dsh-acp-enhanced 0.1.0 → 0.2.0

package/README-en.md

@@ -13,6 +13,7 @@ controls. This project is a drop-in replacement that surfaces what the Web GUI h
 | Surface | ACP mechanism | What you see in Zed |
 |---|---|---|
 | **Block-level streaming** | `agent_message_chunk` per committed text block (`block-end`), grouped by `messageId` per model step | Text appears while the agent works; cancelled/retried blocks never leak torn output |
+| **Reasoning streaming** | reasoning blocks (`blockType: 'reasoning'`) commit the same way, sent as `agent_thought_chunk` | the model's thinking scrolls live in Zed's thinking area |
 | **Token & context telemetry** | standard `usage_update` (`used` = context pressure, `size` = model context window) | Context meter in the agent status bar |
 | **Cache hit rate / TPS / input-output-reasoning tokens / tool timing / turn count** | `usage_update._meta` + `tool_call` / `tool_call_update` `_meta` | Raw numbers every step (the `_meta` extension field carries the full breakdown) |
 | **Tool-call visibility** | `tool_call` carries `rawInput` (parsed arguments) and `kind` (read/edit/execute/…); `tool_call_update` carries `rawOutput` (result preview, capped at 12k chars) | Tool cards expand to show the **exact arguments** (e.g. the bash command) and the **result**, with kind-based icons |
@@ -26,19 +27,20 @@ controls. This project is a drop-in replacement that surfaces what the Web GUI h
 | **Session resume** | `loadSession` capability + `session/load` resumes the persisted agent via `agents.resume` and replays history as `user_message_chunk` / `agent_message_chunk` / `tool_call` | **Continue a previous thread** in Zed (long investigations keep their context) |
 | **Session archive list** | `sessionCapabilities.list/delete`; `session/list` enumerates persisted sessions via `ctx.sessionPersistence.list()` (titles read from `session/title` events in the stored log), `session/delete` disposes the live agent and removes its persisted directory; `session/title` / `turn/end` push live `session_info_update`s | The **thread archive** shows all sessions (titled, sorted by last activity) — click to resume, delete to remove |
 | **Empty-option suppression** | no `reasoning_effort` option is advertised when the routed model exposes no efforts | No empty, unclickable "Reasoning effort" chip |
+| **MCP servers** | `session/new` / `session/load` `mcpServers` mount `@deepseek-ai/dsh-mcp-client` per entry (stdio + streamable HTTP, resolved from the host dsh install for a single module instance); tools join as `mcp__<serverName>__<tool>`; a failing server never takes the session down | tools from any MCP server (database, browser, filesystem, …) land directly in the model's tool list |
 
 ## Preview
 
 After picking **dsh-acp-enhanced** in Zed's AI Agent panel, you get:
 
-![Approval popup, model/reasoning-effort switches, context ring](assets/screenshots/approval-config-context.png)
+<img src="assets/screenshots/approval-config-context.png" alt="Approval popup, model/reasoning-effort switches, context ring" width="560">
 
 - Tool calls that need permission pop a **native approval prompt** (allow-once /
   reject-once); below the input box sit the **model**, **reasoning effort**,
   **permission preset**, and **plan mode** config options plus the **context usage
   ring** (`usage_update` telemetry with cache hit rate, TPS, and more).
 
-![Tool call inputs and outputs, native Zed question form](assets/screenshots/tool-cards-elicitation.png)
+<img src="assets/screenshots/tool-cards-elicitation.png" alt="Tool call inputs and outputs, native Zed question form" width="320">
 
