Skip to main content

Roo Code 3.33.0 Release Notes (2025-11-18)

Gemini 3 Pro Preview is now available in Roo Code and already performing extremely well in real-world coding tasks, alongside 16 tweaks and fixes—thanks to everyone in the Roo Code community who helped shape this release.

Roo Code v3.33.0 Release

Gemini 3 Pro Preview​

Roo Code now supports Google’s Gemini 3 Pro Preview model through direct Gemini, Vertex AI, and aggregator providers like OpenRouter and Requesty (#9357, #9327):

  • 1M-token, reasoning-capable model: Handles very large conversations while providing higher-quality multi-step reasoning on complex coding and refactoring tasks.
  • Strong eval performance: Achieves a 100% score on internal Roo Code evals and 76.2% on SWE-bench Verified, giving more consistent solutions on real-world coding tasks.
  • Reliable tool usage: Executes complex multi-step tool workflows without getting stuck or losing track, especially in long, tool-heavy tasks.
  • Better out-of-the-box defaults: Uses gemini-2.5-pro by default where supported, sets a more natural temperature of 1, cleans up the Gemini model list, and includes reasoning / “thought” tokens in cost reporting so usage numbers better match provider billing.

QOL Improvements​

  • Git status in environment details (#9310): Shows git status information in environment details so agents have more context about untracked, modified, and staged files when reasoning about your workspace.
  • Tool protocol selector in advanced settings (#9324): Lets you choose which tool protocol to use (such as XML vs native) without editing config files, making it easier to experiment with different tool behaviors.
  • Dynamic tool protocol resolution (#9286): Resolves the active tool protocol using a clear precedence hierarchy, so provider defaults, mode settings, and user overrides interact in a predictable way.
  • Improved Modes view toolbar (#9077): Moves Import/Export into the Modes view toolbar and cleans up the Mode edit view, making it easier to manage and share modes from a single place.
  • Cloud agent CTA points to setup page (#9338): Updates the cloud agent call-to-action to link directly to the setup page so new users can get started faster.
  • Roo Code Cloud provider pricing page (#9195): Adds a pricing page and related Cloud provider tweaks so pricing is easier to understand before you enable Roo Code Cloud.

Bug Fixes​

  • Prevent duplicate tool_result blocks in native protocol (#9248): Ensures each native tool call emits a single tool_result block, avoiding 400 errors and duplicated tool executions.
  • Format tool responses for native protocol (#9270): Normalizes the structure of tool responses so native protocol runs are easier for models to follow and less likely to error.
  • Centralize toolProtocol configuration checks (#9279): Uses a single source of truth for toolProtocol configuration, reducing configuration drift and subtle behavior differences.
  • Preserve tool blocks in conversation history (#9319): Keeps native protocol tool blocks intact in history so follow-up turns can reason correctly about prior tool calls.
  • Prevent infinite loops after successful finalization (#9325): Fixes a regression where certain native tool flows could loop after successful completion instead of stopping cleanly.
  • Sync parser state with profile and model changes (#9355): Keeps the conversation parser aligned with the active profile and model so switching models or profiles does not leave the parser in an inconsistent state.
  • Pass tool protocol to truncation errors (#9358): Ensures truncation errors know which tool protocol is active so error handling and messaging stay accurate.
  • VS Code theme-colored outline button borders (#9336): Aligns outline button borders with the current VS Code theme for a more consistent UI.
  • Use shields.io badges instead of badgen.net (#9318): Replaces broken badge URLs with shields.io so badges render reliably again.
  • Cap git status file sampling in evals (#9322): Adds a maximum for git status files in eval settings so evaluations don’t pull excessively large environment details.