0.75.0

Highlights

🚀 Migrated to AI SDK v7

We've finally migrated our agent engine to AI SDK v7. This lays the groundwork for exciting new features coming in future releases. If you maintain an extension that creates a custom provider, you may need to update it for compatibility - check the extension migration guide if you encounter issues.

📡 Tool Call Streaming

Tool call streaming has finally arrived in agent runs! Instead of staring at a "Preparing tool..." message while the LLM streams long file writes or edits, you can now watch the content stream in real-time - just like you're familiar with from Aider's edit mode. This makes long tool calls feel much more responsive and gives you immediate visibility into what the agent is doing. Some tool messages have been intentionally excluded from streaming to avoid weird behavior, and providers/models can individually disable streaming if needed.

🖼️ Improved Image Handling

images generated by tools are now properly supported and displayed inline in tool messages, and we've fixed issues where image content wasn't being sent back to the LLM correctly. You can also now click on images attached to user messages to view them at full size.

📊 Usage Dashboard: Cache Hit Rate & Savings

The usage dashboard now includes cache hit rate and savings charts, giving you better insight into how effectively prompt caching is being utilized across your providers and models.

📏 Resizable Task Sidebar

The task sidebar is now resizable - drag the edge to give yourself more or less space as needed. The working mode switcher has also been updated to use explicit buttons for clearer interaction. Additionally, parent tasks now collapse automatically in the sidebar when a subtask is selected.

Other Changes

• Extension Creator Skill: Improved to have a better understanding of AiderDesk's extension capabilities. Added missing OS and AutonomyMode enums to the extension runtime to prevent undefined errors.
• Workspace: Added a refresh button to the Rules Files section, and you can now view files from other Workspace sections, not just Project Files.
• Handoff: Errors during the handoff action are now handled gracefully with clear feedback to the user.
• Error Handling: Tasks are now automatically set to "Interrupted" state on agent errors.
• Models: Added a warning when a context window is not set for a model, and improved token threshold calculation to handle zero values.
• Bug Fixes:
  • Cancel action on loading messages now works properly.
  • Added missing working directory info to the bash tool message.

🧩 Extensions

• Destructive Command Guard: While the new GPT-5.6 models are SOTA, there have been reports of the model using destructive commands like rm -rf ~/. If you're not working inside a sandboxed environment, we highly recommend installing this extension. It intercepts bash commands and blocks dangerous operations - protecting against accidental data loss, recursive deletion, force pushes, and more.
• OpenAI Codex Auth: The previously separate OpenAI Codex Auth extension is now part of the official extensions list for easy discoverability. It provides an OpenAI Codex provider using ChatGPT Plus/Pro OAuth authentication - no API key required. If you used the previous standalone version, make sure to install this one instead, as it includes compatibility updates for the AI SDK v7 migration.

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

0.74.0

Highlights

🤖 Agent Profiles: System Prompts, Subagent Filters & Redesigned UI

Agent configuration has been significantly enhanced:

• Customizable System Prompts: You can now define a system prompt directly in the agent profile settings, overriding the default without needing to touch template files. For more advanced use cases, the existing Custom Prompts template system remains fully supported.
• Dynamic System Prompt Placeholders: System prompts can now include dynamic placeholders that are resolved at runtime, configurable directly through the editor.
• Subagent Filters: Agent profiles can now specify which subagents are allowed, giving you precise control over the delegation hierarchy.
• Redesigned Agent Settings UI: The agent profile settings have been refactored into a tabbed layout, making it much easier to navigate between configuration groups.
• Agent Selector Updates: The agent selector now includes enabled subagents management, so you can control subagent availability right from the selector.
• Agent Profile Persistence: The agent profile is now stored on the task at the first run, preserving the configuration even if the default profile changes later.

To help you create new agent profiles, the agent-creator skill has been added as a built-in skill. Just ask in Agent mode and your agent will help you set up a new profile.

