0.85.0

Highlights

🛠️ Take Control of Your Skills

Skills just got a serious upgrade in flexibility. You can now disable skills per project, so each task only loads the skills that actually matter for the codebase it's working in. Skill authors get new frontmatter options too: disable-model-invocation prevents the agent from triggering a skill on its own, while user-invocable skills let you decide exactly when a skill should run. To make skills more discoverable, built-in skill names now appear in the / command autocompletion.

👁️ Live File Watching During Tasks

A new file watcher now monitors the task directory while a task is running, keeping task files, worktree status, and autocompletion fresh in real time. When your agent creates, modifies, or deletes files mid-run, the UI reflects those changes immediately - no more stale file lists or outdated worktree information while the work is still in progress.

🔍 Find in File Editor

The file editor now supports find in file with Ctrl+F / Cmd+F. Quickly locate that one function or string in larger files opened from your context without leaving the editor.

🤖 New LLM API Provider

AiderDesk now ships with a built-in LLM API provider. Add your API key (or set the LLMAPI_API_KEY environment variable) and your available models load automatically - with pricing and context information included. One more way to connect the models you want to the agent workflow you already have.

🧩 Extension API: Message Dropdown Menu Actions

Extensions can now attach custom actions to the message dropdown menu via the new task-message-bar-menu UI placement. This opens the door for extensions that operate on individual messages. And to make custom extension commands more powerful, images attached to the prompt are now passed to custom extension commands.

Other Changes

• Prompt Autocompletion: Added a /resume command for quickly resuming previous tasks.
• Command Palette: Added actions to show or hide archived tasks.
• Project Tabs: Added a context menu action to open the project folder directly from project tabs.
• Model Library: Updated default models for built-in providers to their latest versions, and added a low reasoning effort option for the Z.ai plan provider.
• Worktrees: Clarified the worktree to local mode switch dialog for shared worktrees.
• Task Sidebar: The in-progress spinner now propagates to the closest visible ancestor task when parent subtasks are collapsed.
• Code Highlighting: Falls back to plain code when the auto-detected language yields no highlighting matches.

Bug Fixes

• Fixed a crash during message streaming (React maximum update depth exceeded) caused by duplicate message ids.
• Fixed duplicate keys in git branches popup rows.
• Fixed committing deleted files by retrying staging with git add -A -f.
• Fixed partial commits including unrelated staged files - only the selected files are now committed via pathspec mode.
• Fixed in-progress tasks being evicted from the task mount pool when switching tasks.
• Fixed the find in file shortcut being swallowed by inactive background project task views.
• Fixed the wrong scroll offset in the updated files diff modal when tracking headers were not found.
• Fixed provider options leaking into task context messages.

🧩 Extensions

• Cursor SDK: Improved context message persistence and MCP server config retrieval.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.84.0

Highlights

📁 Git Support for Every Project

Projects that aren't git repositories can now join the party! AiderDesk will automatically initialize a git repository for non-git projects, unlocking everything that comes with it: branches, worktrees, and version tracking. No more sitting on the sidelines - all of AiderDesk's git-powered workflow is now available for any project you open.

⚡ Major Performance Improvements

This release is full of speed-ups across the board:

• Workspace file trees are now virtualized (migrated to headless-tree), so large projects render smoothly regardless of how many files and folders they contain.
• Updated files diffs are loaded lazily on demand instead of being computed upfront with the file list - the diff modal opens noticeably faster.
• Background projects stay lightweight: a task is only loaded when you first activate the project, reducing memory usage while you work elsewhere.
• Tooltips render faster across the app.

🛡️ Safer Git & Worktree Workflows

We've hardened the git and worktree experience with a series of reliability improvements:

• Switching to local mode is now blocked when the worktree has uncommitted changes, so you'll never silently lose work again.
• Creating a task worktree whose branch is already checked out in another worktree now creates a unique sibling branch instead of failing.
• Task branches are no longer renamed when a worktree is shared with other tasks.
• Ignored files are now staged with --force on retry, so ignored files no longer fail to commit.
• Uncommitted diffs work in freshly initialized repositories with no commits yet (diffing against the empty tree, including untracked files).

🔐 OpenAI Zero Data Retention Support

A new "Store Responses" setting for OpenAI lets organizations with Zero Data Retention (ZDR) requirements use AiderDesk by running requests in stateless mode. If your organization doesn't allow OpenAI to store request data, this setting keeps you compliant.

