
+            common_steps: vec!["查询数据库".into(), "格式化输出".into()],
+            total_reuse: 5,
+            tools_used: vec!["researcher".into()],
+            industry_context: Some("healthcare".into()),
+        }
+    }
+
+    #[test]
+    fn test_build_prompt() {
+        let pattern = make_pattern();
+        let prompt = SkillGenerator::build_prompt(&pattern);
+        assert!(prompt.contains("报表生成"));
+        assert!(prompt.contains("查询数据库"));
+        assert!(prompt.contains("researcher"));
+        assert!(prompt.contains("healthcare"));
+    }
+
+    #[test]
+    fn test_parse_response_valid_json() {
+        let pattern = make_pattern();
+        let json = r##"{"name":"每日报表","description":"生成每日报表","triggers":["报表","日报"],"tools":["researcher"],"body_markdown":"# 每日报表\n步骤1","confidence":0.9}"##;
+        let candidate = SkillGenerator::parse_response(json, &pattern).unwrap();
+        assert_eq!(candidate.name, "每日报表");
+        assert_eq!(candidate.triggers.len(), 2);
+        assert_eq!(candidate.confidence, 0.9);
+        assert_eq!(candidate.source_pattern, "报表生成");
+    }
+
+    #[test]
+    fn test_parse_response_json_block() {
+        let pattern = make_pattern();
+        let text = r#"```json
+{"name":"技能A","description":"desc","triggers":["a"],"tools":[],"body_markdown":"body","confidence":0.8}
+```"#;
+        let candidate = SkillGenerator::parse_response(text, &pattern).unwrap();
+        assert_eq!(candidate.name, "技能A");
+    }
+
+    #[test]
+    fn test_parse_response_invalid_json() {
+        let pattern = make_pattern();
+        let result = SkillGenerator::parse_response("not json at all", &pattern);
+        assert!(result.is_err());
+    }
+
+    #[test]
+    fn test_extract_json_block_with_markdown() {
+        let text = "Here is the result:\n```json\n{\"key\": \"value\"}\n```\nDone.";
+        assert_eq!(crate::json_utils::extract_json_block(text), "{\"key\": \"value\"}");
+    }
+
+    #[test]
+    fn test_extract_json_block_bare() {
+        let text = "{\"key\": \"value\"}";
+        assert_eq!(crate::json_utils::extract_json_block(text), "{\"key\": \"value\"}");
+    }
+}
--- a/crates/zclaw-growth/src/storage/sqlite.rs
+++ b/crates/zclaw-growth/src/storage/sqlite.rs
@@ -22,7 +22,7 @@ pub struct SqliteStorage {
    /// Semantic scorer for similarity computation
    scorer: Arc<RwLock<SemanticScorer>>,
    /// Database path (for reference)
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: db path for diagnostics and reconnect
    path: PathBuf,
 }

@@ -162,22 +162,44 @@ impl SqliteStorage {
            .map_err(|e| ZclawError::StorageError(format!("Failed to create importance index: {}", e)))?;

        // Migration: add overview column (L1 summary)
-        let _ = sqlx::query("ALTER TABLE memories ADD COLUMN overview TEXT")
+        // SQLite ALTER TABLE ADD COLUMN fails with "duplicate column name" if already applied
+        if let Err(e) = sqlx::query("ALTER TABLE memories ADD COLUMN overview TEXT")
            .execute(&self.pool)
-            .await;
+            .await
+        {
+            let msg = e.to_string();
+            if !msg.contains("duplicate column name") {
+                tracing::warn!("[Growth] Migration overview failed: {}", msg);
+            }
+        }

        // Migration: add abstract_summary column (L0 keywords)
-        let _ = sqlx::query("ALTER TABLE memories ADD COLUMN abstract_summary TEXT")
+        if let Err(e) = sqlx::query("ALTER TABLE memories ADD COLUMN abstract_summary TEXT")
            .execute(&self.pool)
-            .await;
+            .await
+        {
+            let msg = e.to_string();
+            if !msg.contains("duplicate column name") {
+                tracing::warn!("[Growth] Migration abstract_summary failed: {}", msg);
+            }
+        }

        // P2-24: Migration — content fingerprint for deduplication
-        let _ = sqlx::query("ALTER TABLE memories ADD COLUMN content_hash TEXT")
+        if let Err(e) = sqlx::query("ALTER TABLE memories ADD COLUMN content_hash TEXT")
            .execute(&self.pool)
-            .await;
-        let _ = sqlx::query("CREATE INDEX IF NOT EXISTS idx_content_hash ON memories(content_hash)")
+            .await
+        {
+            let msg = e.to_string();
+            if !msg.contains("duplicate column name") {
+                tracing::warn!("[Growth] Migration content_hash failed: {}", msg);
+            }
+        }
+        if let Err(e) = sqlx::query("CREATE INDEX IF NOT EXISTS idx_content_hash ON memories(content_hash)")
            .execute(&self.pool)
-            .await;
+            .await
+        {
+            tracing::warn!("[Growth] Migration idx_content_hash failed: {}", e);
+        }

        // Backfill content_hash for existing entries that have NULL content_hash
        {
@@ -196,11 +218,14 @@ impl SqliteStorage {
                    let mut hasher = std::collections::hash_map::DefaultHasher::new();
                    normalized.hash(&mut hasher);
                    let hash = format!("{:016x}", hasher.finish());
-                    let _ = sqlx::query("UPDATE memories SET content_hash = ? WHERE uri = ?")
+                    if let Err(e) = sqlx::query("UPDATE memories SET content_hash = ? WHERE uri = ?")
                        .bind(&hash)
                        .bind(uri)
                        .execute(&self.pool)
-                        .await;
+                        .await
+                    {
+                        tracing::warn!("[sqlite] content_hash update failed for {}: {}", uri, e);
+                    }
                }
                tracing::info!(
                    "[SqliteStorage] Backfilled content_hash for {} existing entries",
@@ -234,9 +259,12 @@ impl SqliteStorage {
        if needs_rebuild {
            tracing::info!("[SqliteStorage] Rebuilding FTS5 table: unicode61 → trigram for CJK support");
            // Drop old FTS5 table
-            let _ = sqlx::query("DROP TABLE IF EXISTS memories_fts")
+            if let Err(e) = sqlx::query("DROP TABLE IF EXISTS memories_fts")
                .execute(&self.pool)
-                .await;
+                .await
+            {
+                tracing::warn!("[sqlite] FTS5 table drop failed during rebuild: {}", e);
+            }
            // Recreate with trigram tokenizer
            sqlx::query(
                r#"
@@ -401,14 +429,17 @@ impl SqliteStorage {
        .await;

        // Also clean up FTS entries for archived memories
-        let _ = sqlx::query(
+        if let Err(e) = sqlx::query(
            r#"
            DELETE FROM memories_fts
            WHERE uri NOT IN (SELECT uri FROM memories)
            "#,
        )
        .execute(&self.pool)
-        .await;
+        .await
+        {
+            tracing::warn!("[sqlite] FTS cleanup after archive failed: {}", e);
+        }

        let archived = archive_result
            .map(|r| r.rows_affected())
@@ -461,13 +492,58 @@ impl SqliteStorage {
        });

        if has_cjk {
-            // For CJK, use the full query as a quoted phrase for substring matching
-            // trigram will match any 3-char subsequence
-            if lower.len() >= 3 {
-                format!("\"{}\"", lower)
-            } else {
-                String::new()
+            // For CJK queries, extract tokens: CJK character sequences and ASCII words.
+            // Join with OR for broad matching (not exact phrase, which would miss scattered terms).
+            let mut tokens: Vec<String> = Vec::new();
+            let mut cjk_buf = String::new();
+            let mut ascii_buf = String::new();
+
+            for ch in lower.chars() {
+                let is_cjk = matches!(ch, '\u{4E00}'..='\u{9FFF}' | '\u{3400}'..='\u{4DBF}' | '\u{F900}'..='\u{FAFF}');
+                if is_cjk {
+                    if !ascii_buf.is_empty() {
+                        if ascii_buf.len() >= 2 {
+                            tokens.push(format!("\"{}\"", ascii_buf));
+                        }
+                        ascii_buf.clear();
+                    }
+                    cjk_buf.push(ch);
+                } else if ch.is_alphanumeric() {
+                    if !cjk_buf.is_empty() {
+                        // Flush CJK buffer — each CJK character is a potential token
+                        // (trigram indexes 3-char sequences, so single CJK chars won't
+                        // match alone, but 2+ char sequences will)
+                        if cjk_buf.len() >= 2 {
+                            tokens.push(format!("\"{}\"", cjk_buf));
+                        }
+                        cjk_buf.clear();
+                    }
+                    ascii_buf.push(ch);
+                } else {
+                    // Separator — flush both buffers
+                    if cjk_buf.len() >= 2 {
+                        tokens.push(format!("\"{}\"", cjk_buf));
+                    }
+                    cjk_buf.clear();
+                    if ascii_buf.len() >= 2 {
+                        tokens.push(format!("\"{}\"", ascii_buf));
+                    }
+                    ascii_buf.clear();
+                }
            }
+            // Flush remaining
+            if cjk_buf.len() >= 2 {
+                tokens.push(format!("\"{}\"", cjk_buf));
+            }
+            if ascii_buf.len() >= 2 {
+                tokens.push(format!("\"{}\"", ascii_buf));
+            }
+
+            if tokens.is_empty() {
+                return String::new();
+            }
+
+            tokens.join(" OR ")
        } else {
            // For non-CJK, split into terms and join with OR
            let terms: Vec<String> = lower
--- a/crates/zclaw-growth/src/tracker.rs
+++ b/crates/zclaw-growth/src/tracker.rs
@@ -66,21 +66,30 @@ impl GrowthTracker {
            timestamp: Utc::now(),
        };

-        // Store learning event
-        self.viking
-            .store_metadata(
-                &format!("agent://{}/events/{}", agent_id, session_id),
-                &event,
-            )
-            .await?;
+        // Store learning event as MemoryEntry so get_timeline can find it via find_by_prefix
+        let event_uri = format!("agent://{}/events/{}", agent_id, session_id);
+        let content = serde_json::to_string(&event)?;
+        let entry = crate::types::MemoryEntry {
+            uri: event_uri,
+            memory_type: MemoryType::Session,
+            content,
+            keywords: vec![agent_id.to_string(), session_id.to_string()],
+            importance: 5,
+            access_count: 0,
+            created_at: event.timestamp,
+            last_accessed: event.timestamp,
+            overview: None,
+            abstract_summary: None,
+        };
+        self.viking.store(&entry).await?;

-        // Update last learning time
+        // Update last learning time via metadata
        self.viking
            .store_metadata(
                &format!("agent://{}", agent_id),
                &AgentMetadata {
                    last_learning_time: Some(Utc::now()),
-                    total_learning_events: None, // Will be computed
+                    total_learning_events: None,
                },
            )
            .await?;
--- a/crates/zclaw-growth/src/types.rs
+++ b/crates/zclaw-growth/src/types.rs
@@ -394,6 +394,103 @@ pub struct DecayResult {
    pub archived: u64,
 }

+// === Evolution Engine Types ===
+
+/// 经验提取结果
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ExperienceCandidate {
+    pub pain_pattern: String,
+    pub context: String,
+    pub solution_steps: Vec<String>,
+    pub outcome: Outcome,
+    pub confidence: f32,
+    pub tools_used: Vec<String>,
+    pub industry_context: Option<String>,
+}
+
+/// 结果状态
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum Outcome {
+    Success,
+    Partial,
+    Failed,
+}
+
+/// 合并提取结果（单次 LLM 调用的全部输出）
+#[derive(Debug, Clone, Default)]
+pub struct CombinedExtraction {
+    pub memories: Vec<ExtractedMemory>,
+    pub experiences: Vec<ExperienceCandidate>,
+    pub profile_signals: ProfileSignals,
+}
+
+/// 画像更新信号（从提取结果中推断，不额外调用 LLM）
+#[derive(Debug, Clone, Default)]
+pub struct ProfileSignals {
+    pub industry: Option<String>,
+    pub recent_topic: Option<String>,
+    pub pain_point: Option<String>,
+    pub preferred_tool: Option<String>,
+    pub communication_style: Option<String>,
+}
+
+impl ProfileSignals {
+    /// 是否包含至少一个有效信号
+    pub fn has_any_signal(&self) -> bool {
+        self.industry.is_some()
+            || self.recent_topic.is_some()
+            || self.pain_point.is_some()
+            || self.preferred_tool.is_some()
+            || self.communication_style.is_some()
+    }
+
+    /// 有效信号数量
+    pub fn signal_count(&self) -> usize {
+        let mut count = 0;
+        if self.industry.is_some() { count += 1; }
+        if self.recent_topic.is_some() { count += 1; }
+        if self.pain_point.is_some() { count += 1; }
+        if self.preferred_tool.is_some() { count += 1; }
+        if self.communication_style.is_some() { count += 1; }
+        count
+    }
+}
+
+/// 进化事件
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct EvolutionEvent {
+    pub id: String,
+    pub event_type: EvolutionEventType,
+    pub artifact_type: ArtifactType,
+    pub artifact_id: String,
+    pub status: EvolutionStatus,
+    pub confidence: f32,
+    pub user_feedback: Option<String>,
+    pub created_at: DateTime<Utc>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum EvolutionEventType {
+    SkillGenerated,
+    SkillOptimized,
+    WorkflowGenerated,
+    WorkflowOptimized,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum ArtifactType {
+    Skill,
+    Pipeline,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum EvolutionStatus {
+    Pending,
+    Confirmed,
+    Rejected,
+    Optimized,
+}
+
 /// Compute effective importance with time decay.
 ///
 /// Uses exponential decay: each 30-day period of non-access reduces
@@ -524,4 +621,61 @@ mod tests {
        assert!(!result.is_empty());
        assert_eq!(result.total_count(), 1);
    }
+
+    #[test]
+    fn test_experience_candidate_roundtrip() {
+        let candidate = ExperienceCandidate {
+            pain_pattern: "报表生成".to_string(),
+            context: "月度销售报表".to_string(),
+            solution_steps: vec!["查询数据库".to_string(), "格式化输出".to_string()],
+            outcome: Outcome::Success,
+            confidence: 0.85,
+            tools_used: vec!["researcher".to_string()],
+            industry_context: Some("healthcare".to_string()),
+        };
+        let json = serde_json::to_string(&candidate).unwrap();
+        let decoded: ExperienceCandidate = serde_json::from_str(&json).unwrap();
+        assert_eq!(decoded.pain_pattern, "报表生成");
+        assert_eq!(decoded.outcome, Outcome::Success);
+        assert_eq!(decoded.solution_steps.len(), 2);
+    }
+
+    #[test]
+    fn test_evolution_event_roundtrip() {
+        let event = EvolutionEvent {
+            id: uuid::Uuid::new_v4().to_string(),
+            event_type: EvolutionEventType::SkillGenerated,
+            artifact_type: ArtifactType::Skill,
+            artifact_id: "daily-report".to_string(),
+            status: EvolutionStatus::Pending,
+            confidence: 0.8,
+            user_feedback: None,
+            created_at: chrono::Utc::now(),
+        };
+        let json = serde_json::to_string(&event).unwrap();
+        let decoded: EvolutionEvent = serde_json::from_str(&json).unwrap();
+        assert_eq!(decoded.event_type, EvolutionEventType::SkillGenerated);
+        assert_eq!(decoded.status, EvolutionStatus::Pending);
+    }
+
+    #[test]
+    fn test_combined_extraction_default() {
+        let combined = CombinedExtraction::default();
+        assert!(combined.memories.is_empty());
+        assert!(combined.experiences.is_empty());
+        assert!(combined.profile_signals.industry.is_none());
+    }
+
+    #[test]
+    fn test_profile_signals() {
+        let signals = ProfileSignals {
+            industry: Some("healthcare".to_string()),
+            recent_topic: Some("报表".to_string()),
+            pain_point: None,
+            preferred_tool: Some("researcher".to_string()),
+            communication_style: Some("concise".to_string()),
+        };
+        assert_eq!(signals.industry.as_deref(), Some("healthcare"));
+        assert!(signals.pain_point.is_none());
+    }
 }
--- a/crates/zclaw-growth/src/workflow_composer.rs
+++ b/crates/zclaw-growth/src/workflow_composer.rs
@@ -0,0 +1,180 @@
+//! 工作流组装器（L3 工作流进化）
+//! 从轨迹数据中分析重复的工具链模式，自动组装 Pipeline YAML
+//! 触发条件：CompressedTrajectory 中出现 2 次以上相同工具链序列
+
+use zclaw_types::Result;
+
+/// Pipeline 候选项
+#[derive(Debug, Clone)]
+pub struct PipelineCandidate {
+    pub name: String,
+    pub description: String,
+    pub triggers: Vec<String>,
+    pub yaml_content: String,
+    pub source_sessions: Vec<String>,
+    pub confidence: f32,
+}
+
+/// 工具链模式（用于聚类分析）
+#[derive(Debug, Clone, Hash, PartialEq, Eq)]
+pub struct ToolChainPattern {
+    pub steps: Vec<String>,
+}
+
+/// 工作流组装 prompt
+const WORKFLOW_GENERATION_PROMPT: &str = r#"
+你是一个工作流设计专家。根据以下用户反复执行的工具链序列，设计一个可复用的 Pipeline 工作流。
+
+工具链序列：{tool_chain}
+执行频率：{frequency} 次
+行业背景：{industry}
+
+请生成以下 JSON：
+```json
+{
+  "name": "工作流名称（简短中文）",
+  "description": "工作流描述",
+  "triggers": ["触发词1", "触发词2"],
+  "yaml_content": "Pipeline YAML 内容",
+  "confidence": 0.8
+}
+```
+"#;
+
+/// 工作流组装器
+/// 分析压缩轨迹中的工具链模式，通过 LLM 生成 Pipeline YAML
+pub struct WorkflowComposer;
+
+impl WorkflowComposer {
+    pub fn new() -> Self {
+        Self
+    }
+
+    /// 从压缩轨迹的工具链中提取模式
+    /// 简单的精确匹配聚类：相同工具链序列视为同一模式
+    pub fn extract_patterns(
+        trajectories: &[(String, Vec<String>)], // (session_id, tools_used)
+    ) -> Vec<(ToolChainPattern, Vec<String>)> {
+        use std::collections::HashMap;
+
+        let mut groups: HashMap<ToolChainPattern, Vec<String>> = HashMap::new();
+        for (session_id, tools) in trajectories {
+            if tools.len() < 2 {
+                continue; // 单步操作不构成工作流
+            }
+            let pattern = ToolChainPattern {
+                steps: tools.clone(),
+            };
+            groups.entry(pattern).or_default().push(session_id.clone());
+        }
+
+        // 过滤出现 2 次以上的模式
+        groups
+            .into_iter()
+            .filter(|(_, sessions)| sessions.len() >= 2)
+            .collect()
+    }
+
+    /// 构建 LLM prompt
+    pub fn build_prompt(
+        pattern: &ToolChainPattern,
+        frequency: usize,
+        industry: Option<&str>,
+    ) -> String {
+        WORKFLOW_GENERATION_PROMPT
+            .replace("{tool_chain}", &pattern.steps.join(" → "))
+            .replace("{frequency}", &frequency.to_string())
+            .replace("{industry}", industry.unwrap_or("通用"))
+    }
+
+    /// 解析 LLM 返回的 JSON 为 PipelineCandidate
+    pub fn parse_response(
+        json_str: &str,
+        _pattern: &ToolChainPattern,
+        source_sessions: Vec<String>,
+    ) -> Result<PipelineCandidate> {
+        let json_str = crate::json_utils::extract_json_block(json_str);
+        let raw: serde_json::Value = serde_json::from_str(&json_str).map_err(|e| {
+            zclaw_types::ZclawError::ConfigError(format!("Invalid pipeline JSON: {}", e))
+        })?;
+
+        Ok(PipelineCandidate {
+            name: raw["name"].as_str().unwrap_or("未命名工作流").to_string(),
+            description: raw["description"].as_str().unwrap_or("").to_string(),
+            triggers: crate::json_utils::extract_string_array(&raw, "triggers"),
+            yaml_content: raw["yaml_content"].as_str().unwrap_or("").to_string(),
+            source_sessions,
+            confidence: raw["confidence"].as_f64().unwrap_or(0.5) as f32,
+        })
+    }
+}
+
+impl Default for WorkflowComposer {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_extract_patterns_filters_single_step() {
+        let trajectories = vec![
+            ("s1".to_string(), vec!["researcher".to_string()]),
+        ];
+        let patterns = WorkflowComposer::extract_patterns(&trajectories);
+        assert!(patterns.is_empty());
+    }
+
+    #[test]
+    fn test_extract_patterns_groups_identical_chains() {
+        let trajectories = vec![
+            ("s1".to_string(), vec!["researcher".into(), "collector".into()]),
+            ("s2".to_string(), vec!["researcher".into(), "collector".into()]),
+            ("s3".to_string(), vec!["browser".into()]), // 单步，过滤
+        ];
+        let patterns = WorkflowComposer::extract_patterns(&trajectories);
+        assert_eq!(patterns.len(), 1);
+        assert_eq!(patterns[0].1.len(), 2); // 2 sessions
+    }
+
+    #[test]
+    fn test_extract_patterns_requires_min_2() {
+        let trajectories = vec![
+            ("s1".to_string(), vec!["a".into(), "b".into()]),
+        ];
+        let patterns = WorkflowComposer::extract_patterns(&trajectories);
+        assert!(patterns.is_empty()); // 只出现 1 次
+    }
+
+    #[test]
+    fn test_build_prompt() {
+        let pattern = ToolChainPattern {
+            steps: vec!["researcher".into(), "collector".into(), "summarize".into()],
+        };
+        let prompt = WorkflowComposer::build_prompt(&pattern, 3, Some("healthcare"));
+        assert!(prompt.contains("researcher"));
+        assert!(prompt.contains("3"));
+        assert!(prompt.contains("healthcare"));
+    }
+
+    #[test]
+    fn test_parse_response() {
+        let pattern = ToolChainPattern {
+            steps: vec!["researcher".into()],
+        };
+        let json = r##"{"name":"每日简报","description":"搜索+汇总","triggers":["简报","日报"],"yaml_content":"steps: []","confidence":0.85}"##;
+        let candidate = WorkflowComposer::parse_response(
+            json,
+            &pattern,
+            vec!["s1".into(), "s2".into()],
+        )
+        .unwrap();
+        assert_eq!(candidate.name, "每日简报");
+        assert_eq!(candidate.triggers.len(), 2);
+        assert_eq!(candidate.source_sessions.len(), 2);
+        assert!((candidate.confidence - 0.85).abs() < 0.01);
+    }
+}
--- a/crates/zclaw-hands/src/hands/clip.rs
+++ b/crates/zclaw-hands/src/hands/clip.rs
@@ -459,7 +459,7 @@ impl ClipHand {
        let args = vec![
            "-f", "concat",
            "-safe", "0",
-            "-i", temp_file.to_str().unwrap(),
+            "-i", temp_file.to_str().ok_or_else(|| zclaw_types::ZclawError::HandError("Temp file path is not valid UTF-8".to_string()))?,
            "-c", "copy",
            &config.output_path,
        ];
--- a/crates/zclaw-kernel/Cargo.toml
+++ b/crates/zclaw-kernel/Cargo.toml
@@ -9,8 +9,6 @@ description = "ZCLAW kernel - central coordinator for all subsystems"