🎙️ Voice Transcription with Latest Models

We've updated the voice transcription models to use the latest and greatest:

• OpenAI: Added gpt-realtime-whisper as a new voice transcription model and set it as the default for better accuracy and speed.
• Gemini: Updated the voice model to gemini-3.1-flash-live-preview and improved session handling for more stable voice interactions.

🏷️ /subagent Command

A new /subagent command lets you manually invoke a prompt in the selected subagent's context directly from the prompt field. The main agent only receives the final response, keeping your conversation clean while delegating focused tasks. Additionally, the subagents tool now supports a description parameter for dynamic status messages, so you can see what each subagent is working on at a glance.

🌐 New Provider: ClinePass

We've added direct support for the ClinePass provider, which offers a great selection of open-source models at very reasonable subscription price. Give it a try if you're looking for affordable alternatives for your agent workflows.

Other Changes

• Extension Reloads: Added diff-based extension reloads and optimized dependency installation for faster and more efficient extension development.
• Group Messages: Added support for removing group messages (subagent, aider) and new "Fork from here" / "Remove up to" actions. Group message blocks also now support compact view mode.
• Custom Commands: Using the task directory as the working directory for custom command shell execution.
• Telemetry: Added user ID metadata for PostHog when used as an OpenTelemetry tracer, and fixed OpenTelemetry initialization in local development mode.
• Bug Fixes:
  • Correctly preserving discovered model IDs and thinking max suffixes in the Cursor extension proxy.
  • Correctly applying timeout configuration to the proxy agent to prevent premature request timeouts.
  • Correctly compacting subagent messages for better context management.

🧩 Extensions

• My Minions: A showcase extension that demonstrates the new system prompt override capabilities. It adds an Orchestrator/Minion agent pair - the Orchestrator delegates ALL work to a focused execution subagent (the Minion), never doing the work itself. Both profiles are fully customizable in Settings.

🌍 Internationalization

• Added Korean translations (thanks @moduvoice).

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

0.73.0

Highlights

⚡ Virtualized Message Rendering as Default

We've made message virtualization the default rendering mode, bringing a significant performance improvement - especially for long-running tasks with hundreds of messages. This has been a long time coming, but previous solutions had too many quirks to become the default. The current implementation is a solid replacement that dramatically improves task loading times and scrolling performance. If you do experience any issues, you can still enable the legacy mode, but we're confident this will be a smooth upgrade. If you run into any problems, please report them on GitHub so we can address them.

🔌 Agent Client Protocol (ACP) Support

The aiderdesk CLI binary now supports the Agent Client Protocol (ACP) - an open standard for communication between AI coding agents and editor/IDE clients. What does this mean for you? You can now use AiderDesk as a backend agent directly within ACP-compatible editors like IntelliJ IDEA, Zed, and others, while keeping all of AiderDesk's powerful features: subagents, extensions, tool approvals, worktrees, task management, and more. No need to leave your IDE or switch contexts - your agent works right where you are, with full control over tool execution and approval flows. This brings the best of both worlds: the rich editing experience of your favorite IDE combined with AiderDesk's flexible agent capabilities.

Other Changes

• OpenAI-Compatible Providers: Added support for extra body parameters, giving you fine-grained control over provider-specific options that aren't part of the standard API.
• Anthropic-Compatible Providers: Ensured the correct /v1 suffix in the base URL to prevent configuration issues.
• Model Defaults: Updated default models across providers to use the latest versions.
• Bug Fixes:
  • Showing correct compact threshold values from agent profile overrides.
  • Fixed shell PATH extraction when login messages were present in output.
  • Excluded worktree symlink folders from git tracking and diffs during rebase.
  • Fixed usage report not being attached to tool-only responses, which previously prevented context compaction from triggering.

🧩 Extensions

• Cursor: Added health checks and automatic restart to the proxy for better reliability, plus a build phase during installation.
• Cursor SDK: Added fast and not-fast model variants, and removed the cursor-sdk/ prefix from model names for cleaner references.
• Auggie SDK: Removed a duplicated claude-opus-4.7 model entry.

