docs: add architecture & flow diagram (SVG)

docs/assets/architecture.svg:云端规划 → 端侧四层(意图桥/LFM Engine/Policy Guard/Redactor)
→ 私有基础设施,只回传脱敏结果的 loop。中英 README 架构段改用该矢量图(标签精确、GitHub 内联渲染、
配色对齐 banner)。

Author: xiaokhkh

README.md

@@ -105,24 +105,15 @@ the guard + mode, returning **redacted** output), `policy_check`, `local_context
 The server's autonomy is set by `SENTINEL_MODE` (default `readonly`); the MCP client's own
 tool-approval prompt is the human gate for `ask`-tier (mutating) steps.
 
-## Architecture — three-layer filter
+## Architecture & flow
 
-```
-        guard run "diagnose not-ready pods in default"
-                          │
-   ① Intent Bridge   ─────▼─────  natural-language entrypoint (CLI / MCP)
-   ② LFM Engine      ─────▼─────  local inference, data stays on device
-        • Local RAG: reads ~/.kube, ~/.ssh as background (non-secret only)
-        • intent alignment: intent -> shell / kubectl
-        • providers: ollama | llamacpp | mlx | mock (OpenAI-compatible)
-   ③ Policy Guard    ─────▼─────  allow / confirm / block
-        • regex + semantic interception (drop / --all / rm -rf ...)
-        • human-in-the-loop confirmation
-                          │
-                    run / refuse / downgrade
-```
+<p align="center">
+  <img src="docs/assets/architecture.svg" width="760" alt="Sentinel-Agent architecture and flow">
+</p>
 
-Inference backends are decoupled from orchestration: every backend speaks the OpenAI-compatible
+The cloud planner sends only an intent; the on-device layers (Intent Bridge → LFM Engine →
+Policy Guard → Redactor) plan and execute locally, and only the **desensitized** result loops
+back. Inference backends are decoupled from orchestration: every backend speaks the OpenAI-compatible
 protocol, so switching is configuration, not code. See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)
 and the [Tool Call protocol](docs/tool-call-protocol.md).
 

README.zh-CN.md

@@ -99,24 +99,14 @@ go build -o bin/guard ./cmd/guard
 `policy_check`、`local_context`、`list_skills`。服务自治级别由 `SENTINEL_MODE` 决定（默认 `readonly`）；
 对 `ask` 级（变更类）步骤，MCP 客户端自身的工具调用审批弹窗即人工门。
 
-## 三层过滤架构
+## 架构与流程
 
-```
-        guard run "诊断 default 里未就绪的 pod"
-                          │
-   ① 意图接口层  ─────▼─────  自然语言入口 (CLI / MCP)
-   ② 端侧推理芯  ─────▼─────  本地推理，数据不出域
-        • Local RAG：读取 ~/.kube、~/.ssh 作背景（仅非密钥信息）
-        • 意图对齐：意图 -> shell / kubectl
-        • Provider：ollama | llamacpp | mlx | mock（OpenAI 兼容）
-   ③ 安全执行围栏 ─────▼─────  allow / confirm / block
-        • 正则 + 语义拦截（drop / --all / rm -rf …）
-        • Human-in-the-loop 确认
-                          │
-                    执行 / 拒绝 / 降级
-```
+<p align="center">
+  <img src="docs/assets/architecture.svg" width="760" alt="Sentinel-Agent 架构与流程">
+</p>
 
-推理后端与编排层解耦：所有后端走 OpenAI 兼容协议，换后端是改配置而非改代码。
+云端规划者只下发意图；端侧四层（意图接口层 → 端侧推理芯 → 安全围栏 → Redactor）在本地规划并执行，
+只有**脱敏后**的结果回环。推理后端与编排层解耦：所有后端走 OpenAI 兼容协议，换后端是改配置而非改代码。
 详见 [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) 与 [Tool Call 协议](docs/tool-call-protocol.md)。
 
 ## 推理后端