Roo Code 3.26 Release Notes
This document combines all releases in the v3.26 series.
Version 3.26.0 (2025-08-26)
This release introduces Grok Code Fast (formerly Sonic), a built-in /init command, and Qwen Code CLI API support.
Grok Code Fast
As you may have already figured out, our stealth model Sonic has officially been uncloaked! (#7426)
From xAI, this model is optimized for coding tasks and already beloved by the community in Code Mode for its:
- Sharp reasoning capabilities
- Plan execution at scale
- Code suggestions with UI taste and intuition
If you've already been enjoying Sonic in Roo Code Cloud, you'll be transitioned to Grok Code Fast. The model xai/grok-code-fast-1 is also available under the xAI Provider and is not free (for when the free period ends on Aug 28th, 2025).
A massive thank-you to our partners at xAI and to all of you — over 100B tokens (and counting!) ran through Sonic during stealth! Your incredible adoption and helpful feedback shaped Grok Code Fast into the powerful model it is today.
Important: Grok Code Fast remains FREE when accessed through the Roo Code Cloud provider during the promotional period. Using it directly through the xAI provider will incur standard charges once pricing is established.
📚 Documentation: See Roo Code Cloud Provider for free access or xAI Provider for direct configuration.
Built-in /init Command
We've added a new /init slash command for project onboarding (#7381, #7400):
- Automatic Project Analysis: Analyzes your entire codebase and creates comprehensive AGENTS.md files
- AI Assistant Optimization: Generates documentation that enables AI assistants to be immediately productive in your codebase
- Mode-Specific Guidance: Creates tailored documentation for different Roo Code modes (code, debug, architect, etc.)
The /init command helps LLMs understand your project's unique patterns and conventions by documenting project-specific information that isn't obvious from the code structure alone.
Qwen Code CLI API Support
We've integrated with the Qwen Code CLI tool, allowing Roo Code to leverage its free access tier for Alibaba's Qwen3 Coder models (#7380):
- Free Inference: Piggybacks off the Qwen Code CLI's generous free tier (2,000 requests/day and 60 requests/minute with no token limits) via OAuth, available during a promotional period.
- 1M Context Windows: Handle entire codebases in a single conversation.
- Seamless Setup: Works automatically if you've already authenticated the Qwen Code CLI tool.
This integration provides free access to the Qwen3 Coder models by using the local authentication from the Qwen Code CLI.
📚 Documentation: See Qwen Code CLI Provider for setup and configuration.
Provider Updates
- DeepSeek V3.1 on Fireworks: Added support for DeepSeek V3.1 model in the Fireworks AI provider (thanks dmarkey!) (#7375)
- Provider Visibility: Static providers with no models are now hidden from the provider list for a cleaner interface (#7392)
QOL Improvements
- Auto-Approve Toggle UI: The auto-approve toggle now stays at the bottom when expanded, reducing mouse movements (thanks elianiva, kyle-apex!) (#7318)
- OpenRouter Cache Pricing: Cache read and write prices are now displayed for OpenRouter models (thanks chrarnoldus!) (#7176)
- Protected Workspace Files: VS Code workspace configuration files (*.code-workspace) are now protected from accidental modification (thanks thelicato!) (#7403)
Bug Fixes
- Security - Symlink Handling: Fixed security vulnerability where symlinks could bypass rooignore patterns (#7405)
- Security - Default Commands: Removed potentially unsafe commands (
npm test,npm install,tsc) from default allowed list (thanks thelicato, SGudbrandsson!) (#7404) - Command Validation: Fixed handling of substitution patterns in command validation (#7390)
- Follow-up Input Preservation: Fixed issue where user input wasn't preserved when selecting follow-up choices (#7394)
- Mistral Thinking Content: Fixed validation errors when using Mistral models that send thinking content (thanks Biotrioo!) (#7106)
- Requesty Model Listing: Fixed model listing for Requesty provider when using custom base URLs (thanks dtrugman!) (#7378)
- Todo List Setting: Fixed newTaskRequireTodos setting to properly enforce todo list requirements (#7363)