Release v0.28.0

[wladimiiir]

Highlights

✨ Introducing Subagents: Specialized AI Assistants for Enhanced Productivity

Unleash a new level of efficiency with Subagents! You can now create specialized AI assistants from any agent profile, each with its own customized system prompt, tool access, and settings. Subagents are designed to handle specific types of tasks, whether invoked automatically by the main agent or on-demand by you.

Key Benefits:

• Context Preservation: Subagents operate with a focused context, preventing pollution of your main conversation.
• Specialized Expertise: Configure subagents with detailed prompts for specific domains, leading to higher success rates.
• Cost Optimization: Strategically select cost-effective models (e.g., Gemini Flash, Claude Haiku) for routine subagent tasks like code reviews, and premium models for complex work.
• Flexible Permissions: Control tool access and approval settings for each subagent.

By offloading specific tasks, subagents help optimize your main context window and make your AI interactions significantly more cost-efficient. We believe this is a game-changing feature and are eager to see what specialized subagents you create! Expect continuous improvements to this functionality in future releases.

Detailed information with some examples can be found in our documentation.

✅ Confirm Before Edit for Aider Coder

Gain more control over AI-generated code changes! When using Aider in Code mode, you now have the option to confirm before edit. This allows you to review the modifications Aider proposes with a diff viewer before accepting them or providing feedback for further improvements to the model.

🌐 Built-in Web Scraping with fetch Power Tool

Your main agent now has the powerful capability to scrape websites it encounters during a conversation, thanks to the new fetch power tool. We've also removed the external Playwright dependency for web scraping. This means the /web command now performs its scraping internally with the same high quality, simplifying your setup and enhancing performance.

📂 Seamless Drag & Drop for Context Files

Managing your context files just got easier and more intuitive! You can now drag and drop files directly into the Context Files panel from external applications or within your project. Files dropped from outside the current project are added as read-only, while those from within the project are fully editable.

🧠 'Minimal' Reasoning Effort for Next-Gen Models

We've added a 'minimal' reasoning effort option for the OpenAI provider and the /reasoning-effort command when using Aider. This is the preferred setting when working with highly capable models like GPT-5, allowing them to operate more efficiently while maintaining high-quality outputs for coding tasks.

Other Changes

• Added a model selector to the Agent profile settings for more granular control.
• The selected model is now correctly used when compacting the conversation history.
• Improved the scrollbar style in the agent selector for better usability with many profiles.
• Fixed an issue where the application could get stuck when dropping files.
• Updated the temperature range for models to 0-2 with a 0.5 step for finer control.
• The file edit tool now gracefully handles scenarios with no changes, preventing unnecessary errors.
• Added a "scroll to bottom" button to the message area for improved navigation in long conversations.
• The AGENTS.md file is now included in the system prompt by default when using Agent mode, providing better context to the AI.
• Fixed an issue causing duplication in the thinking section of messages with some OpenRouter models.
• Context messages and files are now properly used with context-aware commands (e.g., /tokens).
• Relevant model selectors are now dynamically shown in the Project bar based on the current mode and available tools.

---

This discussion was created from the release 0.28.0.

[michabbb]

Built-in Web Scraping with fetch Power Tool

is that a html only fetcher or does it support js ?

Subagents

it would be amazing if a sub-agent had its own MCP config
today, every sub-agent concept is built in a way that the sub-agent gets all MCPs from its parent, which is a terrible way to fill up the expansive context window

[wladimiiir]

is that a html only fetcher or does it support js ?

It is loading the page inside the Electron's browser, so it's also loading JS scripts there.

it would be amazing if a sub-agent had its own MCP config

Every subagent is agent profile with its own set of tools and MCPs enabled - it does not inherit them from the main agent.