Skip to main content

Roo Code 3.30 Release Notes (2025-11-03)

This release adds OpenRouter embeddings and Bedrock Global Inference, updates provider defaults and limits, and delivers stability and UI improvements.

Roo Code v3.30 Release

OpenRouter Embeddings​

We've added OpenRouter as an embedding provider for codebase indexing in Roo Code (thanks dmarkey!) (#8973).

OpenRouter currently supports 7 embedding models, including the top‑ranking Qwen3 Embedding.

📚 Documentation: See Codebase Indexing and OpenRouter Provider.

QOL Improvements​

  • Terminal settings cleanup with Inline as the default terminal and clearer options; shell integration default is disabled to reduce environment conflicts (#8342)
  • File path tooltips show full paths on hover across reads/edits, directory listings, and search, reducing confusion in narrow layouts (thanks da2ce7!) (#9030)
  • Home screen visuals refined for clarity and space usage (#9057)
  • Chat diff appearance is clearer and easier to scan (#8932)
  • Settings copy clarifies that an Error & Repetition Limit of 0 disables the mechanism (#8965)

Bug Fixes​

  • Prevent message loss during queue drain race conditions to preserve message order and reliable chats (#8955)
  • Requesty OAuth: auto-create a stable "Requesty" profile with a default model so sign-in completes reliably (thanks Thibault00!) (#8699)
  • Cancel during streaming no longer causes flicker; you can resume in place, input stays enabled, and the spinner stops deterministically (#8986)
  • Remove newline-only reasoning blocks from OpenAI-compatible responses for cleaner output and logs (#8990)
  • "Disable Terminal Shell Integration" now links to the correct documentation section (#8997)
  • Correct OpenRouter Mistral embedding dimension to 1536 to prevent vector size errors; existing Qdrant collections are recreated with the correct size for consistent similarity search (#9028)
  • Revert the recent cancel/resume change that caused UI flicker and unreliable resumption, restoring the previous stable behavior (#9032)
  • Eliminate UI flicker during task cancellation; the task view remains stable (#9037)
  • Keep pinned models fixed at the top of long lists for faster access (thanks XiaoYingYo!) (#8813)
  • Ensure Checkpoints commit to Roo’s shadow repository even when GIT_DIR is set in Dev Containers (thanks nonsleepr!) (#8811)
  • Restore correct 32K maxTokens for Claude Opus 4.1 to avoid premature truncation (thanks kaveh-deriv!) (#9046)
  • Fix dynamic provider model validation so provider switches use a valid default model without TypeErrors (thanks NotADev137!) (#9054)
  • Include full user agent details in AWS Bedrock requests for better auditing and troubleshooting (thanks ajjuaire!) (#9043)
  • Auto‑retry on empty assistant response to prevent task cancellations (reporter: Akillatech) (#9083)
  • Use “system” role for non‑streaming OpenAI‑compatible requests to prevent provider errors (reporter: whitfin) (#8216)
  • Prevent notification sound when messages are queued to auto‑continue (#8540)

Misc Improvements​

  • Add preserveReasoning flag to optionally include reasoning in API history so later turns can leverage prior reasoning; off by default and model‑gated (#8934)
  • Optional pre-push full test run via RUN_TESTS_ON_PUSH=true in .env.local; default behavior keeps pushes fast (tests skipped) (#9055)

Provider Updates​

  • Chutes: dynamic/router provider so new models appear automatically; safer error logging and temperature applied only when supported (#8980)
  • OpenAI‑compatible providers: handle <think> reasoning tags in streaming for consistent reasoning chunk handling (#8989)
  • GLM 4.6: capture reasoning content in base OpenAI‑compatible provider during streaming (#8976)
  • Fireworks: add GLM‑4.6 to the model dropdown for stronger coding performance and longer context (thanks mmealman!) (#8754)
  • Fireworks: add MiniMax M2 with 204.8K context and 4K output tokens; correct pricing metadata (thanks dmarkey!) (#8962)
  • Set Claude Sonnet 4.5 as the default where available across Anthropic, Bedrock, OpenRouter, Claude Code, and Unbound (#8922)
  • Cerebras zai‑glm‑4.6: 40,960 max output tokens and a 131,072‑token context window; faster responses (~1,000 tok/s) (thanks sebastiand-cerebras!) (#9024)
  • OpenRouter: add Qwen3‑Embedding‑0.6B (1024‑dim) and 4B (2560‑dim) for code indexing (thanks dmarkey!) (#9060)
  • Moonshot: add kimi‑k2‑thinking (262k context) (#9079)
  • MiniMax: add MiniMax‑M2‑Stable and enable prompt caching for MiniMax models (#9072)