isTargetSleeping v1.29.0

Fixed

• Existing custom activity lists now include the notification indicator after updating, so notifications can use the closed notch's priority slot as well as the open header.
• A permission request no longer opens the notch or takes focus while something is full screen (a game, a video) or game mode is on: it waits in the queue (and in the Agents page) and the card opens by itself when you leave full screen. Before, a pinned notch ignored the full-screen rule and could pull you out of the game.
• «Deny» in the agent permission card is readable in high contrast: the red tint behind its red text is replaced by the theme's plain background.
• Team task merges now work when Git has no user identity configured; the merge uses a local fallback identity.

Added

• Optional connected pet companion: download the separate desktop server from Settings › Notch. It serves the mobile web app privately over Tailscale, works without an ESP32 board, and can connect a board as an optional client.
• Bluetooth connection card in the closed notch (the «Bluetooth headphones» live activity): when a Bluetooth device connects or disconnects it shows its name, «Connected · 50%» (battery when Windows reports it) and its icon, with the battery as a ring around the Bluetooth badge. It now names the device and covers any Bluetooth device (headphones, mouse, keyboard…), not just audio outputs.
• Richer session rows on the Agents page: under each session's name, its git branch (read straight from .git/HEAD, also inside worktrees, with no git process) and how much context it is using, in thousands of tokens, taken from the last answer in its log (Claude Code sessions only; it is the context in use, not a percentage, because the window size is not in the log).
• Circuit breaker for the Team widget's scheduled tasks: a task that fails ten times in a row switches itself off (it keeps its history) instead of burning tokens forever, shows «paused after 10 failures in a row» in the schedule list and raises an alert on the Agents page. A success resets the count, cancelling or missing a run does not count, and turning the task back on by hand starts again from zero. «Every N hours» tasks already keep a minimum of one hour between runs.
• Decision time for permission requests (Settings › General): choose how long the notch waits for your answer (30 s, 1 min, 2 min or the maximum) before leaving the permission to Claude Code's terminal. Claude Code cuts the hook at 5 minutes, so that stays the maximum.
• The Agents page «Now» tab is now a single inbox: pending permissions, questions and plans, the while-you-were-away summary, agent alerts (stuck, looping, usage limits) and routines paused by the circuit breaker, followed by your sessions.
• Control over agent alerts (Settings › General): choose «Everything» (default), «Only what matters» (permissions, questions, plans, serious alerts and the while-you-were-away summary) or «Nothing», and snooze for 30 minutes, 1 hour or until tomorrow morning. It decides whether something chimes, the notch bell rings or the notch opens by itself; requests and alerts always stay on the Agents page (with «Nothing» a permission waits there without opening the notch or making a sound).
• Resume a session in a real terminal: on the Agents page, every Claude Code or Codex session that nobody is using (finished or idle, with a resumable id and an existing project folder) has a «Resume in the terminal» button. It opens Windows Terminal (or PowerShell) in the project folder and runs claude --resume <id> / codex resume <id>, so you carry on yourself. Unlike the read-only question box of the Agents widget, this hands you the real CLI. Sessions that are working are never offered, because two writers would trample the same session file. The Agents page also lists recently finished sessions now.
• «While you were away»: when you come back after 10 minutes or more without touching the keyboard or mouse, the pet sums up what your agents did meanwhile (how many finished, how many permissions were requested, how many alerts went off and how many requests are still waiting for you). It appears as a card on the Agents page, as a pet message in the Chat page and with the notch bell, and it stays quiet in game mode and when nothing happened.
• Notch light for your agents (Settings › General, on by default): a beam of light runs round the notch's edge, entering at the right ear, going down the side, along the base and out at the left ear (it never crosses the top, next to the screen), with a sharp tip and a tail that fades smoothly. Blue while an agent works, amber when something waits for your answer, red when there is an alert (stuck, looping or a usage limit used up) and green for a few seconds when one finishes. If another state arrives while a beam is running, the new colour waits for the current lap to end instead of overlapping it. It stays still (a plain ring) with reduced motion and takes the text colour in high contrast.
• Agent alerts on the Agents page (Now tab), with a bell in the notch the first time each one appears and a short mention in the closed pill: an agent that has shown no signs of life for 2 minutes (it may be stuck or waiting on a long command), one that repeats the same tool on the same file six or more times in a row (a possible loop), and a usage limit (Claude, Codex, Gemini) that is at 90 % or used up, with a countdown to when it comes back. They are computed locally from the session logs and the AI usage widget.
