feat(agents): add Kimi Code CLI support

PRIVACY.md

@@ -60,7 +60,7 @@ Unlike the AI CLIs above, the following network activity is initiated by Paralle
   - **Over Tailscale.** Traffic is carried by your tailnet — typically a direct WireGuard connection between your devices, but Tailscale's coordination service and (when direct connection is not possible) DERP relays may be involved per [Tailscale's network architecture](https://tailscale.com/kb/1257/connection-types). How Tailscale handles that traffic is governed by Tailscale's own policies, not this one.
 - **Sub-task coordinator (MCP)** — when sub-tasks run under a coordinator agent, Parallel Code starts a local token-protected HTTP/WebSocket server so sub-task agents can call back into the app (e.g. to signal completion). No traffic from this feature passes through infrastructure operated by the Parallel Code project.
   - **Bind address.** When MCP starts its own listener, it binds to `127.0.0.1` except on macOS Docker setups, where it binds to `0.0.0.0` so containers can reach it via `host.docker.internal` — this also makes the port reachable from other hosts on your LAN, though access still requires the token. If a Remote Access server is already running when a coordinator starts, the coordinator reuses that listener; because Remote Access binds to `0.0.0.0`, MCP routes inherit that LAN reach on any platform (including Linux), though access still requires the MCP token.
-  - **Where the token can land.** Token-bearing MCP data is written or passed in several places: a worktree `.mcp.json` when a worktree path is available, or a project-root `.mcp.json` otherwise, so the coordinator agent can auto-discover the server (Parallel Code also adds `.mcp.json` to your `.git/info/exclude` so it is not committed); a non-Docker coordinator config in your OS temp directory named `parallel-code-mcp-<coordinatorTaskId>.json`; per-sub-task configs in your OS temp directory for host-mode sub-tasks (`parallel-code-subtask-<taskId>.json`) or under the coordinator's `.parallel-code/` directory for Docker sub-tasks (`subtask-<taskId>.json`); and short-lived `.parallel-code-atomic-<uuid>.tmp` files written next to these configs during atomic-rename steps. These files are written with `0600` permissions where the platform supports it.
+  - **Where the token can land.** Token-bearing MCP data is written or passed in several places: a worktree `.mcp.json` when a worktree path is available, or a project-root `.mcp.json` otherwise, so the coordinator agent can auto-discover the server (Parallel Code also adds `.mcp.json` to your `.git/info/exclude` so it is not committed); an auto-discovered `.kimi-code/mcp.json` (or fallback `.mcp.json`) inside each Kimi sub-task worktree, with both that config and its adjacent `.parallel-code-atomic-*.tmp` files added to `.git/info/exclude` before credentials are written; a non-Docker coordinator config in your OS temp directory named `parallel-code-mcp-<coordinatorTaskId>.json`; per-sub-task configs in your OS temp directory for host-mode sub-tasks (`parallel-code-subtask-<taskId>.json`) or under the coordinator's `.parallel-code/` directory for Docker sub-tasks (`subtask-<taskId>.json`); and short-lived `.parallel-code-atomic-<uuid>.tmp` files written next to these configs during atomic-rename steps. These files are written with `0600` permissions where the platform supports it.
   - **Codex token in the command line.** For Codex specifically, the MCP token is passed as a literal command-line argument (`--config mcp_servers.parallel-code={... env = { PARALLEL_CODE_MCP_TOKEN = "..." }}`). Process command lines are visible to other processes — via `/proc/<pid>/cmdline` on Linux or `ps` on macOS — so any local process that runs concurrently with a Codex sub-task can read that token and call back into the coordinator under its authority until the coordinator exits. Other agents receive the token through a token-protected file or env var instead.
 - **Docker task isolation** — when you enable Docker mode for a task, or when you opt coordinator sub-tasks into Docker-isolated mode, the agent launches in a container via `docker run --network host`. **Docker mode is not a security boundary.** It isolates the filesystem against the worktree, but does not isolate the network or credentials from the agent.
   - **Network.** `--network host` means the container shares the host's network namespace; its outbound reachability is the same as your host's, including loopback services and any LAN address your host can reach.