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

0.72.0

Highlights

🧠 Smarter Context Compaction with Additional Levels

Smart Compaction continues to evolve with new progressive levels that now also trim verbose bash tool outputs and reasoning content. As conversations grow longer and the compaction threshold is reached, the system applies increasingly aggressive strategies—starting with truncating verbose results, then stripping reasoning traces, and summarizing older messages as a last resort. This keeps your context lean and focused without losing the essential information your agent needs.

🧩 Extension API: Worktree Functions

The Extension API now includes a suite of worktree management functions that give extensions full programmatic control over working modes:

• switchToWorktreeWorkingMode — Switch to worktree mode with options to carry over uncommitted changes.
• switchToLocalWorkingMode — Switch to local mode with options to merge before switching and switch all tasks in the worktree.
• getLocalUncommittedFiles — Retrieve unstaged, staged, and untracked files from the main repository.
• applyUncommittedChanges — Apply uncommitted changes from a worktree to a target branch.
• mergeWorktreeToWorktree — Merge commits and uncommitted changes from one worktree to another—useful for multi-model workflows.

🌐 Proxy Support for Dependency Downloads

Proxy support has been extended to cover downloading dependencies (like the probe binary and Python packages). If you're behind a proxy, all AiderDesk network operations now respect your proxy configuration, ensuring smooth setup and operation in restricted network environments.

🔍 New Languages in Semantic Search

The semantic search power tool now supports Solidity, Crystal, and Haskell. This means the probe binary can index and search these languages, giving your agent accurate code search results in more project types.

Other Changes

• Task Memory: Improved task memory management in the renderer process for better performance with many tasks.
• Worktree Reliability: Correctly restoring stashed changes on failures when switching to worktree mode.
• Touch Devices: The compaction slider now works properly on touch devices.
• Model Loading: Added a timeout to provider model loading to prevent indefinite hangs.
• Extension Management:
  • Added a specific error message when trying to install an extension that requires npm while it's not available.
  • Correctly showing "Not supported on OS" for available extensions in the extensions gallery.
• Aider Integration: Clarified the missing API key warning when using local inference via Aider (thanks @zerone0x).
• UI Polish: Optimized all files loading on the renderer to avoid duplicated entries.
• Bug Fixes: Correctly selecting input history items and fixed task archiving from the sidebar menu.

🧩 Extensions

• Cursor: A new, recommended Cursor extension that integrates Cursor models via a local OpenAI-compatible proxy, giving AiderDesk full control over tool execution and approval. Unlike the existing cursor-sdk extension (which runs Cursor agents as a subprocess), this one runs Cursor's models through AiderDesk's native agent loop—with full approval support, MCP integration, and three tool dispatch modes (Reject, Redirect, Native). Note: it communicates with Cursor's internal API, which may change and break compatibility.
• GitHub Copilot: Now supports system-installed Copilot CLI and correctly resolves the CLI binary path (thanks to @batongo for his help).
• Multi-Model Run: Improved usability for worktree switching when running prompts across multiple models.
• Protected Paths: Improved functionality.
• Redact Secrets: Improved functionality.

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

0.71.0

Highlights

💻 New CLI run Command: Automate Prompts from Your Terminal

We've added a powerful new aiderdesk run CLI command that lets you execute prompts on a running AiderDesk server -non-interactively and with a live streaming view. This opens the door to integrating AiderDesk into your CI/CD pipelines, scripts, and automations. Submit a prompt, watch the streamed response in real-time, and programmatically receive the result. Check out the CLI Run documentation to learn more.

🧩 Extension API Updates

The Extension API continues to evolve with several new additions:

• skillsToActivate on the onAgentStarted event: Extensions can now programmatically activate skills when an agent starts, enabling more sophisticated automated workflows.
• images support on the onAgentStarted event: The event now includes any images sent with the prompt, so extensions can react to visual context.
• getElectronApp function: Retrieve additional app information from within your extension.
• Granular Floating Placements: The single floating placement has been replaced with three distinct levels: app-floating, project-floating, and task-floating. This gives you precise control over the lifecycle and scope of your floating panels - whether they should persist at the app level, be available across all tasks in a project, or be tied to a specific task.
• Colons in Command Names: Extension commands can now include colons, enabling namespace-style naming like myext:action.
• Added missing reasoning property to the Response message type.

