Skip to content

Beacon

Jump to bottom
LeonXu edited this page Apr 20, 2026 · 3 revisions

English | 中文

Beacon — Dynamic Island for AI Coding

Beacon is a floating panel that lives above your macOS menu bar, inspired by the iPhone Dynamic Island. It provides real-time project context and AI-powered authorization for your coding tools.

States

Collapsed

A pixel crab mascot sits in the notch area with a pulse indicator. Visible at all times when a coding app is in the foreground.

Collapsed

Expanded

Shows the current project name, AI-generated summary, and a link to open the project in Memora. The crab walks from its collapsed position to the left edge with a leg animation. Since v0.4.1, clicking an empty area collapses the panel so you can see what's behind it; it re-expands automatically on the next window-title change.

Expanded

Smart Approval (v0.4)

When your AI tool needs permission to run a command or edit a file, Beacon shows a semantic action verb (e.g. "🔍 Search files", "💾 Git commit"), key parameters, and — once the Risk Engine finishes — an AI-generated risk level (low/medium/high) and recommendation (allow/ask/deny). While analysis is running, a spinner is shown; if the LLM times out or is unavailable, Beacon falls back to static rules or raw command display so you never see a blank panel.

Authorization

How Authorization Works

Beacon is a Channel — one of several notification surfaces Memora can dispatch approval requests to. See Settings for enabling Lark (remote) channel alongside Beacon.

  1. Hook integration — memora-hook's PreToolUse handler writes a JSON request file to ~/.memora/beacon-requests/<toolUseId>.json and blocks.
  2. Three-layer decision (see Changelog v0.4.0):
    • Hard Deny (regex + context) — obvious dangers (rm -rf /, git push --force, credential paths) → always ask.
    • Hard Allow — project-local Edit/Write and read-only/build/git-day-to-day Bash → response written immediately, no notification.
    • Gray area — dispatched to all enabled channels; the Risk Engine (LLM advisor: DeepSeek / Claude / OpenAI) runs in parallel and pushes its verdict to the panel via beacon:updateNotification.
  3. File-based IPC — when you click Allow/Deny (or Lark responds first, or Unattended Mode auto-allows), Memora writes ~/.memora/beacon-responses/<toolUseId>.json. The hook is watching that directory and exits with the decision. No keystroke injection, no focus stealing.
  4. Auto-dismiss — PostToolUse hook writes ~/.memora/beacon-resolved/latest; any stale panel whose request file has been consumed is cleaned up on the next scan cycle.

Operation Modes

Mode Behavior
Normal (default) Everything outside deny-list / hard-allow prompts. LLM verdict shown as advice.
Unattended (opt-in) LLM judges level != high && rec != deny → auto-approved without asking. Toggle in Settings → Risk Engine.

Unattended Mode requires an LLM provider configured; if the provider is offline, Memora falls back to Normal Mode (prompts you) as a safety net.

Button Layout

Source Tool Buttons
Claude Code Edit, Write 3 (Allow / Allow All / Deny)
Claude Code Bash 2 (Allow / Deny)
Other sources Any 2 (Allow / Deny)

Visual Design

  • Shape: Concave bezier arcs (15px × 30px) create a "growing from the screen edge" effect.
  • Level: NSPanel at statusBar + 1 — always above the menu bar.
  • Background: Pure black NSView (NSVisualEffectView cannot achieve true black).
  • Coordinate system: Paths drawn in visual coordinates (y=0 at top), flipped to CG coordinates via CGAffineTransform.

Clone this wiki locally