@@ -56,7 +56,7 @@ This project grew from a small demo into an indispensable "productivity assistan
 | **Terminal Mode**            | Smooth TUI in terminal                                                               |
 | **Smart Context Management**  | Auto-compact long session context, configurable threshold, millisecond-level token estimation         |
 | **Skills System**             | Skills integration, auto-discover `SKILL.md`, activate by scenario                                    |
-| **Deep MCP Integration**      | Local/remote MCP servers, OAuth login, runtime dynamic switching                                      |
+| **Deep MCP Integration**      | Local/remote MCP servers, OAuth login, per-session dynamic switching                                   |
 | **Enterprise-Grade Security** | Tool approval system (auto-approve/manual-approve), supports once/session/deny modes                  |
 | **OpenAI Compatible**         | Works with any OpenAI-compatible API, flexible multi-Provider configuration                           |
 
@@ -103,19 +103,18 @@ First run will guide you through Provider/Model setup and save config to `~/.mem
 ```
 memo-code/
 ├── packages/
-│   ├── core/          # Core logic: Session state machine, Config handling
-│   ├── tools/         # Tool routing, MCP Client management, built-in tools (exec_command, read_text_file, apply_patch...)
-│   ├── tui/           # Terminal runtime: CLI entry, interactive TUI
-└── docs/              # Technical documentation
+│   ├── core/          # Agent engine: session state machine, LLM/tool loop, built-in tools, MCP client, skills
+│   └── tui/           # Terminal runtime: CLI entry, interactive TUI (Ink)
+└── site/              # Documentation website (Next.js, static export)
 ```
 
 **Technical Highlights:**
 
-- **Architecture**: Clean Core / Tools / TUI separation, state-machine driven session management
-- **Testing**: Core + Tools coverage > 70%, complete unit + integration tests
+- **Architecture**: Core engine with integrated tool routing, thin TUI on top, state-machine driven session management
+- **Testing**: Unit + integration tests, coverage threshold ≥70%
 - **Protocol**: Native MCP (Model Context Protocol) support, can integrate any MCP tool server
 - **Token Estimation**: Real-time context monitoring based on tiktoken, configurable auto-compaction strategy
-- **Distribution**: npm package with hot-reloading without perception
+- **Distribution**: Published to npm with version-driven auto releases via CI
 
 ---
 
@@ -125,9 +124,11 @@ memo-code/
 - `apply_patch` - Structured patch editing (`*** Begin Patch`/`*** End Patch`)
 - `read_text_file` / `read_media_file` / `read_files` / `write_file` / `edit_file` / `list_directory` / `search_files` - Filesystem read/write/search
 - `webfetch` - Paged web fetching with markdown extraction and policy guards
-- MCP resource access - `list_mcp_resources`, `read_mcp_resource`
+- MCP resource access - `list_mcp_resources`, `list_mcp_resource_templates`, `read_mcp_resource`
 - `update_plan` - Structured task progress management
+- `read_skill` - Load skill instructions on demand
 - `get_memory` - Persistent memory reading
+- Agent collaboration - `spawn_agent` / `send_message` / `followup_task` / `wait_agent` / `interrupt_agent` / `list_agents`
 
 ---
 
@@ -148,8 +149,8 @@ base_url = "https://api.openai.com/v1"
 command = "npx"
 args = ["-y", "@modelcontextprotocol/server-github"]
 
