test: automate desktop protocol journey

.github/workflows/ci.yml

@@ -80,6 +80,34 @@ jobs:
       - name: Verify current documentation
         run: node scripts/check-docs.mjs
 
+  desktop-preview:
+    name: Desktop protocol journey
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v6
+      - uses: pnpm/action-setup@v6
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+          cache: 'pnpm'
+      - name: Install dependencies
+        run: pnpm install --frozen-lockfile
+      - name: Install Chromium
+        run: pnpm --filter @deepcode/desktop exec playwright install --with-deps chromium
+      - name: Exercise the desktop protocol fixture
+        run: pnpm --filter @deepcode/desktop test:e2e
+      - name: Upload browser diagnostics
+        if: failure()
+        uses: actions/upload-artifact@v4
+        with:
+          name: desktop-playwright-report
+          path: |
+            apps/desktop/playwright-report
+            apps/desktop/test-results
+          if-no-files-found: ignore
+          retention-days: 7
+
   desktop-rust:
     name: Desktop Rust check + test
     runs-on: macos-latest
@@ -88,6 +116,17 @@ jobs:
       - uses: actions/checkout@v6
       - name: Show Rust toolchain
         run: rustc --version && cargo --version
+      # Tauri validates every externalBin path in its build script. This job
+      # only compiles/tests Rust and never executes or packages the sidecar, so
+      # use a target-correct placeholder instead of copying a 100+ MB runtime.
+      - name: Prepare Tauri sidecar placeholder
+        run: |
+          target="$(rustc -vV | sed -n 's/^host: //p')"
+          runtime="apps/desktop/src-tauri/binaries/deepcode-runtime-${target}"
+          mkdir -p "$(dirname "$runtime")"
+          touch "$runtime"
+          mkdir -p apps/server/dist-sidecar
+          touch apps/server/dist-sidecar/app-server.cjs
       - name: Check and test Tauri backend
         run: |
           cargo check --manifest-path apps/desktop/src-tauri/Cargo.toml --locked

.github/workflows/release.yml

@@ -116,6 +116,19 @@ jobs:
       - name: pnpm install
         run: pnpm install --frozen-lockfile
 
+      - name: Prepare pinned Node sidecar runtime
+        env:
+          NODE_SIDECAR_VERSION: 22.23.1
+          NODE_SIDECAR_SHA256: ef28d8fab2c0e4314522d4bb1b7173270aa3937e93b92cb7de79c112ac1fa953
+        run: |
+          archive="node-v${NODE_SIDECAR_VERSION}-darwin-arm64.tar.xz"
+          curl --fail --location --retry 3 \
+            "https://nodejs.org/dist/v${NODE_SIDECAR_VERSION}/${archive}" \
+            --output "$RUNNER_TEMP/$archive"
+          echo "${NODE_SIDECAR_SHA256}  $RUNNER_TEMP/$archive" | shasum -a 256 --check
+          tar -xJf "$RUNNER_TEMP/$archive" -C "$RUNNER_TEMP"
+          echo "DEEPCODE_NODE_RUNTIME=$RUNNER_TEMP/node-v${NODE_SIDECAR_VERSION}-darwin-arm64/bin/node" >> "$GITHUB_ENV"
+
       - name: Set version
         run: |
           cd apps/desktop

.gitignore

@@ -31,6 +31,8 @@ yarn-error.log
 # Test outputs
 coverage/
 .nyc_output/
+apps/desktop/playwright-report/
+apps/desktop/test-results/
 
 # Electron build outputs (pre-Tauri pivot — kept for historical artifacts)
 apps/desktop/release/
@@ -39,6 +41,8 @@ apps/desktop/dist-electron/
 # Tauri build outputs (Rust target dir is large)
 apps/desktop/src-tauri/target/
 apps/desktop/src-tauri/gen/
+apps/desktop/src-tauri/binaries/
+apps/server/dist-sidecar/
 # Note: Cargo.lock IS committed (best practice for applications)
 