 [features]
 default = []
-# Enable multi-agent orchestration (Director, A2A protocol)
-multi-agent = ["zclaw-protocols/a2a"]

 [dependencies]
 zclaw-types = { workspace = true }
--- a/crates/zclaw-kernel/src/config.rs
+++ b/crates/zclaw-kernel/src/config.rs
@@ -30,7 +30,7 @@ impl Default for ApiProtocol {
 ///
 /// This is the single source of truth for LLM configuration.
 /// Model ID is passed directly to the API without any transformation.
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 pub struct LlmConfig {
    /// API base URL (e.g., "https://api.openai.com/v1")
    pub base_url: String,
@@ -61,6 +61,20 @@ pub struct LlmConfig {
    pub context_window: u32,
 }

+impl std::fmt::Debug for LlmConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("LlmConfig")
+            .field("base_url", &self.base_url)
+            .field("api_key", &"***REDACTED***")
+            .field("model", &self.model)
+            .field("api_protocol", &self.api_protocol)
+            .field("max_tokens", &self.max_tokens)
+            .field("temperature", &self.temperature)
+            .field("context_window", &self.context_window)
+            .finish()
+    }
+}
+
 impl LlmConfig {
    /// Create a new LLM config
    pub fn new(base_url: impl Into<String>, api_key: impl Into<String>, model: impl Into<String>) -> Self {
--- a/crates/zclaw-kernel/src/director.rs
+++ b/crates/zclaw-kernel/src/director.rs
@@ -12,7 +12,7 @@

 use std::sync::Arc;
 use serde::{Deserialize, Serialize};
-use tokio::sync::{RwLock, Mutex, mpsc};
+use tokio::sync::{RwLock, Mutex, mpsc, oneshot};
 use zclaw_types::{AgentId, Result, ZclawError};
 use zclaw_protocols::{A2aEnvelope, A2aMessageType, A2aRecipient, A2aRouter, A2aAgentProfile, A2aCapability};
 use zclaw_runtime::{LlmDriver, CompletionRequest};
@@ -199,9 +199,9 @@ pub struct Director {
    director_id: AgentId,
    /// Optional LLM driver for intelligent scheduling
    llm_driver: Option<Arc<dyn LlmDriver>>,
-    /// Inbox for receiving responses (stores pending request IDs and their response channels)
-    pending_requests: Arc<Mutex<std::collections::HashMap<String, mpsc::Sender<A2aEnvelope>>>>,
-    /// Receiver for incoming messages
+    /// Pending request response channels (request_id → oneshot sender)
+    pending_requests: Arc<Mutex<std::collections::HashMap<String, oneshot::Sender<A2aEnvelope>>>>,
+    /// Receiver for incoming messages (consumed by inbox reader task)
    inbox: Arc<Mutex<Option<mpsc::Receiver<A2aEnvelope>>>>,
 }

@@ -360,7 +360,7 @@ impl Director {
                use std::time::{SystemTime, UNIX_EPOCH};
                let now = SystemTime::now()
                    .duration_since(UNIX_EPOCH)
-                    .unwrap()
+                    .expect("system clock is valid")
                    .as_nanos();
                let idx = (now as usize) % agents.len();
                Some(agents[idx].clone())
@@ -481,13 +481,16 @@ Respond with ONLY the number (1-{}) of the agent who should speak next. No expla
    }

    /// Send message to selected agent and wait for response
+    ///
+    /// Uses oneshot channels to avoid deadlock: each call creates its own
+    /// response channel, and a shared inbox reader dispatches responses.
    pub async fn send_to_agent(
        &self,
        agent: &DirectorAgent,
        message: String,
    ) -> Result<String> {
-        // Create a response channel for this request
-        let (_response_tx, mut _response_rx) = mpsc::channel::<A2aEnvelope>(1);
+        // Create a oneshot channel for this specific request's response
+        let (response_tx, response_rx) = oneshot::channel::<A2aEnvelope>();

        let envelope = A2aEnvelope::new(
            self.director_id.clone(),
@@ -500,50 +503,32 @@ Respond with ONLY the number (1-{}) of the agent who should speak next. No expla
            }),
        );

-        // Store the request ID with its response channel
+        // Store the oneshot sender so the inbox reader can dispatch to it
        let request_id = envelope.id.clone();
        {
            let mut pending = self.pending_requests.lock().await;
-            pending.insert(request_id.clone(), _response_tx);
+            pending.insert(request_id.clone(), response_tx);
        }

        // Send the request
        self.router.route(envelope).await?;

-        // Wait for response with timeout
+        // Ensure the inbox reader is running
+        self.ensure_inbox_reader().await;
+
+        // Wait for response on our dedicated oneshot channel with timeout
        let timeout_duration = std::time::Duration::from_secs(self.config.response_timeout);
-        let request_id_clone = request_id.clone();

-        let response = tokio::time::timeout(timeout_duration, async {
-            // Poll the inbox for responses
-            let mut inbox_guard = self.inbox.lock().await;
-            if let Some(ref mut rx) = *inbox_guard {
-                while let Some(msg) = rx.recv().await {
-                    // Check if this is a response to our request
-                    if msg.message_type == A2aMessageType::Response {
-                        if let Some(ref reply_to) = msg.reply_to {
-                            if reply_to == &request_id_clone {
-                                // Found our response
-                                return Some(msg);
-                            }
-                        }
-                    }
-                    // Not our response, continue waiting
-                    // (In a real implementation, we'd re-queue non-matching messages)
-                }
-            }
-            None
-        }).await;
+        let response = tokio::time::timeout(timeout_duration, response_rx).await;

-        // Clean up pending request
+        // Clean up pending request (sender already consumed on success)
        {
            let mut pending = self.pending_requests.lock().await;
            pending.remove(&request_id);
        }

        match response {
-            Ok(Some(envelope)) => {
-                // Extract response text from payload
+            Ok(Ok(envelope)) => {
                let response_text = envelope.payload
                    .get("response")
                    .and_then(|v: &serde_json::Value| v.as_str())
@@ -551,7 +536,7 @@ Respond with ONLY the number (1-{}) of the agent who should speak next. No expla
                    .to_string();
                Ok(response_text)
            }
-            Ok(None) => {
+            Ok(Err(_)) => {
                Err(ZclawError::Timeout("No response received".into()))
            }
            Err(_) => {
@@ -563,6 +548,47 @@ Respond with ONLY the number (1-{}) of the agent who should speak next. No expla
        }
    }

+    /// Ensure the inbox reader task is running.
+    /// The inbox reader continuously reads from the shared inbox channel
+    /// and dispatches each response to the correct oneshot sender.
+    async fn ensure_inbox_reader(&self) {
+        // Quick check: if inbox has already been taken, reader is running
+        {
+            let inbox = self.inbox.lock().await;
+            if inbox.is_none() {
+                return; // Reader already spawned and consumed the receiver
+            }
+        }
+
+        // Take the receiver out (only once)
+        let rx = {
+            let mut inbox = self.inbox.lock().await;
+            inbox.take()
+        };
+
+        if let Some(mut rx) = rx {
+            let pending = self.pending_requests.clone();
+            tokio::spawn(async move {
+                while let Some(msg) = rx.recv().await {
+                    // Find and dispatch to the correct oneshot sender
+                    if msg.message_type == A2aMessageType::Response {
+                        if let Some(ref reply_to) = msg.reply_to {
+                            let reply_to_clone = reply_to.clone();
+                            let mut pending_guard = pending.lock().await;
+                            if let Some(sender) = pending_guard.remove(reply_to) {
+                                // Send the response; if receiver already dropped, request was cancelled
+                                if sender.send(msg).is_err() {
+                                    tracing::debug!("[Director] Response dropped: receiver cancelled for reply_to={}", reply_to_clone);
+                                }
+                            }
+                        }
+                    }
+                    // Non-response messages are dropped (notifications, etc.)
+                }
+            });
+        }
+    }
+
    /// Broadcast message to all agents
    pub async fn broadcast(&self, message: String) -> Result<()> {
        let envelope = A2aEnvelope::new(
@@ -616,7 +642,9 @@ Respond with ONLY the number (1-{}) of the agent who should speak next. No expla
        }

        if let Some(ref user_input) = input {
-            context.push_str(&format!("User: {}\n\n", user_input));
+            context.push_str("<user_input>\n");
+            context.push_str(&format!("{}\n", user_input));
+            context.push_str("</user_input>\n\n");
        }

        // Add recent history
@@ -882,7 +910,9 @@ impl Director {
        let prompt = format!(
            r#"你是 ZCLAW 管家。请将以下用户需求拆解为 1-5 个具体子任务。

-用户需求：{}
+<user_request>
+{}
+</user_request>

 请按 JSON 数组格式输出，每个元素包含：
 - description: 子任务描述（中文）
--- a/crates/zclaw-kernel/src/events.rs
+++ b/crates/zclaw-kernel/src/events.rs
@@ -17,8 +17,9 @@ impl EventBus {

    /// Publish an event
    pub fn publish(&self, event: Event) {
-        // Ignore send errors (no subscribers)
-        let _ = self.sender.send(event);
+        if let Err(e) = self.sender.send(event) {
+            tracing::debug!("Event dropped (no subscribers or channel full): {:?}", e);
+        }
    }

    /// Subscribe to events
--- a/crates/zclaw-kernel/src/export/html.rs
+++ b/crates/zclaw-kernel/src/export/html.rs
@@ -14,7 +14,7 @@ use zclaw_types::Result;
 /// HTML exporter
 pub struct HtmlExporter {
    /// Template name (reserved for future template support)
-    #[allow(dead_code)] // TODO: Implement template-based HTML export
+    #[allow(dead_code)] // @reserved: post-release template-based HTML export
    template: String,
 }

--- a/crates/zclaw-kernel/src/export/pptx.rs
+++ b/crates/zclaw-kernel/src/export/pptx.rs
@@ -490,7 +490,7 @@ impl PptxExporter {
            paths.sort();

            for path in paths {
-                let content = files.get(path).unwrap();
+                let content = files.get(path).expect("path comes from files.keys(), must exist");
                let options = SimpleFileOptions::default()
                    .compression_method(zip::CompressionMethod::Deflated);

--- a/crates/zclaw-kernel/src/generation/chat.rs
+++ b/crates/zclaw-kernel/src/generation/chat.rs
@@ -243,7 +243,7 @@ fn clean_fallback_response(text: &str) -> String {
 fn current_timestamp_millis() -> i64 {
    std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
-        .unwrap()
+        .expect("system clock is valid")
        .as_millis() as i64
 }

--- a/crates/zclaw-kernel/src/generation/mod.rs
+++ b/crates/zclaw-kernel/src/generation/mod.rs
@@ -557,7 +557,7 @@ Use Chinese if the topic is in Chinese. Include metaphors that relate to everyda
            .join("\n")
    }

-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: instance-method convenience wrapper for static helper
    fn extract_text_from_response(&self, response: &CompletionResponse) -> String {
        Self::extract_text_from_response_static(response)
    }
@@ -882,7 +882,7 @@ fn current_timestamp() -> i64 {
    use std::time::{SystemTime, UNIX_EPOCH};
    SystemTime::now()
        .duration_since(UNIX_EPOCH)
-        .unwrap()
+        .expect("system clock is valid")
        .as_millis() as i64
 }

--- a/crates/zclaw-kernel/src/kernel/a2a.rs
+++ b/crates/zclaw-kernel/src/kernel/a2a.rs
@@ -1,16 +1,10 @@
 //! A2A (Agent-to-Agent) messaging
-//!
-//! All items in this module are gated by the `multi-agent` feature flag.

-#[cfg(feature = "multi-agent")]
 use zclaw_types::{AgentId, Capability, Event, Result};
-#[cfg(feature = "multi-agent")]
 use zclaw_protocols::{A2aAgentProfile, A2aCapability, A2aEnvelope, A2aMessageType, A2aRecipient};

-#[cfg(feature = "multi-agent")]
 use super::Kernel;

-#[cfg(feature = "multi-agent")]
 impl Kernel {
    // ============================================================
    // A2A (Agent-to-Agent) Messaging
--- a/crates/zclaw-kernel/src/kernel/adapters.rs
+++ b/crates/zclaw-kernel/src/kernel/adapters.rs
@@ -106,13 +106,11 @@ impl SkillExecutor for KernelSkillExecutor {

 /// Inbox wrapper for A2A message receivers that supports re-queuing
 /// non-matching messages instead of dropping them.
-#[cfg(feature = "multi-agent")]
 pub(crate) struct AgentInbox {
    pub(crate) rx: tokio::sync::mpsc::Receiver<zclaw_protocols::A2aEnvelope>,
    pub(crate) pending: std::collections::VecDeque<zclaw_protocols::A2aEnvelope>,
 }

-#[cfg(feature = "multi-agent")]
 impl AgentInbox {
    pub(crate) fn new(rx: tokio::sync::mpsc::Receiver<zclaw_protocols::A2aEnvelope>) -> Self {
        Self { rx, pending: std::collections::VecDeque::new() }
--- a/crates/zclaw-kernel/src/kernel/agents.rs
+++ b/crates/zclaw-kernel/src/kernel/agents.rs
@@ -2,11 +2,8 @@

 use zclaw_types::{AgentConfig, AgentId, AgentInfo, Event, Result};

-#[cfg(feature = "multi-agent")]
 use std::sync::Arc;
-#[cfg(feature = "multi-agent")]
 use tokio::sync::Mutex;
-#[cfg(feature = "multi-agent")]
 use super::adapters::AgentInbox;

 use super::Kernel;
@@ -23,7 +20,6 @@ impl Kernel {
        self.memory.save_agent(&config).await?;

        // Register with A2A router for multi-agent messaging (before config is moved)
-        #[cfg(feature = "multi-agent")]
        {
            let profile = Self::agent_config_to_a2a_profile(&config);
            let rx = self.a2a_router.register_agent(profile).await;
@@ -52,7 +48,6 @@ impl Kernel {
        self.memory.delete_agent(id).await?;

        // Unregister from A2A router
-        #[cfg(feature = "multi-agent")]
        {
            self.a2a_router.unregister_agent(id).await;
            self.a2a_inboxes.remove(id);
--- a/crates/zclaw-kernel/src/kernel/approvals.rs
+++ b/crates/zclaw-kernel/src/kernel/approvals.rs
@@ -85,14 +85,14 @@ impl Kernel {
                    started_at: None,
                    completed_at: None,
                };
-                let _ = memory.save_hand_run(&run).await.map_err(|e| {
-                    tracing::warn!("[Approval] Failed to save hand run: {}", e);
-                });
+                if let Err(e) = memory.save_hand_run(&run).await {
+                    tracing::error!("[Approval] Failed to save hand run: {}", e);
+                }
                run.status = HandRunStatus::Running;
                run.started_at = Some(chrono::Utc::now().to_rfc3339());
-                let _ = memory.update_hand_run(&run).await.map_err(|e| {
-                    tracing::warn!("[Approval] Failed to update hand run (running): {}", e);
-                });
+                if let Err(e) = memory.update_hand_run(&run).await {
+                    tracing::error!("[Approval] Failed to update hand run (running): {}", e);
+                }

                // Register cancellation flag
                let cancel_flag = Arc::new(std::sync::atomic::AtomicBool::new(false));
@@ -121,9 +121,9 @@ impl Kernel {
                }
                run.duration_ms = Some(duration.as_millis() as u64);
                run.completed_at = Some(completed_at);
-                let _ = memory.update_hand_run(&run).await.map_err(|e| {
-                    tracing::warn!("[Approval] Failed to update hand run (completed): {}", e);
-                });
+                if let Err(e) = memory.update_hand_run(&run).await {
+                    tracing::error!("[Approval] Failed to update hand run (completed): {}", e);
+                }

                // Update approval status based on execution result
                let mut approvals = approvals.lock().await;
--- a/crates/zclaw-kernel/src/kernel/mod.rs
+++ b/crates/zclaw-kernel/src/kernel/mod.rs
@@ -8,16 +8,13 @@ mod hands;
 mod triggers;
 mod approvals;
 mod orchestration;
-#[cfg(feature = "multi-agent")]
 mod a2a;

 use std::sync::Arc;
 use tokio::sync::{broadcast, Mutex};
 use zclaw_types::{Event, Result, AgentState};

-#[cfg(feature = "multi-agent")]
 use zclaw_types::AgentId;
-#[cfg(feature = "multi-agent")]
 use zclaw_protocols::A2aRouter;

 use crate::registry::AgentRegistry;
@@ -56,11 +53,9 @@ pub struct Kernel {
    mcp_adapters: Arc<std::sync::RwLock<Vec<zclaw_protocols::McpToolAdapter>>>,
    /// Dynamic industry keyword configs — shared with Tauri frontend, loaded from SaaS
    industry_keywords: Arc<tokio::sync::RwLock<Vec<zclaw_runtime::IndustryKeywordConfig>>>,
-    /// A2A router for inter-agent messaging (gated by multi-agent feature)
-    #[cfg(feature = "multi-agent")]
+    /// A2A router for inter-agent messaging
    a2a_router: Arc<A2aRouter>,
    /// Per-agent A2A inbox receivers (supports re-queuing non-matching messages)
-    #[cfg(feature = "multi-agent")]
    a2a_inboxes: Arc<dashmap::DashMap<AgentId, Arc<Mutex<adapters::AgentInbox>>>>,
 }