-# Skills
-active_skills = ["./skills/doc-writing/SKILL.md"]
+# Skills (absolute paths to SKILL.md files)
+active_skills = ["/path/to/skills/doc-writing/SKILL.md"]
 ```
 
 ---
 
@@ -52,7 +52,7 @@ Memo 诞生于一个简单的想法：**我想验证一个最简单的 Agent 是
 | **终端模式**         | 终端 TUI 交互流畅                                                  |
 | **智能上下文管理**  | 自动压缩长会话上下文，支持配置压缩阈值，毫秒级 token 估算               |
 | **Skills 技能系统** | Skills 技能集成，自动发现 `SKILL.md`，支持按场景激活                    |
-| **MCP 深度集成**    | 支持本地/远程 MCP 服务器，OAuth 登录，运行时动态切换                    |
+| **MCP 深度集成**    | 支持本地/远程 MCP 服务器，OAuth 登录，会话级动态切换                    |
 | **企业级安全**      | 工具分级审批机制（自动批准/手动批准），支持单次/会话/拒绝三种模式       |
 | **OpenAI 兼容**     | 支持任意 OpenAI 兼容 API，灵活配置多 Provider 切换                      |
 
@@ -93,29 +93,30 @@ memo
 ```
 memo-code/
 ├── packages/
-│   ├── core/          # 核心逻辑：Session 状态机、Config 处理
-│   ├── tools/         # Tool 路由、MCP Client管理、内置工具实现（exec_command, read_text_file, apply_patch...）
-│   ├── tui/           # 终端运行时：CLI 入口、交互式 TUI
-└── docs/              # 技术文档
+│   ├── core/          # 核心引擎：Session 状态机、LLM/工具循环、内置工具、MCP 客户端、技能
+│   └── tui/           # 终端运行时：CLI 入口、交互式 TUI (Ink)
+└── site/              # 文档网站（Next.js 静态导出）
 ```
 
 **技术亮点：**
 
-- **架构**：清晰的 Core / Tools / TUI 分层，状态机驱动会话管理
-- **测试**：Core + Tools 覆盖率 > 70%，完整的单元 + 集成测试
+- **架构**：核心引擎内置工具路由，TUI 薄壳，状态机驱动会话管理
+- **测试**：单元 + 集成测试，覆盖率门槛 ≥70%
 - **协议**：原生支持 MCP (Model Context Protocol)，可接入任意 MCP 工具服务器
 - **Token 估算**：基于 tiktoken 的实时上下文监控，支持可配置的自动压缩策略
-- **分发**：npm 包热加载无感知
+- **分发**：发布至 npm，CI 版本驱动自动发版
 
 ## 🔧 内置工具
 
 - `exec_command` / `write_stdin` - 执行 Shell 命令
-- `apply_patch` - 字符串级代码编辑（单文件/批量）
+- `apply_patch` - 结构化补丁编辑（`*** Begin Patch`/`*** End Patch`）
 - `read_text_file` / `read_media_file` / `read_files` / `write_file` / `edit_file` / `list_directory` / `search_files` - 文件系统读写与检索
 - `webfetch` - 支持分页、Markdown 提取与策略防护的网页抓取
-- MCP 资源访问 - `list_mcp_resources`, `read_mcp_resource`
+- MCP 资源访问 - `list_mcp_resources`, `list_mcp_resource_templates`, `read_mcp_resource`
 - `update_plan` - 结构化任务进度管理
+- `read_skill` - 按需加载技能指令
 - `get_memory` - 持久化记忆读取
+- Agent 协作 - `spawn_agent` / `send_message` / `followup_task` / `wait_agent` / `interrupt_agent` / `list_agents`
 
 ## ⚙️ 配置示例
 
@@ -134,8 +135,8 @@ base_url = "https://api.openai.com/v1"
 command = "npx"
 args = ["-y", "@modelcontextprotocol/server-github"]
 
-# Skills
-active_skills = ["./skills/doc-writing/SKILL.md"]
+# Skills（SKILL.md 的绝对路径）
+active_skills = ["/path/to/skills/doc-writing/SKILL.md"]
 ```
 
 ---
 
@@ -23,9 +23,9 @@
         "start": "tsx packages/tui/src/cli.tsx",
         "build": "tsup",
         "dev": "tsup --watch",
-        "site:dev": "pnpm --filter @memo-code/site dev",
-        "site:build": "pnpm --filter @memo-code/site build",
-        "site:start": "pnpm --filter @memo-code/site start",
+        "site:dev": "pnpm --dir site dev",
+        "site:build": "pnpm --dir site build",
+        "site:start": "pnpm --dir site start",
         "format": "oxfmt packages site package.json pnpm-workspace.yaml tsconfig.json tsup.config.ts vitest.config.ts vitest.setup.ts .oxfmtrc.json",
         "format:check": "oxfmt --check packages site package.json pnpm-workspace.yaml tsconfig.json tsup.config.ts vitest.config.ts vitest.setup.ts .oxfmtrc.json",
         "lint": "oxlint packages site",
@@ -67,7 +67,7 @@
         "string-width": "^7.2.0",
         "toml": "^3.0.0",
         "turndown": "^7.2.2",
-        "undici": "^6.23.0",
+        "undici": "^6.28.0",
         "zod": "^4.3.6"
     }
 }
@@ -26,7 +26,8 @@
         "diff": "^8.0.3",
         "ignore": "^7.0.5",
         "js-tiktoken": "^1.0.21",
-        "minimatch": "^10.0.1",
+        "minimatch": "^10.2.6",
+        "undici": "^6.28.0",
         "zod": "^4.3.6"
     },
     "devDependencies": {
 