 # Release artifacts — too large for git; CI uploads to GitHub Releases instead

README.md

@@ -63,6 +63,7 @@ Mac 客户端（v1 即将发布）：拖入 Applications → 首启完成 onboar
 | [docs/DEVELOPMENT_PLAN.md](docs/DEVELOPMENT_PLAN.md)                         | 整体开发方案 v0.5（1500+ 行 / §3 模块 / §6 里程碑） |
 | [docs/VISUAL_DESIGN.html](docs/VISUAL_DESIGN.html)                           | 视觉设计 v0.4（11 屏 mockup）                       |
 | [docs/security-model.md](docs/security-model.md)                             | 威胁模型 + 防御层 + 攻击向量测试 + 已知缺口         |
+| [docs/design/session-format-v1.md](docs/design/session-format-v1.md)         | 统一 session JSONL、旧格式迁移与 writer ownership   |
 | [docs/design/sandbox-plan-worktree.md](docs/design/sandbox-plan-worktree.md) | sandbox × plan mode × worktree 关系矩阵             |
 | [docs/design/plugin-security.md](docs/design/plugin-security.md)             | plugin 信任 ladder + sandbox 子进程                 |
 | [docs/design/effort-levels.md](docs/design/effort-levels.md)                 | 5 档 effort 到 DeepSeek API 参数映射                |

apps/cli/package.json

@@ -21,6 +21,7 @@
     "start": "node ./dist/cli.js"
   },
   "dependencies": {
+    "@deepcode/app-server": "workspace:*",
     "@deepcode/core": "workspace:*",
     "@deepcode/shared-ui": "workspace:*"
   },

apps/cli/src/cli.ts

@@ -4,6 +4,7 @@
 // M2: onboarding + REPL + slash commands + settings + permissions matcher.
 
 import { CredentialsStore, VERSION, redact } from '@deepcode/core';
+import { runAppServer } from '@deepcode/app-server';
 import { homedir } from 'node:os';
 import { resolve } from 'node:path';
 import { runHeadless } from './headless.js';
@@ -80,6 +81,14 @@ async function main(): Promise<number> {
       errOutput: process.stderr,
     });
   }
+  if (args.positional[0] === 'app-server') {
+    await runAppServer({
+      input: process.stdin,
+      output: process.stdout,
+      home: process.env.DEEPCODE_HOME ?? resolve(homedir(), '.deepcode'),
+    });
+    return 0;
+  }
   if (args.positional[0] === 'trust') {
     return runTrustCommand(args.positional.slice(1), {
       cwd: process.cwd(),

apps/cli/src/completion.ts

@@ -53,6 +53,7 @@ const SUBCOMMANDS = [
   'doctor',
   'upgrade',
   'mcp',
+  'app-server',
   'trust',
   'plugins',
   'skills',

apps/cli/src/headless.ts

@@ -21,6 +21,7 @@ import {
   EFFORT_PARAMS,
   HookDispatcher,
   ReadTool,
+  RuntimeHost,
   SessionManager,
   ToolRegistry,
   WebFetchTool,
@@ -39,7 +40,6 @@ import {
   loadSkills,
   makeSkillTool,
   resolveCredentials,
-  runAgent,
   wirePlugins,
   collectPluginContributions,
   type AgentEvent,
@@ -271,25 +271,28 @@ export async function runHeadless(opts: HeadlessOpts): Promise<number> {
   }
   let exitCode = 0;
   try {
-    const result = await runAgent({
+    const runtime = new RuntimeHost({
       provider,
       tools,
+      cwd,
+      mode,
+      permissions: settings.permissions,
+      hooks,
+      pluginDirs: pluginContrib.dirs,
+      autoMode: settings.autoMode,
+      sandboxConfig: settings.sandbox,
+    });
+    const result = await runtime.run({
       systemPrompt,
       userMessage,
       history: [],
       model,
       maxTokens,
       temperature,
       maxTurns,
-      cwd,
+      signal: ctrl.signal,
       session: { manager: sessions, id: session.id },
-      mode,
-      permissions: settings.permissions,
-      hooks,
-      pluginDirs: pluginContrib.dirs,
       autoCompact: { contextWindow: contextWindowFor(model), threshold: 0.8 },
-      autoMode: settings.autoMode,
-      sandboxConfig: settings.sandbox,
       // In headless mode there's no human to ask: auto-deny anything that
       // would normally need approval. Users wanting auto-yes should pass
       // --mode dontAsk or --mode bypassPermissions (gated by trust).
@@ -441,6 +444,7 @@ function formatEventText(out: Writable, e: AgentEvent): void {
       return;
     case 'usage':
     case 'thinking_delta':
+    case 'model_step_complete':
     case 'turn_complete':
       return;
   }

apps/cli/src/parse-args.ts

@@ -285,6 +285,7 @@ USAGE
   deepcode cron <cmd>                   Scheduled tasks: install/uninstall/list/status
   deepcode scheduler run                Run due scheduled jobs (invoked by launchd)
   deepcode mcp serve                    Expose DeepCode tools as an MCP server (stdio)
+  deepcode app-server                  Run the experimental lifecycle server (JSONL stdio)
   deepcode trust [--plan-only]          Trust this directory's project config (hooks/MCP/...)
   deepcode plugins list [--json]        List installed plugins
   deepcode plugins install <spec>       Install a plugin (gh:owner/repo | name@npm | ./path)