✅ Bulk Tool Approval for MCP Servers

Approving MCP tools one by one is history: a new "All" select in the server tools list lets you bulk-approve tools per MCP server. Combined with the new keyboard navigation in the git branches dropdown (arrow keys, submenu controls, and Escape to close), power workflows just got a lot smoother.

Other Changes

• Task Management: Added an ephemeral indicator for transient message groups, and task IDs are now stored in the messages database, enabling per-task usage tracking.
• Agent Reliability: The agent no longer stops prematurely when a model finishes with stop but produces no messages (or only reasoning content) - the step is retried automatically.
• Updated Files Diff Modal: Files are now ordered to match the sidebar tree, commit errors show clean, selectable messages (no more IPC wrapper noise), and full old/new file contents are fetched on demand.
• Markdown Rendering: Numbered lists starting mid-paragraph now render as proper lists instead of plain text.
• In-Chat Search: The messages scroll container is now reliably detected via a dedicated callback.
• Model Library: The temperature column now fits its values.
• UI Polish: Toast notifications got a more visible border (errors use the error color), and the extensions search field is auto-focused when opening extensions settings.
• Extension API: projectDir is now passed to extension component wrappers on header placements.
• Platform: Updated Electron to 44.3.0.

🧩 Extensions

• Git Log: Added per-file diff sections with scroll-to-file, a commit context menu, and smarter project directory selection that prioritizes the current project.
• Task Scheduler: Added one-time scheduling - schedule a task to run once at a specific time.
• SearXNG Search: Now talks to the SearXNG API directly instead of spawning a CLI process - more reliable and faster.
• tps-counter: Improved streaming and subagent token rates (Thanks @pcfreak30)
• Cursor SDK: Fixed the vendored ripgrep path configuration.
• Reasoning Selector: Added an extension icon.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.83.0

Highlights

🖥️ Terminal Sessions That Stick With You

Terminal sessions now persist when you switch between tasks or projects. Come back to a task and your session is reattached instantly with full output replay, so you never lose track of what happened. Sessions restart automatically after they exit, and touch input is now supported - a much more reliable terminal experience all around.

⚡ Instant Task and Project Switching

Recently active tasks now stay mounted in the background, making switching between tasks and projects nearly instantaneous. No more waiting for everything to reload every time you jump around your workspace.

📝 Bash Command Descriptions

Modern models love using shell commands for all kinds of actions, and sometimes it can be hard to understand what a command actually does. Bash tool commands can now carry an optional description shown right in the tool message, so you can see at a glance what the command is supposed to do.

🔐 Interactive Git Authentication

Git operations that require authentication now prompt you interactively for credentials and passphrases instead of failing silently. Whether it's cloning a private repository, pushing, or pulling - when Git needs credentials, AiderDesk asks.

🛡️ Automatic Renderer Crash Recovery

If the renderer process crashes, AiderDesk now recovers automatically by reloading, with crash loop protection to prevent endless restart cycles. Event delivery has also been hardened to gracefully handle sending to destroyed or crashed windows. A crash should no longer take your whole session down with it.

🌳 Worktree Post-Create Command

You can now configure an optional post-create command that runs in a freshly created worktree, with project, worktree, task, and branch environment variables available. Perfect for initializing submodules, installing dependencies, or any other setup your project needs before work begins.

⌨️ Task Management in the Command Palette

The command palette is now a task management hub: create subtasks and sibling tasks, rename tasks, change their state, and copy task IDs (Thanks @pcfreak30) - all without leaving the keyboard.