@@ -0,0 +1,19 @@
+import { describe, expect, test } from 'vitest'
+import { InputQueue } from './communication'
+
+describe('InputQueue', () => {
+    test('wakes event-driven waiters without polling', async () => {
+        const queue = new InputQueue()
+        const waiting = queue.waitForActivity(1_000)
+        queue.enqueue({ author: '/root/a', recipient: '/root', content: 'done', triggerTurn: false })
+        await expect(waiting).resolves.toBe('mailbox')
+    })
+
+    test('only drains a triggered batch when it contains trigger work', () => {
+        const queue = new InputQueue()
+        queue.enqueue({ author: '/root', recipient: '/root/a', content: 'note', triggerTurn: false })
+        expect(queue.drainTriggeredBatch()).toEqual([])
+        queue.enqueue({ author: '/root', recipient: '/root/a', content: 'continue', triggerTurn: true })
+        expect(queue.drainTriggeredBatch().map((message) => message.content)).toEqual(['note', 'continue'])
+    })
+})
@@ -0,0 +1,89 @@
+export type InterAgentCommunication = {
+    author: string
+    recipient: string
+    content: string
+    triggerTurn: boolean
+}
+
+export type InputQueueActivity = 'mailbox' | 'timeout' | 'closed' | 'aborted'
+
+type ActivityWaiter = (activity: InputQueueActivity) => void
+
+export class InputQueue {
+    private messages: InterAgentCommunication[] = []
+    private waiters = new Set<ActivityWaiter>()
+    private closed = false
+
+    enqueue(communication: InterAgentCommunication): void {
+        if (this.closed) throw new Error('agent mailbox is closed')
+        this.messages.push(communication)
+        this.publish('mailbox')
+    }
+
+    hasMessages(): boolean {
+        return this.messages.length > 0
+    }
+
+    hasTrigger(): boolean {
+        return this.messages.some((message) => message.triggerTurn)
+    }
+
+    drainAll(): InterAgentCommunication[] {
+        if (this.messages.length === 0) return []
+        const drained = this.messages
+        this.messages = []
+        return drained
+    }
+
+    drainTriggeredBatch(): InterAgentCommunication[] {
+        if (!this.hasTrigger()) return []
+        return this.drainAll()
+    }
+
+    async waitForActivity(timeoutMs: number, signal?: AbortSignal): Promise<InputQueueActivity> {
+        if (this.hasMessages()) return 'mailbox'
+        if (this.closed) return 'closed'
+        if (signal?.aborted) return 'aborted'
+
+        return new Promise<InputQueueActivity>((resolve) => {
+            let settled = false
+            const finish = (activity: InputQueueActivity) => {
+                if (settled) return
+                settled = true
+                clearTimeout(timer)
+                signal?.removeEventListener('abort', onAbort)
+                this.waiters.delete(finish)
+                resolve(activity)
+            }
+            const onAbort = () => finish('aborted')
+            const timer = setTimeout(() => finish('timeout'), timeoutMs)
+
+            this.waiters.add(finish)
+            signal?.addEventListener('abort', onAbort, { once: true })
+
+            // Recheck after subscribing so an enqueue cannot be missed at the boundary.
+            if (this.hasMessages()) finish('mailbox')
+            else if (this.closed) finish('closed')
+        })
+    }
+
+    close(): void {
+        if (this.closed) return
+        this.closed = true
+        this.publish('closed')
+    }
+
+    private publish(activity: InputQueueActivity): void {
+        for (const waiter of this.waiters) waiter(activity)
+    }
+}
+
+export function formatInterAgentCommunication(message: InterAgentCommunication): string {
+    return `<agent_message author="${escapeAttribute(message.author)}" recipient="${escapeAttribute(
+        message.recipient,
+    )}">\n${message.content}\n</agent_message>`
+}
+
+function escapeAttribute(value: string): string {
+    return value.replace(/&/g, '&amp;').replace(/"/g, '&quot;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
+}
@@ -35,8 +35,9 @@ describe('compact_prompt', () => {
         const prompt = buildCompactionUserPrompt(messages)
         assert.ok(prompt.includes('[0] ASSISTANT (tool_calls: exec_command)'))
         assert.ok(prompt.includes('[1] TOOL (exec_command)'))
-        // Truncation keeps the tail (tool result/error at the end carries the info).
-        assert.ok(prompt.includes(`...${'x'.repeat(4_000)}`))
+        // Long tool output is NOT truncated per-message: the budget selector is
+        // the only truncation point, so the compaction model sees full outputs.
+        assert.ok(prompt.includes(longToolOutput))
         assert.ok(prompt.includes('Return only the summary body in plain text. Do not add markdown fences.'))
     })
 