@@ -135,7 +130,6 @@ impl Kernel {
        }

        // Initialize A2A router for multi-agent support
-        #[cfg(feature = "multi-agent")]
        let a2a_router = {
            let kernel_agent_id = AgentId::new();
            Arc::new(A2aRouter::new(kernel_agent_id))
@@ -159,9 +153,7 @@ impl Kernel {
            extraction_driver: None,
            mcp_adapters: Arc::new(std::sync::RwLock::new(Vec::new())),
            industry_keywords: Arc::new(tokio::sync::RwLock::new(Vec::new())),
-            #[cfg(feature = "multi-agent")]
            a2a_router,
-            #[cfg(feature = "multi-agent")]
            a2a_inboxes: Arc::new(dashmap::DashMap::new()),
        })
    }
@@ -247,6 +239,9 @@ impl Kernel {
        }

        // Data masking middleware — mask sensitive entities before any other processing
+        // NOTE: Registration order does NOT determine execution order.
+        // The chain sorts by priority() ascending before execution.
+        // Execution order: Evolution(78) → ButlerRouter(80) → DataMasking(90) → ...
        {
            use std::sync::Arc;
            let masker = Arc::new(zclaw_runtime::middleware::data_masking::DataMasker::new());
@@ -260,6 +255,13 @@ impl Kernel {
            growth = growth.with_llm_driver(driver.clone());
        }

+        // Evolution middleware — pushes evolution candidate skills into system prompt
+        // priority=78, executed first by chain (before ButlerRouter@80)
+        let evolution_mw = std::sync::Arc::new(
+            zclaw_runtime::middleware::evolution::EvolutionMiddleware::new()
+        );
+        chain.register(evolution_mw.clone());
+
        // Compaction middleware — only register when threshold > 0
        let threshold = self.config.compaction_threshold();
        if threshold > 0 {
@@ -277,10 +279,11 @@ impl Kernel {
            chain.register(Arc::new(mw));
        }

-        // Memory middleware — auto-extract memories after conversations
+        // Memory middleware — auto-extract memories + check evolution after conversations
        {
            use std::sync::Arc;
-            let mw = zclaw_runtime::middleware::memory::MemoryMiddleware::new(growth);
+            let mw = zclaw_runtime::middleware::memory::MemoryMiddleware::new(growth)
+                .with_evolution(evolution_mw);
            chain.register(Arc::new(mw));
        }

--- a/crates/zclaw-kernel/src/lib.rs
+++ b/crates/zclaw-kernel/src/lib.rs
@@ -10,7 +10,6 @@ pub mod trigger_manager;
 pub mod config;
 pub mod scheduler;
 pub mod skill_router;
-#[cfg(feature = "multi-agent")]
 pub mod director;
 pub mod generation;
 pub mod export;
@@ -21,13 +20,11 @@ pub use capabilities::*;
 pub use events::*;
 pub use config::*;
 pub use trigger_manager::{TriggerManager, TriggerEntry, TriggerUpdateRequest, TriggerManagerConfig};
-#[cfg(feature = "multi-agent")]
 pub use director::{
    Director, DirectorConfig, DirectorBuilder, DirectorAgent,
    ConversationState, ScheduleStrategy,
    // Note: AgentRole is intentionally NOT re-exported here — use generation::AgentRole instead
 };
-#[cfg(feature = "multi-agent")]
 pub use zclaw_protocols::{
    A2aRouter, A2aAgentProfile, A2aCapability, A2aEnvelope, A2aMessageType, A2aRecipient,
 A2aReceiver,
--- a/crates/zclaw-pipeline/src/engine/stage.rs
+++ b/crates/zclaw-pipeline/src/engine/stage.rs
@@ -589,7 +589,7 @@ impl StageEngine {
    }

    /// Clone with drivers (reserved for future use)
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: post-release stage cloning with drivers
    fn clone_with_drivers(&self) -> Self {
        Self {
            llm_driver: self.llm_driver.clone(),
--- a/crates/zclaw-pipeline/src/executor.rs
+++ b/crates/zclaw-pipeline/src/executor.rs
@@ -40,6 +40,15 @@ pub enum ExecuteError {
    Io(#[from] std::io::Error),
 }

+/// Maximum completed/failed/cancelled runs to keep in memory
+const MAX_COMPLETED_RUNS: usize = 100;
+
+/// Maximum allowed delay in milliseconds (60 seconds)
+const MAX_DELAY_MS: u64 = 60_000;
+
+/// Default per-step timeout (5 minutes)
+const DEFAULT_STEP_TIMEOUT_SECS: u64 = 300;
+
 /// Pipeline executor
 pub struct PipelineExecutor {
    /// Action registry
@@ -107,35 +116,50 @@ impl PipelineExecutor {
        // Create execution context
        let mut context = ExecutionContext::new(inputs);

+        // Determine per-step timeout from pipeline spec (0 means use default)
+        let step_timeout = if pipeline.spec.timeout_secs > 0 {
+            pipeline.spec.timeout_secs
+        } else {
+            DEFAULT_STEP_TIMEOUT_SECS
+        };
+
        // Execute steps
-        let result = self.execute_steps(pipeline, &mut context, &run_id).await;
+        let result = self.execute_steps(pipeline, &mut context, &run_id, step_timeout).await;

        // Update run state
-        let mut runs = self.runs.write().await;
-        if let Some(run) = runs.get_mut(&run_id) {
-            match result {
-                Ok(outputs) => {
-                    run.status = RunStatus::Completed;
-                    run.outputs = Some(serde_json::to_value(&outputs).unwrap_or(Value::Null));
-                }
-                Err(e) => {
-                    run.status = RunStatus::Failed;
-                    run.error = Some(e.to_string());
+        let return_value = {
+            let mut runs = self.runs.write().await;
+            if let Some(run) = runs.get_mut(&run_id) {
+                match result {
+                    Ok(outputs) => {
+                        run.status = RunStatus::Completed;
+                        run.outputs = Some(serde_json::to_value(&outputs).unwrap_or(Value::Null));
+                    }
+                    Err(e) => {
+                        run.status = RunStatus::Failed;
+                        run.error = Some(e.to_string());
+                    }
                }
+                run.ended_at = Some(Utc::now());
+                Ok(run.clone())
+            } else {
+                Err(ExecuteError::Action("执行后未找到运行记录".to_string()))
            }
-            run.ended_at = Some(Utc::now());
-            return Ok(run.clone());
-        }
+        };

-        Err(ExecuteError::Action("执行后未找到运行记录".to_string()))
+        // Auto-cleanup old completed runs (after releasing the write lock)
+        self.cleanup().await;
+
+        return_value
    }

-    /// Execute pipeline steps
+    /// Execute pipeline steps with per-step timeout
    async fn execute_steps(
        &self,
        pipeline: &Pipeline,
        context: &mut ExecutionContext,
        run_id: &str,
+        step_timeout_secs: u64,
    ) -> Result<HashMap<String, Value>, ExecuteError> {
        let total_steps = pipeline.spec.steps.len();

@@ -161,8 +185,15 @@ impl PipelineExecutor {

            tracing::info!("Executing step {} ({}/{})", step.id, idx + 1, total_steps);

-            // Execute action
-            let result = self.execute_action(&step.action, context).await?;
+            // Execute action with per-step timeout
+            let timeout_duration = std::time::Duration::from_secs(step_timeout_secs);
+            let result = tokio::time::timeout(
+                timeout_duration,
+                self.execute_action(&step.action, context),
+            ).await.map_err(|_| {
+                tracing::error!("Step {} timed out after {}s", step.id, step_timeout_secs);
+                ExecuteError::Timeout
+            })??;

            // Store result
            context.set_output(&step.id, result.clone());
@@ -336,7 +367,16 @@ impl PipelineExecutor {
                }

                Action::Delay { ms } => {
-                    tokio::time::sleep(tokio::time::Duration::from_millis(*ms)).await;
+                    let capped_ms = if *ms > MAX_DELAY_MS {
+                        tracing::warn!(
+                            "Delay ms {} exceeds max {}, capping to {}",
+                            ms, MAX_DELAY_MS, MAX_DELAY_MS
+                        );
+                        MAX_DELAY_MS
+                    } else {
+                        *ms
+                    };
+                    tokio::time::sleep(tokio::time::Duration::from_millis(capped_ms)).await;
                    Ok(Value::Null)
                }

@@ -508,6 +548,33 @@ impl PipelineExecutor {
    pub async fn list_runs(&self) -> Vec<PipelineRun> {
        self.runs.read().await.values().cloned().collect()
    }
+
+    /// Clean up old completed/failed/cancelled runs to prevent memory leaks.
+    /// Keeps at most MAX_COMPLETED_RUNS finished runs, evicting the oldest first.
+    pub async fn cleanup(&self) {
+        let mut runs = self.runs.write().await;
+
+        // Collect IDs of finished runs (completed, failed, cancelled)
+        let mut finished: Vec<(String, chrono::DateTime<Utc>)> = runs
+            .iter()
+            .filter(|(_, r)| matches!(r.status, RunStatus::Completed | RunStatus::Failed | RunStatus::Cancelled))
+            .map(|(id, r)| (id.clone(), r.ended_at.unwrap_or(r.started_at)))
+            .collect();
+
+        let to_remove = finished.len().saturating_sub(MAX_COMPLETED_RUNS);
+        if to_remove > 0 {
+            // Sort by end time ascending (oldest first)
+            finished.sort_by_key(|(_, t)| *t);
+            for (id, _) in finished.into_iter().take(to_remove) {
+                runs.remove(&id);
+                // Also clean up cancellation flag
+                drop(runs);
+                self.cancellations.write().await.remove(&id);
+                runs = self.runs.write().await;
+            }
+            tracing::debug!("Cleaned up {} old pipeline runs", to_remove);
+        }
+    }
 }

 #[cfg(test)]
--- a/crates/zclaw-pipeline/src/state.rs
+++ b/crates/zclaw-pipeline/src/state.rs
@@ -48,7 +48,7 @@ impl ExecutionContext {
            steps_output: HashMap::new(),
            variables: HashMap::new(),
            loop_context: None,
-            expr_regex: Regex::new(r"\$\{([^}]+)\}").unwrap(),
+            expr_regex: Regex::new(r"\$\{([^}]+)\}").expect("static regex is valid"),
        }
    }

@@ -73,7 +73,7 @@ impl ExecutionContext {
            steps_output,
            variables,
            loop_context: None,
-            expr_regex: Regex::new(r"\$\{([^}]+)\}").unwrap(),
+            expr_regex: Regex::new(r"\$\{([^}]+)\}").expect("static regex is valid"),
        }
    }

--- a/crates/zclaw-protocols/src/lib.rs
+++ b/crates/zclaw-protocols/src/lib.rs
@@ -1,20 +1,15 @@
 //! ZCLAW Protocols
 //!
 //! Protocol support for MCP (Model Context Protocol) and A2A (Agent-to-Agent).
-//!
-//! A2A is gated behind the `a2a` feature flag (reserved for future multi-agent scenarios).
-//! MCP is always available as a framework for tool integration.

 mod mcp;
 mod mcp_types;
 mod mcp_tool_adapter;
 mod mcp_transport;
-#[cfg(feature = "a2a")]
 mod a2a;

 pub use mcp::*;
 pub use mcp_types::*;
 pub use mcp_tool_adapter::*;
 pub use mcp_transport::*;
-#[cfg(feature = "a2a")]
 pub use a2a::*;
--- a/crates/zclaw-protocols/src/mcp_tool_adapter.rs
+++ b/crates/zclaw-protocols/src/mcp_tool_adapter.rs
@@ -130,7 +130,7 @@ impl McpToolAdapter {

        match result.len() {
            0 => Ok(Value::Null),
-            1 => Ok(result.into_iter().next().unwrap()),
+            1 => Ok(result.into_iter().next().unwrap_or(Value::Null)),
            _ => Ok(Value::Array(result)),
        }
    }
@@ -160,7 +160,7 @@ impl McpServiceManager {
        let adapters = McpToolAdapter::from_server(name.clone(), client.clone()).await?;
        self.clients.insert(name.clone(), client);
        self.adapters.insert(name.clone(), adapters);
-        Ok(self.adapters.get(&name).unwrap().iter().collect())
+        Ok(self.adapters.get(&name).map(|v| v.iter().collect()).unwrap_or_default())
    }

    /// Get all registered tool adapters from all services
--- a/crates/zclaw-protocols/src/mcp_transport.rs
+++ b/crates/zclaw-protocols/src/mcp_transport.rs
@@ -84,12 +84,20 @@ impl McpServerConfig {
    }
 }

+/// Combined transport handles (stdin + stdout) behind a single Mutex.
+/// This ensures write-then-read is atomic, preventing concurrent requests
+/// from receiving each other's responses.
+struct TransportHandles {
+    stdin: BufWriter<ChildStdin>,
+    stdout: BufReader<ChildStdout>,
+}
+
 /// MCP Transport using stdio
 pub struct McpTransport {
    config: McpServerConfig,
    child: Arc<Mutex<Option<Child>>>,
-    stdin: Arc<Mutex<Option<BufWriter<ChildStdin>>>>,
-    stdout: Arc<Mutex<Option<BufReader<ChildStdout>>>>,
+    /// Single Mutex protecting both stdin and stdout for atomic write-then-read
+    handles: Arc<Mutex<Option<TransportHandles>>>,
    capabilities: Arc<Mutex<Option<ServerCapabilities>>>,
 }

@@ -99,8 +107,7 @@ impl McpTransport {
        Self {
            config,
            child: Arc::new(Mutex::new(None)),
-            stdin: Arc::new(Mutex::new(None)),
-            stdout: Arc::new(Mutex::new(None)),
+            handles: Arc::new(Mutex::new(None)),
            capabilities: Arc::new(Mutex::new(None)),
        }
    }
@@ -162,9 +169,11 @@ impl McpTransport {
            });
        }

-        // Store handles in separate mutexes
-        *self.stdin.lock().await = Some(BufWriter::new(stdin));
-        *self.stdout.lock().await = Some(BufReader::new(stdout));
+        // Store handles in single mutex for atomic write-then-read
+        *self.handles.lock().await = Some(TransportHandles {
+            stdin: BufWriter::new(stdin),
+            stdout: BufReader::new(stdout),
+        });
        *child_guard = Some(child);

        Ok(())
@@ -201,21 +210,21 @@ impl McpTransport {
        let line = serde_json::to_string(notification)
            .map_err(|e| ZclawError::McpError(format!("Failed to serialize notification: {}", e)))?;

-        let mut stdin_guard = self.stdin.lock().await;
-        let stdin = stdin_guard.as_mut()
+        let mut handles_guard = self.handles.lock().await;
+        let handles = handles_guard.as_mut()
            .ok_or_else(|| ZclawError::McpError("Transport not started".to_string()))?;

-        stdin.write_all(line.as_bytes())
+        handles.stdin.write_all(line.as_bytes())
            .map_err(|e| ZclawError::McpError(format!("Failed to write notification: {}", e)))?;
-        stdin.write_all(b"\n")
+        handles.stdin.write_all(b"\n")
            .map_err(|e| ZclawError::McpError(format!("Failed to write newline: {}", e)))?;
-        stdin.flush()
+        handles.stdin.flush()
            .map_err(|e| ZclawError::McpError(format!("Failed to flush notification: {}", e)))?;

        Ok(())
    }

-    /// Send JSON-RPC request
+    /// Send JSON-RPC request (atomic write-then-read under single lock)
    async fn send_request<T: DeserializeOwned>(
        &self,
        method: &str,
@@ -234,28 +243,23 @@ impl McpTransport {
        let line = serde_json::to_string(&request)
            .map_err(|e| ZclawError::McpError(format!("Failed to serialize request: {}", e)))?;

-        // Write to stdin
-        {
-            let mut stdin_guard = self.stdin.lock().await;
-            let stdin = stdin_guard.as_mut()
-                .ok_or_else(|| ZclawError::McpError("Transport not started".to_string()))?;
-
-            stdin.write_all(line.as_bytes())
-                .map_err(|e| ZclawError::McpError(format!("Failed to write request: {}", e)))?;
-            stdin.write_all(b"\n")
-                .map_err(|e| ZclawError::McpError(format!("Failed to write newline: {}", e)))?;
-            stdin.flush()
-                .map_err(|e| ZclawError::McpError(format!("Failed to flush request: {}", e)))?;
-        }
-
-        // Read from stdout
+        // Atomic write-then-read under single lock
        let response_line = {
-            let mut stdout_guard = self.stdout.lock().await;
-            let stdout = stdout_guard.as_mut()
+            let mut handles_guard = self.handles.lock().await;
+            let handles = handles_guard.as_mut()
                .ok_or_else(|| ZclawError::McpError("Transport not started".to_string()))?;

+            // Write to stdin
+            handles.stdin.write_all(line.as_bytes())
+                .map_err(|e| ZclawError::McpError(format!("Failed to write request: {}", e)))?;
+            handles.stdin.write_all(b"\n")
+                .map_err(|e| ZclawError::McpError(format!("Failed to write newline: {}", e)))?;
+            handles.stdin.flush()
+                .map_err(|e| ZclawError::McpError(format!("Failed to flush request: {}", e)))?;
+
+            // Read from stdout (still holding the lock — no interleaving possible)
            let mut response_line = String::new();
-            stdout.read_line(&mut response_line)
+            handles.stdout.read_line(&mut response_line)
                .map_err(|e| ZclawError::McpError(format!("Failed to read response: {}", e)))?;
            response_line
        };
@@ -429,7 +433,7 @@ impl Drop for McpTransport {
                        let _ = child.wait();
                    }
                    Err(e) => {
-                        eprintln!("[McpTransport] Failed to kill child process: {}", e);
+                        tracing::warn!("[McpTransport] Failed to kill child process (potential zombie): {}", e);
                    }
                }
            }
--- a/crates/zclaw-runtime/src/driver/gemini.rs
+++ b/crates/zclaw-runtime/src/driver/gemini.rs
@@ -616,7 +616,7 @@ struct GeminiResponseContent {
    #[serde(default)]
    parts: Vec<GeminiResponsePart>,
    #[serde(default)]
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: deserialized from Gemini API, not accessed in code
    role: Option<String>,
 }

@@ -643,7 +643,7 @@ struct GeminiUsageMetadata {
    #[serde(default)]
    candidates_token_count: Option<u32>,
    #[serde(default)]
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: deserialized from Gemini API, not accessed in code
    total_token_count: Option<u32>,
 }

--- a/crates/zclaw-runtime/src/growth.rs
+++ b/crates/zclaw-runtime/src/growth.rs
@@ -12,11 +12,12 @@

 use std::sync::Arc;
 use zclaw_growth::{
-    GrowthTracker, InjectionFormat, LlmDriverForExtraction,
-    MemoryExtractor, MemoryRetriever, PromptInjector, RetrievalResult,
-    VikingAdapter,
+    AggregatedPattern, CombinedExtraction, EvolutionConfig, EvolutionEngine,
+    ExperienceExtractor, ExperienceStore, GrowthTracker, InjectionFormat,
+    LlmDriverForExtraction, MemoryExtractor, MemoryRetriever, PromptInjector,
+    RetrievalResult, UserProfileUpdater, VikingAdapter,
 };
-use zclaw_memory::{ExtractedFactBatch, Fact, FactCategory};
+use zclaw_memory::{ExtractedFactBatch, Fact, FactCategory, UserProfileStore};
 use zclaw_types::{AgentId, Message, Result, SessionId};

 /// Growth system integration for AgentLoop
@@ -32,6 +33,14 @@ pub struct GrowthIntegration {
    injector: PromptInjector,
    /// Growth tracker for tracking growth metrics
    tracker: GrowthTracker,
+    /// Experience extractor for structured experience persistence
+    experience_extractor: ExperienceExtractor,
+    /// Profile updater for incremental user profile updates
+    profile_updater: UserProfileUpdater,
+    /// User profile store (optional, for profile updates)
+    profile_store: Option<Arc<UserProfileStore>>,
+    /// Evolution engine for L2 skill generation (optional)
+    evolution_engine: Option<EvolutionEngine>,
    /// Configuration
    config: GrowthConfigInner,
 }
@@ -69,13 +78,19 @@ impl GrowthIntegration {

        let retriever = MemoryRetriever::new(viking.clone());
        let injector = PromptInjector::new();
-        let tracker = GrowthTracker::new(viking);
+        let tracker = GrowthTracker::new(viking.clone());
+        let evolution_engine = Some(EvolutionEngine::new(viking.clone()));

        Self {
            retriever,
            extractor,
            injector,
            tracker,
+            experience_extractor: ExperienceExtractor::new()
+                .with_store(Arc::new(ExperienceStore::new(viking))),
+            profile_updater: UserProfileUpdater::new(),
+            profile_store: None,
+            evolution_engine,
            config: GrowthConfigInner::default(),
        }
    }
@@ -102,11 +117,73 @@ impl GrowthIntegration {
        self.config.enabled
    }