Other Changes

• Updated Files Diff View: You can now select which changed files to include when committing, and file selection is properly restricted to uncommitted files.
• Copy as Image: Chat messages can be copied as an image to the clipboard.
• Custom Commands: Queued custom commands now run directly from the queue instead of appearing as user prompts, and they queue correctly while another prompt is running.
• Chat Auto-Scroll: Auto-scroll now pauses only when you scroll upward and resumes reliably when you return to the bottom - no more false pauses from internal scroll adjustments.
• Worktree Reliability:
  • Fixed worktree rebase silently dropping commits when the tracked base commit was stale.
  • Using merge-base for worktree flat diffs to avoid reverse diffs.
  • Uncommitted changes are carried over when switching to local after a checkoutless merge.
  • Fixed a subtask inheriting the parent worktree when its working mode was overridden to local.
  • The branches dropdown refreshes correctly after a branch switch or rebase completion.
  • The git branches tooltip now shows icons and a rebase hint when a worktree is behind its base branch.
• Git: Adding a file to Git now updates the file state optimistically without refetching, and checkout no longer detaches the main repository branch.
• Prompt Field: The Escape key no longer interrupts task processing while you're typing.
• UI Polish: Tool message status icons no longer shrink when message content is long, and password fields got improved toggle alignment with new visibility translations.
• Performance: Reduced the diffs worker pool size and AST cache, lowering memory usage.

🧩 Extensions

• BMAD: Fully rewritten around bmad-method v6, bringing the latest BMAD methodology workflows to your projects (Thanks @777marvin). Also fixed the default Welcome message showing outside of BMAD mode.
• OpenAI Codex: Added a 5-hour usage window display, and token storage was moved to the extension folder with legacy migration.
• Cursor SDK: Updated @cursor/sdk to the latest version and improved tool message error handling and result formatting.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.82.0

Highlights

🌿 Full Git Branch Management, Safer Pushes & AI-Powered Recovery

This release turns AiderDesk into a complete git workstation. The new branch management UI lets you work with branches without ever touching the terminal - and when something goes wrong, your AI agent is one click away from fixing it.

• Branch management: list, create, checkout, rename, and delete branches directly from the UI.
• Safer pushes: a confirmation dialog now shows you exactly which commits are about to be pushed, with an opt-in force push option when you really need it.
• Smarter pulls: when branches have diverged, you can choose the pull strategy that fits your workflow.
• Automatic upstream setup: pushing a branch without a remote tracking branch now sets up the upstream automatically.
• Checkoutless merge/squash for worktrees: merge or squash worktree changes into the target branch without ever checking it out - faster and safer for worktree-based workflows.
• Resolve with AI: when a git operation fails (conflicts, interrupted rebases, and friends), a new "Resolve with AI" action hands the problem to your agent to analyze and fix the error for you.
• Command palette integration: all git actions are now available in the command palette with configurable hotkeys in Settings.
• Reliability fixes: worktree commit tracking and operations now correctly use the base branch, integration checks no longer fail when a worktree directory no longer exists, rebasing onto your own branch is prevented, and git state refreshes automatically.

🔍 Search Gitignored Files with the grep Tool

The grep power tool now accepts an ignoreGitignore input parameter, letting your agent search inside gitignored files. When the answer hides in build outputs, vendored dependencies, or generated code, your agent no longer comes back empty-handed.

🧠 Delete Stored Memories on the Spot

Agents build up memories as they work with you - and now you have immediate control over what sticks. When the store_memory tool saves something you don't want to keep, you can delete it right from the tool message the moment it's created. No digging through settings, no letting bad memories linger.

Other Changes

• Task Management: Improved the create_task tool description and worktree inheritance for new tasks.
• Subtasks: Distinct "Create subtask" label when creating subtasks via the create task tool message.
• Reliability:
  • Atomic context saving with corrupt recovery, so a crash can no longer corrupt your task context. (Thanks @pcfreak30)
  • Automatic retry for transient network errors during window load prevents the dreaded blank window.
• Performance: Sped up loading of available extensions. (Thanks @pcfreak30)
• Providers: Added the x-opencode-session header to the OpenCode Go provider.
• Bug Fixes: Fixed the working mode switch dialog staying open and appearing frozen while switching.

🧩 Extensions

• OpenAI Codex: Dedicated sign-in UI with device flow authentication for a smoother Codex account setup, plus the new gpt-6-astra model. (Thanks @pcfreak30)
• TPS Counter: Diagnostics logs moved to debug level to reduce noise. (Thanks @pcfreak30)

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.81.0

Highlights

🌳 Nested Subtasks

Complex work no longer has to be flattened into a single level of subtasks. Subtasks can now have subtasks of their own, letting you (and your agent) break work down into a true hierarchy - epics, phases, and individual steps - all visible as a tree in the task sidebar. In agent mode, the task creation tool supports nesting natively, so your agent can organize bigger endeavors on its own.