 - **Tool cards** expand to show each call's full arguments (e.g. the exact bash
   command) and the result preview (`rawInput` / `rawOutput`); when dsh needs your
@@ -206,6 +208,8 @@ dsh plugin --profile acp-enhanced add "link:/absolute/path/to/dsh-acp-enhanced/p
 ```sh
 node scripts/acp-client.mjs           # end-to-end smoke test (needs a routable provider)
 node scripts/acp-client-tools.mjs     # client-tool tests (mocks Zed fs/terminal/elicitation/plan)
+node scripts/acp-mcp-test.mjs         # MCP mount test (mounts a minimal stdio MCP server, no model calls)
+node scripts/acp-smoke-keyless.mjs    # keyless boot smoke (CI: auto-creates a profile → initialize → session/new)
 node scripts/acp-resume-test.mjs      # resume tests (two processes: create+persist → load+replay → continue)
 ACP_DEBUG=1 dsh --profile acp-enhanced   # lifecycle trace on stderr
 ```
@@ -218,8 +222,11 @@ switching, and `session/cancel`. `acp-client-tools.mjs` uses the SDK's
 `zed_*` tool calls arrive as `fs/write_text_file`, `fs/read_text_file`,
 `terminal/create` requests, that `ask_user_question` arrives as an `elicitation/create`
 form (with enum options), that the `plan_mode` boolean toggle emits ACP `plan` updates
-(on → entry, off → cleared), and that routes without reasoning efforts no longer
-advertise an empty `reasoning_effort`.
+(on → entry, off → cleared), and that the `reasoning_effort` option is route-conditional
+(present with non-empty options on routes with efforts, suppressed on routes without).
+`acp-mcp-test.mjs` uses `scripts/fixtures/mcp-echo-server.mjs` to verify that
+`session/new` `mcpServers` are really mounted (the server receives initialize and
+tools/list) and that an identical list is reused, not re-mounted.
 
 ## Design notes
 
@@ -327,8 +334,9 @@ advertise an empty `reasoning_effort`.
 - **Modes**: permission presets are presented as ACP session modes, so Zed's mode
   switcher drives the sandbox/approval presets.
 - **Known limitations** (inherited from the official bridge): baseline prompts only
-  (no image/audio/MCP attachments), no `additionalDirectories`/MCP server attachment,
-  committed text streams at block granularity, and one in-flight prompt per session.
+  (no image/audio attachments), no `additionalDirectories`, committed text streams at
+  block granularity, and one in-flight prompt per session. MCP servers (stdio +
+  streamable HTTP) are supported; legacy SSE and `acp` transports are not advertised.
   Session resume and the archive list are supported (see above), but
   `session/close` / `session/fork` / `session/resume` are not implemented (the
   capabilities are not declared, so conforming clients do not call them);

package/README.md

@@ -13,6 +13,7 @@
 | 能力 | ACP 机制 | 你在 Zed 中看到的 |
 |---|---|---|
 | **块级流式输出** | 每个已提交文本块（`block-end`）发送一条 `agent_message_chunk`，按每个模型 step 的 `messageId` 分组 | agent 工作时文本实时出现；被取消/重试的块不会残留撕裂的半截输出 |
+| **推理流式** | reasoning 块（`blockType: 'reasoning'`）同样按块提交，发送 `agent_thought_chunk` | 模型的思考过程实时滚动（Zed 的 thinking 区域） |
 | **Token 与上下文遥测** | 标准 `usage_update`（`used` = 上下文压力，`size` = 模型上下文窗口） | agent 状态栏中的上下文仪表 |
 | **缓存命中率 / TPS / 输入-输出-推理 token / 工具耗时 / 轮次计数** | `usage_update._meta` + `tool_call` / `tool_call_update` 的 `_meta` | 每一步都有原始数字（`_meta` 扩展字段携带完整明细） |
 | **工具调用可见性** | `tool_call` 携带 `rawInput`（解析后的参数对象）与 `kind`（read/edit/execute/…）；`tool_call_update` 携带 `rawOutput`（结果预览，最多 12k 字符） | 工具卡片能展开看到**具体参数**（如 bash 执行的命令）与**执行结果**，并按工具类型渲染图标 |
@@ -26,18 +27,19 @@
 | **会话恢复（resume）** | 声明 `loadSession` 能力 + `session/load` 走 `agents.resume` 加载持久化会话，并把历史回放为 `user_message_chunk` / `agent_message_chunk` / `tool_call` | 在 Zed 里可以**继续之前的对话线程**（长排查不丢上下文） |
 | **会话归档列表** | 声明 `sessionCapabilities.list/delete`；`session/list` 从持久化存储（`ctx.sessionPersistence.list()`）枚举会话（标题从存储日志的 `session/title` 事件读取），`session/delete` 释放在线 agent 并删除其持久化目录；`session/title` / `turn/end` 实时推送 `session_info_update` | Zed 的**历史线程归档**能看到本项目的全部会话（带标题、按更新时间排序），可点击恢复，也可删除 |
 | **空选项抑制** | 当前模型路由无 reasoning efforts 时不广播 `reasoning_effort` 配置项 | 不再出现一个空的、点不动的"Reasoning effort"chip |
+| **MCP servers** | `session/new` / `session/load` 的 `mcpServers` 逐项挂载 `@deepseek-ai/dsh-mcp-client`（stdio + streamable HTTP，从宿主 dsh 安装解析、保证单实例）；工具注册为 `mcp__<serverName>__<tool>`；失败的 server 不会拖垮会话 | 任意 MCP server 的工具（数据库、浏览器、文件系统…）直接进入模型工具列表 |
 
