[0.8.3] - 2026-02-17
Added
- ✏️ Model edit shortcut. Users can now edit models directly from the model selector dropdown menu, making it faster to modify model settings without navigating to separate admin or workspace pages. Commit
- 🎨 Image edit API background support. The image edit API now supports the background parameter for OpenAI's gpt-image-1 model, enabling background transparency control ("transparent", "opaque", "auto") when the feature is exposed in the UI. #21459
- ⚡ Faster model filtering. Model access control filtering no longer makes a redundant database query to re-fetch model info that is already available in memory, reducing latency when loading model lists for non-admin users. Commit
- 🔧 Tool call display improvements. Tool call results now display arguments in a cleaner key-value format instead of raw JSON, with a responsive layout that shows only the tool name on narrow screens and the full label on wider screens, preventing text wrapping to multiple lines. Commit
- 🔄 General improvements. Various improvements were implemented across the application to enhance performance, stability, and security.
- 🌐 Translations for Portuguese (Brazil), Simplified Chinese, and Traditional Chinese were enhanced and expanded.
Fixed
- 📧 USER_EMAIL variable fix. The {{USER_EMAIL}} template variable now correctly returns the user's email address instead of "Unknown" in prompts. #21479, #21465
- 🖼️ Image and file attachment handling fixes. Uploaded images are now correctly sent to vision-enabled models, and file attachments now work even when no user text is entered alongside a system prompt. This fixes two issues where the backend was not properly processing file attachments: images weren't converted to the expected format for API requests, and file context was dropped when the user sent only a file without accompanying text. Commit, #21477, #21457
- 🛡️ Missing function error handling. Models that reference deleted functions no longer cause the entire /api/models endpoint to crash; instead, the missing functions are skipped and logged, allowing the rest of the models to load successfully. #21476, #21464
- 🚀 Startup model pre-fetch error handling. If model pre-fetching fails during app startup, the application now logs a warning and continues instead of crashing entirely. Commit
- ⚙️ Function module loading error handling. Function modules that fail to load during startup or model processing are now caught and logged, preventing crashes when models reference functions with loading errors. Commit
- 🗄️ PostgreSQL group query fix. The '/api/v1/groups/' endpoint no longer fails with a GROUP BY error when using PostgreSQL; member counts are now calculated using correlated subqueries for better database compatibility. #21458, #21467