README.md

@@ -14,7 +14,7 @@
 </p>
 
 <p align="center">
-  Works with Claude Code, Codex, and Gemini · Every change isolated in its own git worktree · Free, open source, no extra platform fee
+  Works with Claude Code, Codex, Gemini, and Docker-pinned Kimi Code · Every change isolated in its own git worktree · Free, open source, no extra platform fee
 </p>
 
 <p align="center">
@@ -45,7 +45,7 @@
 
 ## Why Parallel Code?
 
-- **Use the AI coding tools you already trust** — [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [Codex CLI](https://github.com/openai/codex), [Gemini CLI](https://github.com/google-gemini/gemini-cli), and [Copilot CLI](https://docs.github.com/en/copilot/concepts/agents/about-copilot-cli) — all from one interface.
+- **Use the AI coding tools you already trust** — [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [Codex CLI](https://github.com/openai/codex), [Gemini CLI](https://github.com/google-gemini/gemini-cli), Docker-pinned [Kimi Code CLI](https://github.com/MoonshotAI/kimi-code), and [Copilot CLI](https://docs.github.com/en/copilot/concepts/agents/about-copilot-cli) — all from one interface.
 - **Free and open source** — no extra subscription required. MIT licensed.
 - **Keep every change isolated and reviewable** — each task gets its own git branch and worktree automatically.
 - **Run agents in parallel, not in sequence** — five agents on five features at the same time, zero conflicts.
@@ -115,10 +115,23 @@ When you're happy with the result, merge the branch back to main from the sideba
    - **macOS** — `.dmg` (universal)
    - **Linux** — `.AppImage` or `.deb`
 
-2. **Install at least one AI coding CLI:** [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [Codex CLI](https://github.com/openai/codex), [Gemini CLI](https://github.com/google-gemini/gemini-cli), [Antigravity CLI](https://antigravity.google/), or [Copilot CLI](https://docs.github.com/en/copilot/concepts/agents/about-copilot-cli)
+2. **Install at least one AI coding CLI:** [Claude Code](https://docs.anthropic.com/en/docs/claude-code), [Codex CLI](https://github.com/openai/codex), [Gemini CLI](https://github.com/google-gemini/gemini-cli), [Antigravity CLI](https://antigravity.google/), or [Copilot CLI](https://docs.github.com/en/copilot/concepts/agents/about-copilot-cli). Kimi Code is currently supported through the bundled Docker image instead of a native installation.
 
 3. **Open Parallel Code**, point it at a git repo, and start dispatching tasks.
 
+<details>
+<summary><strong>Kimi Code: use Docker mode</strong></summary>
+
+Parallel Code's Kimi integration writes an auto-discovered project MCP config into each fresh
+task worktree. Kimi Code 0.33 added a workspace-trust prompt, and 0.36 defaults to declining
+project MCP launch targets in that prompt. A current native Kimi installation can therefore
+pause on every new worktree instead of starting the task unattended.
+
+Use Docker-isolated Kimi tasks for now. The bundled image pins Kimi Code 0.32, before the
+workspace-trust gate. Native Kimi support is not currently claimed.
+
+</details>
+
 <details>
 <summary><strong>Antigravity CLI: run natively, not in Docker isolation</strong></summary>
 

docker/Dockerfile

@@ -50,7 +50,8 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 RUN ln -sf "$(command -v fdfind)" /usr/local/bin/fd 2>/dev/null || true
 
 # AI agent CLIs — must be present so Docker-mode tasks can execute them
-RUN npm install -g @anthropic-ai/claude-code @openai/codex @google/gemini-cli opencode-ai
+# Keep Kimi below 0.33: newer releases block fresh worktrees on workspace trust.
+RUN npm install -g @anthropic-ai/claude-code @openai/codex @google/gemini-cli opencode-ai @moonshot-ai/kimi-code@0.32.0
 
