Skip to main content
July 17, 2026 v0.38.0 plus the big panel 0.9.0.

​
Kimi K3 (Moonshot) as a first-class provider

Moonshot’s Kimi K3 joins the provider roster (#233) — API-key based, alongside the existing Claude / ChatGPT / Gemini / Grok / GLM / Copilot / local lineup. Full story in the Kimi K3 post.

​
CivitAI browser: browse, like, load, recreate

Panel 0.9.0 makes the in-panel CivitAI browser a full loop:
  • Load workflow onto canvas (a Discord community request) — a lightbox post with an embedded UI-format graph gets a “Load onto canvas” action through the same undoable path the agent uses; model versions grow a per-file load button. CivitAI’s .zip workflow wrappers (780 of 844 live versions) are unpacked in the browser under zip-bomb caps, downloads stream through a same-origin proxy with an SSRF guard, gated files return a clean 401 with a one-click sign-in, and every outcome surfaces as a toast + inline status — never a silent no-op.
  • Favorites shows ALL your likes — the feed reads your likes collection (the one the website’s ❤ saves into), not just in-panel reactions, and pagination no longer drops one item per page or dead-ends on short pages.
  • Creator filter + search qualifiers — a top-creators leaderboard when empty, debounced username search, removable creator pill across every feed, and GitHub-style “@name terms” search (#86).
  • Like/unlike hearts on cards and in the lightbox, mirroring into a default likes collection; full-screen lightbox with details, prompt/parameters, and share-with-agent / save-workflow actions; debounced search on every tab.

​
Panel-owned sessions (the new default)

The conversation and agent memory now persist while you switch, save, rename, or create workflows — the agent is mechanically told which canvas it operates on, and the live agent instance is rebound across tab ids (no respawn), so even in-memory local backends keep their history. Settings → General → “Conversation follows the panel” restores the legacy per-workflow mode. Why and how on the blog.

​
Also in panel 0.9.0

  • LoRA Training modal (coming-soon preview, mobile-app parity): six training flows as cards with base-model dropdowns under a Local/Cloud switch — Local shows your rig’s real GPU. (The trainer itself lands July 20.)
  • The iOS TestFlight and Android beta buttons are now live invite links.
  • The agent-feed gate is renamed Deafen (ear / slashed-ear icons) — it stops the agent hearing canvas events; your typed messages still go through.
  • Provider chips, connect help, and tooltips no longer hardcode Claude.