Glossarion v9.10.9 - Thinking fixes

🔧 Fixes & Changes

• Fixed an issue where thinking was hardcoded to low for Kimi K3 on authnd/ endpoints.
• Fixed a generic 400 error on authnd/ endpoints caused by exceeding output-token limits. Glossarion will now probe the site for its actual token limit and clamp the output-token value during 400-error retries.
• Added retry handling for thinking-related errors on OpenAI, OpenAI-compatible, NVIDIA NIM, and authnd/ endpoints.
• Added a prompt-profile system for glossary refinement.

• Updated the glossary refinement “Send all entry types” fallback logic. When all entries cannot fit within the token budget, it now prioritizes them in the following order:

  1. Characters and surnames
  2. Gendered entry types
  3. Non-gendered entry types

• Updated the default glossary refinement prompt. Click “Reset to Default” to apply the new default.

---

📦 Packages

• L_Glossarion.v9.10.9.exe
  Full-featured translation build (Excluding Manga Translation). Most optimal build for Novel translation.

• L_Glossarion_Lite.v9.10.9.exe
  Excludes the EPUB Reader, SDLXLIFF Notepad mode, and 'authnd/' prefix routing.

• L_Glossarion_TurboLite.v9.10.9.exe
  Excludes the EPUB Reader, SDLXLIFF Notepad mode, 'authnd/' prefix routing, Vertex AI SDK, and PDF generation.

• N_Glossarion_NoCuda.v9.10.9.exe
  Full-featured build with Manga Translation, and some experimental novel translation features (e.g., Silent truncation detection, and Argos translate fallback to google translate). Most optimal build for Manga translation.

• N_Glossarion_Heavy.v9.10.2.exe
  Includes CUDA support for the manga translator - ⚠️ May be unstable.
  Google Drive: Download****