🌐 NoProxy Support in Network Settings

For environments where certain hosts should bypass the proxy, we've added noProxy support to the network settings. This is essential for setups where internal services or local development servers need direct access while everything else goes through the proxy.

🧠 Reasoning Effort Settings for Z.ai & NeuralWatt

We've added reasoning effort settings to both the Z.ai Coding Plan and NeuralWatt providers. This lets you control how much "thinking" the models do before responding—choose higher effort for complex reasoning tasks or lower effort for faster, more economical responses. Note that Z.ai Coding Plan supports only High and Max effort levels.

📋 Move to Top in Task Sidebar

A handy new "Move to top" action is now available in the task sidebar context menu. Quickly bring an older task back to the top of your list without recreating it—perfect for resuming work on something that got buried.

🔍 Streaming Grep Output

The power grep tool now streams its output instead of waiting to collect all results before displaying anything. For large codebase searches, you'll start seeing matches immediately as they're found, making the tool feel much more responsive.

Other Changes

• Smart Compaction UI: Added special UI elements for edited files in the assistant message after smart compaction, so you can still see what was changed even in compacted view.
• MCP Server: Extracted the integrated MCP server into a separate @aiderdesk/mcp-server package for easier maintenance and external integration. If you were using the MCP server, install the standalone package going forward.
• Probe Binary: Added verification and automatic download of the probe binary on runner startup.
• UI Polish: Truncated long model IDs in the Model Library table, styling improvements for compaction snippet blocks.
• Bug Fixes:
  • Using shell option when installing extension dependencies to fix issues on Windows.
  • Correctly placing skill messages after the user prompt when used from custom commands.
  • Base commit validation and state updates on rebase continue.
  • Correct prompt field history menu items order.
  • Removed unnecessary .aider-desk from ripgrep mock in unit tests.
  • Cleaned up React hooks dependencies and removed eslint-disable overrides.

⚠️ Gemini CLI Provider Removed: Due to Google ending subscription support for the Gemini CLI, the internal Gemini CLI provider has been removed. Please use the standard Google Gemini provider with API keys instead.

🧩 Extensions

• GitHub Copilot: Use your GitHub Copilot subscription (Individual, Business, or Enterprise) in Agent mode. The extension bundles the official Copilot SDK and gives access to GPT-5+ and Claude models at no extra cost.
• DOOM: Yes, really. A playable version of DOOM that runs directly in AiderDesk using a floating UI component. Because why not? 🎮

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

0.70.0

Highlights

⚡ Rendering Performance Overhaul

The recent additions of the Extension UI component system introduced some rendering performance issues, and this release tackles them head-on. We've made several rendering optimizations to ensure the app rerenders properly during streaming messages, implemented lazy loading for heavy dependencies so they're only loaded when needed, and optimized Zustand store subscriptions using module-level actions to eliminate unnecessary re-renders. The result is a snappier experience, especially during long agent runs with many tool calls and UI extensions active.

🧩 Extension System: Custom Messages & New Capabilities

The Extension system continues to unlock new customization possibilities:

• task-message UI Placement: A powerful new placement that lets extensions completely customize how messages are rendered in the conversation. This opens the door to tailoring what's visible in your message list—hide verbose tool outputs, highlight important results, or create entirely custom layouts for specific message types.
• generateObject Function: A new addition to the extension task context that allows extensions to generate structured objects from LLM responses—perfect for extensions that need to extract structured data or make AI-powered decisions programmatically.

💬 Custom Commands with Quoted Arguments

Custom commands now support quoted arguments, enabling more complex command definitions where arguments may contain spaces. For example: /generate-tests "src/utils/helper.ts" vitest — the quoted path is correctly parsed as a single argument even with spaces, making custom commands more flexible and reliable.

