What's Changed
- feat: Add automatic context compaction, with editable summary templates, per-conversation settings, and undo. OpenAI native compaction can be enabled separately.
- feat: Add a fixed context start, so history is sent from a chosen message without rolling truncation by message count.
- feat: Add Standard, Fast, and UltraFast response speed controls in the reasoning picker, saved per conversation and model.
- feat: Add a per-conversation control for image recall, plus local budgets for image size, image count, and request body.
- feat: Detect vision, tools, and thinking support on local Ollama models.
- feat: Let each assistant choose which memory types are always injected in full.
- feat: Add custom model icons, including color and monochrome LobeHub styles.
- feat: Add an optional legacy code block style.
- feat: Quote selected reasoning text into the composer.
- feat: Render Markdown and LaTeX in in-app translation results.
- feat: Add a custom balance query, with its own path, method, headers, and body.
- feat: Add Russian, Korean, and Turkish.
- fix: Limit a browser agent pause to the conversation that started it.
- fix: Recover orphaned Android workspace processes.
- fix: Preserve Codex sessions and Claude thinking across retries and tool continuations, and avoid signing out when a usage check fails.
- fix: Keep follow-up tool requests aligned with the current round after a provider or model change.
- fix: Limit Mermaid rendering memory, and close Windows WebViews before the plugin is torn down.
- fix: Preserve the iOS browser scroll position when the safe area changes.
- fix: Show full model names in lists and settings.
- fix: Keep desktop send shortcuts available in narrow windows.
- fix: Show reasoning activity while the summary is still empty, and restore context usage estimates for chats with images.
- fix: Allow keyboard suggestions on ordinary provider fields while keeping API keys on the secure keyboard.
- fix: Fix several other known bugs and UI issues.
Full Changelog: v1.3.4...v1.3.5