• Answer Claude Code's questions and review its plans from the notch. A question (AskUserQuestion) shows its options as buttons, one choice per question or several when the agent allows it, and «Send answer» replies with them; a plan (ExitPlanMode) shows its text with «Approve the plan» and «Ask for changes» (which tells the agent you want changes). «Decide in the terminal» and Esc leave either one to the terminal. Claude Code only accepts these answers from a PreToolUse hook, so turning on «Approve Claude Code permissions from the notch» now also installs a PreToolUse hook filtered to AskUserQuestion|ExitPlanMode (it never touches other tools). Questions and plans are never saved as «Always» rules, and the Y key does not answer a question.
• Agent center: an «Agents» page in the notch dock (shown while approving permissions from the notch) with three tabs. Now lists every permission request waiting for you, with Allow / Always / Deny right in the page, plus your latest agent sessions. History is a local log of what each agent asked and what was answered (allowed, denied, always, automatic or left to the terminal), saved as decisions.jsonl, trimmed on its own and clearable. Rules lists every «Always» rule so you can revoke any of them.
• Sensitive-command guard: recursive deletes, force pushes, git reset --hard / clean, downloaded scripts piped to a shell, sudo, disk formatting, dropping databases, shutdown, bulk permission or registry changes, and writes to credentials or sensitive settings now show a red warning on the card. They cannot be allowed with the Y key, cannot be saved as «Always» and no saved rule lets them through on its own; the warning is also kept in the history.
• Voice for your pet's chat, fully local: a microphone button on the Chat page listens for one phrase at a time (press to talk, never always-on) with the speech recognizer installed in Windows, and an optional switch makes the pet read its answers aloud with a Windows voice. No audio leaves the PC. Needs a Windows speech recognizer for your language; without one the microphone is hidden. Uses the System.Speech package.
• Weekly summary on the Chat page: messages, agents that finished, permissions granted and denied, Pomodoros and model naps this week, counted locally (week.json) and reset every Monday.
• Ambient chatter (Settings › General, off by default; Never / Rarely / Sometimes): now and then the pet adds a short, kind line to the Chat page and rings the notch bell. It does not use the model, picks the line from its mood and bond, keeps a minimum gap (50 min or 2 h) and stays quiet in game mode, with a permission pending or while listening.
• Bond and mood for your pet, shown on the Chat page. The bond grows with healthy use (chatting, an agent finishing, finishing a Pomodoro), with a daily cap per kind so it cannot be farmed, never drops when you are away, is stored locally (bond.json), can be reset, and does not reward freed RAM. The mood follows the context (worried when memory is critical, alert when an agent waits for your permission, happy for a while after something good, sleepy at night, calm otherwise) and, with the bond level, shapes how the pet talks in the chat.
• Chat with your pet (Settings › General, off by default): a «Chat» page in the notch dock where the local model you already have loaded in Ollama answers in short, friendly messages. It never loads a model on its own (if none is loaded it says so), nothing leaves the PC, the conversation only lives in memory, and answers stream in with a Stop button. Each pet talks in its own character (Mira dry and watchful, the llama proud and dramatic, the capybara calm, the orange cat playful), the same one its animations already have.
• Approve Claude Code permissions from the notch (Settings › General, off by default). Turning it on adds PermissionRequest, Stop, Notification, UserPromptSubmit and SessionStart hooks to ~/.claude/settings.json (with a timestamped backup, never touching other hooks) and turning it off removes them. A permission request opens the notch with a card showing the agent, project and exact command, and Allow / Always / Deny buttons (Y, N, Esc on the keyboard; Esc leaves the decision to the terminal). Several requests wait their turn. «Always» is a rule of its own for Claude Code only (same project, same tool and, for Bash, the identical command) kept in hook-rules.json. The hook talks to the app through a named pipe restricted to your user (no network); if the app is not running or nobody answers within 280 s the hook exits quietly and Claude Code asks in its terminal as usual.
• Short, soft chimes when an agent asks for permission or finishes (synthesized by the app, switchable in Settings › General), and the taskbar pet sparkles when an agent needs you and jumps when it finishes.

Changed

• Add a Processes tab to the app window; it reuses the existing process explorer and keeps Activity › Processes in the notch and tray panel.
• Animate the notch hover control dock as it opens and closes, including an exit transition before its space collapses.
• Smooth the notch's widget page transitions and avoid rebuilding the Home and widget folders when navigating.
• Adopt the sleeping slime as the app icon, with an updated animated tray icon that keeps status and the optional RAM percentage.
• Use native Windows notifications with the full-resolution logo, an Activity button for sleeping models and progress for detected download layers. Keep the tray notification as a fallback.