What changed in each release of Ghast Workbench for Mac and Windows. Latest: 0.1.29, 15 September 2026.
0.1.29
Better Teams. Richer agents, group collaboration, and shared tasks.
More connections. Easier plugin setup and support for more tools.
Faster and smoother. Improved startup, long sessions, voice, and background tasks.
Smarter web search. Automatic search and better webpage understanding, no API key required.
Stronger continuity. Better memory and context across long or interrupted tasks.
Simpler Settings. Key options are easier to find and manage.
Less disk usage. Smarter cleanup keeps Ghast lighter over time.
0.1.28
Added Teams mode, with 26 expert agents available at launch. Supports group conversations, shared tasks, and private chats.
Redesigned the plugin store with brand‑style icons. Plugins can now be selected mid‑chat using the @ symbol for installed plugins.
Optimized prompt‑cache reuse mechanism. Enables stable team context, cache breakpoint configuration, and consistent service‑provider routing.
Fixed compatibility issues between GLM tool‑calling and Think Mode. Improved recovery capability after task interruptions.
Fixed installed plugins remaining unavailable after approving updated Skill content.
Fixed Coding searches for text beginning with a dash and file rollback in projects that ignore Ghast workspace metadata.
Fixed roundtable summaries pausing while trying to assign new experts after the original participants had finished.
Fixed expert model selections reverting to the conversation default after saving.
Clarified that roundtable leaders must answer independently before summarizing the other participants.
0.1.27
Added a new plugin system for installing and managing compatible tools, Skills, commands, and connections.
Fixed GPT-family models showing tool instructions in replies, losing tool calls, or stopping before giving a final answer.
Improved research, planning, and long tasks so Ghast checks stronger sources, follows through more reliably, and keeps important context after cleanup.
Improved prompt caching and 0G provider continuity to reduce repeated input and unnecessary cost.
Made learned Skills and memories more reliable across projects and conversations, including selection, reuse, and removal.
Made approval prompts smaller and easier to reach, and added custom date ranges to Usage.
Improved long-running tasks and message delivery across Discord, Slack, Telegram, and WeChat.
0.1.26
Refreshed Chat and Coding Mode with more consistent controls, colors, Tool activity, and clearer command details.
Fixed new replies, Plans, and streaming output sometimes falling behind the latest message, and made pending approvals easier to reach.
Improved model and reasoning selection with search and model families, and fixed selecting or removing multiple Skills from the message box.
Made Coding Mode more reliable when reading, editing, and patching files, running commands, or working across related project folders.
Improved complex tasks, research, and long conversations so Ghast checks stronger evidence, recovers from temporary failures, and keeps completed work settled.
Made installed and learned Skills easier to discover and reuse, with safer personalization and more reliable context recovery.
Improved startup speed, context caching, and 0G provider selection to reduce unnecessary waiting and cost.
0.1.25
Fixed Coding Mode on Windows blocking common commands, SSH connections, local development servers, and work across different project folders.
Made Full Access work as expected, with fewer unnecessary permission interruptions during development tasks.
Improved longer and more complex tasks so Ghast follows the latest request, checks its work, and recovers more reliably after an interruption.
Fixed long replies falling behind the latest output or occasionally repeating themselves.
Made learned Skills more useful, easier to reuse, and less likely to include generic or duplicate suggestions.
Fixed working animations on Windows and added response speed to conversation activity details.
0.1.24
Fixed Coding Mode reading files inside WSL on Windows.
Made Ghast's working status clearer and more lively while it prepares a reply.
MiniMax-H3 now follows the requested video length instead of stopping after only a few seconds.
0.1.23
Made Ghast clearer while it works, with brief progress updates and more compact Tool activity.
Added support for reading more document formats before sending their contents to a model.
Made safe, read-only project work require fewer permission confirmations in Coding Mode.
Improved long and complex tasks so Ghast keeps the right tools available, verifies its work, and recovers without losing progress.
Fixed replies sometimes disappearing while they were still being generated, and corrected context usage after reopening a conversation.
Improved tool use with GPT-family models, including more reliable results after long output or an unsuccessful first attempt.
Fixed Coding Mode commands failing after installing or updating Ghast on Windows.
0.1.22
Coding Mode can run commands on Windows again, while important private files stay protected.
Safe project reads now need fewer permission confirmations.
Complex tasks and longer research stay on track more reliably, including after using tools.
Fixed replies sometimes stopping unexpectedly and reduced background work that could make the app feel less responsive.
0.1.21
Made long conversations and complex tasks more reliable, with fewer interruptions and less lost context.
Improved web research so results are broader, more useful, and less likely to crash the app.
Made Tool and Skill activity easier to understand, including when Ghast recovers from a temporary problem.