 ## 效果预览
 
 在 Zed 的 AI Agent 面板中选择 **dsh-acp-enhanced** 后，你会看到：
 
-![审批弹窗与模型/推理强度切换、上下文环](assets/screenshots/approval-config-context.png)
+<img src="assets/screenshots/approval-config-context.png" alt="审批弹窗与模型/推理强度切换、上下文环" width="560">
 
 - 工具调用需要许可时弹出**原生审批弹窗**（allow-once / reject-once）；输入框下方是
   **模型**、**推理强度**、**权限预设**、**Plan mode** 配置项与**上下文用量环**
   （`usage_update` 遥测，含缓存命中率、TPS 等明细）。
 
-![工具调用入参与输出、Zed 原生提问表单](assets/screenshots/tool-cards-elicitation.png)
+<img src="assets/screenshots/tool-cards-elicitation.png" alt="工具调用入参与输出、Zed 原生提问表单" width="320">
 
 - **工具卡片**可展开查看每次调用的完整入参（如 bash 执行的命令）与结果预览
   （`rawInput` / `rawOutput`）；DSH 需要你确认或选择时，以 **Zed 原生表单**弹出
@@ -191,6 +193,8 @@ dsh plugin --profile acp-enhanced add "link:/absolute/path/to/dsh-acp-enhanced/p
 node scripts/acp-client.mjs           # 端到端冒烟测试（需要 DEEPSEEK_API_KEY）
 DEEPSEEK_API_KEY=... node scripts/acp-client.mjs
 node scripts/acp-client-tools.mjs     # 客户端工具测试（模拟 Zed 的 fs/terminal/elicitation/plan 能力）
+node scripts/acp-mcp-test.mjs         # MCP 挂载测试（挂载一个最小 stdio MCP server，无模型调用）
+node scripts/acp-smoke-keyless.mjs    # keyless 冒烟（CI：自动建 profile → initialize → session/new）
 node scripts/acp-resume-test.mjs      # 会话恢复测试（两个进程：创建持久化 → 加载回放 → 续聊）
 ACP_DEBUG=1 dsh --profile acp-enhanced   # stderr 上的详细生命周期 trace
 ```
@@ -201,7 +205,10 @@ ACP_DEBUG=1 dsh --profile acp-enhanced   # stderr 上的详细生命周期 trace
 `fs.readTextFile/writeTextFile/terminal/elicitation` 能力，验证模型调用 `zed_*` 工具时请求以
 `fs/write_text_file`、`fs/read_text_file`、`terminal/create` 正确到达客户端，`ask_user_question`
 以 `elicitation/create` 表单（含 enum 选项）到达客户端，`plan_mode` 布尔开关触发 ACP `plan`
-update（开→条目、关→清空），并验证无 reasoning efforts 的路由不再广播空 `reasoning_effort`。
+update（开→条目、关→清空），并验证 `reasoning_effort` 选项按路由条件出现（有 efforts 的路由
+出现、无 efforts 的路由抑制）。`acp-mcp-test.mjs` 用 `scripts/fixtures/mcp-echo-server.mjs`
+验证 `session/new` 的 `mcpServers` 被真实挂载（server 收到 initialize 与 tools/list）且相同
+列表不重复挂载。
 
 ## 设计说明
 
@@ -283,8 +290,9 @@ update（开→条目、关→清空），并验证无 reasoning efforts 的路
   配置项——Zed 就不会渲染一个空的、无法操作的"Reasoning effort"chip。切到带 efforts 的
   模型后该选项自动重新出现（每次切换都会重播 `config_option_update`）。
 - **模式**：权限预设被呈现为 ACP 会话模式，因此 Zed 的模式切换器驱动 sandbox/approval 预设。
-- **已知限制**（继承自官方桥接器）：仅 baseline prompt（无图片/音频/MCP 附件）、不支持
-  `additionalDirectories`/MCP server 附加、已确认文本按块粒度流式，且每个会话同时只能有一个
-  in-flight prompt。会话恢复与归档列表已支持（见上），但 `session/close` / `session/fork` /
-  `session/resume` 未实现（不声明能力，合规客户端不会调用）；`session/delete` 因 dsh 持久化
-  面没有官方删除 API，采用直接删除后端目录的方式。
+- **已知限制**（继承自官方桥接器）：仅 baseline prompt（无图片/音频附件）、不支持
+  `additionalDirectories`、已确认文本按块粒度流式，且每个会话同时只能有一个
+  in-flight prompt。MCP servers（stdio + streamable HTTP）已支持，但 legacy SSE
+  传输与 `acp` 传输不声明。`session/close` / `session/fork` / `session/resume` 未实现
+  （不声明能力，合规客户端不会调用）；`session/delete` 因 dsh 持久化面没有官方删除
+  API，采用直接删除后端目录的方式。

package/lib/index.js

@@ -341,9 +341,14 @@ export function apply(ctx, config) {
     switch (chunk.type) {
       case 'block-start':
         if (chunk.blockType === 'text') record.buffer[chunk.index] = ''
+        else if (chunk.blockType === 'reasoning') record.thoughtBuffer[chunk.index] = ''
         break
       case 'text-delta':
