Skip to main content

Roo Code 3.32 Release Notes (2025-11-14)

Roo Code 3.32 combines GPT-5.1 model support and OpenAI Responses prompt caching with reliability fixes for OpenAI Native and native tool protocol.

Roo Code v3.32 Release

GPT-5.1 models and OpenAI prompt caching​

  • GPT-5.1 in Roo Code (#9252, #9258, #9261, #9259)
    Adds the new GPT-5.1 models across supported providers, with extended 24‑hour prompt caching support in the OpenAI Native provider and improved overall software‑engineering performance.

Bug Fixes​

  • Share button works reliably again (#9253)
    Fixes the Share button so you can open the share popover and send shared tasks and messages without failures.

  • Keep OpenAI Native reasoning stable during condensing (#9263)
    Keeps encrypted reasoning blocks paired with the assistant messages that produced them so OpenAI Native condensing no longer throws errors and long-running tasks with reasoning stay stable even as history is trimmed.

  • Prevent duplicate tool_result blocks for read_file in native protocol (#9272)
    Ensures read_file emits exactly one tool_result per tool_call_id in native protocol, preventing 400 errors while still surfacing clear error messages to both the agent and the user.

  • Fix duplicate tool blocks causing "tool has already been used" errors (#9275)
    Stops XML protocol runs from merging duplicate tool blocks into assistant messages so each tool block executes only once per turn and tool-driven tasks no longer get stuck behind false “tool has already been used” errors.

  • Make OpenAI Native cancellation stop streaming immediately (#9276)
    Adds an abort controller so cancelling an OpenAI Native request stops streaming promptly, avoiding extra tokens after cancellation and reducing wasted usage.

  • Disable XML parser for native tool protocol (#9277)
    Bypasses XML parsing when toolProtocol is set to native and relies on structured tool_call chunks instead, avoiding mis-parsed XML-like snippets and making native tool runs more reliable—especially for models that sometimes mix formats.