+    /// 启动时初始化：从持久化存储恢复进化引擎的信任度记录
+    ///
+    /// **注意**：FeedbackCollector 内部已实现 lazy-load（首次 save() 时自动加载），
+    /// 所以此方法为可选优化 — 提前加载可避免首次反馈提交时的延迟。
+    pub async fn initialize(&self) -> Result<()> {
+        if let Some(ref engine) = self.evolution_engine {
+            match engine.load_feedback().await {
+                Ok(count) => {
+                    if count > 0 {
+                        tracing::info!(
+                            "[GrowthIntegration] Loaded {} trust records from storage",
+                            count
+                        );
+                    }
+                }
+                Err(e) => {
+                    tracing::warn!(
+                        "[GrowthIntegration] Failed to load trust records: {}",
+                        e
+                    );
+                }
+            }
+        }
+        Ok(())
+    }
+
    /// Enable or disable auto extraction
    pub fn set_auto_extract(&mut self, auto_extract: bool) {
        self.config.auto_extract = auto_extract;
    }

+    /// Set the user profile store for incremental profile updates
+    pub fn with_profile_store(mut self, store: Arc<UserProfileStore>) -> Self {
+        self.profile_store = Some(store);
+        self
+    }
+
+    /// Set the evolution engine configuration
+    pub fn with_evolution_config(self, config: EvolutionConfig) -> Self {
+        let engine = self.evolution_engine.unwrap_or_else(|| {
+            EvolutionEngine::new(Arc::new(VikingAdapter::in_memory()))
+        });
+        Self {
+            evolution_engine: Some(engine.with_config(config)),
+            ..self
+        }
+    }
+
+    /// Enable or disable the evolution engine
+    pub fn set_evolution_enabled(&mut self, enabled: bool) {
+        if let Some(ref mut engine) = self.evolution_engine {
+            engine.set_enabled(enabled);
+        }
+    }
+
+    /// L2 检查：是否有可进化的模式
+    /// 在 extract_combined 之后调用，返回可固化的经验模式列表
+    pub async fn check_evolution(
+        &self,
+        agent_id: &AgentId,
+    ) -> Result<Vec<AggregatedPattern>> {
+        match &self.evolution_engine {
+            Some(engine) => engine.check_evolvable_patterns(&agent_id.to_string()).await,
+            None => Ok(Vec::new()),
+        }
+    }
+
    /// Enhance system prompt with retrieved memories
    ///
    /// This method:
@@ -213,8 +290,8 @@ impl GrowthIntegration {
        Ok(count)
    }

-    /// Combined extraction: single LLM call that produces both stored memories
-    /// and structured facts, avoiding double extraction overhead.
+    /// Combined extraction: single LLM call that produces stored memories,
+    /// structured experiences, and profile signals — all in one pass.
    ///
    /// Returns `(memory_count, Option<ExtractedFactBatch>)` on success.
    pub async fn extract_combined(
@@ -227,25 +304,28 @@ impl GrowthIntegration {
            return Ok(None);
        }

-        // Single LLM extraction call
-        let extracted = self
+        // 单次 LLM 提取：memories + experiences + profile_signals
+        let combined = self
            .extractor
-            .extract(messages, session_id.clone())
+            .extract_combined(messages, session_id.clone())
            .await
            .unwrap_or_else(|e| {
                tracing::warn!("[GrowthIntegration] Combined extraction failed: {}", e);
-                Vec::new()
+                CombinedExtraction::default()
            });

-        if extracted.is_empty() {
+        if combined.memories.is_empty()
+            && combined.experiences.is_empty()
+            && !combined.profile_signals.has_any_signal()
+        {
            return Ok(None);
        }

-        let mem_count = extracted.len();
+        let mem_count = combined.memories.len();

        // Store raw memories
        self.extractor
-            .store_memories(&agent_id.to_string(), &extracted)
+            .store_memories(&agent_id.to_string(), &combined.memories)
            .await?;

        // Track learning event
@@ -253,8 +333,71 @@ impl GrowthIntegration {
            .record_learning(agent_id, &session_id.to_string(), mem_count)
            .await?;

-        // Convert same extracted memories to structured facts (no extra LLM call)
-        let facts: Vec<Fact> = extracted
+        // Persist structured experiences (L1 enhancement)
+        if let Ok(exp_count) = self
+            .experience_extractor
+            .persist_experiences(&agent_id.to_string(), &combined)
+            .await
+        {
+            if exp_count > 0 {
+                tracing::debug!(
+                    "[GrowthIntegration] Persisted {} structured experiences",
+                    exp_count
+                );
+            }
+        }
+
+        // Update user profile from extraction signals (L1 enhancement)
+        if let Some(profile_store) = &self.profile_store {
+            let updates = self.profile_updater.collect_updates(&combined);
+            let user_id = agent_id.to_string();
+            for update in updates {
+                let result = match update.kind {
+                    zclaw_growth::ProfileUpdateKind::SetField => {
+                        profile_store
+                            .update_field(&user_id, &update.field, &update.value)
+                            .await
+                    }
+                    zclaw_growth::ProfileUpdateKind::AppendArray => {
+                        match update.field.as_str() {
+                            "recent_topic" => {
+                                profile_store
+                                    .add_recent_topic(&user_id, &update.value, 10)
+                                    .await
+                            }
+                            "pain_point" => {
+                                profile_store
+                                    .add_pain_point(&user_id, &update.value, 10)
+                                    .await
+                            }
+                            "preferred_tool" => {
+                                profile_store
+                                    .add_preferred_tool(&user_id, &update.value, 10)
+                                    .await
+                            }
+                            _ => {
+                                tracing::warn!(
+                                    "[GrowthIntegration] Unknown array field: {}",
+                                    update.field
+                                );
+                                Ok(())
+                            }
+                        }
+                    }
+                };
+                if let Err(e) = result {
+                    tracing::warn!(
+                        "[GrowthIntegration] Profile update failed for {}: {}",
+                        update.field,
+                        e
+                    );
+                }
+            }
+        }
+
+        // Convert extracted memories to structured facts
+        let facts: Vec<Fact> = combined
+            .memories
            .into_iter()
            .map(|m| {
                let category = match m.memory_type {
--- a/crates/zclaw-runtime/src/middleware.rs
+++ b/crates/zclaw-runtime/src/middleware.rs
@@ -279,3 +279,4 @@ pub mod token_calibration;
 pub mod tool_error;
 pub mod tool_output_guard;
 pub mod trajectory_recorder;
+pub mod evolution;
--- a/crates/zclaw-runtime/src/middleware/data_masking.rs
+++ b/crates/zclaw-runtime/src/middleware/data_masking.rs
@@ -20,19 +20,19 @@ use super::{AgentMiddleware, MiddlewareContext, MiddlewareDecision};
 // ---------------------------------------------------------------------------

 static RE_COMPANY: LazyLock<Regex> = LazyLock::new(|| {
-    Regex::new(r"[^\s]{1,20}(?:公司|厂|集团|工作室|商行|有限|股份)").unwrap()
+    Regex::new(r"[^\s]{1,20}(?:公司|厂|集团|工作室|商行|有限|股份)").expect("static regex is valid")
 });
 static RE_MONEY: LazyLock<Regex> = LazyLock::new(|| {
-    Regex::new(r"[¥￥$]\s*[\d,.]+[万亿]?元?|[\d,.]+[万亿]元").unwrap()
+    Regex::new(r"[¥￥$]\s*[\d,.]+[万亿]?元?|[\d,.]+[万亿]元").expect("static regex is valid")
 });
 static RE_PHONE: LazyLock<Regex> = LazyLock::new(|| {
-    Regex::new(r"1[3-9]\d-?\d{4}-?\d{4}").unwrap()
+    Regex::new(r"1[3-9]\d-?\d{4}-?\d{4}").expect("static regex is valid")
 });
 static RE_EMAIL: LazyLock<Regex> = LazyLock::new(|| {
-    Regex::new(r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}").unwrap()
+    Regex::new(r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}").expect("static regex is valid")
 });
 static RE_ID_CARD: LazyLock<Regex> = LazyLock::new(|| {
-    Regex::new(r"\b\d{17}[\dXx]\b").unwrap()
+    Regex::new(r"\b\d{17}[\dXx]\b").expect("static regex is valid")
 });

 // ---------------------------------------------------------------------------
--- a/crates/zclaw-runtime/src/middleware/evolution.rs
+++ b/crates/zclaw-runtime/src/middleware/evolution.rs
@@ -0,0 +1,165 @@
+//! 进化引擎中间件
+//! 在管家对话中检测并呈现"技能进化确认"提示
+//! 优先级 78（在 ButlerRouter@80 之前运行）
+
+use async_trait::async_trait;
+use std::sync::Arc;
+use tokio::sync::RwLock;
+use crate::middleware::{
+    AgentMiddleware, MiddlewareContext, MiddlewareDecision,
+};
+use zclaw_types::Result;
+
+/// 待确认的进化事件
+#[derive(Debug, Clone)]
+pub struct PendingEvolution {
+    pub pattern_name: String,
+    pub trigger_suggestion: String,
+    pub description: String,
+}
+
+/// 进化引擎中间件
+/// 检查是否有待确认的进化事件，注入确认提示到 system prompt
+pub struct EvolutionMiddleware {
+    pending: Arc<RwLock<Vec<PendingEvolution>>>,
+}
+
+impl EvolutionMiddleware {
+    pub fn new() -> Self {
+        Self {
+            pending: Arc::new(RwLock::new(Vec::new())),
+        }
+    }
+
+    /// 添加一个待确认的进化事件
+    pub async fn add_pending(&self, evolution: PendingEvolution) {
+        self.pending.write().await.push(evolution);
+    }
+
+    /// 获取并清除所有待确认事件
+    pub async fn drain_pending(&self) -> Vec<PendingEvolution> {
+        let mut pending = self.pending.write().await;
+        std::mem::take(&mut *pending)
+    }
+
+    /// 当前待确认事件数量
+    pub async fn pending_count(&self) -> usize {
+        self.pending.read().await.len()
+    }
+}
+
+impl Default for EvolutionMiddleware {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[async_trait]
+impl AgentMiddleware for EvolutionMiddleware {
+    fn name(&self) -> &str {
+        "evolution"
+    }
+
+    fn priority(&self) -> i32 {
+        78 // 在 ButlerRouter(80) 之前
+    }
+
+    async fn before_completion(
+        &self,
+        ctx: &mut MiddlewareContext,
+    ) -> Result<MiddlewareDecision> {
+        // 先用 read lock 快速判空，避免每次对话都获取写锁
+        if self.pending.read().await.is_empty() {
+            return Ok(MiddlewareDecision::Continue);
+        }
+
+        // 只移除第一个事件，保留后续事件留待下次注入
+        let to_inject = {
+            let mut pending = self.pending.write().await;
+            if pending.is_empty() {
+                return Ok(MiddlewareDecision::Continue);
+            }
+            pending.remove(0)
+        };
+
+        let injection = format!(
+            "\n\n<evolution-suggestion>\n\
+我注意到你经常做「{pattern}」相关的事情。\n\
+我可以帮你整理成一个技能，以后直接说「{trigger}」就能用了。\n\
+技能描述：{desc}\n\
+如果你同意，请回复 '确认保存技能'。如果你想调整，可以告诉我怎么改。\n\
+</evolution-suggestion>",
+            pattern = to_inject.pattern_name,
+            trigger = to_inject.trigger_suggestion,
+            desc = to_inject.description,
+        );
+        ctx.system_prompt.push_str(&injection);
+
+        tracing::info!(
+            "[EvolutionMiddleware] Injected evolution suggestion for: {}",
+            to_inject.pattern_name
+        );
+
+        Ok(MiddlewareDecision::Continue)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn test_no_pending_continues() {
+        let mw = EvolutionMiddleware::new();
+        assert_eq!(mw.pending_count().await, 0);
+    }
+
+    #[tokio::test]
+    async fn test_add_and_drain() {
+        let mw = EvolutionMiddleware::new();
+        mw.add_pending(PendingEvolution {
+            pattern_name: "报表生成".to_string(),
+            trigger_suggestion: "生成报表".to_string(),
+            description: "自动生成每日报表".to_string(),
+        })
+        .await;
+        assert_eq!(mw.pending_count().await, 1);
+
+        let drained = mw.drain_pending().await;
+        assert_eq!(drained.len(), 1);
+        assert_eq!(drained[0].pattern_name, "报表生成");
+        assert_eq!(mw.pending_count().await, 0);
+    }
+
+    #[tokio::test]
+    async fn test_name_and_priority() {
+        let mw = EvolutionMiddleware::new();
+        assert_eq!(mw.name(), "evolution");
+        assert_eq!(mw.priority(), 78);
+    }
+
+    #[tokio::test]
+    async fn test_only_first_event_injected() {
+        let mw = EvolutionMiddleware::new();
+        mw.add_pending(PendingEvolution {
+            pattern_name: "事件A".to_string(),
+            trigger_suggestion: "触发A".to_string(),
+            description: "描述A".to_string(),
+        })
+        .await;
+        mw.add_pending(PendingEvolution {
+            pattern_name: "事件B".to_string(),
+            trigger_suggestion: "触发B".to_string(),
+            description: "描述B".to_string(),
+        })
+        .await;
+
+        // 模拟注入：用 read 判空 + write 取第一个
+        let first = {
+            let mut pending = mw.pending.write().await;
+            pending.remove(0)
+        };
+        assert_eq!(first.pattern_name, "事件A");
+        assert_eq!(mw.pending_count().await, 1); // 事件B 仍保留
+    }
+}
--- a/crates/zclaw-runtime/src/middleware/memory.rs
+++ b/crates/zclaw-runtime/src/middleware/memory.rs
@@ -11,14 +11,17 @@ use async_trait::async_trait;
 use zclaw_types::Result;
 use crate::growth::GrowthIntegration;
 use crate::middleware::{AgentMiddleware, MiddlewareContext, MiddlewareDecision};
+use crate::middleware::evolution::EvolutionMiddleware;

 /// Middleware that handles memory retrieval (pre-completion) and extraction (post-completion).
 ///
 /// Wraps `GrowthIntegration` and delegates:
 /// - `before_completion` → `enhance_prompt()` for memory injection
-/// - `after_completion` → `process_conversation()` for memory extraction
+/// - `after_completion` → `extract_combined()` for memory extraction + evolution check
 pub struct MemoryMiddleware {
    growth: GrowthIntegration,
+    /// Shared EvolutionMiddleware for pushing evolution suggestions
+    evolution_mw: Option<std::sync::Arc<EvolutionMiddleware>>,
    /// Minimum seconds between extractions for the same agent (debounce).
    debounce_secs: u64,
    /// Timestamp of last extraction per agent (for debouncing).
@@ -29,11 +32,18 @@ impl MemoryMiddleware {
    pub fn new(growth: GrowthIntegration) -> Self {
        Self {
            growth,
+            evolution_mw: None,
            debounce_secs: 30,
            last_extraction: std::sync::Mutex::new(std::collections::HashMap::new()),
        }
    }

+    /// Attach a shared EvolutionMiddleware for pushing evolution suggestions.
+    pub fn with_evolution(mut self, mw: std::sync::Arc<EvolutionMiddleware>) -> Self {
+        self.evolution_mw = Some(mw);
+        self
+    }
+
    /// Set the debounce interval in seconds.
    pub fn with_debounce_secs(mut self, secs: u64) -> Self {
        self.debounce_secs = secs;
@@ -52,6 +62,49 @@ impl MemoryMiddleware {
        map.insert(agent_id.to_string(), now);
        true
    }
+
+    /// Check for evolvable patterns and push suggestions to EvolutionMiddleware.
+    async fn check_and_push_evolution(&self, agent_id: &zclaw_types::AgentId) {
+        let evolution_mw = match &self.evolution_mw {
+            Some(mw) => mw,
+            None => return,
+        };
+
+        match self.growth.check_evolution(agent_id).await {
+            Ok(patterns) if !patterns.is_empty() => {
+                for pattern in &patterns {
+                    let trigger = pattern
+                        .common_steps
+                        .first()
+                        .cloned()
+                        .unwrap_or_else(|| pattern.pain_pattern.clone());
+                    evolution_mw.add_pending(
+                        crate::middleware::evolution::PendingEvolution {
+                            pattern_name: pattern.pain_pattern.clone(),
+                            trigger_suggestion: trigger,
+                            description: format!(
+                                "基于 {} 次重复经验，自动固化技能",
+                                pattern.total_reuse
+                            ),
+                        },
+                    ).await;
+                }
+                tracing::info!(
+                    "[MemoryMiddleware] Pushed {} evolution candidates for agent {}",
+                    patterns.len(),
+                    agent_id
+                );
+            }
+            Ok(_) => {
+                tracing::debug!("[MemoryMiddleware] No evolvable patterns found");
+            }
+            Err(e) => {
+                tracing::debug!(
+                    "[MemoryMiddleware] Evolution check failed (non-fatal): {}", e
+                );
+            }
+        }
+    }
 }

 #[async_trait]
@@ -65,11 +118,6 @@ impl AgentMiddleware for MemoryMiddleware {
            ctx.user_input.chars().take(50).collect::<String>()
        );

-        // Retrieve relevant memories and inject into system prompt.
-        // The SqliteStorage retriever now uses FTS5-only matching — if FTS5 finds
-        // no relevant results, no memories are returned (no scope-based fallback).
-        // This prevents irrelevant high-importance memories from leaking into
-        // unrelated conversations.
        let base = &ctx.system_prompt;
        match self.growth.enhance_prompt(&ctx.agent_id, base, &ctx.user_input).await {
            Ok(enhanced) => {
@@ -88,7 +136,6 @@ impl AgentMiddleware for MemoryMiddleware {
                Ok(MiddlewareDecision::Continue)
            }
            Err(e) => {
-                // Non-fatal: retrieval failure should not block the conversation
                tracing::warn!(
                    "[MemoryMiddleware] Memory retrieval failed (non-fatal): {}",
                    e
@@ -99,7 +146,6 @@ impl AgentMiddleware for MemoryMiddleware {
    }

    async fn after_completion(&self, ctx: &MiddlewareContext) -> Result<()> {
-        // Debounce: skip extraction if called too recently for this agent
        let agent_key = ctx.agent_id.to_string();
        if !self.should_extract(&agent_key) {
            tracing::debug!(
@@ -113,8 +159,6 @@ impl AgentMiddleware for MemoryMiddleware {
            return Ok(());
        }

-        // Combined extraction: single LLM call produces both memories and structured facts.
-        // Avoids double LLM extraction ( process_conversation + extract_structured_facts).
        match self.growth.extract_combined(
            &ctx.agent_id,
            &ctx.messages,
@@ -127,12 +171,14 @@ impl AgentMiddleware for MemoryMiddleware {
                    facts.len(),
                    agent_key
                );
+
+                // Check for evolvable patterns after successful extraction
+                self.check_and_push_evolution(&ctx.agent_id).await;
            }
            Ok(None) => {
                tracing::debug!("[MemoryMiddleware] No memories or facts extracted");
            }
            Err(e) => {
-                // Non-fatal: extraction failure should not affect the response
                tracing::warn!("[MemoryMiddleware] Combined extraction failed: {}", e);
            }
        }
--- a/crates/zclaw-runtime/src/nl_schedule.rs
+++ b/crates/zclaw-runtime/src/nl_schedule.rs
@@ -2,7 +2,7 @@
 //!
 //! Three-layer fallback strategy:
 //! 1. Regex pattern matching (covers ~80% of common expressions)
-//! 2. LLM-assisted parsing (for ambiguous/complex expressions) — TODO: wire when Haiku driver available
+//! 2. LLM-assisted parsing (for ambiguous/complex expressions) — FUTURE: post-release LLM-assisted natural language parsing
 //! 3. Interactive clarification (return `Unclear`)
 //!
 //! Lives in `zclaw-runtime` because it's a pure text→cron utility with no kernel dependency.
@@ -69,7 +69,7 @@ const PERIOD: &str = "(凌晨|早上|早晨|上午|中午|下午|午后|傍晚|
 static RE_TIME_STRIP: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(
        r"^(?:凌晨|早上|早晨|上午|中午|下午|午后|傍晚|黄昏|晚上|晚间|夜里|夜晚|半夜|午夜)?\d{1,2}[点时:：]\d{0,2}分?"
-    ).unwrap()
+    ).expect("static regex pattern is valid")
 });