         if (record.buffer[chunk.index] !== undefined) record.buffer[chunk.index] += chunk.text
+        else if (record.thoughtBuffer[chunk.index] !== undefined) record.thoughtBuffer[chunk.index] += chunk.text
+        break
+      case 'reasoning-delta':
+        if (record.thoughtBuffer[chunk.index] !== undefined) record.thoughtBuffer[chunk.index] += chunk.text
         break
       case 'block-end': {
         const text = record.buffer[chunk.index]
@@ -358,6 +363,18 @@ export function apply(ctx, config) {
             },
           })
         }
+        const thought = record.thoughtBuffer[chunk.index]
+        delete record.thoughtBuffer[chunk.index]
+        if (chunk.block.type === 'reasoning' && thought !== undefined && thought.length > 0) {
+          notify({
+            sessionId: record.agent.session.id,
+            update: {
+              sessionUpdate: 'agent_thought_chunk',
+              messageId: record.messageId,
+              content: { type: 'text', text: thought },
+            },
+          })
+        }
         break
       }
       case 'usage':
@@ -886,6 +903,104 @@ export function apply(ctx, config) {
     }
   }
 
+  // ── per-session MCP servers (dsh-mcp-client mounts) ──────────────────────
+
+  const mcpMounts = new Map() // serverName → { configJson, fiber }
+  let mcpClientModule
+  let warnedNoMcpClient = false
+
+  function sanitizeMcpServerName(raw) {
+    const cleaned = String(raw ?? 'server').replace(/[^A-Za-z0-9_-]/g, '_').slice(0, 32)
+    return cleaned.length > 0 ? cleaned : 'server'
+  }
+
+  /** Project one ACP McpServer onto an mcp-client config, if supported. */
+  function mcpConfigFor(server, cwd, serverName) {
+    const envList = (list) => Object.fromEntries(
+      (Array.isArray(list) ? list : []).map((entry) => [entry.name, entry.value]),
+    )
+    if (typeof server.command === 'string') {
+      return {
+        transport: 'stdio',
+        serverName,
+        command: server.command,
+        args: Array.isArray(server.args) ? server.args : [],
+        env: envList(server.env),
+        cwd: typeof server.cwd === 'string' ? server.cwd : cwd,
+        // A dead or misconfigured server must not take the session down;
+        // its tools simply stay out of the model's tool list.
+        failOnStartupError: false,
+      }
+    }
+    if (server.type === 'http' && typeof server.url === 'string') {
+      return {
+        transport: 'streamable-http',
+        serverName,
+        url: server.url,
+        headers: envList(server.headers),
+        failOnStartupError: false,
+      }
+    }
+    return undefined
+  }
+
+  /**
+   * Mount/reuse/replace mcp-client instances for a session's server list.
+   * Mounts are connection-scoped: the latest session's list wins, so tools
+   * registered under `mcp__<serverName>__<tool>` are available to every
+   * session on this connection. `ctx.loader.import` resolves the module from
+   * the host dsh installation (single module instance, never a profile copy).
+   */
+  async function syncMcpServers(servers, cwd) {
+    if (servers === undefined || servers.length === 0) return