⌨️ Activate Skills with the /skill: Command

Skills are now just as easy to trigger as any other command. Type /skill: in the prompt field and pick the skill you want from the autocomplete list, optionally followed by a prompt to run it with. No more digging through menus - your skills are a few keystrokes away.

🔐 Clone Repositories over SSH

Adding a project from a remote repository now supports SSH URLs alongside HTTPS. If you work with private repos or already have SSH keys set up, cloning is now seamless - no tokens required.

🧩 Discover Popular Extensions

The extensions gallery now shows install counts for available extensions, so you can see at a glance which extensions the community relies on most. You can also sort by most installed or by name - perfect for finding proven favorites or discovering new tools worth trying.

🖼️ Queued Prompts with Images

Queuing prompts while your agent is busy keeps the momentum going - and queued prompts can now include images too. Attach screenshots or mockups, and they'll show up with an image indicator in the queued prompts list so you always know what's waiting in line.

Other Changes

• Worktree Reliability: Improved worktree pruning, and fixed worktree merges failing on in-progress rebases or colliding with untracked stash files (Thanks @sbodese).
• Auto-Compaction: Fixed auto-compaction not triggering for models without a known context window - it now falls back to the configured token threshold. For Ollama models, the context window is now read from the model's reported context_length, enabling the token usage bar and percentage-based compaction (Thanks @sbodese).
• Task Management:
  • The task creation tool now has clearer create only, wait for finish, and run in background modes.
  • The workspace sidebar size is now stored on the project level instead of per task.
  • Subtasks of archived parents are now hidden at any nesting depth when archived tasks are hidden.
  • Added loading indicators for workspace sections.
  • Added a highlighted background for the task sidebar filter and search icons when active.
• Voice: Added support for the gemini-3.5-transcribe-live transcription model, now the default for Gemini voice.
• Extensions: Project started events are now re-delivered to extensions that load or reload after a project has started, fixing a startup race; fixed queued prompts for extension custom agent runs.
• Bug Fixes:
  • The MCP servers initialization loading message is no longer shown without prompt context.
  • The agent loop no longer silently re-sends the same prompt when step processing fails.
  • Usage reports no longer fail when provider models fail to load.
  • Fixed context size undercounting that could skip compaction when tool results were appended after the last usage report.

🧩 Extensions

• Learn: A new extension that adds a /learn command to create reusable skills from any source - directories, URLs, conversation history, or pasted notes.
• Provider Quota: Added OpenCode Go provider quota support.
• Cursor SDK: Agent run cancellation now uses an abort signal for cleaner stops.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.80.0

Highlights

🔐 TLS Certificate Settings for OpenAI & Anthropic Compatible Endpoints

Working behind a corporate proxy or with self-hosted models that use a custom certificate authority? You can now configure TLS certificate settings directly for OpenAI Compatible and Anthropic Compatible provider endpoints. No more workarounds - just secure, reliable connections to your own model deployments straight from the provider configuration.

📥 Clone Git Projects Directly from the Open Project Dialog

Getting started with a new project just got simpler. The Open Project dialog now lets you clone a git repository right where you would normally add it as a project - no more switching to a terminal, running git clone, and then hunting for the folder. Paste the repository URL, pick a location, and you're ready to start coding with AiderDesk.

🚀 OpenCode Go as a Direct Provider

OpenCode Go joins the family of directly supported providers. Add your credentials in Settings -> Providers and use its models natively - no proxy setup required.

🧩 Extension API: Disposable Resource Management

Extensions can now manage their resources cleanly with context.addDisposable() - register a setup function that returns a cleanup function, and it runs automatically when the extension unloads. Cleanups execute in reverse registration order and support async functions, so no more leaked timers, watchers, or child processes. We also fixed extension lifecycle handling: disabled and uninstalled extensions no longer keep resources running in the background, and toggling an extension on/off now properly calls onLoad/onUnload.

Other Changes

• Settings UI: Responsive design, so Settings now adapts nicely to smaller windows and screens.
• Tools: Added result limits to the glob tool to prevent excessive output.
• Message Presentation: Improved rendering of multiline bash tool commands.
• Extension Development: Added a validator script for extension UI components, used by the extension-creator skill.
• Bug Fixes:
  • Fixed extraction of images and safe serialization for non-text data in tool results - images and other binary content coming back from tools are no longer corrupted by truncation.
  • Fixed occasional virtualized messages not showing up.
  • Fixed the providers API not returning data after an update, and added runtime array checks for provider data.
  • Fixed the task loading state not being reset on error.
  • Fixed syntax highlighting while tool messages are streaming.
  • Fixed excessive bottom padding causing empty space under messages in the virtualized list.
  • Disabled router transitions to prevent UI issues.
  • Correctly handling the project tab selection when the path ends with /.