apps/cli/src/repl.ts

@@ -8,6 +8,7 @@ import {
   EFFORT_PARAMS,
   HookDispatcher,
   ReadTool,
+  RuntimeHost,
   SessionManager,
   TaskManager,
   ToolRegistry,
@@ -37,7 +38,6 @@ import {
   contextWindowFor,
   makeSkillTool,
   resolveCredentials,
-  runAgent,
   settingsPaths,
   wirePlugins,
   collectPluginContributions,
@@ -429,6 +429,17 @@ export async function startRepl(opts: ReplOpts): Promise<number> {
   }
 
   let history: StoredMessage[] = resolved.seededHistory;
+  const runtime = new RuntimeHost({
+    provider,
+    tools,
+    cwd,
+    mode,
+    permissions: settings.permissions,
+    hooks,
+    pluginDirs: pluginContrib.dirs,
+    autoMode: settings.autoMode,
+    sandboxConfig: settings.sandbox,
+  });
   const ctx: SessionContext = {
     cwd,
     model,
@@ -471,25 +482,20 @@ export async function startRepl(opts: ReplOpts): Promise<number> {
   // reading ctx.model/ctx.mode live so /model and /mode switches are honored.
   const tasks = new TaskManager((spec) => {
     const ac = new AbortController();
-    const done = runAgent({
-      provider,
-      tools,
-      systemPrompt,
-      userMessage: spec.prompt,
-      model: ctx.model,
-      maxTokens,
-      temperature,
-      cwd: ctx.cwd,
-      signal: ac.signal,
-      mode: ctx.mode as Mode,
-      permissions: settings.permissions,
-      hooks,
-      pluginDirs: pluginContrib.dirs,
-      sandboxConfig: settings.sandbox,
-      autoMode: settings.autoMode,
-      subAgentDepth: 1,
-      systemReminders: false,
-    }).then((r) => assistantText(r.history));
+    const done = runtime
+      .run({
+        systemPrompt,
+        userMessage: spec.prompt,
+        model: ctx.model,
+        maxTokens,
+        temperature,
+        cwd: ctx.cwd,
+        signal: ac.signal,
+        modeOverride: ctx.mode as Mode,
+        subAgentDepth: 1,
+        systemReminders: false,
+      })
+      .then((r) => assistantText(r.history));
     return { done, abort: () => ac.abort() };
   });
   ctx.tasks = tasks;
@@ -649,9 +655,7 @@ export async function startRepl(opts: ReplOpts): Promise<number> {
     }
 
     // Otherwise: send to agent (with mode/permission/hooks gating from M3b)
-    const result = await runAgent({
-      provider,
-      tools,
+    const result = await runtime.run({
       systemPrompt,
       userMessage: userInput,
       history,
@@ -663,13 +667,8 @@ export async function startRepl(opts: ReplOpts): Promise<number> {
       // ctx.sessionId (not the launch `session.id`) so a live `/resume <id>`
       // switch redirects new messages to the resumed session.
       session: { manager: sessions, id: ctx.sessionId },
-      mode: ctx.mode as Mode,
-      permissions: settings.permissions,
-      hooks,
-      pluginDirs: pluginContrib.dirs,
+      modeOverride: ctx.mode as Mode,
       autoCompact: { contextWindow: contextWindowFor(ctx.model), threshold: 0.8 },
-      autoMode: settings.autoMode,
-      sandboxConfig: settings.sandbox,
       // Session-scoped manager: the agent's TaskCreate calls land here too, so
       // background tasks persist across turns and show up in /tasks.
       taskManager: tasks,
@@ -762,6 +761,7 @@ function formatEvent(out: Writable, e: AgentEvent): void {
       else out.write(`    ✓ ${truncate(e.result.content, 200)}\n`);
       return;
     case 'usage':
+    case 'model_step_complete':
       return;
     case 'error':
       out.write(`\n  ✕ ${e.error}\n`);

apps/cli/tsconfig.json

@@ -9,5 +9,9 @@
   },
   "include": ["src/**/*"],
   "exclude": ["node_modules", "dist"],
-  "references": [{ "path": "../../packages/core" }, { "path": "../../packages/shared-ui" }]
+  "references": [
+    { "path": "../../packages/core" },
+    { "path": "../../packages/shared-ui" },
+    { "path": "../server" }
+  ]
 }