 # Antigravity CLI (agy) — distributed as a Go binary via the official installer
 # (not on npm). The installer's `--dir` flag drops the binary straight into a

electron/ipc/agents.test.ts

@@ -8,4 +8,8 @@ describe('getSkipPermissionsArgs', () => {
 
     expect(getSkipPermissionsArgs('claude')).toEqual(['--dangerously-skip-permissions']);
   });
+
+  it('returns Kimi Code skip-permission args', () => {
+    expect(getSkipPermissionsArgs('kimi')).toEqual(['--yolo']);
+  });
 });

electron/ipc/agents.ts

@@ -44,6 +44,15 @@ const DEFAULT_AGENTS: AgentDef[] = [
     skip_permissions_args: ['--yolo'],
     description: "Google's Gemini CLI agent",
   },
+  {
+    id: 'kimi',
+    name: 'Kimi Code CLI',
+    command: 'kimi',
+    args: [],
+    resume_args: ['--continue'],
+    skip_permissions_args: ['--yolo'],
+    description: "Moonshot AI's Kimi Code CLI agent",
+  },
   {
     id: 'opencode',
     name: 'OpenCode',

electron/ipc/pty.test.ts

@@ -433,6 +433,7 @@ describe('spawnAgent docker mode', () => {
       ['opencode', '.config/opencode'],
       ['copilot', '.config/github-copilot'],
       ['agy', '.gemini/antigravity-cli'],
+      ['kimi', '.kimi-code'],
     ])(
       '%s bind-mounts a user-owned host directory when shareDockerAgentAuth is enabled',
       (command, relDir) => {

electron/ipc/pty.ts

@@ -835,6 +835,7 @@ const AGENT_CONFIG_DIRS: Record<string, string[]> = {
   opencode: ['.config/opencode'],
   copilot: ['.config/github-copilot'],
   agy: ['.gemini/antigravity-cli'],
+  kimi: ['.kimi-code'],
 };
 
 // Config files (not directories) each agent CLI uses for auth, relative to HOME.

electron/ipc/register.ts

@@ -1561,6 +1561,7 @@ export function registerAllHandlers(win: BrowserWindow): void {
           landingSummary?: string;
           landedMetadata?: import('../mcp/types.js').LandedMetadata;
           mcpConfigPath?: string;
+          autoDiscoveredMcpConfig?: import('../mcp/types.js').AutoDiscoveredMcpConfigState;
           agentCommand?: string;
           preambleFileExistedBefore?: boolean;
           initialPrompt?: string;
@@ -1599,6 +1600,7 @@ export function registerAllHandlers(win: BrowserWindow): void {
           landingSummary: args.landingSummary,
           landedMetadata: args.landedMetadata,
           mcpConfigPath: args.mcpConfigPath,
+          autoDiscoveredMcpConfig: args.autoDiscoveredMcpConfig,
           agentCommand: args.agentCommand,
           preambleFileExistedBefore: args.preambleFileExistedBefore,
           initialPrompt: args.initialPrompt,

electron/mcp/agent-args.test.ts

@@ -6,6 +6,7 @@ import {
   isAntigravityCommand,
   isCodexCommand,
   isCopilotCommand,
+  isKimiCommand,
 } from './agent-args.js';
 
 const config = {
@@ -67,6 +68,16 @@ describe('MCP agent launch args', () => {
     expect(buildMcpLaunchArgs('agy', '/tmp/config.json', config)).toEqual([]);
   });
 
+  it('detects Kimi commands by executable name', () => {
+    expect(isKimiCommand('kimi')).toBe(true);
+    expect(isKimiCommand('/home/agent/.local/bin/kimi')).toBe(true);
+    expect(isKimiCommand('claude')).toBe(false);
+  });
+
+  it('emits no --mcp-config for Kimi Code', () => {
+    expect(buildMcpLaunchArgs('kimi', '/tmp/config.json', config)).toEqual([]);
+  });
+
   it('detects copilot commands by executable name', () => {
     expect(isCopilotCommand('copilot')).toBe(true);
     expect(isCopilotCommand('/opt/homebrew/bin/copilot')).toBe(true);

electron/mcp/agent-args.ts

@@ -18,6 +18,10 @@ export function isAntigravityCommand(command: string): boolean {
   return command.split('/').pop() === 'agy';
 }
 
