v4.3.0

What's Changed

Prompt templates grow up: Jinja2 logic, richer variables, and in-app reminders
that point you to what's new.

Added

• Jinja2 prompt templates. Prompts now support Jinja2 syntax — {% if %} /
  {% for %}, field access ({{ user.name }}), filters ({{ x | upper }}), and
  an @include("path/to/file") macro to inline another file's contents.
• New variable types. Alongside plain Text, add JSON variables
  (parsed at render into an object you can branch and iterate over; a bare
  {{ data }} prints pretty JSON) and File content variables (pick a file
  via the native OS picker or type a path; its contents are inlined at run time).

Changed

• The model selector shows its parameter buttons by default and selects the model
  name on focus for quicker editing.

Fixed

• External links are restricted to safe URL schemes.
• Root directory creation is confined to the expected location.
• Thread files are written atomically to avoid corruption on interrupted writes.

Performance

• The trace sidebar is lazy-loaded.
• The RPC stream queue is now amortized O(1).

Install

Two editions. Install either, or both — they share the same ~/.llm-space data, so switching keeps your threads and settings.

• LLM Space — uses the system WebView. Small download (~27 MB), light on memory and battery.
• LLM Space Performance — embeds its own rendering engine (~130 MB). Rendering stays consistent across macOS versions, and usually performs better.

Download a DMG below. Existing installs update in place.

Full changelog: https://github.com/deer-flow/llm-space/blob/v4.3.0/CHANGELOG.md