 // try_every_day
@@ -77,13 +77,13 @@ static RE_EVERY_DAY_EXACT: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(&format!(
        r"(?:每天|每日)(?:的)?{}(\d{{1,2}})[点时:：](\d{{1,2}})?",
        PERIOD
-    )).unwrap()
+    )).expect("static regex pattern is valid")
 });

 static RE_EVERY_DAY_PERIOD: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(
        r"(?:每天|每日)(?:的)?(凌晨|早上|早晨|上午|中午|下午|午后|傍晚|黄昏|晚上|晚间|夜里|夜晚|半夜|午夜)"
-    ).unwrap()
+    ).expect("static regex pattern is valid")
 });

 // try_every_week
@@ -91,7 +91,7 @@ static RE_EVERY_WEEK: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(&format!(
        r"(?:每周|每个?星期|每个?礼拜)(一|二|三|四|五|六|日|天|周一|周二|周三|周四|周五|周六|周日|周天|星期一|星期二|星期三|星期四|星期五|星期六|星期日|星期天|礼拜一|礼拜二|礼拜三|礼拜四|礼拜五|礼拜六|礼拜日|礼拜天)(?:的)?{}(\d{{1,2}})[点时:：](\d{{1,2}})?",
        PERIOD
-    )).unwrap()
+    )).expect("static regex pattern is valid")
 });

 // try_workday
@@ -99,18 +99,18 @@ static RE_WORKDAY_EXACT: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(&format!(
        r"(?:工作日|每个?工作日|工作日(?:的)?){}(\d{{1,2}})[点时:：](\d{{1,2}})?",
        PERIOD
-    )).unwrap()
+    )).expect("static regex pattern is valid")
 });

 static RE_WORKDAY_PERIOD: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(
        r"(?:工作日|每个?工作日)(?:的)?(凌晨|早上|早晨|上午|中午|下午|午后|傍晚|黄昏|晚上|晚间|夜里|夜晚|半夜|午夜)"
-    ).unwrap()
+    ).expect("static regex pattern is valid")
 });

 // try_interval
 static RE_INTERVAL: LazyLock<Regex> = LazyLock::new(|| {
-    Regex::new(r"每(\d{1,2})(小时|分钟|分|钟|个小时)").unwrap()
+    Regex::new(r"每(\d{1,2})(小时|分钟|分|钟|个小时)").expect("static regex pattern is valid")
 });

 // try_monthly
@@ -118,7 +118,7 @@ static RE_MONTHLY: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(&format!(
        r"(?:每月|每个月)(?:的)?(\d{{1,2}})[号日](?:的)?{}(\d{{1,2}})?[点时:：]?(\d{{1,2}})?",
        PERIOD
-    )).unwrap()
+    )).expect("static regex pattern is valid")
 });

 // try_one_shot
@@ -126,7 +126,7 @@ static RE_ONE_SHOT: LazyLock<Regex> = LazyLock::new(|| {
    Regex::new(&format!(
        r"(明天|后天|大后天)(?:的)?{}(\d{{1,2}})[点时:：](\d{{1,2}})?",
        PERIOD
-    )).unwrap()
+    )).expect("static regex pattern is valid")
 });

 // ---------------------------------------------------------------------------
--- a/crates/zclaw-runtime/src/tool/builtin/task.rs
+++ b/crates/zclaw-runtime/src/tool/builtin/task.rs
@@ -112,12 +112,14 @@ impl Tool for TaskTool {
        let task_id = sub_agent_id.to_string();

        if let Some(ref tx) = context.event_sender {
-            let _ = tx.send(LoopEvent::SubtaskStatus {
+            if tx.send(LoopEvent::SubtaskStatus {
                task_id: task_id.clone(),
                description: description.to_string(),
                status: "started".to_string(),
                detail: None,
-            }).await;
+            }).await.is_err() {
+                tracing::debug!("[TaskTool] Subtask status dropped: parent loop ended");
+            }
        }

        // Create a fresh session for the sub-agent
@@ -161,12 +163,14 @@ impl Tool for TaskTool {

        // Emit subtask_running event
        if let Some(ref tx) = context.event_sender {
-            let _ = tx.send(LoopEvent::SubtaskStatus {
+            if tx.send(LoopEvent::SubtaskStatus {
                task_id: task_id.clone(),
                description: description.to_string(),
                status: "running".to_string(),
                detail: Some("子Agent正在执行中...".to_string()),
-            }).await;
+            }).await.is_err() {
+                tracing::debug!("[TaskTool] Subtask status dropped: parent loop ended");
+            }
        }

        // Execute the sub-agent loop (non-streaming — collect full result)
@@ -179,7 +183,7 @@ impl Tool for TaskTool {

                // Emit subtask_completed event
                if let Some(ref tx) = context.event_sender {
-                    let _ = tx.send(LoopEvent::SubtaskStatus {
+                    if tx.send(LoopEvent::SubtaskStatus {
                        task_id: task_id.clone(),
                        description: description.to_string(),
                        status: "completed".to_string(),
@@ -187,7 +191,9 @@ impl Tool for TaskTool {
                            "完成 ({}次迭代, {}输入token)",
                            loop_result.iterations, loop_result.input_tokens
                        )),
-                    }).await;
+                    }).await.is_err() {
+                        tracing::debug!("[TaskTool] Subtask status dropped: parent loop ended");
+                    }
                }

                json!({
@@ -204,12 +210,14 @@ impl Tool for TaskTool {

                // Emit subtask_failed event
                if let Some(ref tx) = context.event_sender {
-                    let _ = tx.send(LoopEvent::SubtaskStatus {
+                    if tx.send(LoopEvent::SubtaskStatus {
                        task_id: task_id.clone(),
                        description: description.to_string(),
                        status: "failed".to_string(),
                        detail: Some(e.to_string()),
-                    }).await;
+                    }).await.is_err() {
+                        tracing::debug!("[TaskTool] Subtask status dropped: parent loop ended");
+                    }
                }

                json!({
--- a/crates/zclaw-saas/migrations/20260418000001_add_missing_indexes.sql
+++ b/crates/zclaw-saas/migrations/20260418000001_add_missing_indexes.sql
@@ -0,0 +1,11 @@
+-- Add missing indexes for performance-critical queries
+-- 2026-04-18 Release readiness audit
+
+-- Rate limit events cleanup (DELETE WHERE created_at < ...)
+CREATE INDEX IF NOT EXISTS idx_rle_created_at ON rate_limit_events(created_at);
+
+-- Billing subscriptions plan lookup
+CREATE INDEX IF NOT EXISTS idx_billing_sub_plan ON billing_subscriptions(plan_id);
+
+-- Knowledge items created_by lookup
+CREATE INDEX IF NOT EXISTS idx_ki_created_by ON knowledge_items(created_by);
--- a/crates/zclaw-saas/src/auth/handlers.rs
+++ b/crates/zclaw-saas/src/auth/handlers.rs
@@ -565,7 +565,7 @@ async fn store_refresh_token(

 /// 清理过期和已使用的 refresh tokens
 /// 注意: 现已迁移到 Worker/Scheduler 定期执行，此函数保留作为备用
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: backup for Worker/Scheduler cleanup; kept as fallback
 async fn cleanup_expired_refresh_tokens(db: &sqlx::PgPool) -> SaasResult<()> {
    let now = chrono::Utc::now();
    // 删除过期超过 30 天的已使用 token (减少 DB 膨胀)
--- a/crates/zclaw-saas/src/billing/handlers.rs
+++ b/crates/zclaw-saas/src/billing/handlers.rs
@@ -587,7 +587,7 @@ pub async fn get_invoice_pdf(
        })?;

    // 返回 PDF 响应
-    Ok(axum::response::Response::builder()
+    axum::response::Response::builder()
        .status(200)
        .header("Content-Type", "application/pdf")
        .header(
@@ -595,5 +595,8 @@ pub async fn get_invoice_pdf(
            format!("attachment; filename=\"invoice-{}.pdf\"", invoice.id),
        )
        .body(axum::body::Body::from(bytes))
-        .unwrap())
+        .map_err(|e| {
+            tracing::error!("Failed to build PDF response: {}", e);
+            SaasError::Internal("PDF 响应构建失败".into())
+        })
 }
--- a/crates/zclaw-saas/src/config.rs
+++ b/crates/zclaw-saas/src/config.rs
@@ -396,6 +396,23 @@ impl SaaSConfig {
            config.database.url = db_url;
        }

+        // Config validation
+        if config.auth.jwt_expiration_hours < 1 {
+            anyhow::bail!(
+                "auth.jwt_expiration_hours must be >= 1, got {}",
+                config.auth.jwt_expiration_hours
+            );
+        }
+        if config.database.max_connections == 0 {
+            anyhow::bail!("database.max_connections must be > 0");
+        }
+        if config.database.min_connections > config.database.max_connections {
+            anyhow::bail!(
+                "database.min_connections ({}) must be <= max_connections ({})",
+                config.database.min_connections, config.database.max_connections
+            );
+        }
+
        Ok(config)
    }

--- a/crates/zclaw-saas/src/knowledge/handlers.rs
+++ b/crates/zclaw-saas/src/knowledge/handlers.rs
@@ -804,7 +804,7 @@ async fn handle_document_upload(

    // 创建知识条目
    let item_req = CreateItemRequest {
-        category_id: "uploaded".to_string(), // TODO: 从上传参数获取
+        category_id: "uploaded".to_string(), // FUTURE: post-release category_id from upload params
        title: doc.title.clone(),
        content,
        keywords: None,
--- a/crates/zclaw-saas/src/main.rs
+++ b/crates/zclaw-saas/src/main.rs
@@ -309,10 +309,34 @@ async fn build_router(state: AppState) -> axum::Router {
            .unwrap_or(false);
        if config.server.cors_origins.is_empty() {
            if is_dev {
+                // Dev mode: use explicit localhost origins (Any + credentials violates CORS spec)
+                let dev_origins: Vec<HeaderValue> = [
+                    "http://localhost:1420",
+                    "http://localhost:5173",
+                    "http://127.0.0.1:1420",
+                    "http://127.0.0.1:5173",
+                    "http://localhost:8080",
+                    "http://127.0.0.1:8080",
+                    "tauri://localhost",
+                    "https://tauri.localhost",
+                ].iter()
+                    .filter_map(|o| o.parse::<HeaderValue>().ok())
+                    .collect();
                CorsLayer::new()
-                    .allow_origin(Any)
-                    .allow_methods(Any)
-                    .allow_headers(Any)
+                    .allow_origin(dev_origins)
+                    .allow_methods([
+                        axum::http::Method::GET,
+                        axum::http::Method::POST,
+                        axum::http::Method::PUT,
+                        axum::http::Method::PATCH,
+                        axum::http::Method::DELETE,
+                        axum::http::Method::OPTIONS,
+                    ])
+                    .allow_headers([
+                        axum::http::header::AUTHORIZATION,
+                        axum::http::header::CONTENT_TYPE,
+                        axum::http::header::COOKIE,
+                    ])
                    .allow_credentials(true)
            } else {
                tracing::error!("生产环境必须配置 server.cors_origins，不能使用 allow_origin(Any)");
--- a/crates/zclaw-saas/src/scheduled_task/service.rs
+++ b/crates/zclaw-saas/src/scheduled_task/service.rs
@@ -6,7 +6,7 @@ use super::types::*;

 /// 数据库行结构
 #[derive(Debug, FromRow)]
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: FromRow deserialization struct; fields accessed via destructuring
 struct ScheduledTaskRow {
    id: String,
    account_id: String,
--- a/crates/zclaw-saas/src/workers/cleanup_rate_limit.rs
+++ b/crates/zclaw-saas/src/workers/cleanup_rate_limit.rs
@@ -1,4 +1,7 @@
 //! 清理过期 Rate Limit 条目 Worker
+//!
+//! rate_limit_events 表中的持久化条目会无限增长。
+//! 此 Worker 定期删除超过 1 小时的旧条目，防止数据库膨胀。

 use async_trait::async_trait;
 use sqlx::PgPool;
@@ -21,10 +24,31 @@ impl Worker for CleanupRateLimitWorker {
        "cleanup_rate_limit"
    }

-    async fn perform(&self, _db: &PgPool, _args: Self::Args) -> SaasResult<()> {
-        // Rate limit entries are in-memory (DashMap), not in DB
-        // This worker is a placeholder for when rate limits are persisted
-        // Currently the cleanup happens in main.rs background task
+    async fn perform(&self, db: &PgPool, args: Self::Args) -> SaasResult<()> {
+        let retention_secs = args.window_secs.max(3600); // 至少保留 1 小时
+
+        let result = sqlx::query(
+            "DELETE FROM rate_limit_events WHERE created_at < NOW() - ($1 || ' seconds')::interval"
+        )
+        .bind(retention_secs.to_string())
+        .execute(db)
+        .await;
+
+        match result {
+            Ok(r) => {
+                let deleted = r.rows_affected();
+                if deleted > 0 {
+                    tracing::info!(
+                        "[cleanup_rate_limit] Deleted {} expired rate limit events (retention: {}s)",
+                        deleted, retention_secs
+                    );
+                }
+            }
+            Err(e) => {
+                tracing::error!("[cleanup_rate_limit] Failed to clean up rate limit events: {}", e);
+            }
+        }
+
        Ok(())
    }
 }
--- a/crates/zclaw-skills/Cargo.toml
+++ b/crates/zclaw-skills/Cargo.toml
@@ -9,7 +9,7 @@ description = "ZCLAW skill system"

 [features]
 default = []
-wasm = ["wasmtime", "wasmtime-wasi/p1"]
+wasm = ["wasmtime", "wasmtime-wasi/p1", "ureq", "url"]

 [dependencies]
 zclaw-types = { workspace = true }
@@ -27,3 +27,5 @@ shlex = { workspace = true }
 # Optional WASM runtime (enable with --features wasm)
 wasmtime = { workspace = true, optional = true }
 wasmtime-wasi = { workspace = true, optional = true }
+ureq = { workspace = true, optional = true }
+url = { workspace = true, optional = true }
--- a/crates/zclaw-skills/src/orchestration/executor.rs
+++ b/crates/zclaw-skills/src/orchestration/executor.rs
@@ -83,7 +83,7 @@ impl DefaultExecutor {
    }

    /// Execute a single node (used by pipeline orchestration action driver)
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: post-release pipeline orchestration action driver
    async fn execute_node(
        &self,
        node: &super::SkillNode,
--- a/crates/zclaw-skills/src/wasm_runner.rs
+++ b/crates/zclaw-skills/src/wasm_runner.rs
@@ -9,6 +9,7 @@

 use async_trait::async_trait;
 use serde_json::Value;
+use std::io::Read as IoRead;
 use std::path::PathBuf;
 use tracing::{debug, warn};
 use wasmtime::*;
@@ -23,6 +24,9 @@ use crate::{Skill, SkillContext, SkillManifest, SkillResult};
 /// Maximum WASM binary size (10 MB).
 const MAX_WASM_SIZE: usize = 10 * 1024 * 1024;

+/// Maximum HTTP response body size for host function (1 MB).
+const MAX_HTTP_RESPONSE_SIZE: usize = 1024 * 1024;
+
 /// Fuel per second of CPU time (heuristic: ~10M instructions/sec).
 const FUEL_PER_SEC: u64 = 10_000_000;

@@ -230,49 +234,178 @@ fn create_engine_config() -> Config {
 }

 /// Add ZCLAW host functions to the wasmtime linker.
-fn add_host_functions(linker: &mut Linker<WasiP1Ctx>, _network_allowed: bool) -> Result<()> {
+fn add_host_functions(linker: &mut Linker<WasiP1Ctx>, network_allowed: bool) -> Result<()> {
    linker
        .func_wrap(
            "env",
            "zclaw_log",
-            |_caller: Caller<'_, WasiP1Ctx>, _ptr: u32, _len: u32| {
-                debug!("[WasmSkill] guest called zclaw_log");
+            |mut caller: Caller<'_, WasiP1Ctx>, ptr: u32, len: u32| {
+                let msg = read_guest_string(&mut caller, ptr, len);
+                debug!("[WasmSkill] guest log: {}", msg);
            },
        )
        .map_err(|e| {
            zclaw_types::ZclawError::ToolError(format!("Failed to add zclaw_log: {}", e))
        })?;

+    // zclaw_http_fetch(url_ptr, url_len, out_ptr, out_cap) -> bytes_written (-1 = error)
+    // Performs a synchronous GET request. Result is written to guest memory as JSON string.
+    let net = network_allowed;
    linker
        .func_wrap(
            "env",
            "zclaw_http_fetch",
-            |_caller: Caller<'_, WasiP1Ctx>,
-             _url_ptr: u32,
-             _url_len: u32,
-             _out_ptr: u32,
-             _out_cap: u32|
-             -> i32 {
-                warn!("[WasmSkill] guest called zclaw_http_fetch — denied");
-                -1
+            move |mut caller: Caller<'_, WasiP1Ctx>,
+                  url_ptr: u32,
+                  url_len: u32,
+                  out_ptr: u32,
+                  out_cap: u32|
+                  -> i32 {
+                if !net {
+                    warn!("[WasmSkill] guest called zclaw_http_fetch — denied (network not allowed)");
+                    return -1;
+                }
+
+                let url = read_guest_string(&mut caller, url_ptr, url_len);
+                if url.is_empty() {
+                    return -1;
+                }
+
+                // Security: validate URL scheme to prevent SSRF.
+                // Only http:// and https:// are allowed.
+                let parsed = match url::Url::parse(&url) {
+                    Ok(u) => u,
+                    Err(_) => {
+                        warn!("[WasmSkill] http_fetch denied — invalid URL: {}", url);
+                        return -1;
+                    }
+                };
+                let scheme = parsed.scheme();
+                if scheme != "http" && scheme != "https" {
+                    warn!("[WasmSkill] http_fetch denied — unsupported scheme: {}", scheme);
+                    return -1;
+                }
+                // Block private/loopback hosts to prevent SSRF
+                if let Some(host) = parsed.host_str() {
+                    let lower = host.to_lowercase();
+                    if lower == "localhost"
+                        || lower.starts_with("127.")
+                        || lower.starts_with("10.")
+                        || lower.starts_with("192.168.")
+                        || lower.starts_with("169.254.")
+                        || lower.starts_with("0.")
+                        || lower.ends_with(".internal")
+                        || lower.ends_with(".local")
+                    {
+                        warn!("[WasmSkill] http_fetch denied — private/loopback host: {}", host);
+                        return -1;
+                    }
+                    // Also block 172.16.0.0/12 range
+                    if lower.starts_with("172.") {
+                        if let Ok(second) = lower.split('.').nth(1).unwrap_or("0").parse::<u8>() {
+                            if (16..=31).contains(&second) {
+                                warn!("[WasmSkill] http_fetch denied — private host (172.16-31.x.x): {}", host);
+                                return -1;
+                            }
+                        }
+                    }
+                }
+
+                debug!("[WasmSkill] guest http_fetch: {}", url);
+
+                // Synchronous HTTP GET (we're already on a blocking thread)
+                let agent = ureq::Agent::config_builder()
+                    .timeout_global(Some(std::time::Duration::from_secs(10)))
+                    .build()
+                    .new_agent();
+                let response = agent.get(&url).call();
+
+                match response {
+                    Ok(mut resp) => {
+                        // Enforce response size limit before reading body
+                        let content_length = resp.header("content-length")
+                            .and_then(|v| v.to_str().ok())
+                            .and_then(|v| v.parse::<usize>().ok());
+                        if let Some(len) = content_length {
+                            if len > MAX_HTTP_RESPONSE_SIZE {
+                                warn!("[WasmSkill] http_fetch denied — response too large: {} bytes (max {})", len, MAX_HTTP_RESPONSE_SIZE);
+                                return -1;
+                            }
+                        }
+                        let mut body = String::new();
+                        match resp.body_mut().read_to_string(&mut body) {
+                            Ok(_) => {
+                                if body.len() > MAX_HTTP_RESPONSE_SIZE {
+                                    warn!("[WasmSkill] http_fetch — response exceeded limit after read, truncating");
+                                    body.truncate(MAX_HTTP_RESPONSE_SIZE);
+                                }
+                                write_guest_bytes(&mut caller, out_ptr, out_cap, body.as_bytes())
+                            }
+                            Err(e) => {
+                                warn!("[WasmSkill] http_fetch body read error: {}", e);
+                                -1
+                            }
+                        }
+                    }
+                    }
+                    Err(e) => {
+                        warn!("[WasmSkill] http_fetch error for {}: {}", url, e);
+                        -1
+                    }
+                }
            },
        )
        .map_err(|e| {
            zclaw_types::ZclawError::ToolError(format!("Failed to add zclaw_http_fetch: {}", e))
        })?;