+export function isKimiCommand(command: string): boolean {
+  return command.split('/').pop() === 'kimi';
+}
+
 export function isCopilotCommand(command: string): boolean {
   return command.split('/').pop() === 'copilot';
 }
@@ -53,6 +57,11 @@ export function buildMcpLaunchArgs(
   if (isAntigravityCommand(command)) {
     return [];
   }
+  // Kimi Code auto-discovers user and project MCP config files and does not
+  // accept the generic `--mcp-config` flag.
+  if (isKimiCommand(command)) {
+    return [];
+  }
   // Copilot has no `--mcp-config` flag — passing it makes Copilot exit immediately
   // with "unknown option" before the prompt is ever sent (#146). It accepts
   // `--additional-mcp-config <@file|json>` (and also auto-discovers a workspace

electron/mcp/coordinator-test-harness.ts

@@ -25,6 +25,7 @@ const enoent = () => Object.assign(new Error('ENOENT'), { code: 'ENOENT' });
 
 const mocks = vi.hoisted(() => {
   const mockExecFile = vi.fn();
+  const mockSpawnSync = vi.fn();
   const mockWriteFileSync = vi.fn();
   const mockReadFileSync = vi.fn();
   const mockExistsSync = vi.fn();
@@ -37,6 +38,7 @@ const mocks = vi.hoisted(() => {
   const mockFsMkdir = vi.fn();
   const mockAtomicWriteFileSync = vi.fn();
   const mockAtomicWriteFile = vi.fn();
+  const mockAppendGitInfoExcludeBlock = vi.fn();
   const mockNotifyRenderer = vi.fn();
   const mockLogInfo = vi.fn();
   const mockLogWarn = vi.fn();
@@ -59,6 +61,7 @@ const mocks = vi.hoisted(() => {
 
   return {
     mockExecFile,
+    mockSpawnSync,
     mockWriteFileSync,
     mockReadFileSync,
     mockExistsSync,
@@ -71,6 +74,7 @@ const mocks = vi.hoisted(() => {
     mockFsMkdir,
     mockAtomicWriteFileSync,
     mockAtomicWriteFile,
+    mockAppendGitInfoExcludeBlock,
     mockNotifyRenderer,
     mockLogInfo,
     mockLogWarn,
@@ -95,6 +99,7 @@ const mocks = vi.hoisted(() => {
 
 vi.mock('child_process', () => ({
   execFile: mocks.mockExecFile,
+  spawnSync: mocks.mockSpawnSync,
 }));
 
 vi.mock('fs', () => ({
@@ -118,6 +123,10 @@ vi.mock('./atomic.js', () => ({
   atomicWriteFile: mocks.mockAtomicWriteFile,
 }));
 
+vi.mock('../ipc/git-exclude.js', () => ({
+  appendGitInfoExcludeBlock: mocks.mockAppendGitInfoExcludeBlock,
+}));
+
 vi.mock('../shared/prompt-detect.js', () => ({
   stripAnsi: (s: string) =>
     s.replace(
@@ -238,6 +247,7 @@ vi.mock('../log.js', () => ({
 
 export const {
   mockExecFile,
+  mockSpawnSync,
   mockWriteFileSync,
   mockReadFileSync,
   mockExistsSync,
@@ -250,6 +260,7 @@ export const {
   mockFsMkdir,
   mockAtomicWriteFileSync,
   mockAtomicWriteFile,
+  mockAppendGitInfoExcludeBlock,
   mockNotifyRenderer,
   mockLogInfo,
   mockLogWarn,
@@ -303,6 +314,8 @@ export function resetCoordinatorMocks(): void {
       return { on: vi.fn() };
     },
   );
+  mockSpawnSync.mockReset();
+  mockSpawnSync.mockReturnValue({ status: 1, error: undefined, stderr: Buffer.alloc(0) });
 
   mockWriteFileSync.mockReset();
   mockReadFileSync.mockReset();
@@ -326,6 +339,8 @@ export function resetCoordinatorMocks(): void {
   mockAtomicWriteFileSync.mockReset();
   mockAtomicWriteFile.mockReset();
   mockAtomicWriteFile.mockResolvedValue(undefined);
+  mockAppendGitInfoExcludeBlock.mockReset();
+  mockAppendGitInfoExcludeBlock.mockReturnValue('appended');
 
   mockNotifyRenderer.mockReset();
   mockLogInfo.mockReset();