Tempest v0.1.12

Ships the two carried-forward commitments from v0.1.11: live agent quota in the title bar and scheduled Automations. Both were promised as "next" for two releases running.

What's in this release

Live agent quota island

The title bar now carries a per-provider quota readout. Each supported CLI has its own reader that talks to the provider's own usage endpoint using the credentials that CLI already stored on disk — no separate login, no keys to paste. Four providers ship in this cut: Claude, Codex, Copilot, Cursor. The seven other agents have no per-user quota surface to read; they are absent from the island rather than faked.

The island itself is a dot at rest, expands at 75% used to name the fullest window inline, and reveals a full per-provider panel on hover with the raw numbers, reset time, and a manual refresh. It polls every five minutes in the background; the OAuth-backed providers (Claude, Codex) refresh their tokens on a 401 and persist the new pair atomically so the readout survives a session expiring mid-poll.

The point is that you can see "you have 12% of your Claude window left" without leaving the app, without a browser tab, and without a second auth flow.

Automations — RRULE-scheduled local agent runs

The visual builder from v0.1.10 has been replaced with a lightweight native scheduler. An automation is now three fields — an agent, a schedule (RRULE), and a markdown prompt — and a Rust background thread wakes every 60s to dispatch anything due as a real PTY session through the existing agent pipeline. Every fired run is a normal session in the sidebar; there is nothing new to learn to reason about what an automation did.

• RRULE schedules. Standard iCalendar recurrence, driven by the rrule npm package on the composer side. Preset buttons cover the common cases (hourly, daily at 9am, weekdays 9-5, etc.) with a raw RRULE field and live validation for anything custom.
• Prompt versioning. The prompt editor debounce-saves in the background and rolls a new version every ten minutes of active edits. The detail page shows the version history and lets you compare or restore.
• Run log per automation. Every dispatch is tracked with its start time, exit status, and a link to the session it opened.
• Twelve starter templates. The Create dialog offers a gallery next to the blank-compose option — daily standup summaries, weekly repo health, on-call handoff, etc. — so a new user has something running in under a minute.

Under the hood the schema is three tables (automations, automation_runs, automation_prompt_versions) with a migration guard that detects the old builder schema (the slug column) and rebuilds cleanly. The old builder — sixteen files, ~7 kLOC of React Flow node/edge editors — is gone; the Automations page is now about 530 lines of CSS instead of 1,960.

Per-automation model override

The Create dialog and detail sidebar gain an optional Model picker. It only appears when the chosen agent declares --model support — today that is Claude Code (Haiku / Sonnet / Opus, with a Default option that omits the flag entirely). The choice is stored as a nullable model column and threaded through the scheduler into the same --model path openSession already uses for manual launches, so a scheduled Sonnet run is indistinguishable from one you'd start by hand.

Alongside: polish on the Automations list, the schedule picker, and the shared SpSelect control that the new pickers use.

Platform Support

Windows requires WebView2 (bundled automatically when necessary).

Database Branches (from v0.1.5) additionally require Docker; without it the feature stays dormant and everything else works normally.

Atlas semantic search (from v0.1.10, fixed in v0.1.11) downloads a ~25 MB embedding model on first opt-in.

What's next

Live quota and scheduled Automations were the standing commitments; both are in this release. Next cycle picks up the Automations run log (currently minimal) and expands --model overrides to the other agents that support it.