🔄 Enhanced Mode Switch Dialogs

Switching between Worktree and Local modes is now more flexible and clear:

• When switching from Worktree to Local, you can now choose to switch all tasks in the worktree at once—no more switching them one by one.
• The mode switch dialogs have been enhanced with clearer options, making it easy to understand exactly what will happen with your changes, branches, and uncommitted work before you commit.

📝 Edit Any User Message

You can now edit any user message in your conversation history—not just the last one. Click edit on an earlier message, make your changes, and the conversation replays from that point. This makes it easy to refine your intent mid-conversation without starting over.

Other Changes

• Auto-Scrolling: Improved scrolling behavior while messages are being processed for a smoother experience.
• Crash Recovery: Added an error boundary and crash fallback screen, so if something goes wrong you'll see a helpful message instead of a blank window.
• Bash Tool: Long bash commands now break across lines to avoid horizontal scrolling.
• Mistral: Deduplicated loaded Mistral models for a cleaner model list.
• Configuration: Added support for AIDER_DESK_DIR and AIDER_DESK_HOME_DIR environment variables for custom data directories.
• Internal: Added restart action for the Aider Connector, Playwright testing framework support with initial onboarding tests, and Zustand store migration for settings data.
• Bug Fixes:
  • Added missing /settings/versions endpoint.
  • Updated usage report logic to avoid duplicate attachments.
  • Hide todo window when there are no todo items.

🧩 Extensions

• Providers Quota: Added support for showing Deepseek balance information alongside existing Synthetic, Z.AI, and Neuralwatt providers.
• UI Placement Demo: Updated to showcase all available placements including the new task-message placement, with improved organization and message filter support.

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

0.69.0

Highlights

🔌 Claude Agent SDK & Auggie SDK Move to Extensions

Two providers have been moved from AiderDesk's core to standalone extensions:

Claude Agent SDK is now a community extension due to Anthropic's policy change—starting June 15, 2026, Claude subscription plans no longer include Agent SDK usage as part of the regular subscription. Instead, plans receive a separate monthly Agent SDK credit (e.g., $20 for Pro, $100-$200 for Max). Once exhausted, you're charged at standard API rates. For this reason, we strongly recommend using the standard Anthropic provider with API keys for production—it offers direct API access, better cost control, lower latency, and full mode support. The extension remains available for educational and experimental purposes, but it only works in Agent mode and has some limitations around message editing and token tracking.

Auggie SDK has been moved to an extension because its underlying library uses an older version of the AI SDK, which was blocking our migration to v6. Both extensions are expected to work as before, but if you encounter any issues, please file a report on our GitHub.

If you were previously using either provider, simply install the corresponding extension from Settings → Extensions → Available to continue where you left off.

🔍 New Extensions: CodeGraph & SearXNG

The extension ecosystem continues to grow with two powerful new tools:

• CodeGraph: A local-first knowledge graph that builds a semantic index of your codebase. It provides your agent with powerful tools for codebase exploration—understanding call paths, finding callers and callees, tracing symbol dependencies, and more. It's like giving your agent a mental map of your entire project.
• SearXNG: A web search tool powered by SearXNG, giving your agent the ability to search the web for documentation, solutions, and up-to-date information.

🛠️ Extension Tool Management via UI

With the growing number of extension-provided tools, we've added a better way to manage them. Head to Settings → Agents → Tools to easily enable, disable, and configure approval requirements for extension tools on a per-agent basis. This gives you fine-grained control over exactly which tools each agent profile has access to.

🪟 Floating UI Placement for Extensions

Extensions can now use a new floating placement that renders UI components in a moveable and resizable dialog. This is perfect for extensions that need more screen real estate or want to persist alongside the main conversation—think dashboards, monitoring panels, or interactive tools.

📦 Carry Over Uncommitted Changes When Switching Modes

Switching from Local mode to Worktree mode? You can now carry over your uncommitted changes in the same action—just like the existing merge option when switching from Worktree to Local. This is handy when you want to postpone some WIP changes to a worktree for later, or simply reorganize your in-progress work across modes.