🧩 Extensions

• SearXNG Search: Correct parsing of results when using the CLI.
• OpenAI Codex: Correctly loading usage information.
• Cursor SDK: Added a shell cwd fallback patch.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.79.0

Highlights

⌨️ Command Palette

AiderDesk now has a centralized Command Palette - press Ctrl/Cmd+Shift+P (customizable in Hotkeys settings) from anywhere to quickly jump between projects and tasks, open settings, or trigger actions without reaching for the mouse. Under the hood it's powered by a new UI actions system that centralizes commands across the app, so actions are consistent everywhere they appear - in the palette, in context menus, and in the header. Extensions providing UI actions are picked up by the palette automatically, making it the one place to reach everything AiderDesk can do.

📝 File Editor Replaces File Viewer

The read-only File Viewer is gone, replaced by a full File Editor. You can now edit files directly in AiderDesk instead of jumping out to an external editor, and the editor supports multiple files at once with tabs, so reviewing agent changes across several files feels natural. Open it from any file listing in the Workspace, or use the new Open editor action straight from the Command Palette.

🧩 Extension UI API: New Components

Extension authors get two new building blocks for richer custom messages: CodeBlock for syntax-highlighted code rendering and ExpandableMessageBlock for collapsible long content. Together with this release's other UI improvements, custom extension messages can now look and feel exactly like native AiderDesk components.

🌿 Untracked Files in Updated Files

The Updated Files panel now also shows untracked files created by your agent, so nothing new slips through unnoticed. Each untracked file comes with an Add to git action, letting you stage new files into version control without leaving AiderDesk.

⚡ Performance and Stability

Several under-the-hood improvements make AiderDesk leaner and more reliable: tool outputs are now bounded and large file reads are streamed, preventing memory exhaustion during long tasks with verbose tools or big files; project and task switching is faster; and task state determination is optimized for tasks showing a proceed (y/n) confirmation.

Other Changes

• UX: The prompt field is automatically focused after closing overlay dialogs, so you can keep typing without clicking.
• Themes: Added a new crimson theme.
• Projects: Simplified active project selection by deriving it from the URL parameter for more reliable project routing.
• Dependencies: Updated Electron to v43 along with other libraries, and fixed a missing OpenTelemetry trace exporter dependency.

🧩 Extensions

• Git Log: A new extension with an IntelliJ IDEA-style commit history viewer - browse the git log of any open project with a virtualized commit list, project and branch selectors, search by message/author/hash, and full commit and per-file diffs.
• Programmatic Tool Calls: Added a custom UI message for better visibility of the executed code and its result.
• OpenAI Codex Auth: Improved caching for faster and more reliable authentication.
• Providers Quota: Correctly loading usage info for the Z.ai provider.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.78.0

Highlights

🗂️ File-Based MCP Server Configuration + Dedicated Settings Tab

MCP server management has been completely reimagined. Server configurations are now stored in JSON files on disk instead of application settings, with support for both global and project-scoped servers:

• Global servers — stored in ~/.aider-desk/mcp-servers.json, available to every project.
• Project servers — stored in <projectDir>/.aider-desk/mcp-servers.json, scoped to that project and committed alongside your code.

Project servers take precedence over global ones with the same name, and the two are merged automatically. Files are watched live, so any changes - whether made through the UI, an external editor, or a git pull - are picked up instantly. This also means MCP Servers now has its own dedicated tab in Settings, with a context switcher to manage global vs. project servers side by side. Under the hood, the MCP integration was migrated to the AI SDK's native MCP client for a more robust and maintainable foundation.

🔐 OAuth Authentication for MCP Servers

Remote MCP servers that require OAuth (such as Sentry, Atlassian, and others) are now supported natively. No more mcp-remote workarounds - just click Connect in the MCP Servers tab, complete the authorization flow in your browser, and the connection is stored automatically. For servers with non-standard auth flows, the mcp-remote proxy option remains available as a fallback.