Fixed PDFs and unsupported files breaking every later message in the same conversation.
Added MiniMax-H3 video creation from text or an image, with in-app playback and downloads.
Improved response quality, personalization, cost efficiency, and everyday reliability across macOS and Windows.
0.1.20
Fixed sudden context spikes in long conversations and made automatic cleanup follow your chosen limit.
Fixed memory saving on Windows, including on new devices and after interrupted downloads.
Fixed copy buttons for generated code and removed duplicate tool sections.
Improved interactive content in light and dark modes.
Added clearer TeeML verification when selecting supported models.
0.1.19
Added reply reports and product feedback with clearer consent, delivery status, and diagnostics that help investigate problems.
Made startup and conversation navigation faster by reducing unnecessary background work and loading less data up front.
Made longer web research more reliable, with automatic fallback for blocked searches, lower browser memory use, and fewer repeated searches.
Improved recovery when MCP tools disconnect or a long-running reply is interrupted, especially after the app has been idle.
Improved Coding Mode commands, formatting, file updates, language services, and terminal behavior across Windows and macOS.
Refined tool cards, failure states, table copy and export controls, and official language icons for code files.
Hardened macOS and Windows release packages by excluding incompatible native components and preserving differential updates.
0.1.18
Fixed cases where installed Skills could be skipped or fail to run in conversations.
Improved image and app screenshot support for compatible 0G vision models.
Improved Skill selection and tool reliability across Windows and macOS.
0.1.17
Made Coding Mode more reliable when reading projects, editing files, running commands, and checking its work.
Made web, X, and Rednote searches faster, broader, and less likely to repeat the same search.
Improved recovery when a reply, tool task, or long conversation is interrupted or appears stuck.
Improved Computer Use across desktop apps, with more reliable controls and better Windows support.
Improved memory quality and multilingual recall while reducing unnecessary background work.
Improved tool approvals, reasoning display, model compatibility, and several chat interface details.
0.1.16
Made conversations smoother, clearer, and more natural during long or tool-heavy tasks.
Added smarter suggestions, interactive previews, and app context capture with an optional global shortcut.
Improved chat controls with slash commands, reasoning levels, and better layouts in narrow windows.
Improved reliability when stopping or resuming tasks, reconnecting terminals, and using interactive content.
Added clearer 0G Storage backup and restore progress with more complete data syncing.
Improved profiles, invitations, referral points, point history, and reward redemption.
0.1.15
Fixed 0G wallet connectivity and restored recent Chat and Coding conversations.
Simplified 0G Compute settings and added Anthropic model grouping.
0.1.14
Simplified settings and 0G Router setup.
Improved reliability across tools, ACP, MCP, Hooks, memory, and long chats.
Added account profile editing and fixed image attachments.
Fixed several bugs and improved the user experience.
0.1.13
Improved startup speed.
Improved signed-in user reopening reliability.
Hardened pre-login routes and database migrations.
0.1.12
Fixed 0G model catalog refresh and display issues.
Added 0G model family filters.
Updated tool security policies.
0.1.11
Fixed desktop app model save failure.
0.1.10
Fixed Windows shortcut refresh after app updates.
Fixed Windows send button and Enter-to-send behavior.
0.1.9
Improved Loop Engineering stability and recovery for long tasks.
Added user-approved crash diagnostic uploads.
Strengthened WebSearch safety with stable fallback paths.
Raised the Skill learning threshold to reduce low-value Skills.
Added personal AGENTS instructions for chat and coding tasks.
0.1.8
Improved MiniMax / DeepSeek compatibility
Hardened managed Skill execution environment
Fixed RedNote login state cleanup
Improved settings and preference persistence
Fixed 0G message recovery after interruption
0.1.7
Configuration for 0G Compute is now automated, including the provisioning of Management Keys and inference API keys.
Added 0G Pay top-up and automatic Payment Layer funding.
Improved 0G Compute settings with cleaner funds, model, menu, icon, and dark-mode controls.
Fixed provider icon mapping and theme-aware model logos for 0G routed models.
0.1.6
Web search is more stable and no longer crashes the app when the built-in browser or Google path fails.
Google remains the default search engine, with automatic fallback when a search page is blocked or unavailable.
Skills run more reliably on clean devices, even when required command-line tools are missing at first.
Third-party skill installation and loading now work more consistently across macOS and Windows.
Long tool tasks recover better after interruptions instead of leaving the chat stuck.
Windows screen capture failures are handled safely without restarting the app.
Activity Recorder permission refresh now shows a real check result instead of appearing to do nothing.
Settings and release checks now cover more of the search, skill, and permission paths before publishing.
0.1.5
Ghast runs more reliably on long tasks and no longer exits prematurely.
Failed/interrupted tool runs recover smoothly with clearer follow-up steps.
