v0.0.17

0.0.17 — 2026-10-22

• Model indexing: a scan only lists files and reads headers, so models are usable at once, and
  hashing runs afterwards in a paced, low-priority pool that pauses during generation and that an
  admin can slow down or pause; a first-run choice (index all, pick models or later) is asked after
  setup instead of over the wizard, onboarding never waits for hashing, and recipes verify at most
  two same-size files before downloading.
• Prompts: if / else if / otherwise conditions pick text per image from a prompt's variables, edited
  in a wide condition editor (typing << opens it) with a per-image preview and warnings for
  unfinished branches; Number variables join Choice variables; variables are kept on the server per
  user, where PotionAI and MCP can read and change them; PotionAI understands conditions and can add
  or delete segments; a literal $, % or # is plain text instead of silently switching off every
  choice; Save as prompt keeps every variable.
• Generate: a context toolbar under the tabs holds preset, mode, variant and session on the left and
  Save plus the tools on the right; the preset dropdown has search, Recent, Pinned and families, and
  a preset without a backend can be picked with Generate disabled and the reason shown; the
  generation panel becomes a now-playing deck with a result thumbnail and a soft glow of the last
  result; tabs show a dot for unsaved session changes, sessions can be reloaded from the last
  snapshot, and a required model slot fills itself when one installed model matches.
• History: cards show run time; /history?generation=<id> opens a generation's Details; deleting an
  X/Y compare removes the whole compare after one confirm; History, Library, Prompts, Inspirations
  and Models share a redesigned collections sidebar with search, Include subcollections and a phone
  sheet; prompt search is faster.
• Model Tools: a new admin-only preset for the local native backend quantizes to FP8, INT8, INT8
  ConvRot, NVFP4 or INT4 ConvRot, bakes LoRA and LoKr into a model and merges LoRAs, writing the
  result next to its source as a model card in the Workbench and History; layers a format cannot
  hold stay in source precision and are reported; FLUX 3 Klein loads fused-QKV INT8 ConvRot
  checkpoints, LoRAs included, and INT4 ConvRot files load too.
• Native engine: MOSS-Image arrives with SFT and Base presets, a LoRA tab, a starter recipe and an
  optional step cache; Krea-2 gets an experimental Text Enhancer (off by default) for stronger
  prompt adherence and a Text Scale slider; torch.compile also runs on fp8-scaled and streamed DiTs;
  a streamed MiniMax-H3 pins 19.5 GB of system memory instead of 30.6 GB and keeps more of long
  clips on the GPU.
• Chat: a configuration's Show thinking to users switch (off by default) shows the model's reasoning
  in a foldable block above each reply, live with a timer and never sent back as context; Ollama's
  Thinking Mode loses Automatic; LLM configurations get one Context window field with a Detect
  button, detected from the server or checkpoint when unset; chat sends shorter prompts and only the
  tools a mode and window need.
• Sign-in: the OIDC plugin can send signed-out visitors straight to the identity provider (off by
  default), with /login?local=1 as the local recovery form; every login returns to the page you
  were on; registration accepts local-domain emails; MCP generation tools return a details link per
  generation, built from a new Public address setting or the request's origin.
• Automations and plugins: an admin-only HTTP request action with a test panel, a public-address
  guard and encrypted secret headers; Auto-organize can file generations by Media Tagging tags;
  plugin pages call the app through an authenticated host API; Admin > Presets keeps its open tab in
  the URL; the account menu links to the community.
• Phones and keyboard: Chat sits in the bottom nav on every page, the deck sits above the tab bar,
  and headers, Generation Details and confirm dialogs fit a phone; focus rings show only while the
  keyboard is in use, tooltips open on keyboard focus, Escape closes only the topmost popover, and
  shortcut hints show the command key on macOS.
• Pinokio: media loads when PotionUI is embedded in Pinokio's window.
• Reliability: the backend keeps serving past 10,000 requests; History queries are indexed; heavy
  calls no longer block the server; unexpected errors no longer leak text or paths to regular users.
• Fixes: Krea-2 Enhance no longer leaves noise; cancelling a cloud generation on a provider that
  cannot cancel shows the calm notice; approving a prompt-variable change in chat reports what was
  applied; the $, /, @ and # pickers insert at the typed character in a multi-line prompt; typing in
  a choice group's options no longer triggers global shortcuts; MiniMax-H3 finds the bundled ffmpeg
  and fails before sampling rather than after; Model Browser previews load without a referrer.
• Upgrading: Ollama configurations on Thinking Mode Automatic (off during tool calls, on otherwise)
  become Disabled, and thinking stays hidden until Show thinking to users is on; installs where any
  model already had a hash keep verifying in the background, while one with none and a large library
  (over 20 files or 64 GiB) is asked once and hashes nothing until it answers, models being usable
  either way; prompt variables kept in the browser move to the server on first load; a stored S3
  secret key and plugin credentials saved in clear are encrypted on first start; the installer needs
  Node.js 22 or newer.