🧩 Extension API: getMcpServers

Extensions can now retrieve MCP server configurations programmatically via getMcpServers(projectDir?), which returns a merged map of global and project-scoped server configs. This enables extensions to interact with or expose MCP server state in new ways.

🔀 Improved Worktree Conflict Detection

Worktree merge and rebase operations are now more reliable thanks to improved conflict detection. The system better identifies conflicting changes during these operations, so you'll get clearer feedback before things go wrong. The rebase button label has also been renamed to more accurately describe the action.

Other Changes

• Browser Connectivity: Prevented duplicate project/task initialization and improved socket connection handling when connecting from a browser client.
• Browser Compatibility: Improved OS detection for browser-based access.
• Subagents: Fixed missing context files being omitted from subagent runs.
• Touch Devices: Prevented auto-scroll pause when scrolling up on touch devices.
• Input History: Correctly loading input history for all opened projects, not just the active one.
• Aider Integration: Removed an unnecessary log message when a file is git-added outside a repo or ignored.

🧩 Extensions

• Skill Gallery: Added infinite scroll for large skill lists, plus the new reverse-skill entry.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.77.0

Highlights

🧠 Optimized System Prompt for Latest Models

We've optimized the default system prompt and init prompt to be more concise and token-efficient. As described in The New Rules of Context Engineering, the latest frontier models no longer need the same level of hand-holding in their instructions, and the more verbose prompts were adding unnecessary token overhead. This optimization reduces your initial token usage per conversation.

If you're relying on less capable or smaller models and notice a negative impact, you can restore the previous verbose system prompt by installing the Legacy System Prompt extension from Settings -> Extensions -> Available.

📜 Read-Only Server Mode

We're introducing Read-Only Server Mode, which allows you to start the AiderDesk server in a view-only state. You can either run the entire server instance in read-only mode, or keep your Electron app running normally while the server is read-only. This lets you safely expose your AiderDesk instance to other devices to check task progress, or share it with friends and colleagues to showcase your workflows - all without risking accidental modifications. Extension UI components can also be toggled in read-only mode. Learn more in the documentation.

🎙️ Live Transcription with gpt-live-transcribe

OpenAI's latest live transcription model, gpt-live-transcribe, has been added and is now the default voice transcription model when you have the OpenAI provider configured and voice enabled in Settings -> Voice. Live dictation is a very effective way of inputting your prompts - if you haven't tried it yet, it's available directly from the prompt field or via the ALT + V shortcut.

📜 Improved Auto-Scrolling During Streaming

After making virtualized message rendering the default, there were some inconsistencies with auto-scrolling - in some cases with longer runs, the scroll would stick to the end of the messages unexpectedly. The new scrolling implementation fixes these issues and also removes a small scroll flickering that was visible during streaming.

🧩 Extension API Updates

Two new additions to the Extension API:

• reloadTasks function: Reloads tasks from disk into the project, allowing extensions to detect and sync external changes to the task directory.
• onTaskDeleted hook: Fired when a task is deleted, enabling extensions to clean up resources or trigger follow-up actions.

Other Changes

• Task Creation: The create_task tool now inherits the agent profile and selected model as defaults, so subagent-created tasks match the parent's configuration.
• List Tasks Tool: Added additional search options for more flexible task filtering.
• CLI Run: Added an option to autostart the AiderDesk server when using the run command.
• Task Management: Added commit progress visibility and cancel action. Added support for saving edited prompts.
• Loading UX: Added a loading projects overlay instead of showing the Open Project page while projects are loading.
• Requesty: Correctly calculating cached tokens for provider models.
• Bug Fixes:
  • Skip adding Todo tool messages when streaming tools.
  • Added --clear flag to uv venv creation.
  • Properly using AIDER_DESK_TASKS_DIR constant when loading tasks in project.

