v0.1.9
Tempest v0.1.9
Two features that have been on the "Still coming" list since v0.1.5 ship in this release: Automations and the CLI-agent chat backend. Alongside those: semantic search lands in Atlas, opt-in telemetry wires up, and the welcome screen gets a new look.
What's in this release
Automations
Workflows you draw once and run without sitting in front of them. An Automations canvas lives in the sidebar under its own nav entry. On the canvas you place nodes and connect them: triggers that start a run, tool nodes that call your CLI agents, and flow-control edges that route between them. The visual builder is n8n-style with a floating palette, click-to-add or drag-to-drop, and a model picker so each automation step can target a different model.
The runtime is Eve, a new agent runtime with its own DB schema and Tauri command surface. Runs and history land in the same SQLite database everything else already uses. The builder ships with a Beta badge because the trigger executor and persistent scheduling are the next pieces; what ships now is the full canvas authoring experience (flow editor, tool editor, trigger editor, model picker, wiring) and the Eve runtime underneath it.
CLI Agent Harness for Chat Nodes
Chat nodes now have a second backend option: a connected CLI agent. Before this release, a chat node always used a BYOK API key routed direct to a model provider. Now you can switch a node to run against Claude Code, letting the conversation harness its tool-use loop, file access, and permissions model rather than a raw completion. The data.backend schema hook was already in place; this release delivers the transport.
Atlas Semantic Search
Atlas gains hybrid semantic retrieval. When you opt in, a MiniLM embedding model downloads locally on first use (consent-gated; nothing downloads until you confirm). After that, every Atlas query runs a two-pass retrieval: keyword search as before, plus a semantic nearest-neighbor pass over local embeddings, merged with RRF fusion. Results surface more of what you mean, not just what you typed.
The download and opt-in are in Settings. No model download, no index build, no data leaves your machine.
Opt-in Telemetry
Tempest now collects anonymous usage events through PostHog, gated behind an explicit opt-in prompt on first launch. A durable anonymous identity is stored in SQLite so events are consistent across sessions without identifying you. Telemetry is off by default and can be toggled at any time in Settings. No personally identifying information is collected; the implementation is open in the codebase.
New Welcome Screen
The welcome screen is redesigned with a liquid-metal CTA replacing the old layout. The footer consolidates the license and version into one row. Settings card copy is tightened. An environment check screen is added to the onboarding flow so environment problems surface early rather than mid-session.
Fixes
- xterm.js rendering corrected a bug that caused display artifacts in the terminal renderer.
- Automations palette drop position and click-to-add target are corrected.
Platform Support
| Platform | Status |
|---|---|
| Windows 11 (x64) | Supported. Pre-built binary available. MSI and NSIS installers. |
| macOS | Supported (unsigned in this release; expect a Gatekeeper warning on first launch — right-click → Open) |
| Linux | Supported. x86_64 and ARM64 AppImage + .deb |
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.
About installer size
Installers in this release are larger than v0.1.8 — roughly 40 MB on Windows/Linux and ~24 MB on macOS. The portable tempest.exe on Windows is still 9 MB; the rest of the installer is the Atlas semantic-search runtime.
Atlas 1.2.0 ships @xenova/transformers (with onnxruntime-web) as a bundled resource so that when you opt into semantic search, the embedding model can run entirely on your machine with no additional runtime download step. Only the ~25 MB model weights download at opt-in; the WASM inference runtime is included in the installer itself, which is the source of the size increase versus v0.1.8.
A follow-up will move the runtime download to opt-in time (mirroring the model-weights flow) so installers return to their prior size for users who don't use semantic search. Tracking this against v0.1.10.
What's next
- Live quota. Promised since v0.1.5. The title-bar island and its formatting exist; the reader that pulls real rate-limit windows from your installed CLIs is what remains.
- Automations triggers and scheduling. The canvas authoring experience is in; persistent trigger execution and run scheduling are the next Automations milestone.
- Installer diet. Move the semantic-search WASM runtime to a consent-gated download alongside the model weights, so installers shrink back to v0.1.8 size for keyword-search-only users.