+    let module = mcpClientModule
+    if (module === undefined) {
+      try {
+        module = await ctx.loader.import('@deepseek-ai/dsh-mcp-client')
+        mcpClientModule = module
+      } catch (error) {
+        if (!warnedNoMcpClient) {
+          warnedNoMcpClient = true
+          logger.warn(`acp-enhanced: ignoring ${servers.length} MCP server(s): @deepseek-ai/dsh-mcp-client is not available: ${String(error.message ?? error)}`)
+        }
+        return
+      }
+    }
+    const taken = new Set()
+    for (const entry of servers) {
+      const base = sanitizeMcpServerName(entry.name)
+      let serverName = base
+      for (let n = 2; taken.has(serverName); n += 1) serverName = `${base.slice(0, 28)}_${n}`
+      const cfg = mcpConfigFor(entry, cwd, serverName)
+      if (cfg === undefined) {
+        logger.warn(`acp-enhanced: skipping MCP server "${serverName}": unsupported transport`)
+        continue
+      }
+      const configJson = JSON.stringify(cfg)
+      const existing = mcpMounts.get(serverName)
+      if (existing !== undefined) {
+        if (existing.configJson === configJson) {
+          taken.add(serverName)
+          continue
+        }
+        try {
+          void existing.fiber.dispose()
+        } catch (error) {
+          logger.warn(`acp-enhanced: disposing MCP server "${serverName}": ${String(error)}`)
+        }
+        mcpMounts.delete(serverName)
+      }
+      try {
+        const fiber = ctx.plugin(module, cfg)
+        mcpMounts.set(serverName, { configJson, fiber })
+        taken.add(serverName)
+        logger.info(`acp-enhanced: mounted MCP server "${serverName}" (${cfg.transport})`)
+      } catch (error) {
+        logger.warn(`acp-enhanced: failed to mount MCP server "${serverName}": ${String(error)}`)
+      }
+    }
+  }
+
   // ── session records + history replay ─────────────────────────────────────
 
   /** Build the bridge-owned protocol record for a fresh or resumed agent. */
@@ -922,6 +1037,7 @@ export function apply(ctx, config) {
       lastActivityAt: undefined,
       toolStats: { count: 0, totalMs: 0, lastCallAt: undefined, lastName: undefined },
       buffer: {},
+      thoughtBuffer: {},
       contextWindow: undefined,
       lastUsage: undefined,
       selection,
@@ -973,6 +1089,15 @@ export function apply(ctx, config) {
     return parts.join('\n')
   }
 
+  /** Extract reasoning text from assistant content blocks (replay only). */
+  function thoughtFromBlocks(content) {
+    const parts = []
+    for (const block of content ?? []) {
+      if (block?.type === 'reasoning' && typeof block.text === 'string') parts.push(block.text)
+    }
+    return parts.join('\n')
+  }
+
   /**
    * Replay a resumed session's conversation to the client as ACP session
    * notifications. Zed inserts the thread before the load RPC completes, so
@@ -1001,12 +1126,22 @@ export function apply(ctx, config) {
           }
           case 'assistant/message': {
             const text = textFromBlocks(event.data.message.content)
-            if (text.trim().length === 0) break
-            await notifyNow(record, {
