[MasterYoav]

Agentic Mode

The iPhone gets a second mode. Settings › Mode: Developer / Agentic. Developer is today's app. Agentic is a Telegram-like chat with your crew: HeadMaster pinned on top, one thread per agent. The agents control xBot and the Mac through xBot's own MCP server.

Spec: docs/superpowers/specs/2026-10-13-agentic-mode-design.md.

iPhone

• Chats: the crew list with still pixel faces, previews, unread marks, "typing…" and search.
• Threads:
  • accent bubbles with Markdown and code;
  • one activity line that updates as the agent works, folding to "Did N things ›" when it's done;
  • cards for screenshots (full-screen viewer), approvals (Allow / Don't Allow), hand-offs, finished work, notes and lessons.
• Composer: grows with the text; on-device dictation; photos (≤1600 px JPEG); long-press for Copy, Retry, or Edit and resend.
• Performance:
  • the last 200 messages per thread;
  • pictures decoded at display size, with a 12 MB cache;
  • no timers or world animation.

Mac

• Threads: one chat per agent (chats.thread, schema 7), in full access, in the Inbox. Photos are saved and attached by path. A lost session continues in a new thread that starts with recent messages.
• XBot --mcp: a headless MCP server (JSON-RPC over stdio). It reaches the running app over a 0600 Unix socket, with a per-turn token.
  • 28 tools: xBot (projects, chats, crew, hand-offs, automations, notes, settings, worlds, xbot_ask, xbot_remember, xbot_lessons) and the Mac (mac_screenshot via ScreenCaptureKit, mac_ui via Accessibility, click, type, keys, scroll, open, apps, quit).
• Attached per CLI:
  • Claude Code: --mcp-config, plus a PreToolUse hook (XBot --hook) that also covers its own shell;
  • Codex: -c mcp_servers.xbot.*;
  • Copilot: --additional-mcp-config;
  • Antigravity: agy mcp add, once.
• RiskRules + ApprovalCenter: delete, send, push, purchase, quit and System Settings changes ask on the phone (the Mac when no phone is attached). No answer within 10 minutes counts as Don't allow.
• Experience (Harness2's cue → lesson):
  • agents call xbot_remember;
  • their best lessons (12 at most, 1,500 characters) go into their next turns;
  • low-confidence, unused lessons are dropped after 30 days;
  • an Experience section in the agent editor shows and edits them.
• Permissions: Accessibility and Screen Recording are in the onboarding Files step, with the floating guide, and asked again when a tool needs them.

Tests

• Package: 688 tests pass. New suites: RiskRules, XBotTools, ToolHost, Threads, AgenticWire, RemoteHostAgentic, Experience, ApprovalCenter.
• Live, phone simulator → sandboxed Mac → real Codex:
  • testAgenticMode: HeadMaster replies pong;
  • testAgenticRealTools: a real mac_screenshot card arrives on the phone; a mac_quit approval comes to the phone, is refused, and the agent carries on;
  • testAgenticCards;
  • testAgenticActivity;
  • Developer Mode testNewChatRunsOnTheMac.
• Live, real Claude Code turn: xbot_crew, mac_apps and TextEdit control; the hook blocks git push when refused.

Gaps

• Codex, Copilot and Antigravity have no per-run hook, so risky commands in their own shell aren't caught; only xBot's tools are checked, and the agents are told to ask first.
• Copilot and Antigravity attachment are unit-tested only.
• Photo picking and dictation weren't driven live.
• Pending approvals expire rather than being cancelled when a turn is stopped.