@@ -93,13 +94,80 @@ describe('compact_prompt', () => {
         )
     })
 
-    test('selectCompactionMessages keeps the newest message even when it exceeds the budget', () => {
+    test('selectCompactionMessages truncates a newest message that alone exceeds the budget', () => {
         const messages: ChatMessage[] = [
             { role: 'user', content: 'old' },
             { role: 'assistant', content: 'x'.repeat(500) },
         ]
         const selected = selectCompactionMessages(messages, 10, (text) => text.length)
-        assert.deepStrictEqual(selected, [messages[1]])
+        assert.strictEqual(selected.length, 1)
+        const content = selected[0]?.content
+        assert.strictEqual(typeof content, 'string')
+        // Tail kept (with the truncation marker), head dropped.
+        assert.ok(String(content).startsWith('...'))
+        assert.ok(String(content).endsWith('x'))
+        assert.ok(String(content).length < 500)
+    })
+
+    test('plan from update_plan tool results survives into the compaction prompt', () => {
+        const planJson = JSON.stringify({
+            message: 'Plan updated',
+            plan: [
+                { step: 'Implement the parser', status: 'in_progress' },
+                { step: 'Wire up the CLI flag', status: 'pending' },
+                { step: 'Add tests for edge cases', status: 'pending' },
+            ],
+        })
+        const messages: ChatMessage[] = [
+            { role: 'user', content: 'Refactor the parser' },
+            {
+                role: 'assistant',
+                content: [{ type: 'text', text: 'Let me update the plan' }],
+            },
+            {
+                role: 'tool',
+                content: [
+                    {
+                        type: 'tool-result',
+                        toolCallId: 'plan-1',
+                        toolName: 'update_plan',
+                        output: { type: 'text', value: planJson },
+                    },
+                ],
+            },
+            { role: 'user', content: 'Continue with step two' },
+        ]
+
+        const selected = selectCompactionMessages(messages, 10_000, (text) => text.length)
+        const prompt = buildCompactionUserPrompt(selected)
+        assert.ok(prompt.includes('Implement the parser'), 'plan steps must reach the compaction model')
+        assert.ok(prompt.includes('Wire up the CLI flag'))
+        assert.ok(prompt.includes('update_plan'))
+    })
+
+    test('selectCompactionMessages truncates oversized tool results keeping the tail', () => {
+        const messages: ChatMessage[] = [
+            {
+                role: 'tool',
+                content: [
+                    {
+                        type: 'tool-result',
+                        toolCallId: 'call-1',
+                        toolName: 'exec_command',
+                        output: { type: 'text', value: `head-noise\n${'y'.repeat(1_000)}` },
+                    },
+                ],
+            },
+        ]
+        const selected = selectCompactionMessages(messages, 20, (text) => text.length)
+        assert.strictEqual(selected.length, 1)
+        const part = selected[0]?.content
+        assert.ok(Array.isArray(part))
+        const value = part?.[0]?.type === 'tool-result' ? part[0].output.value : ''
+        assert.ok(String(value).startsWith('...'))
+        assert.ok(String(value).endsWith('y'))
+        assert.ok(String(value).length < 1_000)
+        assert.ok(!String(value).includes('head-noise'), 'tool output head is dropped')
     })
 
     test('selectCompactionMessages returns empty for an empty array', () => {