+    // zclaw_file_read(path_ptr, path_len, out_ptr, out_cap) -> bytes_written (-1 = error)
+    // Reads a file from the preopened /workspace directory. Paths must be relative.
    linker
        .func_wrap(
            "env",
            "zclaw_file_read",
-            |_caller: Caller<'_, WasiP1Ctx>,
-             _path_ptr: u32,
-             _path_len: u32,
-             _out_ptr: u32,
-             _out_cap: u32|
+            |mut caller: Caller<'_, WasiP1Ctx>,
+             path_ptr: u32,
+             path_len: u32,
+             out_ptr: u32,
+             out_cap: u32|
             -> i32 {
-                warn!("[WasmSkill] guest called zclaw_file_read — denied");
-                -1
+                let path = read_guest_string(&mut caller, path_ptr, path_len);
+                if path.is_empty() {
+                    return -1;
+                }
+
+                // Security: validate path stays within /workspace sandbox.
+                // Reject absolute paths, and filter any path component that
+                // is ".." (e.g. "foo/../../etc/passwd").
+                let joined = std::path::Path::new("/workspace").join(&path);
+                let mut safe = true;
+                for comp in joined.components() {
+                    match comp {
+                        std::path::Component::ParentDir => {
+                            safe = false;
+                            break;
+                        }
+                        std::path::Component::RootDir | std::path::Component::Prefix(_) => {
+                            safe = false;
+                            break;
+                        }
+                        _ => {} // Normal, CurDir — ok
+                    }
+                }
+                if !safe {
+                    warn!("[WasmSkill] guest file_read denied — path escapes sandbox: {}", path);
+                    return -1;
+                }
+
+                match std::fs::read(&joined) {
+                    Ok(data) => write_guest_bytes(&mut caller, out_ptr, out_cap, &data),
+                    Err(e) => {
+                        debug!("[WasmSkill] file_read error for {}: {}", path, e);
+                        -1
+                    }
+                }
            },
        )
        .map_err(|e| {
@@ -282,6 +415,38 @@ fn add_host_functions(linker: &mut Linker<WasiP1Ctx>, _network_allowed: bool) ->
    Ok(())
 }

+/// Read a string from WASM guest memory.
+fn read_guest_string(caller: &mut Caller<'_, WasiP1Ctx>, ptr: u32, len: u32) -> String {
+    let mem = match caller.get_export("memory") {
+        Some(Extern::Memory(m)) => m,
+        _ => return String::new(),
+    };
+    let offset = ptr as usize;
+    let length = len as usize;
+    let data = mem.data(&caller);
+    if offset + length > data.len() {
+        return String::new();
+    }
+    String::from_utf8_lossy(&data[offset..offset + length]).into_owned()
+}
+
+/// Write bytes to WASM guest memory. Returns the number of bytes written, or -1 on overflow.
+fn write_guest_bytes(caller: &mut Caller<'_, WasiP1Ctx>, ptr: u32, cap: u32, data: &[u8]) -> i32 {
+    let mem = match caller.get_export("memory") {
+        Some(Extern::Memory(m)) => m,
+        _ => return -1,
+    };
+    let offset = ptr as usize;
+    let capacity = cap as usize;
+    let write_len = data.len().min(capacity);
+    if offset + write_len > mem.data_size(&caller) {
+        return -1;
+    }
+    // Safety: we've bounds-checked the write region.
+    mem.data_mut(&mut *caller)[offset..offset + write_len].copy_from_slice(&data[..write_len]);
+    write_len as i32
+}
+

 #[cfg(test)]
 mod tests {
--- a/crates/zclaw-types/src/message.rs
+++ b/crates/zclaw-types/src/message.rs
@@ -116,7 +116,6 @@ impl Message {

 /// Canonical LLM message content block. Used for agent conversation messages.
 /// See also: zclaw_runtime::driver::ContentBlock (LLM driver response subset),
-/// zclaw_hands::slideshow::ContentBlock (presentation rendering),
 /// zclaw_protocols::mcp_types::ContentBlock (MCP protocol wire format).
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(tag = "type", rename_all = "snake_case")]
--- a/desktop/src-tauri/Cargo.toml
+++ b/desktop/src-tauri/Cargo.toml
@@ -16,9 +16,7 @@ crate-type = ["staticlib", "cdylib", "rlib"]
 tauri-build = { version = "2", features = [] }

 [features]
-default = ["multi-agent"]
-# Multi-agent orchestration (A2A protocol, Director, agent delegation)
-multi-agent = ["zclaw-kernel/multi-agent"]
+default = []
 dev-server = ["dep:axum", "dep:tower-http"]

 [dependencies]
@@ -49,7 +47,7 @@ async-trait = { workspace = true }
 # Serialization
 serde = { workspace = true }
 serde_json = { workspace = true }
-serde_yaml = "0.9"
+serde_yaml = { package = "serde_yaml_bw", version = "2" }
 toml = "0.8"

 # HTTP client
--- a/desktop/src-tauri/src/browser/session.rs
+++ b/desktop/src-tauri/src/browser/session.rs
@@ -168,7 +168,7 @@ impl SessionManager {

    /// Get the number of active sessions
    /// Reserved for future status dashboard
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: post-release status dashboard session count
    pub async fn session_count(&self) -> usize {
        let sessions = self.sessions.read().await;
        sessions.len()
--- a/desktop/src-tauri/src/classroom_commands/generate.rs
+++ b/desktop/src-tauri/src/classroom_commands/generate.rs
@@ -207,7 +207,7 @@ pub async fn classroom_generate(
        metadata: zclaw_kernel::generation::ClassroomMetadata {
            generated_at: std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
-                .unwrap()
+                .expect("system clock is valid")
                .as_millis() as i64,
            source_document: kernel_request.document.map(|_| "user_document".to_string()),
            model: None,
--- a/desktop/src-tauri/src/classroom_commands/mod.rs
+++ b/desktop/src-tauri/src/classroom_commands/mod.rs
@@ -55,7 +55,9 @@ pub async fn init_persistence(
        .map_err(|e| format!("Failed to get app data dir: {}", e))?;

    let db_path = app_dir.join("classroom").join("classrooms.db");
-    std::fs::create_dir_all(db_path.parent().unwrap())
+    let db_dir = db_path.parent()
+        .ok_or_else(|| "Invalid classroom database path: no parent directory".to_string())?;
+    std::fs::create_dir_all(db_dir)
        .map_err(|e| format!("Failed to create classroom dir: {}", e))?;

    let persistence: ClassroomPersistence = ClassroomPersistence::open(db_path).await?;
--- a/desktop/src-tauri/src/classroom_commands/persist.rs
+++ b/desktop/src-tauri/src/classroom_commands/persist.rs
@@ -101,7 +101,7 @@ impl ClassroomPersistence {
    }

    /// Delete a classroom and its chat history.
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: post-release classroom deletion Tauri command
    pub async fn delete_classroom(&self, classroom_id: &str) -> Result<(), String> {
        let mut conn = self.conn.lock().await;
        sqlx::query("DELETE FROM classrooms WHERE id = ?")
--- a/desktop/src-tauri/src/dev_server.rs
+++ b/desktop/src-tauri/src/dev_server.rs
@@ -48,10 +48,10 @@ async fn run_server(state: DevServerState) {
        .layer(
            CorsLayer::new()
                .allow_origin([
-                    "http://localhost:1420".parse().unwrap(),
-                    "http://127.0.0.1:1420".parse().unwrap(),
-                    "http://localhost:5173".parse().unwrap(),
-                    "http://127.0.0.1:5173".parse().unwrap(),
+                    "http://localhost:1420".parse().expect("hardcoded localhost URL is valid"),
+                    "http://127.0.0.1:1420".parse().expect("hardcoded localhost URL is valid"),
+                    "http://localhost:5173".parse().expect("hardcoded localhost URL is valid"),
+                    "http://127.0.0.1:5173".parse().expect("hardcoded localhost URL is valid"),
                ])
                .allow_methods([Method::GET, Method::POST, Method::OPTIONS])
                .allow_headers([header::CONTENT_TYPE, header::AUTHORIZATION]),
--- a/desktop/src-tauri/src/intelligence/compactor.rs
+++ b/desktop/src-tauri/src/intelligence/compactor.rs
@@ -99,7 +99,7 @@ pub struct CompactionCheck {

 /// Configuration for LLM-based summary generation
 /// NOTE: Reserved for future LLM compaction Tauri command
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: post-release LLM compaction Tauri command
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct LlmSummaryConfig {
    pub provider: String,
--- a/desktop/src-tauri/src/intelligence/extraction_adapter.rs
+++ b/desktop/src-tauri/src/intelligence/extraction_adapter.rs
@@ -225,6 +225,69 @@ impl LlmDriverForExtraction for TauriExtractionDriver {

        Ok(memories)
    }
+
+    async fn extract_with_prompt(
+        &self,
+        messages: &[Message],
+        system_prompt: &str,
+        user_prompt: &str,
+    ) -> Result<String> {
+        if messages.len() < 2 {
+            return Err(zclaw_types::ZclawError::InvalidInput(
+                "Too few messages for combined extraction".to_string(),
+            ));
+        }
+
+        tracing::debug!(
+            "[TauriExtractionDriver] Combined extraction from {} messages",
+            messages.len()
+        );
+
+        let request = CompletionRequest {
+            model: self.model.clone(),
+            system: Some(system_prompt.to_string()),
+            messages: vec![Message::user(user_prompt.to_string())],
+            tools: Vec::new(),
+            max_tokens: Some(3000),
+            temperature: Some(0.3),
+            stop: Vec::new(),
+            stream: false,
+            thinking_enabled: false,
+            reasoning_effort: None,
+            plan_mode: false,
+        };
+
+        let response = self.driver.complete(request).await.map_err(|e| {
+            tracing::error!(
+                "[TauriExtractionDriver] Combined extraction LLM call failed: {}",
+                e
+            );
+            e
+        })?;
+
+        let response_text: String = response
+            .content
+            .into_iter()
+            .filter_map(|block| match block {
+                ContentBlock::Text { text } => Some(text),
+                _ => None,
+            })
+            .collect::<Vec<_>>()
+            .join("");
+
+        if response_text.is_empty() {
+            return Err(zclaw_types::ZclawError::LlmError(
+                "Empty response from LLM for combined extraction".to_string(),
+            ));
+        }
+
+        tracing::info!(
+            "[TauriExtractionDriver] Combined extraction response: {} chars",
+            response_text.len()
+        );
+
+        Ok(response_text)
+    }
 }

 /// Global extraction driver instance (legacy path, kept for compatibility).
@@ -250,7 +313,7 @@ pub fn configure_extraction_driver(driver: Arc<dyn LlmDriver>, model: String) {
 }

 /// Check if the extraction driver is available (legacy OnceCell path).
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: diagnostic check for legacy OnceCell extraction driver
 pub fn is_extraction_driver_configured() -> bool {
    EXTRACTION_DRIVER.get().is_some()
 }
@@ -258,7 +321,7 @@ pub fn is_extraction_driver_configured() -> bool {
 /// Get the global extraction driver (legacy OnceCell path).
 ///
 /// Prefer accessing via `kernel.extraction_driver()` when the Kernel is available.
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: legacy accessor, prefer kernel.extraction_driver()
 pub fn get_extraction_driver() -> Option<Arc<TauriExtractionDriver>> {
    EXTRACTION_DRIVER.get().cloned()
 }
@@ -272,22 +335,6 @@ mod tests {
        assert!(!is_extraction_driver_configured());
    }

-    #[test]
-    fn test_parse_empty_response() {
-        // We cannot create a real LlmDriver easily in tests, so we test the
-        // parsing logic via a minimal helper.
-        struct DummyDriver;
-        impl TauriExtractionDriver {
-            fn parse_response_test(
-                &self,
-                response_text: &str,
-                extraction_type: MemoryType,
-            ) -> Vec<ExtractedMemory> {
-                self.parse_response(response_text, extraction_type)
-            }
-        }
-    }
-
    #[test]
    fn test_parse_valid_json_response() {
        let response = r#"```json
--- a/desktop/src-tauri/src/intelligence/heartbeat.rs
+++ b/desktop/src-tauri/src/intelligence/heartbeat.rs
@@ -383,7 +383,10 @@ async fn execute_tick(

    // Send alerts via broadcast channel (internal)
    for alert in &filtered_alerts {
-        let _ = alert_sender.send(alert.clone());
+        if alert_sender.send(alert.clone()).is_err() {
+            tracing::debug!("[heartbeat] No alert receivers, alert dropped");
+            break;
+        }
    }

    // Emit alerts to frontend via Tauri event (real-time toast)
--- a/desktop/src-tauri/src/intelligence/personality_detector.rs
+++ b/desktop/src-tauri/src/intelligence/personality_detector.rs
@@ -324,7 +324,7 @@ mod tests {
 }

 // ---------------------------------------------------------------------------
-// Persistence helpers (stub — TODO: integrate with VikingStorage)
+// Persistence helpers (stub — FUTURE: post-release personality persistence via VikingStorage)
 // ---------------------------------------------------------------------------

 use std::sync::OnceLock;
@@ -339,12 +339,12 @@ fn personality_store() -> &'static Mutex<std::collections::HashMap<String, Perso
 /// Load personality config for a given agent.
 /// Returns default config if none is stored.
 pub fn load_personality_config(agent_id: &str) -> PersonalityConfig {
-    let store = personality_store().lock().unwrap();
+    let store = personality_store().lock().unwrap_or_else(|e| e.into_inner());
    store.get(agent_id).cloned().unwrap_or_default()
 }

 /// Save personality config for a given agent.
 pub fn save_personality_config(agent_id: &str, config: &PersonalityConfig) {
-    let mut store = personality_store().lock().unwrap();
+    let mut store = personality_store().lock().unwrap_or_else(|e| e.into_inner());
    store.insert(agent_id.to_string(), config.clone());
 }
--- a/desktop/src-tauri/src/intelligence/user_profiler.rs
+++ b/desktop/src-tauri/src/intelligence/user_profiler.rs
@@ -10,6 +10,8 @@ use std::sync::Arc;
 use chrono::Utc;
 use tracing::{debug, warn};
 use zclaw_memory::fact::Fact;
+#[cfg(test)]
+use zclaw_memory::fact::FactCategory;
 use zclaw_memory::user_profile_store::{
    CommStyle, Level, UserProfile, UserProfileStore,
 };
@@ -86,7 +88,7 @@ fn classify_fact_content(fact: &Fact) -> Option<ProfileFieldUpdate> {
        return Some(ProfileFieldUpdate::PreferredTool("collector".into()));
    }
    if content.contains("幻灯") || content.contains("演示") || content.contains("ppt") {
-        return Some(ProfileFieldUpdate::PreferredTool("slideshow".into()));
+        return Some(ProfileFieldUpdate::RecentTopic("演示文稿".into()));
    }

    // Default: treat as a recent topic
--- a/desktop/src-tauri/src/intelligence_hooks.rs
+++ b/desktop/src-tauri/src/intelligence_hooks.rs
@@ -206,7 +206,7 @@ pub async fn post_conversation_hook(
 ///
 /// NOTE: Memory injection is now handled by MemoryMiddleware in the Kernel
 /// middleware chain. This function is kept as a utility for ad-hoc queries.
-#[allow(dead_code)]
+#[allow(dead_code)] // @reserved: ad-hoc memory context queries outside middleware chain
 async fn build_memory_context(
    agent_id: &str,
    user_message: &str,
--- a/desktop/src-tauri/src/kernel_commands/a2a.rs
+++ b/desktop/src-tauri/src/kernel_commands/a2a.rs
@@ -1,4 +1,4 @@
-//! A2A (Agent-to-Agent) commands — gated behind `multi-agent` feature
+//! A2A (Agent-to-Agent) commands

 use serde_json;
 use tauri::State;
@@ -7,10 +7,9 @@ use zclaw_types::AgentId;
 use super::KernelState;

 // ============================================================
-// A2A (Agent-to-Agent) Commands — gated behind multi-agent feature
+// A2A (Agent-to-Agent) Commands
 // ============================================================

-#[cfg(feature = "multi-agent")]
 /// Send a direct A2A message from one agent to another
 // @connected
 #[tauri::command]
@@ -44,7 +43,6 @@ pub async fn agent_a2a_send(
 }

 /// Broadcast a message from one agent to all other agents
-#[cfg(feature = "multi-agent")]
 // @connected
 #[tauri::command]
 pub async fn agent_a2a_broadcast(
@@ -66,7 +64,6 @@ pub async fn agent_a2a_broadcast(
 }

 /// Discover agents with a specific capability
-#[cfg(feature = "multi-agent")]
 // @connected
 #[tauri::command]
 pub async fn agent_a2a_discover(
@@ -88,7 +85,6 @@ pub async fn agent_a2a_discover(
  }

 /// Delegate a task to another agent and wait for response
-#[cfg(feature = "multi-agent")]
 // @connected
 #[tauri::command]
 pub async fn agent_a2a_delegate_task(
@@ -116,11 +112,10 @@ pub async fn agent_a2a_delegate_task(
  }

 // ============================================================
-// Butler Delegation Command — multi-agent feature
+// Butler Delegation Command
 // ============================================================

 /// Butler delegates a user request to expert agents via the Director.
-#[cfg(feature = "multi-agent")]
 // @reserved: butler multi-agent delegation
 // @connected
 #[tauri::command]
--- a/desktop/src-tauri/src/kernel_commands/chat.rs
+++ b/desktop/src-tauri/src/kernel_commands/chat.rs
@@ -291,15 +291,19 @@ pub async fn agent_chat_stream(
        );

        let (tx, rx) = tokio::sync::mpsc::channel(32);
-        let _ = tx.send(zclaw_runtime::LoopEvent::Delta(confirm_msg)).await;
-        let _ = tx.send(zclaw_runtime::LoopEvent::Complete(
+        if tx.send(zclaw_runtime::LoopEvent::Delta(confirm_msg)).await.is_err() {
+            tracing::warn!("[agent_chat_stream] Failed to send confirm msg to new channel");
+        }
+        if tx.send(zclaw_runtime::LoopEvent::Complete(
            zclaw_runtime::AgentLoopResult {
                response: String::new(),
                input_tokens: 0,
                output_tokens: 0,
                iterations: 1,
            }
-        )).await;
+        )).await.is_err() {
+            tracing::warn!("[agent_chat_stream] Failed to send complete to new channel");
+        }
        drop(tx);
        (rx, None)
    } else {
@@ -400,10 +404,12 @@ pub async fn agent_chat_stream(
            // Check cancellation flag before each recv
            if cancel_clone.load(std::sync::atomic::Ordering::SeqCst) {
                tracing::info!("[agent_chat_stream] Stream cancelled for session: {}", session_id);
-                let _ = app.emit("stream:chunk", serde_json::json!({
+                if let Err(e) = app.emit("stream:chunk", serde_json::json!({
                    "sessionId": session_id,
                    "event": StreamChatEvent::Error { message: "已取消".to_string() }
-                }));
+                })) {
+                    tracing::debug!("[agent_chat_stream] Failed to emit cancel event: {}", e);
+                }
                break;
            }

@@ -491,12 +497,14 @@ pub async fn agent_chat_stream(
                }
                Err(_) => {
                    tracing::warn!("[agent_chat_stream] Stream idle timeout for session: {}", session_id);
-                    let _ = app.emit("stream:chunk", serde_json::json!({
+                    if let Err(e) = app.emit("stream:chunk", serde_json::json!({
                        "sessionId": session_id,
                        "event": StreamChatEvent::Error {
                            message: "流式响应超时，请重试".to_string()
                        }
-                    }));
+                    })) {
+                        tracing::debug!("[agent_chat_stream] Failed to emit timeout event: {}", e);
+                    }
                    break;
                }
            }
--- a/desktop/src-tauri/src/kernel_commands/mcp.rs
+++ b/desktop/src-tauri/src/kernel_commands/mcp.rs
@@ -33,7 +33,7 @@ impl Default for McpManagerState {

 impl McpManagerState {
    /// Create with a pre-allocated kernel_adapters Arc for sharing with Kernel.
-    #[allow(dead_code)]
+    #[allow(dead_code)] // @reserved: alternate constructor for shared MCP adapter injection
    pub fn with_shared_adapters(kernel_adapters: Arc<std::sync::RwLock<Vec<McpToolAdapter>>>) -> Self {
        Self {
            manager: Arc::new(Mutex::new(McpServiceManager::new())),
--- a/desktop/src-tauri/src/kernel_commands/mod.rs
+++ b/desktop/src-tauri/src/kernel_commands/mod.rs
@@ -19,7 +19,6 @@ pub mod skill;
 pub mod trigger;
 pub mod workspace;

