open-design-v0.24.0
gohugoio/hugoopen-design-v0.24.0Sep 22, 2026by github-actions[bot]
Upgrading?
AI Summary
This release focuses on improving the robustness of the application's conversational state and preview functionality, ensuring that work and context are preserved during reloads and interruptions. It introduces Novita AI as a new Bring Your Own Key (BYOK) provider and resolves several issues with Windows installations and rich answer rendering.
Key Highlights
- Conversations retain their thread and context across reloads and continuations, preventing lost work.
- Previews remain stable and are not overwritten by stale file reads or blank snapshots.
- Novita AI is added as a BYOK provider with model catalog support (DeepSeek, MiniMax, Qwen, etc.).
- Rich answers are decoded safely to prevent raw data spilling into the conversation.
- Windows installations now correctly connect to the local engine on first launch.
New Features
- Novita AI BYOK provider integration
- Conversation state persistence for planning and continuations
- Stable preview snapshots for real pages
- Safe decoding of thinking-stream cards
- Fixed Windows desktop handoff
Full Release Notes
### 🌟 Codename: *Every Run, Kept Intact* 🧭 **`26 PRs` · `13 contributors` · 1 day** — **A run should survive the real world, not just the happy path.** Reloads, follow-up questions, preview changes and first launches used to be the moments work could lose its place; now the project keeps hold of it. 🚀 ## 🔥 Highlights - 🧵 **Conversations keep their thread when the work keeps moving.** Planning answers, live task continuations, blocked-task decisions and recovered output stay with the task that produced them—even after a reload or a follow-up. Forked chats start clean instead of inheriting another conversation’s work. A planning-only reply now ends as the answer you received, not as a red failed run. (#8322, #8195, #8008, #8007, #8172, #8029, #8158) Thanks @itscheems, @lefarcen. - 🖼️ **Your current preview stays the preview you chose.** Real pages no longer turn into blank snapshots, and a stale file read cannot steal the chat preview while you are inspecting something else. (#7125, #8067) Thanks @huynextlevel. - 🃏 **Rich answers arrive as answers, not markup.** Open Design decodes cards in the thinking stream and safely drops a malformed card instead of spilling its raw data into the conversation. (#8258, #8264) Thanks @lefarcen. - 🪟 **A fresh Windows install gets to the app, first time.** The packaged app no longer mistakes its initial payload state for a failed handoff, so the web interface can connect to its local engine on first launch. (#7520) Thanks @lorenzozanee. - 🔑 **Novita AI is ready in the BYOK picker.** Choose Novita AI, bring your key, and pick from its current DeepSeek, MiniMax, Qwen, GLM, Kimi and GPT-OSS models without hand-building the provider setup. (#6327) Thanks @jax-novita. > 📥 **Download:** Tag `open-design-v0.24.0`. > > | Platform | Architecture | Assets | > |---|---|---| > | macOS | Apple silicon | [DMG](https://github.com/nexu-io/open-design/releases/download/open-design-v0.24.0/open-design-0.24.0-mac-arm64.dmg) | > | macOS | Intel | [DMG](https://github.com/nexu-io/open-design/releases/download/open-design-v0.24.0/open-design-0.24.0-mac-x64.dmg) | > | Windows | x64 | [Installer](https://github.com/nexu-io/open-design/releases/download/open-design-v0.24.0/open-design-0.24.0-win-x64-setup.exe) | ## ✨ Added ### 🔑 BYOK, models and media - Novita AI joins the BYOK provider presets with its model catalog and API-key management link. (#6327) Thanks @jax-novita. ## 🔁 Changed ### 🧠 Agents, runtimes and sandbox - Workspace subscriptions reset cleanly when you change profiles, keeping the projects and account state in front of you in sync with the workspace you actually selected. (#8020) Thanks @AmyShang-alt. - When both are installed, local Agent detection prefers the direct Copilot CLI over VS Code’s bootstrap copy. (#7974) Thanks @tony-box. - `od automation create` and `od automation update` can again accept comma-separated skill context without failing before the command runs. (#7612) Thanks @johnkattenhorn. - `od media generate --help` now shows help without trying to start a generation. (#8066) ### 🎨 Studio, editing and canvas - Manifest-backed Markdown artifacts now count as real run output, and late cover rendering reaches a clear final state instead of leaving a misleading result behind. (#7586, #8061) Thanks @arccat-114. ## 🐛 Fixed ### 🧠 Agents, runtimes and conversations - Recovered conversations keep the right task verdicts, outputs and planning intent across reloads and continuations. (#8195, #8008, #8007, #8172) - Forking a conversation no longer carries its source run artifacts into the new chat. (#8029) Thanks @lefarcen. - A continued conversation collapses an unanswered question without pretending it was answered. (#8158) - A planning-stage refusal that still gave you a useful reply keeps the normal Done state instead of showing a failed-run card; genuine production failures still surface as failures. (#8322) Thanks @itscheems. ### 🎨 Studio, previews and export - Preview snapshots retain real page content, and delayed file reads cannot replace the preview you are actively using. (#7125, #8067) Thanks @huynextlevel. - Valid thinking-stream cards render as cards; invalid ones remain safely out of the conversation. (#8258, #8264) Thanks @lefarcen. ### 🌍 Localization - French copy is complete across the Design System flow, Library and Memory. (#7906, #7756) Thanks @davezfr. ### 🖥️ Desktop - Fresh Windows installations complete the local desktop handoff and can load the app normally. (#7520) Thanks @lorenzozanee. ## 🙏 Thanks to everyone who shipped 0.24.0 @AmyShang-alt · @arccat-114 · @davezfr · @huynextlevel · @itscheems · @jax-novita · @johnkattenhorn · @lefarcen · @lorenzozanee · @maoxin1234 · @nettee · @Siri-Ray · @tony-box
Related Repositories
Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
A fast reverse proxy to help you expose a local server behind a NAT or firewall to the internet.
RAGFlow is a leading open-source Retrieval-Augmented Generation (RAG) engine that fuses cutting-edge RAG with Agent capabilities to create a superior context layer for LLMs
:cherry_blossom: A command-line fuzzy finder
Fast and extensible multi-platform HTTP/1-2-3 web server with automatic HTTPS