-              sessionUpdate: 'agent_message_chunk',
-              messageId: randomUUID(),
-              content: { type: 'text', text },
-            })
+            const thought = thoughtFromBlocks(event.data.message.content)
+            if (text.trim().length === 0 && thought.trim().length === 0) break
+            if (text.trim().length > 0) {
+              await notifyNow(record, {
+                sessionUpdate: 'agent_message_chunk',
+                messageId: randomUUID(),
+                content: { type: 'text', text },
+              })
+            }
+            if (thought.trim().length > 0) {
+              await notifyNow(record, {
+                sessionUpdate: 'agent_thought_chunk',
+                messageId: randomUUID(),
+                content: { type: 'text', text: thought },
+              })
+            }
             break
           }
           case 'tool/call':
@@ -1067,11 +1202,14 @@ export function apply(ctx, config) {
         syncClientTools()
         return Promise.resolve({
           protocolVersion: PROTOCOL_VERSION,
-          agentInfo: { name: 'deepseek-harness-acp-enhanced', version: '0.1.0' },
+          agentInfo: { name: 'deepseek-harness-acp-enhanced', version: '0.2.0' },
           agentCapabilities: {
             loadSession: true,
             sessionCapabilities: { list: {}, delete: {} },
             promptCapabilities: { image: false, audio: false, embeddedContext: false },
+            // Stdio MCP servers always work; streamable HTTP maps onto
+            // dsh-mcp-client's second transport. Legacy SSE does not.
+            mcpCapabilities: { http: true, sse: false },
           },
           authMethods: [],
         })
@@ -1099,6 +1237,7 @@ export function apply(ctx, config) {
         }
         const record = makeRecord(handle)
         sessions.set(sessionId, record)
+        await syncMcpServers(params.mcpServers, params.cwd)
         const permission = permissionPresets()
         return {
           sessionId,
@@ -1125,7 +1264,7 @@ export function apply(ctx, config) {
         if (params.additionalDirectories !== undefined && params.additionalDirectories.length > 0) {
           throw invalidParams('additionalDirectories is not supported')
         }
-        if (params.mcpServers.length > 0) throw invalidParams('mcpServers is not supported')
+        await syncMcpServers(params.mcpServers, params.cwd)
         const sessionId = SessionId(params.sessionId)
         const live = sessions.get(sessionId)
         if (live !== undefined) {
@@ -1385,6 +1524,14 @@ export function apply(ctx, config) {
       record.agent.cancel({ kind: 'user' })
       settlePrompt(record, 'cancelled')
     }
+    for (const { fiber } of mcpMounts.values()) {
+      try {
+        void fiber.dispose()
+      } catch (error) {
+        logger.warn(`acp-enhanced: disposing MCP server mount failed: ${String(error)}`)
+      }
+    }
+    mcpMounts.clear()
     quiescing = (async () => {
       const subagents = ctx.get('subagents')
       if (subagents?.drainContinuableDescendants !== undefined) {
@@ -1425,5 +1572,4 @@ function validateSessionParams(params) {
