Skip to main content

Roo Code 3.26.7 Release Notes (2025-09-05)

This release brings enhanced Kimi K2 models with massive context windows, OpenAI service tier selection, and DeepInfra as a new provider offering 100+ models.

Kimi K2-0905: Moonshot's Latest Open Source Model is Live in Roo Code​

We've upgraded to the latest Kimi K2-0905 models across multiple providers (thanks CellenLee!) (#7663, #7693):

K2-0905 comes with three major upgrades:

  • 256K Context Window: Massive context supporting up to 256K-262K tokens, doubling the previous limit for processing much larger documents and conversations
  • Improved Tool Calling: Enhanced function calling and tool use capabilities for better agentic workflows
  • Enhanced Front-end Development: Superior HTML, CSS, and JavaScript generation with modern framework support

Available through Groq, Moonshot, and Fireworks providers. These models excel at handling large codebases, long conversations, and complex multi-file operations.

OpenAI Service Tiers​

We've added support for OpenAI's new Responses API service tiers (#7646):

  • Standard Tier: Default tier with regular pricing
  • Flex Tier: 50% discount with slightly longer response times for non-urgent tasks
  • Priority Tier: Faster response times for time-critical operations

Select your preferred tier directly in the UI based on your needs and budget. This gives you more control over costs while maintaining access to OpenAI's powerful models.

📚 Documentation: See OpenAI Provider Guide for detailed tier comparison and pricing.

DeepInfra Provider​

DeepInfra is now available as a model provider (thanks Thachnh!) (#7677):

  • 100+ Models: Access to a vast selection of open-source and frontier models
  • Competitive Pricing: Very cost-effective rates compared to other providers
  • Automatic Prompt Caching: Built-in prompt caching for supported models like Qwen3 Coder
  • Fast Inference: Optimized infrastructure for quick response times

DeepInfra is an excellent choice for developers looking for variety and value in their AI model selection.

📚 Documentation: See DeepInfra Provider Setup to get started.

QOL Improvements​

  • Shell Security: Added shell executable allowlist validation with platform-specific fallbacks for improved command execution safety (#7681)

Bug Fixes​

  • MCP Tool Validation: Roo now validates MCP tool existence before execution and shows helpful error messages with available tools (thanks R-omk!) (#7632)
  • OpenAI API Key Errors: Clear error messages now display when API keys contain invalid characters instead of cryptic ByteString errors (thanks A0nameless0man!) (#7586)
  • Follow-up Questions: Fixed countdown timer incorrectly reappearing in task history for already answered follow-up questions (thanks XuyiK!) (#7686)
  • Moonshot Token Limit: Resolved issue where Moonshot models were incorrectly limited to 1024 tokens, now properly respects configured limits (thanks wangxiaolong100, greyishsong!) (#7673)
  • Zsh Command Safety: Improved handling of zsh process substitution and glob qualifiers to prevent auto-execution of potentially dangerous commands (#7658, #7667)
  • Traditional Chinese Localization: Fixed typo in zh-TW locale text (thanks PeterDaveHello!) (#7672)