⚡ Improved Extension UI Rendering Performance

We've optimized how extension UI components are rendered, which should be noticeably faster when loading UI extensions or when using message placement extensions that appear many times throughout a conversation.

Other Changes

• Smart Compaction: Improved compaction for bash tool results and fixed progressive level increases.
• Agent Tools: Added coercion for boolean and number inputs, making tool calls more forgiving of parameter types.
• Extension API: Added additional parameters to createLlm for better custom LLM provider support.
• Read-Only Files: Added indicators and instructions for read-only files, so the agent knows when it can't modify a file.
• Model Updates: Added MiniMax-M3 to the MiniMax provider.
• Bug Fixes:
  • Correctly handling compaction when the first message is a skill activation message.

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

0.68.0

Highlights

🧠 Progressive Smart Compaction

Smart Compaction now works in progressive levels for more nuanced context management. Instead of a single compaction pass, the system applies increasingly aggressive compaction strategies—starting with truncating verbose tool results, then summarizing older messages, and only condensing recent context as a last resort. This preserves the most relevant information at each level while keeping token usage under control, giving you better results in long-running conversations.

🏷️ Checkpoint Extension: Instant Code Rewind

The new Checkpoints extension automatically creates Git snapshots of your working tree before every file edit or write operation. A "Reset to checkpoint" button appears above each edit in your conversation—click it to instantly revert your entire codebase to the state before that specific edit. It works in any Git repository with zero configuration, survives restarts, and fully supports Worktree mode. No more manually undoing a chain of changes you're not happy with—just rewind to any point.

📋 Kanban Board Extension — Powered by External UI Libraries

The new Kanban extension adds a visual task board to AiderDesk, displaying all your project tasks organized by state across 8 columns. Click any task card to jump to it in the sidebar, and the board auto-updates as tasks are created and completed. This is a foundational implementation—PRs and suggestions for additional features like drag-and-drop, filtering, or custom columns are very welcome!

Behind the scenes, Kanban also showcases a powerful new capability: extension UI components can now use external libraries loaded via CDN (like esm.sh). Kanban uses react-kanban-kit loaded externally, and you can do the same with any React component library—charts, rich text editors, or whatever your extension needs.

Other Changes

• Extension API: Added toolCallId to onToolCalled and onToolFinished hook events for precise tracking of tool executions.
• New UI Placements: Added tasks-sidebar-actions-left and tasks-sidebar-actions-right placements for adding buttons to the Tasks sidebar header.
• Extension UI: Added activateTask action to UI component props, and UI extension placements now render correctly in compact display mode.
• OS-Aware Extensions: Extensions can now declare which operating systems they support, showing a clear info message on incompatible platforms.
• UI Polish: Added icons to archive/unarchive and delete buttons in the task actions panel.
• Bug Fixes:
  • Correctly placing the unmaximized app window on startup.

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

0.67.0

Highlights

🎛️ New Autonomy Modes (Breaking Change)

We've replaced the simple Auto Approve toggle with three distinct Autonomy modes that better reflect how your agent interacts with you:

Mode	Behavior
Guided (default)	Tools run without approval, but the agent asks for plan approval before executing changes
Manual	Every tool call requires your approval
Autonomous	No confirmations or approvals - the agent runs without interruption

⚠️ Breaking Change for Extensions & Automations: If you use autoApprove: true in your extensions, automations, or REST API calls, replace it with autonomyMode: 'autonomous'. The old autoApprove property is no longer supported.

🧩 New Extension Context Functions

Two new functions are now available in the Extension context:

• getProviders: Fetch all configured providers from within your extension, enabling extensions that adapt their behavior based on which providers and models are available.
• truncateToolResult: Use AiderDesk's built-in truncation logic to handle large tool results in your own extensions, keeping context usage under control.

🧠 Smart Compaction Improvements

Smart Compaction continues to get smarter with additional truncation of tool results at the compaction point. Non-power tool results (including MCP tools) and read file tool results are now truncated during compaction, preserving the most relevant context while significantly reducing token usage in long-running conversations.