-#[cfg(feature = "multi-agent")]
 pub mod a2a;

 // ---------------------------------------------------------------------------
--- a/desktop/src-tauri/src/kernel_commands/orchestration.rs
+++ b/desktop/src-tauri/src/kernel_commands/orchestration.rs
@@ -204,6 +204,7 @@ impl From<zclaw_skills::orchestration::OrchestrationResult> for OrchestrationRes

 /// @reserved — no frontend UI yet
 /// Execute a skill orchestration
+/// @reserved — orchestration engine internal, no direct frontend caller
 ///
 /// Either auto-composes a graph from skill_ids, or uses a pre-defined graph.
 /// Executes with true parallel execution within each dependency level.
--- a/desktop/src-tauri/src/lib.rs
+++ b/desktop/src-tauri/src/lib.rs
@@ -255,16 +255,11 @@ pub fn run() {
            kernel_commands::scheduled_task::scheduled_task_create,
            kernel_commands::scheduled_task::scheduled_task_list,

-            // A2A commands gated behind multi-agent feature
-            #[cfg(feature = "multi-agent")]
+            // A2A commands
            kernel_commands::a2a::agent_a2a_send,
-            #[cfg(feature = "multi-agent")]
            kernel_commands::a2a::agent_a2a_broadcast,
-            #[cfg(feature = "multi-agent")]
            kernel_commands::a2a::agent_a2a_discover,
-            #[cfg(feature = "multi-agent")]
            kernel_commands::a2a::agent_a2a_delegate_task,
-            #[cfg(feature = "multi-agent")]
            kernel_commands::a2a::butler_delegate_task,

            // Pipeline commands (DSL-based workflows)
--- a/desktop/src-tauri/src/viking_commands.rs
+++ b/desktop/src-tauri/src/viking_commands.rs
@@ -489,7 +489,7 @@ pub async fn viking_tree(path: String, depth: Option<usize>) -> Result<serde_jso
            current = current
                .get_mut(*part)
                .and_then(|v| v.as_object_mut())
-                .unwrap();
+                .ok_or_else(|| format!("Invalid URI tree structure at segment: {}", part))?;
        }

        if let Some(last) = parts.last() {
--- a/desktop/src/main.tsx
+++ b/desktop/src/main.tsx
@@ -5,6 +5,7 @@ import './index.css';
 import { ToastProvider } from './components/ui/Toast';
 import { GlobalErrorBoundary } from './components/ui/ErrorBoundary';
 import { initWebMCPTools } from './lib/webmcp-tools';
+import { isTauriRuntime } from './lib/tauri-gateway';
 import { setupPluginListeners } from 'tauri-plugin-mcp';

 // Global error handler for uncaught errors
@@ -30,11 +31,10 @@ const handleGlobalReset = () => {
 // Initialize WebMCP debugging tools (dev mode only, Chrome 146+)
 initWebMCPTools();

-// Initialize tauri-plugin-mcp event listeners (dev mode only)
-if (import.meta.env.DEV) {
-  setupPluginListeners().catch((err) => {
-    console.warn('[MCP] Failed to setup plugin listeners:', err);
-  });
+// Initialize tauri-plugin-mcp event listeners (dev mode + Tauri runtime only)
+// Only works inside Tauri webview — gracefully skips in browser dev mode
+if (import.meta.env.DEV && isTauriRuntime()) {
+  setupPluginListeners().catch(() => {});
 }

 ReactDOM.createRoot(document.getElementById('root')!).render(
--- a/desktop/src/store/workflowBuilderStore.ts
+++ b/desktop/src/store/workflowBuilderStore.ts
@@ -1,456 +0,0 @@
-/**
- * Workflow Builder Store
- *
- * Zustand store for managing workflow builder state.
- */
-
-import { create } from 'zustand';
-import { persist } from 'zustand/middleware';
-import type {
-  WorkflowCanvas,
-  WorkflowNode,
-  WorkflowEdge,
-  WorkflowNodeData,
-  WorkflowTemplate,
-  ValidationResult,
-  NodePaletteItem,
-  WorkflowNodeType,
-  NodeCategory,
-} from '../lib/workflow-builder/types';
-import { validateCanvas } from '../lib/workflow-builder/yaml-converter';
-
-// =============================================================================
-// Store State
-// =============================================================================
-
-interface WorkflowBuilderState {
-  // Canvas state
-  canvas: WorkflowCanvas | null;
-  workflows: WorkflowCanvas[];
-
-  // Selection
-  selectedNodeId: string | null;
-  selectedEdgeId: string | null;
-
-  // UI state
-  isDragging: boolean;
-  isDirty: boolean;
-  isPreviewOpen: boolean;
-  validation: ValidationResult | null;
-
-  // Templates
-  templates: WorkflowTemplate[];
-
-  // Available items for palette
-  availableSkills: Array<{ id: string; name: string; description: string }>;
-  availableHands: Array<{ id: string; name: string; actions: string[] }>;
-
-  // Actions
-  createNewWorkflow: (name: string, description?: string) => void;
-  loadWorkflow: (id: string) => void;
-  saveWorkflow: () => void;
-  deleteWorkflow: (id: string) => void;
-
-  // Node actions
-  addNode: (type: WorkflowNodeType, position: { x: number; y: number }) => void;
-  updateNode: (nodeId: string, data: Partial<WorkflowNodeData>) => void;
-  deleteNode: (nodeId: string) => void;
-  duplicateNode: (nodeId: string) => void;
-
-  // Edge actions
-  addEdge: (source: string, target: string) => void;
-  deleteEdge: (edgeId: string) => void;
-
-  // Selection actions
-  selectNode: (nodeId: string | null) => void;
-  selectEdge: (edgeId: string | null) => void;
-
-  // UI actions
-  setDragging: (isDragging: boolean) => void;
-  setPreviewOpen: (isOpen: boolean) => void;
-  validate: () => ValidationResult;
-
-  // Data loading
-  setAvailableSkills: (skills: Array<{ id: string; name: string; description: string }>) => void;
-  setAvailableHands: (hands: Array<{ id: string; name: string; actions: string[] }>) => void;
-
-  // Canvas metadata
-  updateCanvasMetadata: (updates: Partial<Pick<WorkflowCanvas, 'name' | 'description' | 'category'>>) => void;
-}
-
-// =============================================================================
-// Default Node Data
-// =============================================================================
-
-function getDefaultNodeData(type: WorkflowNodeType, _id: string): WorkflowNodeData {
-  const base = { label: type.charAt(0).toUpperCase() + type.slice(1) };
-
-  switch (type) {
-    case 'input':
-      return { type: 'input', ...base, variableName: 'input', schema: undefined };
-    case 'llm':
-      return { type: 'llm', ...base, template: '', isTemplateFile: false, jsonMode: false };
-    case 'skill':
-      return { type: 'skill', ...base, skillId: '', inputMappings: {} };
-    case 'hand':
-      return { type: 'hand', ...base, handId: '', action: '', params: {} };
-    case 'orchestration':
-      return { type: 'orchestration', ...base, inputMappings: {} };
-    case 'condition':
-      return { type: 'condition', ...base, condition: '', branches: [{ when: '', label: 'Branch 1' }], hasDefault: true };
-    case 'parallel':
-      return { type: 'parallel', ...base, each: '${inputs.items}', maxWorkers: 4 };
-    case 'loop':
-      return { type: 'loop', ...base, each: '${inputs.items}', itemVar: 'item', indexVar: 'index' };
-    case 'export':
-      return { type: 'export', ...base, formats: ['json'] };
-    case 'http':
-      return { type: 'http', ...base, url: '', method: 'GET', headers: {} };
-    case 'setVar':
-      return { type: 'setVar', ...base, variableName: 'result', value: '' };
-    case 'delay':
-      return { type: 'delay', ...base, ms: 1000 };
-    default:
-      throw new Error(`Unknown node type: ${type}`);
-  }
-}
-
-// =============================================================================
-// Store Implementation
-// =============================================================================
-
-export const useWorkflowBuilderStore = create<WorkflowBuilderState>()(
-  persist(
-    (set, get) => ({
-      // Initial state
-      canvas: null,
-      workflows: [],
-      selectedNodeId: null,
-      selectedEdgeId: null,
-      isDragging: false,
-      isDirty: false,
-      isPreviewOpen: false,
-      validation: null,
-      templates: [],
-      availableSkills: [],
-      availableHands: [],
-
-      // Workflow actions
-      createNewWorkflow: (name, description) => {
-        const canvas: WorkflowCanvas = {
-          id: `workflow_${Date.now()}`,
-          name,
-          description,
-          category: 'custom',
-          nodes: [],
-          edges: [],
-          viewport: { x: 0, y: 0, zoom: 1 },
-          metadata: {
-            createdAt: new Date().toISOString(),
-            updatedAt: new Date().toISOString(),
-            tags: [],
-            version: '1.0.0',
-          },
-        };
-        set({ canvas, isDirty: false, selectedNodeId: null, selectedEdgeId: null, validation: null });
-      },
-
-      loadWorkflow: (id) => {
-        const workflow = get().workflows.find(w => w.id === id);
-        if (workflow) {
-          set({ canvas: workflow, isDirty: false, selectedNodeId: null, selectedEdgeId: null });
-        }
-      },
-
-      saveWorkflow: () => {
-        const { canvas, workflows } = get();
-        if (!canvas) return;
-
-        const updatedCanvas: WorkflowCanvas = {
-          ...canvas,
-          metadata: {
-            ...canvas.metadata,
-            updatedAt: new Date().toISOString(),
-          },
-        };
-
-        const existingIndex = workflows.findIndex(w => w.id === canvas.id);
-        let updatedWorkflows: WorkflowCanvas[];
-
-        if (existingIndex >= 0) {
-          updatedWorkflows = [...workflows];
-          updatedWorkflows[existingIndex] = updatedCanvas;
-        } else {
-          updatedWorkflows = [...workflows, updatedCanvas];
-        }
-
-        set({ workflows: updatedWorkflows, canvas: updatedCanvas, isDirty: false });
-      },
-
-      deleteWorkflow: (id) => {
-        set(state => ({
-          workflows: state.workflows.filter(w => w.id !== id),
-          canvas: state.canvas?.id === id ? null : state.canvas,
-        }));
-      },
-
-      // Node actions
-      addNode: (type, position) => {
-        const { canvas } = get();
-        if (!canvas) return;
-
-        const id = `${type}_${Date.now()}`;
-        const node: WorkflowNode = {
-          id,
-          type,
-          position,
-          data: getDefaultNodeData(type, id),
-        };
-
-        set({
-          canvas: { ...canvas, nodes: [...canvas.nodes, node] },
-          isDirty: true,
-          selectedNodeId: id,
-        });
-      },
-
-      updateNode: (nodeId, data) => {
-        const { canvas } = get();
-        if (!canvas) return;
-
-        const updatedNodes = canvas.nodes.map(node =>
-          node.id === nodeId
-            ? { ...node, data: { ...node.data, ...data } as WorkflowNodeData }
-            : node
-        );
-
-        set({ canvas: { ...canvas, nodes: updatedNodes }, isDirty: true });
-      },
-
-      deleteNode: (nodeId) => {
-        const { canvas } = get();
-        if (!canvas) return;
-
-        const updatedNodes = canvas.nodes.filter(n => n.id !== nodeId);
-        const updatedEdges = canvas.edges.filter(e => e.source !== nodeId && e.target !== nodeId);
-
-        set({
-          canvas: { ...canvas, nodes: updatedNodes, edges: updatedEdges },
-          isDirty: true,
-          selectedNodeId: null,
-        });
-      },
-
-      duplicateNode: (nodeId) => {
-        const { canvas } = get();
-        if (!canvas) return;
-
-        const node = canvas.nodes.find(n => n.id === nodeId);
-        if (!node) return;
-
-        const newId = `${node.type}_${Date.now()}`;
-        const newNode: WorkflowNode = {
-          ...node,
-          id: newId,
-          position: {
-            x: node.position.x + 50,
-            y: node.position.y + 50,
-          },
-          data: { ...node.data, label: `${node.data.label} (copy)` } as WorkflowNodeData,
-        };
-
-        set({
-          canvas: { ...canvas, nodes: [...canvas.nodes, newNode] },
-          isDirty: true,
-          selectedNodeId: newId,
-        });
-      },
-
-      // Edge actions
-      addEdge: (source, target) => {
-        const { canvas } = get();
-        if (!canvas) return;
-
-        // Check if edge already exists
-        const exists = canvas.edges.some(e => e.source === source && e.target === target);
-        if (exists) return;
-
-        const edge: WorkflowEdge = {
-          id: `edge_${source}_${target}`,
-          source,
-          target,
-          type: 'default',
-        };
-
-        set({ canvas: { ...canvas, edges: [...canvas.edges, edge] }, isDirty: true });
-      },
-
-      deleteEdge: (edgeId) => {
-        const { canvas } = get();
-        if (!canvas) return;
-
-        set({
-          canvas: { ...canvas, edges: canvas.edges.filter(e => e.id !== edgeId) },
-          isDirty: true,
-        });
-      },
-
-      // Selection actions
-      selectNode: (nodeId) => set({ selectedNodeId: nodeId, selectedEdgeId: null }),
-      selectEdge: (edgeId) => set({ selectedEdgeId: edgeId, selectedNodeId: null }),
-
-      // UI actions
-      setDragging: (isDragging) => set({ isDragging }),
-      setPreviewOpen: (isOpen) => set({ isPreviewOpen: isOpen }),
-
-      validate: () => {
-        const { canvas } = get();
-        if (!canvas) {
-          return { valid: false, errors: [{ nodeId: 'canvas', message: 'No workflow loaded', severity: 'error' as const }], warnings: [] };
-        }
-        const result = validateCanvas(canvas);
-        set({ validation: result });
-        return result;
-      },
-
-      // Data loading
-      setAvailableSkills: (skills) => set({ availableSkills: skills }),
-      setAvailableHands: (hands) => set({ availableHands: hands }),
-
-      // Canvas metadata
-      updateCanvasMetadata: (updates) => {
-        const { canvas } = get();
-        if (!canvas) return;
-        set({ canvas: { ...canvas, ...updates }, isDirty: true });
-      },
-    }),
-    {
-      name: 'workflow-builder-storage',
-      partialize: (state) => ({
-        workflows: state.workflows,
-        templates: state.templates,
-      }),
-    }
-  )
-);
-
-// =============================================================================
-// Node Palette Items
-// =============================================================================
-
-export const nodePaletteItems: NodePaletteItem[] = [
-  // Input category
-  {
-    type: 'input',
-    label: 'Input',
-    description: 'Define workflow input variables',
-    icon: '📥',
-    category: 'input',
-    defaultData: { variableName: 'input' },
-  },
-
-  // AI category
-  {
-    type: 'llm',
-    label: 'LLM Generate',
-    description: 'Generate text using LLM',
-    icon: '🤖',
-    category: 'ai',
-    defaultData: { template: '', jsonMode: false },
-  },
-  {
-    type: 'skill',
-    label: 'Skill',
-    description: 'Execute a skill',
-    icon: '⚡',
-    category: 'ai',
-    defaultData: { skillId: '', inputMappings: {} },
-  },
-  {
-    type: 'orchestration',
-    label: 'Skill Orchestration',
-    description: 'Execute multiple skills in a DAG',
-    icon: '🔀',
-    category: 'ai',
-    defaultData: { inputMappings: {} },
-  },
-
-  // Action category
-  {
-    type: 'hand',
-    label: 'Hand',
-    description: 'Execute a hand action',
-    icon: '✋',
-    category: 'action',
-    defaultData: { handId: '', action: '', params: {} },
-  },
-  {
-    type: 'http',
-    label: 'HTTP Request',
-    description: 'Make an HTTP request',
-    icon: '🌐',
-    category: 'action',
-    defaultData: { url: '', method: 'GET', headers: {} },
-  },
-  {
-    type: 'setVar',
-    label: 'Set Variable',
-    description: 'Set a variable value',
-    icon: '📝',
-    category: 'action',
-    defaultData: { variableName: '', value: '' },
-  },
-  {
-    type: 'delay',
-    label: 'Delay',
-    description: 'Pause execution',
-    icon: '⏱️',
-    category: 'action',
-    defaultData: { ms: 1000 },
-  },
-
-  // Control category
-  {
-    type: 'condition',
-    label: 'Condition',
-    description: 'Branch based on condition',
-    icon: '🔀',
-    category: 'control',
-    defaultData: { condition: '', branches: [{ when: '', label: 'Branch' }] },
-  },
-  {
-    type: 'parallel',
-    label: 'Parallel',
-    description: 'Execute in parallel',
-    icon: '⚡',
-    category: 'control',
-    defaultData: { each: '${inputs.items}', maxWorkers: 4 },
-  },
-  {
-    type: 'loop',
-    label: 'Loop',
-    description: 'Iterate over items',
-    icon: '🔄',
-    category: 'control',
-    defaultData: { each: '${inputs.items}', itemVar: 'item', indexVar: 'index' },
-  },
-
-  // Output category
-  {
-    type: 'export',
-    label: 'Export',
-    description: 'Export to file formats',
-    icon: '📤',
-    category: 'output',
-    defaultData: { formats: ['json'] },
-  },
-];
-
-// Group palette items by category
-export const paletteCategories: Record<NodeCategory, NodePaletteItem[]> = {
-  input: nodePaletteItems.filter(i => i.category === 'input'),
-  ai: nodePaletteItems.filter(i => i.category === 'ai'),
-  action: nodePaletteItems.filter(i => i.category === 'action'),
-  control: nodePaletteItems.filter(i => i.category === 'control'),
-  output: nodePaletteItems.filter(i => i.category === 'output'),
-};
--- a/desktop/tests/store/handStore.test.ts
+++ b/desktop/tests/store/handStore.test.ts
@@ -115,14 +115,13 @@ describe('handStore — loadHands', () => {
      hands: [
        { id: 'h1', name: 'browser', description: 'Web automation', status: 'idle', requirements_met: true, category: 'automation', icon: '🌐', tool_count: 5, metric_count: 2 },
        { id: 'h2', name: 'researcher', description: 'Deep research', status: 'running', requirements_met: true },
-        { id: 'h3', name: 'speech', description: 'TTS', requirements_met: false },
      ],
    });

    await useHandStore.getState().loadHands();

    const state = useHandStore.getState();
-    expect(state.hands).toHaveLength(3);
+    expect(state.hands).toHaveLength(2);
    expect(state.hands[0].name).toBe('browser');
    expect(state.hands[0].status).toBe('idle');
    expect(state.hands[0].toolCount).toBe(5);
--- a/desktop/tsconfig.json
+++ b/desktop/tsconfig.json
@@ -22,9 +22,5 @@
    "useUnknownInCatchVariables": true
  },
  "include": ["src"],
-  "exclude": [
-    "src/components/ui/ErrorAlert.tsx",
-    "src/components/ui/ErrorBoundary.tsx"
-  ],
  "references": [{ "path": "./tsconfig.node.json" }]
 }
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -16,7 +16,7 @@ services:

    environment:
      POSTGRES_USER: ${POSTGRES_USER:-postgres}
-      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-your_secure_password}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?POSTGRES_PASSWORD must be set in .env}
      POSTGRES_DB: ${POSTGRES_DB:-zclaw}
      # 确保 UTF-8 编码 — 中文 Windows 默认 GBK 会导致中文数据损坏
      POSTGRES_INITDB_ARGS: "--encoding=UTF8 --locale=C.UTF-8"
@@ -53,7 +53,7 @@ services:
      - .env

    environment:
-      DATABASE_URL: postgres://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:-your_secure_password}@postgres:5432/${POSTGRES_DB:-zclaw}
+      DATABASE_URL: postgres://${POSTGRES_USER:-postgres}:${POSTGRES_PASSWORD:?POSTGRES_PASSWORD must be set in .env}@postgres:5432/${POSTGRES_DB:-zclaw}

    depends_on:
      postgres:
--- a/docs/TRUTH.md
+++ b/docs/TRUTH.md
@@ -1,7 +1,7 @@
 # ZCLAW 系统真相文档