   if (params.additionalDirectories !== undefined && params.additionalDirectories.length > 0) {
     throw invalidParams('additionalDirectories is not supported')
   }
-  if (params.mcpServers.length > 0) throw invalidParams('mcpServers is not supported')
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-acp-enhanced",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Enhanced ACP server for DeepSeek Harness: block-level streaming, usage/stat telemetry (cache hit rate, token speed, input/output tokens, context length, turns, tool timing), model & reasoning-effort switching, and permission-preset control over the ACP wire (Zed-friendly)",
   "keywords": [
     "dsh",
@@ -32,7 +32,26 @@
     "test:client": "node scripts/acp-client.mjs"
   },
   "dependencies": {
-    "@agentclientprotocol/sdk": "0.25.1",
+    "@agentclientprotocol/sdk": "1.3.0",
+    "@deepseek-ai/schemastery": "^3.18.1",
+    "zod": "^4.4.3"
+  },
+  "peerDependencies": {
+    "@deepseek-ai/cordis": "^4.0.1-rc.1",
+    "@deepseek-ai/cordis-plugin-include": "^1.0.6-rc.1",
+    "@deepseek-ai/cordis-plugin-loader": "^1.0.2-rc.1",
+    "@deepseek-ai/dsh-agent": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-agent-instructions": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-invariants": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-llm": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-mcp-client": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-permission-presets": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-session": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-session-query": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-tools": "^0.1.0-rc.6",
+    "@deepseek-ai/dsh-user-approval": "^0.1.0-rc.6"
+  },
+  "devDependencies": {
     "@deepseek-ai/cordis": "4.0.1-rc.4",
     "@deepseek-ai/cordis-plugin-include": "1.0.6-rc.4",
     "@deepseek-ai/cordis-plugin-loader": "1.0.2-rc.4",
@@ -40,23 +59,22 @@
     "@deepseek-ai/dsh-agent-instructions": "0.1.0-rc.6",
     "@deepseek-ai/dsh-invariants": "0.1.0-rc.6",
     "@deepseek-ai/dsh-llm": "0.1.0-rc.6",
+    "@deepseek-ai/dsh-mcp-client": "0.1.0-rc.6",
     "@deepseek-ai/dsh-permission-presets": "0.1.0-rc.6",
     "@deepseek-ai/dsh-session": "0.1.0-rc.6",
     "@deepseek-ai/dsh-session-query": "0.1.0-rc.6",
     "@deepseek-ai/dsh-tools": "0.1.0-rc.6",
-    "@deepseek-ai/dsh-user-approval": "0.1.0-rc.6",
-    "@deepseek-ai/schemastery": "^3.18.1",
-    "zod": "^4.4.3"
-  },
-  "peerDependencies": {
-    "@deepseek-ai/cordis": "^4.0.1-rc.1",
-    "@deepseek-ai/dsh-agent": "^0.0.1-rc.1",
-    "@deepseek-ai/dsh-llm": "^0.0.1-rc.1",
-    "@deepseek-ai/dsh-permission-presets": "^0.0.1-rc.1",
-    "@deepseek-ai/dsh-session": "^0.0.1-rc.1",
-    "@deepseek-ai/dsh-user-approval": "^0.0.1-rc.1"
+    "@deepseek-ai/dsh-user-approval": "0.1.0-rc.6"
   },
   "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/grunmin/dsh-acp-enhanced.git"
+  },
+  "homepage": "https://github.com/grunmin/dsh-acp-enhanced",
+  "bugs": {
+    "url": "https://github.com/grunmin/dsh-acp-enhanced/issues"
+  },
   "dsh": {
     "bundle": {
       "patch": "./cordis.patch.yml"