apps/desktop/README.md

@@ -15,13 +15,14 @@ src/                 renderer（React + Vite，无 Tailwind，手写设计系统
   screens/           About / MCPManager / Onboarding / Permissions /
                      Plugins / Repl / Sessions / Settings / Skills
   components/        Sidebar / InspectorRail / ToolCard / UpdateBanner …
-  lib/               tauri-api（renderer↔Rust IPC 封装）· mac-agent ·
-                     mac-tools · repl-stream · updater …
+  lib/               tauri-api（renderer↔Rust IPC 封装）· protocol-client ·
+                     protocol-agent · repl-stream · updater …
 src-tauri/           Rust 主进程
+  src/app_server.rs  bundled runtime 启停、stdio 与 crash event
   src/commands.rs    #[tauri::command] —— renderer 通过 invoke() 调用
-  src/credentials.rs 凭据读写（原子写入）
+  src/credentials.rs 凭据保存与无密钥状态查询
   src/settings.rs    设置持久化
-  src/tools.rs       工具实现
+  src/tools.rs       legacy native helpers（renderer 仅暴露只读 file read）
   src/lib.rs         Tauri builder / 插件注册
   tauri.conf.json    窗口 + 构建 + 打包配置
   capabilities/      权限能力声明
@@ -31,6 +32,11 @@ src-tauri/           Rust 主进程
 renderer ↔ Rust 的 IPC 边界由 `src/lib/tauri-api.ts` 封装，契约测试见
 `src/lib/tauri-api.test.ts`（#84）。
 
+app-server 由 Tauri 作为 target-specific sidecar 监督。`apps/server` 会被打成单个
+`app-server.cjs` resource，Node runtime 通过 `bundle.externalBin` 进入 `.app`；renderer 只能通过
+Rust commands 与版本化协议通信，不能直接使用 shell plugin。provider、agent loop、tools、权限、
+session materialization 和凭证明文都只存在于 sidecar；renderer 不再带有第二套运行时。
+
 ## 开发
 
 依赖在 monorepo 根 `pnpm install` 一次装好；Rust 工具链 + Tauri CLI 见下。
@@ -40,25 +46,28 @@ renderer ↔ Rust 的 IPC 边界由 `src/lib/tauri-api.ts` 封装，契约测试
 | `pnpm dev`                   | 仅 Vite dev server（5173）—— 一般由 Tauri 自动拉起 |
 | `pnpm tauri:dev`             | 完整 app：Tauri 启 dev server + 原生窗口，热重载   |
 | `pnpm build`                 | `tsc -b` + `vite build` → `dist/`（renderer 产物） |
-| `pnpm tauri:build`           | 当前架构的 .app / .dmg                             |
+| `pnpm tauri:build`           | 构建包含 runtime + app-server 的 `.app`            |
 | `pnpm tauri:build:universal` | universal-apple-darwin 通用二进制                  |
 | `pnpm typecheck`             | `tsc -b`                                           |
 | `pnpm test`                  | `vitest run`（lib 单测 + IPC 契约测试）            |
 
-`tauri.conf.json` 里 `beforeDevCommand` / `beforeBuildCommand` 分别接
-`pnpm dev` / `pnpm build`，所以平时只跑 `pnpm tauri:dev` 即可。
+`tauri.conf.json` 的 dev/build hooks 会先生成 app-server bundle 和目标 runtime，再启动 Vite 或
+Tauri release build，所以平时只跑 `pnpm tauri:dev` 即可。
 
 ### 前置工具
 
 - Node ≥ 22、pnpm
 - Rust 工具链（`rustup`）—— Tauri 主进程是 Rust
 - 通用构建需 `rustup target add aarch64-apple-darwin x86_64-apple-darwin`
+- 通用构建还要求 `DEEPCODE_NODE_RUNTIME` 指向同时含 arm64/x86_64 的通用 Node binary
 
 ## 打包 / 签名
 
 - 产物配置在 `src-tauri/tauri.conf.json`，公证 entitlements 在
   `src-tauri/Entitlements.plist`。
 - 签名 + 公证需要 Apple Developer ID 证书，以及 `APPLE_ID` /
   `APPLE_APP_SPECIFIC_PASSWORD` 等环境变量（CI 走 secrets）。
+- release CI 固定 Node 22.23.1，校验官方 SHA256 后才进入 Tauri 打包；nested runtime 先签，outer
+  `.app` 后签，再做 strict deep verification 与 notarization。
 
 详见 `docs/DEVELOPMENT_PLAN.md` §4 / §4a / §4b。