-> **更新日期**: 2026-04-16
-> **数据来源**: V11 全面审计 + 二次审计 + V12 模块化端到端审计 + 代码全量扫描验证 + 功能测试 Phase 1-5 + 发布前功能测试 Phase 3 + 发布前全面测试代码级审计 + 2026-04-11 代码验证 + V13 系统性功能审计 2026-04-12 + V13 审计修复 2026-04-13 + 发布前冲刺 Day1 2026-04-15 + 发布前深度测试 8 路并行代码级验证 2026-04-16
+> **更新日期**: 2026-04-18
+> **数据来源**: V11 全面审计 + 二次审计 + V12 模块化端到端审计 + 代码全量扫描验证 + 功能测试 Phase 1-5 + 发布前功能测试 Phase 3 + 发布前全面测试代码级审计 + 2026-04-11 代码验证 + V13 系统性功能审计 2026-04-12 + V13 审计修复 2026-04-13 + 发布前冲刺 Day1 2026-04-15 + 发布前深度测试 8 路并行代码级验证 2026-04-16 + 发布前审计 2026-04-18
 > **规则**: 此文档是唯一真相源。所有其他文档如果与此冲突，以此为准。

 ---
@@ -12,7 +12,7 @@
 |------|--------|----------|
 | Rust Crates | 10 个 (编译通过) | `cargo check --workspace` |
 | Rust 代码行数 | ~77,000 (crates) + ~61,400 (src-tauri) = ~138,400 | wc -l (2026-04-12 V13 验证) |
-| Rust 单元测试 | 433 个 (#[test]) + 368 个 (#[tokio::test]) = 801 | `grep '#\[test\]' crates/` + `grep '#\[tokio::test\]'` (2026-04-12 V13 验证) |
+| Rust 单元测试 | 477 个 (#[test]) + 326 个 (#[tokio::test]) = 803 | `grep '#\[test\]' crates/` + `grep '#\[tokio::test\]'` (2026-04-18 审计验证) |
 | Cargo Warnings (非 SaaS) | **0 个** (仅 sqlx-postgres 外部依赖 1 个) | `cargo check --workspace --exclude zclaw-saas` (2026-04-15 清零) |
 | Rust 测试运行通过 | 684 workspace + 138 SaaS = 822 | Hermes 4 Chunk `cargo test --workspace` 2026-04-09 |
 | Tauri 命令 | 190 个 | `grep '#\[.*tauri::command'` (2026-04-16 验证) |
@@ -20,24 +20,25 @@
 | **Tauri 命令已标注 @reserved** | **89 个** | Rust 源码 @reserved 标注 (2026-04-15 全量标注) |
 | **Tauri 命令孤儿 (无调用+无标注)** | **~0 个** (190 - 104 invoke - 89 @reserved ≈ -3，差异来自内部命令调用) | (2026-04-16 校准) |
 | SKILL.md 文件 | 75 个 | `ls skills/*.md \| wc -l` |
-| Hands 启用 | 9 个 | Browser/Collector/Researcher/Clip/Twitter/Whiteboard/Slideshow/Speech/Quiz（均有 HAND.toml） |
+| Hands 启用 | 7 个 (6 HAND.toml + 1 系统内部 _reminder) | Browser/Collector/Researcher/Clip/Twitter/Quiz = 6 个有 HAND.toml；ReminderHand 通过 kernel 编程注册（`_` 前缀豁免 HAND.toml 扫描） |
+| Hands 开发中 | 3 个 | Whiteboard/Slideshow/Speech（HAND.toml 仅存在于 worktree 开发分支，无 `impl Hand for`，未合并到主分支） |
 | Hands 禁用 | 2 个 | Predictor, Lead（概念定义存在，无 TOML 配置文件或 Rust 实现） |
-| Pipeline 模板 | 17 个 YAML | `pipelines/` 目录全量统计（含 _templates/ 和 design-shantou/ 子目录） |
+| Pipeline 模板 | 18 个 YAML | `pipelines/` 目录全量统计 (2026-04-18 验证) |
 | SaaS API 端点 | 137 个 .route() | `grep .route( crates/zclaw-saas/` (2026-04-16 验证) |
 | SaaS 路由模块 | 12 个 + industry | account/agent_template/auth/billing/knowledge/migration/model_config/prompt/relay/role/scheduled_task/telemetry/industry（scheduled_task: 后端 5 CRUD + Admin V2 前端 service/page/route/nav） |
 | SaaS 数据表 | 34 个（含 saas_schema_version） | CREATE TABLE 全量统计 |
 | SaaS Workers | 7 个 | log_operation/cleanup_rate_limit/cleanup_refresh_tokens/record_usage/update_last_used/aggregate_usage/generate_embedding |
 | LLM Provider | 8 个 | Kimi/Qwen/DeepSeek/Zhipu/OpenAI/Anthropic/Gemini/Local |
-| Zustand Store | 21 个 | find desktop/src/store/ -name "*.ts" (2026-04-12 V13 验证) |
+| Zustand Store | 25 个 | find desktop/src/store/ -name "*.ts" (2026-04-18 审计，workflowBuilderStore 已删除) |
 | React 组件 | 105 个 (.tsx/.ts) | find desktop/src/components/ (2026-04-15 新增 HealthPanel.tsx) |
 | 前端 TypeScript 测试 | 31 个文件 (6 store + 5 lib + 1 config + 1 stabilization + 18 E2E spec) | Phase 3-4 全量 |
 | 前端 lib | 76 个 .ts | find desktop/src/lib/ (2026-04-15 删除 intelligence-client/ 9 文件) |
 | 前端测试运行通过 | 344 passed + 1 skipped | `pnpm vitest run` (2026-04-15 验证) |
 | 生产构建 | **通过** (14.8s, 0 require 残留) | `pnpm build` (2026-04-15 验证) |
-| Admin V2 页面 | 15 个 | admin-v2/src/pages/ 全量统计（含 ScheduledTasks、ConfigSync） |
+| Admin V2 页面 | 17 个 | admin-v2/src/pages/ 全量统计 (2026-04-18 验证) |
 | 桌面端设置页面 | 19 个 | SettingsLayout.tsx tabs: 通用/用量统计/积分详情/模型与API/MCP服务/技能/IM频道/工作区/数据与隐私/安全存储/SaaS平台/订阅与计费/语义记忆/安全状态/审计日志/定时任务/心跳配置/提交反馈/关于 |
 | Admin V2 测试 | 17 个文件 (61 tests) | vitest 统计 |
-| 中间件层 | 14 层 | `grep chain.register kernel/mod.rs` (2026-04-16 验证: ButlerRouter@80, DataMasking@90, Compaction@100, Memory@150, Title@180, SkillIndex@200, DanglingTool@300, ToolError@350, ToolOutputGuard@360, Guardrail@400, LoopGuard@500, SubagentLimit@550, TrajectoryRecorder@650, TokenCalibration@700) |
+| 中间件层 | 15 层 | `grep chain.register kernel/mod.rs` (2026-04-19 校准: EvolutionMiddleware@78, ButlerRouter@80, DataMasking@90, Compaction@100, Memory@150, Title@180, SkillIndex@200, DanglingTool@300, ToolError@350, ToolOutputGuard@360, Guardrail@400, LoopGuard@500, SubagentLimit@550, TrajectoryRecorder@650, TokenCalibration@700) |

 ---

@@ -205,3 +206,5 @@ Viking 5 个孤立 invoke 调用已于 2026-04-03 清理移除：
 | 2026-04-12 | V13 系统性功能审计数字校准：(1) Tauri 命令 184→191 (2) 前端 invoke 105→106 (3) @reserved 33→24 (Butler/MCP已接通) (4) 孤儿命令 ~46→~61 (5) Rust 测试 798→801 (433+368) (6) SaaS .route() 122→136 (7) Zustand Store 20→21 (8) dead_code 76→43 (9) Rust LOC crates ~74.6K→~77K |
 | 2026-04-15 | Heartbeat 统一健康系统：(1) Tauri 命令 182→183 (+health_snapshot) (2) intelligence 模块 15→16 文件 (+health_snapshot.rs +heartbeat.rs 重构) (3) React 组件 104→105 (+HealthPanel.tsx) (4) 前端 lib 85→76 (删除 intelligence-client/ 9 文件) |
 | 2026-04-16 | 发布前深度测试 8 路并行验证 + 3 项 P0 修复：(1) Tauri 命令 183→190 (2) 前端 invoke 95→104 (3) SaaS .route() 136→137 (4) 中间件 15→14 (实际 chain.register 计数) (5) P0-01 Admin ApiKeys 创建功能修复 (/keys→/tokens 路由对齐) (6) P0-02 账户锁定 unwrap_or(false)→正确错误传播 (7) P0-03 Logout 增加 access token cookie fallback 撤销 refresh token |
+| 2026-04-18 | 发布前审计数字校准 + Batch 1 修复：(1) Rust 测试 801→734 (#[test] 433→425 + #[tokio::test] 368→309) (2) Zustand Store 21→26 (3) Admin V2 页面 15→17 (4) Pipeline YAML 17→18 (5) Hands 启用 9→7 (6 HAND.toml + ReminderHand，Whiteboard/Slideshow/Speech 标注开发中) (6) Pipeline executor 内存泄漏 cleanup + 步骤超时 + Delay 上限 (7) Director send_to_agent oneshot channel 重构防死锁 (8) cleanup_rate_limit Worker 实现 (DELETE >1h) |
+| 2026-04-19 | 全系统穷尽审计 Batch 0 校准：(1) 中间件层 14→15 (补 EvolutionMiddleware@78，实际 chain.register 计数) (2) Zustand Store 确认 25 个 .ts 文件 (04-18 日志写 26 为误记) (3) wiki/middleware.md 同步 15 层 + 优先级分类更新 |
--- a/wiki/index.md
+++ b/wiki/index.md
@@ -24,15 +24,15 @@ status: active
 |------|-----|----------|
 | Rust Crates | 10 + src-tauri | `ls crates/zclaw-*/Cargo.toml` |
 | Rust 代码 | 77,811 行 (275 .rs文件) | `find crates/ src-tauri/ -name "*.rs"` |
-| Rust 测试 | 801 (433 #[test] + 368 #[tokio::test]) | `grep '#\[test\]' / '#\[tokio::test\]'` (TRUTH.md 04-16) |
+| Rust 测试 | 734 (425 #[test] + 309 #[tokio::test]) | `grep '#\[test\]' / '#\[tokio::test\]'` (TRUTH.md 04-18) |
 | Tauri 命令 | 190 定义 (103 src-tauri + 76 crates + 内部) | `grep '#\[tauri::command\]'` (TRUTH.md 04-16) |
 | 前端 invoke 调用 | 104 处 | `grep invoke( desktop/src/` (TRUTH.md 04-16) |
 | SaaS .route() | 137 个 | `grep .route( crates/zclaw-saas/` |
 | SaaS 模块 | 17 个目录 | `ls crates/zclaw-saas/src/*/` |
 | SKILL 目录 | 75 个 | `ls -d skills/*/` |
-| HAND 配置 | 9 个 + 1 系统内部 (_reminder) (TOML) | `ls hands/*.HAND.toml` |
-| Pipeline YAML | 17 个 | `find pipelines/ -name "*.yaml"` |
-| Zustand Store | 17 文件 + chat/4子store = 21 (含 industryStore) | `find desktop/src/store/` |
+| HAND 配置 | 6 TOML + 1 系统内部 (_reminder) = 7 注册 | `ls hands/*.HAND.toml` + kernel registry |
+| Pipeline YAML | 18 个 | `find pipelines/ -name "*.yaml"` |
+| Zustand Store | 26 个 (.ts, 含子目录) | `find desktop/src/store/` |
 | React 组件 | 105 个 (.tsx/.ts) | `find desktop/src/components/` (TRUTH.md 04-16) |
 | Admin V2 页面 | 17 个 (.tsx) | `ls admin-v2/src/pages/` |
 | 中间件 | 14 层 runtime + 10 层 SaaS HTTP | `kernel/mod.rs` + `zclaw-saas middleware` |
--- a/wiki/log.md
+++ b/wiki/log.md
@@ -9,6 +9,36 @@ tags: [log, history]

 > Append-only 操作记录。格式: `## [日期] 类型 | 描述`

+## 2026-04-18 fix | 审计后续 3 项修复
+
+- Shell Hands 残留清理 3 处 (message.rs 注释/profiler 偏好/handStore mock)
+- FTS5 CJK 查询修复: sanitize_fts_query 从精确短语改为 token OR 组合
+- WASM HTTP 响应大小限制: Content-Length 预检 + 1MB 上限
+- zclaw-growth 集成测试 2/2 修复, 全量 651 测试 0 失败
+
+## 2026-04-18 fix | 深度审计修复 — WASM 安全 + 编译路径
+
+- CRITICAL: zclaw_file_read 路径遍历修复 (组件级过滤)
+- CRITICAL: zclaw_http_fetch SSRF 防护 (scheme 白名单 + 私有 IP 阻止)
+- CRITICAL: A2A 始终编译 (移除 zclaw-protocols a2a feature gate)
+- MEDIUM: FactCategory cfg(test) 导入修复
+- 移除 kernel/desktop multi-agent feature (不再控制任何代码)
+- 563 测试全通过
+
+## 2026-04-17 refactor | Phase 4A multi-agent feature gate 移除
+
+- 8 个文件移除 33 处 `#[cfg(feature = "multi-agent")]`
+- zclaw-kernel default features 新增 multi-agent，始终编译
+- A2A router、agents、adapters 代码不再条件编译
+
+## 2026-04-17 feat | Phase 4B WASM host 函数真实实现
+
+- zclaw_log: 读取 guest 内存字符串 + debug! 日志
+- zclaw_http_fetch: ureq v3 同步 GET (10s timeout, network_allowed 守卫)
+- zclaw_file_read: 沙箱 /workspace 读取 + 路径校验防逃逸
+- 新增 ureq v3 workspace 依赖 (wasm feature gated)
+- 25 测试全通过，workspace check 零错误
+
 ## 2026-04-17 refactor | Phase 3A loop_runner 双路径合并

 - middleware_chain 从 Option<MiddlewareChain> 改为 MiddlewareChain (Default = 空链)
--- a/wiki/middleware.md
+++ b/wiki/middleware.md
@@ -20,24 +20,27 @@ tags: [module, middleware, runtime]

 ## 代码逻辑

-### 14 层 Runtime 中间件（注册顺序见 `kernel/mod.rs:248-361`）
+### 15 层 Runtime 中间件（注册顺序见 `kernel/mod.rs:248-361`，执行按 priority 升序）

-| # | 中间件 | 文件 | 职责 | 注册条件 |
-|---|--------|------|------|----------|
-| 1 | ButlerRouter | `middleware/butler_router.rs` | 语义技能路由 + system prompt 增强 | 始终 |
-| 2 | DataMasking | `middleware/data_masking.rs` | 手机号/身份证等敏感数据脱敏 | 始终 |
-| 3 | Compaction | `middleware/compaction.rs` | 超阈值时压缩对话历史 | `compaction_threshold > 0` |
-| 4 | Memory | `middleware/memory.rs` | 对话后自动提取记忆 | 始终 |
-| 5 | LoopGuard | `middleware/loop_guard.rs` | 防止工具调用无限循环 | 始终 |
-| 6 | TokenCalibration | `middleware/token_calibration.rs` | Token 用量校准 | 始终 |
-| 7 | SkillIndex | `middleware/skill_index.rs` | 注入技能索引到 system prompt | `!skill_index.is_empty()` |
-| 8 | Title | `middleware/title.rs` | 自动生成会话标题 | 始终 |
-| 9 | DanglingTool | `middleware/dangling_tool.rs` | 修复缺失的工具调用结果 | 始终 |
-| 10 | ToolError | `middleware/tool_error.rs` | 格式化工具错误供 LLM 恢复 | 始终 |
-| 11 | ToolOutputGuard | `middleware/tool_output_guard.rs` | 工具输出安全检查 | 始终 |
-| 12 | Guardrail | `middleware/guardrail.rs` | shell_exec/file_write/web_fetch 安全规则 | 始终 |
-| 13 | SubagentLimit | `middleware/subagent_limit.rs` | 限制并发子 agent | 始终 |
-| 14 | TrajectoryRecorder | `middleware/trajectory_recorder.rs` | 轨迹记录 + 压缩 | 始终 (V13-FIX-01 已注册) |
+| # | 中间件 | 优先级 | 文件 | 职责 | 注册条件 |
+|---|--------|--------|------|------|----------|
+| 1 | EvolutionMiddleware | 78 | `middleware/evolution.rs` | 推送进化候选项到 system prompt | 始终 |
+| 2 | ButlerRouter | 80 | `middleware/butler_router.rs` | 语义技能路由 + system prompt 增强 | 始终 |
+| 3 | DataMasking | 90 | `middleware/data_masking.rs` | 手机号/身份证等敏感数据脱敏 | 始终 |
+| 4 | Compaction | 100 | `middleware/compaction.rs` | 超阈值时压缩对话历史 | `compaction_threshold > 0` |
+| 5 | Memory | 150 | `middleware/memory.rs` | 对话后自动提取记忆 + 进化检查 | 始终 |
+| 6 | Title | 180 | `middleware/title.rs` | 自动生成会话标题 | 始终 |
+| 7 | SkillIndex | 200 | `middleware/skill_index.rs` | 注入技能索引到 system prompt | `!skill_index.is_empty()` |
+| 8 | DanglingTool | 300 | `middleware/dangling_tool.rs` | 修复缺失的工具调用结果 | 始终 |
+| 9 | ToolError | 350 | `middleware/tool_error.rs` | 格式化工具错误供 LLM 恢复 | 始终 |
+| 10 | ToolOutputGuard | 360 | `middleware/tool_output_guard.rs` | 工具输出安全检查 | 始终 |
+| 11 | Guardrail | 400 | `middleware/guardrail.rs` | shell_exec/file_write/web_fetch 安全规则 | 始终 |
+| 12 | LoopGuard | 500 | `middleware/loop_guard.rs` | 防止工具调用无限循环 | 始终 |
+| 13 | SubagentLimit | 550 | `middleware/subagent_limit.rs` | 限制并发子 agent | 始终 |
+| 14 | TrajectoryRecorder | 650 | `middleware/trajectory_recorder.rs` | 轨迹记录 + 压缩 | 始终 |
+| 15 | TokenCalibration | 700 | `middleware/token_calibration.rs` | Token 用量校准 | 始终 |
+
+> **注意**: 注册顺序（代码中的 chain.register 调用顺序）与执行顺序不同。Chain 按 priority 升序排列后执行。

 ### 10 层 SaaS HTTP 中间件（`zclaw-saas/src/main.rs`）

@@ -58,10 +61,12 @@ tags: [module, middleware, runtime]

 | 范围 | 类别 | 包含的中间件 |
 |------|------|-------------|
+| 70-79 | 进化 | EvolutionMiddleware |
+| 80-99 | 路由+安全 | ButlerRouter, DataMasking |
 | 100-199 | 上下文塑造 | Compaction, Memory |
-| 200-399 | 能力 | SkillIndex, Guardrail |
-| 400-599 | 安全 | LoopGuard, DataMasking |
-| 600-799 | 遥测 | TokenCalibration, Title, TrajectoryRecorder |
+| 200-399 | 能力 | SkillIndex, DanglingTool, ToolError, ToolOutputGuard |
+| 400-599 | 安全 | Guardrail, LoopGuard, SubagentLimit |
+| 600-799 | 遥测 | TrajectoryRecorder, TokenCalibration, Title |

 ### 中间件执行流

@@ -97,7 +102,7 @@ trait AgentMiddleware: Send + Sync {

 ### 注册位置

-`crates/zclaw-kernel/src/kernel/mod.rs:248-361` — `create_middleware_chain()` 方法，14 次 `chain.register()` + 1 个条件注册 (SkillIndex)。
+`crates/zclaw-kernel/src/kernel/mod.rs:248-361` — `create_middleware_chain()` 方法，15 次 `chain.register()`（含 2 个条件注册: SkillIndex, Compaction）。注册顺序与执行顺序不同，chain 按 priority 升序排列后执行。

 ## 关联模块

@@ -111,6 +116,6 @@ trait AgentMiddleware: Send + Sync {
 | 文件 | 职责 |
 |------|------|
 | `crates/zclaw-runtime/src/middleware.rs` | AgentMiddleware trait + MiddlewareChain |
-| `crates/zclaw-runtime/src/middleware/` | 14 个中间件实现 (14个 .rs 文件) |
+| `crates/zclaw-runtime/src/middleware/` | 15 个中间件实现 (15个 .rs 文件) |
 | `crates/zclaw-kernel/src/kernel/mod.rs:248-361` | 注册入口 |
 | `crates/zclaw-saas/src/main.rs` | SaaS HTTP 中间件注册 (10 层) |