Skip to content

Open Dungeon 0.4.0: streaming narration

Choose a tag to compare

View all tags
@newideas99 newideas99 released this 06 Jul 04:30
· 4 commits to main since this release
v0.4.0
e232735

The narrator now streams, and the whole app reads better while it does.

  • Narrator text streams in real time. Passages appear token by token as the model writes them — on the bundled Ollama models and on custom backends (LM Studio, llama.cpp, OpenRouter, …) alike. Timeouts reset per chunk, so long passages flow while a stalled backend still gets cut off. (#7, #10)
  • No more <|channel>thought leakage. Reasoning tags that some models leak into their replies are stripped — both live during streaming and from the saved passage. (#7)
  • New passages start at the top. A finished (or streaming) passage pins its first line to the top of the story view instead of dropping you at the bottom of a wall of text. (#10)
  • The typing box grows with you. Write a few paragraphs and the composer expands up to about 40% of the window, then shrinks back. (#12)
  • New stories keep your backend. Saving Text Model settings on any story now also sets the app-wide default, so a fresh story no longer falls back to Ollama when you use LM Studio or another server. (#9)
  • Sidebar cleanup. Panels act as an accordion with every header always visible, Chats is now Stories with New story inside the panel, and characters list as compact rows that expand one at a time for editing. (#8)
  • Native AMD GPU images on Windows. Launch-Windows.bat now detects Radeons covered by AMD's PyTorch-on-Windows wheels (RX 9070 XT / 9070 / 9060 XT, RX 7900 XTX, RX 7700, AI PRO R9700, PRO W7900), sets up Python 3.12 + ROCm automatically, and runs the same FLUX backends on the GPU — no more 15–30 minute CPU images. Needs the Adrenalin 26.2.2+ driver; OPEN_DUNGEON_ROCM=0/1 overrides. (#13)
  • Clearer image-worker status. The Start button no longer says "already running" while the panel says "Worker stopped" — it now explains when the worker is up but its backend isn't installed, and points at the setup docs. llama.cpp backends without an mmproj no longer 500 on turns with character portraits; the turn retries text-only. (#11)

Install:

  • macOS Apple Silicon: open the DMG, drag Open Dungeon to Applications, and open it (right-click then Open the first time, since it is unsigned).
  • Windows: download Open-Dungeon-0.4.0-Windows.zip, unzip it, then run Launch-Windows.bat.

Validation: Windows smoke passed on the release commit, local lint/build passed, streaming verified end-to-end against a real Ollama model and an OpenAI-compatible server, and the DMG verified with hdiutil verify. The AMD ROCm path is code-reviewed and validated but not yet exercised on AMD hardware — reports welcome on #13.

Full changelog: v0.3.3...v0.4.0

Assets 4
Loading