🌐 New Provider: Neuralwatt

We've added direct support for Neuralwatt, a great and affordable provider for open-source weight models including GLM-5.1, Kimi K2.6, Qwen3.6, MiniMax M2.5, and more. The providers-quota extension has also been updated to display Neuralwatt quota information.

⚡ Reduced Idle CPU Usage with File Watch Mode

We've changed how AiderDesk detects file changes, significantly reducing idle CPU usage. The app now automatically determines the best detection method, but in some scenarios (network disks, Docker environments), you may need to switch back to the polling mechanism. You can configure this in Settings → General → Notification & Files → File Watch Mode.

📎 New /subtask Command

A new /subtask command lets you quickly create and switch to a subtask from within your conversation. It's a fast way to delegate work to a subtask without leaving the chat or navigating the sidebar.

Other Changes

• New Task Tool: Added run_prompt to the task tools, allowing agents to run a prompt programmatically.
• Context Messages: Added timestamp property to context messages for better temporal tracking.
• Skills Deduplication: Skills are now deduplicated with clear priority: extension skills → project skills → global skills.
• Grep Tool: Fixed context rendering around match lines for accurate and readable output.
• Cursor SDK Extension: Added support for MCP servers and automatic Git staging of files created by Cursor tools.
• Stability: Replaced the sharp native dependency with a stub to prevent SIGILL crashes on older CPUs.
• Bug Fixes:
  • Properly handling "Always" approval with tool calls.
  • Properly focusing search input in the task sidebar.
  • Hide redundant task directory path in semantic search tool messages.

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

0.66.0

Highlights

🔄 Undo for Clear Messages & Compaction

Ever accidentally cleared your conversation or triggered compaction and immediately regretted it? Now you can undo both actions. Whether you cleared messages or the context was compacted, a simple undo restores your previous conversation state - no more lost context from misclicks or premature compaction.

🖼️ Images in Agent Mode Conversations

You can now send images directly within your user messages in Agent mode. Paste a screenshot of a UI bug, share a design mockup, or show an error dialog - your agent will be able to see and reason about the visual content alongside your text prompt. This brings visual context to your AI conversations, making it much easier to communicate UI-related issues and design feedback.

🔍 Integrated Ripgrep for Lightning-Fast Searches

We've integrated ripgrep as the engine behind the grep power tool, delivering dramatically faster and more accurate code searches. Ripgrep is purpose-built for searching code repositories and respects .gitignore rules out of the box, so your searches are both speedy and relevant.

♾️ Infinite Agent Iterations

As models get better at agentic coding and can now run for much longer, the current max iteration limit was not enough in some cases. You can now configure infinite agent iterations, removing the cap entirely and letting your agent run as long as it needs to complete the job - perfect for large-scale refactors, multi-file implementations, or any task where you want the agent to work without artificial limits.

🛡️ Enhanced Extension Blocking & Context

The extension event system has been updated with richer context: events now include the agent profile information and support blocking reasons. This means extensions that block operations can communicate why they're blocking - giving users clear feedback and enabling more sophisticated extension workflows.

Other Changes

• Updated Files: Added a sticky file path header that stays visible as you scroll through large diffs, so you always know which file you're reviewing.
• Process Management: Improved process termination on POSIX systems by using process groups, preventing orphaned processes.
• Requesty Provider: Updated to use costs directly from the API response for more accurate usage tracking.
• OpenAI-Compatible Providers: Added an option to manage token usage tracking for providers that support it.
• Task Tools: list_tasks results are now sorted by descending updated date, and subtask agents correctly use the parent task when creating nested subtasks.
• Git Handling: Improved handling of missing git configuration and empty repositories for a smoother first-run experience.
• Localization: Added Russian (ru) locale support (thanks @AlexZander85)
• Bug Fixes:
  • Fixed missing gap between messages in virtualized message rendering.
  • Properly handling existing files when adding context files.
  • Correctly installing repo-root extensions.

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