Roo Code 3.39.0 Release Notes (2026-01-08)

This release adds image file @mentions, makes provider profiles stick to tasks, and updates key UI labels.

Image file @mentions

You can now @mention image files to include them as inline images in your message, making it easier to share screenshots and UI mockups without manually attaching files (thanks hannesrudolph!) (#10189).

📚 Documentation: See Context Mentions for usage details.

Sticky provider profiles

Tasks now remember the provider profile (API configuration) they started with, so switching profiles elsewhere doesn’t affect running tasks or resumed tasks (thanks hannesrudolph!) (#10018).

📚 Documentation: See API Configuration Profiles for details.

YOLO → BRRRRRRRRRR

The auto-approve mode label has been renamed from “YOLO” to “BRRRRRRRRRR” across the UI (thanks app/roomote!) (#10507).

📚 Documentation: See Auto-Approving Actions for details.

QOL Improvements

• The @ file picker now respects .rooignore, reducing noise in large workspaces and helping you avoid accidentally attaching ignored/generated files (thanks app/roomote, jerrill-johnson-bitwerx!) (#10174)
• Adds debug-only proxy routing settings so you can inspect extension network traffic while running under the VS Code debugger (F5) (thanks hannesrudolph, SleeperSmith!) (#10467)
• Improves the follow-up suggestion mode badge styling for better readability (thanks mrubens!) (#9260)
• Clarifies in the native read_file tool description that image formats are supported when the model supports vision (thanks app/roomote, nabilfreeman!) (#10442)

Bug Fixes

• Fixes an issue where conversations could fail after condensation due to missing/mismatched tool call IDs, improving reliability in longer chats (thanks daniel-lxs!) (#10471)
• Fixes an issue where duplicate tool_result blocks could cause provider API errors (including Anthropic “duplicate toolResult” failures), improving reliability in tool-heavy workflows (thanks daniel-lxs!) (#10497)
• Fixes an edge case where switching terminals mid-run could produce duplicate tool results and trigger protocol errors, reducing unattended-mode soft-locks (thanks app/roomote, nabilfreeman!) (#10466)
• Fixes an issue where Roo could generate the wrong command chaining syntax on Windows, making suggested terminal commands more likely to work without edits (thanks app/roomote, AlexNek!) (#10434)
• Fixes an issue where chat requests could fail on Windows systems without PowerShell in PATH (“spawnSync powershell ENOENT”) (thanks app/roomote, Yang-strive!) (#9897)
• Fixes a rare edge case where an API rate limit setting could be ignored when provider state is temporarily unavailable (thanks app/roomote!) (#10266)
• Fixes validation failures in nightly builds by adding missing setting descriptions for debug proxy configuration (thanks app/roomote!) (#10505)

Provider Updates

• Adds the kimi-k2-thinking model to the Fireworks AI provider (thanks app/roomote, kavehsfv!) (#9202)
• Adds an “Extra High (xhigh)” reasoning effort option for OpenAI-compatible providers when supported by the endpoint/model (thanks app/roomote, Soorma718!) (#10061)
• Adds zai-glm-4.7 to the Cerebras model list (thanks sebastiand-cerebras!) (#10500)
• Removes legacy Claude 2 models from AWS Bedrock to prevent selecting unsupported options (thanks app/roomote, KevinZhao!) (#10501)
• Improves compatibility with OpenAI’s Responses API for MCP tool schemas, reducing tool-calling failures when schemas omit additionalProperties: false (thanks daniel-lxs!) (#10472)

Misc Improvements

• Adds a one-line CLI installer workflow to simplify installing/upgrading the Roo Code CLI (thanks cte!) (#10474)
• Adds early-stage CLI support used by eval tooling, including a new “CLI” execution option for eval runs (thanks cte!) (#10452, #10456)