Skip to main content

Roo Code 3.37 Release Notes (2025-12-22)

Roo Code 3.37 introduces experimental custom tools, adds new provider capabilities, and improves tool-call reliability.

Roo Code v3.37.0 Release

New models​

Z.ai GLM-4.7 (thinking mode)​

GLM-4.7 is now available directly through the Z.ai provider in Roo Code, as well as via the Roo Code Cloud provider (and other provider routes that surface Z.ai). It’s a strong coding model for agentic workflows, with improved multilingual coding, terminal tasks, tool use, and complex reasoning compared to GLM-4.6 (#10282).

GLM-4.7 benchmark performance

MiniMax M2.1 improvements​

MiniMax M2.1 is now available directly through the MiniMax provider in Roo Code, as well as via the Roo Code Cloud provider (and other provider routes that surface MiniMax). It’s a strong pick for agentic coding workflows, with better tool use, instruction following, and long-horizon planning for multi-step tasks—and it’s fast (#10284).

MiniMax M2.1 Benchmark Performance

Experimental custom tools​

You can now define and use custom tools so Roo can call your project- or team-specific actions like built-in tools. This makes it easier to standardize workflows across a team by shipping tool schemas alongside your project, instead of repeatedly re-prompting the same steps (#10083).

Bug Fixes​

  • Fixes an issue where Roo could appear stuck after a tool call with some OpenAI-compatible providers when streaming ended at the tool-calls boundary (thanks torxeon!) (#10280)
  • Fixes an issue where Roo could appear stuck after a tool call with some OpenAI-compatible providers by ensuring final tool-call completion events are emitted (#10293)
  • Fixes an issue where MCP tools could break under strict schema mode when optional parameters were treated as required (#10220)
  • Fixes an issue where the built-in read_file tool could fail on some models due to invalid schema normalization for optional array parameters (#10276)
  • Fixes an issue where search_replace / search_and_replace could miss matches on CRLF files, improving cross-platform search-and-replace reliability (#10288)
  • Fixes an issue where Requesty’s Refresh Models could leave the model list stale by not including credentials in the refresh flow (thanks requesty-JohnCosta27!) (#10273)
  • Fixes an issue where Chutes model loading could fail if the provider returned malformed model entries (#10279)
  • Fixes an issue where reasoning_details could be merged/ordered incorrectly during streaming, improving reliability for providers that depend on strict reasoning serialization (#10285)
  • Fixes an issue where DeepSeek-reasoner could error after condensation if the condensed summary lacked required reasoning fields (#10292)

Misc Improvements​

  • Cleaner eval logs: Deduplicates repetitive message log entries so eval traces are easier to read (#10286)

QOL Improvements​

  • New tasks now default to native tool calling on models that support it, reducing the need for manual tool protocol selection (#10281)

Provider Updates​

  • Improves Z.ai thinking model message formatting by attaching environment_details to tool results instead of emitting separate system messages (#10289)
  • LiteLLM no longer sends parallel_tool_calls, improving tool-call compatibility (thanks farazoman!) (#10274)