rjyo/homebrew-moshi v0.2.76

moshi-hook v0.2.76

on GitHub

What's Changed

• Asking an agent a question no longer makes the terminal scroll. While waiting for your answer, Moshi re-reads the pane several times a second to keep track of the prompt, and each read was reaching into scrollback it never needed — so the view jumped up and snapped back, again and again. It now reads only what is on screen.
• Chat View now supports Hermes. Hermes keeps its conversation history in a database, so this needs the sqlite3 command available on your machine; moshi install tells you if it is missing. Hooks and approvals work either way — only Chat View needs it.
• Claude sessions started from claude agents now show up on their own. Agent view runs its sessions in a background daemon rather than in the terminal you are looking at, so Moshi could not tell which pane they belonged to and they never appeared. It now asks Claude directly which session is running where. Running claude normally is unaffected.
• The Codex usage card keeps up with plan changes. When Codex sign-in expired, the card froze on whatever plan name and reset date it had last seen, so an upgrade never showed — it looked stuck on the old plan indefinitely. It now falls back to fresher data on disk instead of trusting a stale copy.
• New setting: suppress-nested-agent-push, off by default. Turn it on and agents launched by another agent — a Codex you started from inside a Claude session, say — stop sending notifications, so one piece of work does not notify you twice. It hides their approval requests too, which means a nested agent waiting on you is invisible from your phone, so it is opt-in: moshi-hook set suppress-nested-agent-push on, then restart the daemon.
• Workspace switcher: tabs and windows now open up to show the panes inside them, with the focused pane marked and any agent running in it shown alongside. Tap a pane to jump straight to it; if it has moved or closed since the list loaded, you land on its tab instead of nowhere.
• Herdr panes are named the way herdr's own navigator names them (claude, p452-ios, …) rather than whatever the agent is currently working on, so the list stays readable and stops shifting under you.
• moshi-hook usage now also reports how many tokens each agent actually burned today and over the last seven days, broken down by model, with an estimated cost beside it. That figure is what the same usage would cost at public API rates — not what your plan charged you. Scanning is incremental and capped so it stays out of the way; a machine with a large history takes a few hours of background passes to fill in. This is in the CLI only for now and does not appear in the app yet.