🧩 Extensions

Updates

• Legacy System Prompt: Restores the verbose pre-optimization system prompt for less capable models that need more prescriptive guidance.
• Reasoning Selector: Now reuses the reasoning setting per provider/model and supports full reasoning model selection.
• Cursor SDK: Added conversion of subagent messages and cost usage, plus full reasoning model selection.
• OpenAI Codex Auth: Updated model costs.
• Skill Gallery: Added new skill sources from Addy Osmani and the superpowers repository.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀

0.76.0

Highlights

🧩 Extension System API Updates

The Extension API continues to grow with several new capabilities:

• resumeTask method on TaskContext: Extensions can now programmatically resume a paused or interrupted task, enabling more sophisticated automation workflows.
• getOpenProjects on the Extension Context: Retrieve all currently open projects, useful for extensions that need cross-project awareness.
• modelCallSettings on onAgentStarted: Override AI SDK model call parameters at runtime - perfect for dynamically adjusting temperature, max tokens, or other provider-specific settings per task.
• task-sidebar-item-badges UI placement: Extensions can now render custom badges on task items in the sidebar, enabling at-a-glance status indicators.
• task-state-actions-all UI placement: Extracted as a chainable component placement, allowing multiple extensions to add actions to the task state actions area.
• Editable Extension Provider Models: Models coming from extension providers can now be edited directly in the Model Library, so you can fine-tune token limits and other settings without waiting for extension updates.

📎 Drag and Drop Tasks to Organize Your Workflow

You can now drag and drop tasks onto other tasks to make them subtasks, making it much easier to organize your task hierarchy. Reorganize your workflow on the fly without recreating tasks or losing context. Subtasks can also be dragged out to promote them back to top-level tasks.

⏰ Task Scheduler (Extension)

You can now schedule your tasks to run automatically on a cron or periodic basis with the Task Scheduler extension. Whether you need nightly code reviews, periodic test runs, dependency update checks, or daily project summaries - just write your prompt, set the schedule, and let it run. It supports standard cron expressions, periodic intervals, and even natural language one-time schedules like "tomorrow at 9am". Scheduled tasks are duplicated and run automatically, with pause/resume and run-now controls. A must-have for any recurring development workflow.

🧠 Reasoning Selector (Extension)

The new Reasoning Selector extension adds a reasoning-effort selector next to the model selector in Agent Mode, letting you quickly switch between None, Minimal, Low, Medium, High, and xHigh reasoning levels. New tasks inherit the selected reasoning effort from the most recently updated task in the same project, so you can set it once and forget it.

📎 Binary Files Tool (Extension)

The Binary Files Tool extension adds the read_binary_file tool, which reads local binary files and attaches them to the next model request as multimodal content - perfect for asking the AI to inspect images, PDFs, Office documents, audio recordings, or videos. Just mention the file path and what you want to learn about it in your prompt.

🗂️ Skill Gallery (Extension)

The Skill Gallery extension provides a browsable gallery of community skills that can be discovered and installed directly from the UI. Browse available skills from various sources, install them with a click, and they're ready to use in your agent conversations.

🔧 Shell Initialization for Bash Tool

The bash power tool now includes shell initialization that loads environment configs (like .bashrc, .zshrc, or profile files) before executing commands. This ensures your PATH, aliases, and environment variables are correctly set up, so commands that depend on your shell configuration work as expected.

Other Changes

• Saved Prompts: Custom and extension commands can now be used in saved prompts, making your prompt templates even more powerful.
• Model Handling: Falling back to the default agent profile when opening a project with an agent profile that no longer exists. Also correctly handling multiple reasoning message parts.
• UI Improvements:
  • Added Copy link context menu action to links in messages.
  • Correctly rendering text on the same line as numbers in ordered lists in message markdown.
  • Prevented accidental task sidebar resize on click.
  • Prevented project tab close button from overlapping the scroll arrow (Thanks @marceli1404).
  • Avoided showing the password dialog on startup in some cases.
  • Added a gap between extension components with task-top-bar-left placement.
• Bug Fixes:
  • Correctly showing Updated Files in repos without any commits.
  • Prevented crash from circular references in error logging (Thanks @marceli1404).
  • Correctly resolving the base branch on worktree tasks after additional commits on the base branch.
  • Cleared exited aider process state (Thanks @ag-linden).
  • Sending the failed event when the aider process exits.

🧩 Extensions

• wigolo: Local-first web search, fetch, crawl, and research tools powered by wigolo. All intelligence runs on your machine with no API keys required and $0 per query.
• OpenAI Codex Auth: Added remaining usage info and correctly using systemPrompt.
• Cursor SDK: Updated to use the new stream processor with token usage support, and added image support in user prompts.

💡 We'd love to hear your thoughts! Your feedback helps us improve AiderDesk - feel free to share any suggestions or issues. Happy coding! 🚀