Skip to main content

Roo Code 3.28 Combined Release Notes

This version introduces the Supernova model to Roo Code Cloud, Task Sync for monitoring tasks from your phone, plus Roo Code Cloud integration features, enhanced chat editing, conversation preservation improvements, and multiple bug fixes.

Supernova Model - FREE Access​

We've added the new roo/code-supernova stealth model to the Roo Code Cloud provider - completely FREE during the promotional period (#8175):

  • FREE ACCESS: No API keys, no costs - completely free through Roo Code Cloud
  • Advanced Capabilities: Powerful AI model with image support for multimodal coding tasks
  • Not So Massive Context: 200,000 token context window with 16,384 max output
  • Stealth Performance: Experience cutting-edge model performance with the code-supernova variant
  • Zero Setup: Just select it from your Roo Code Cloud provider - no configuration needed

Select the FREE code-supernova model from the Roo provider to start using this powerful new capability at no cost.

Documentation: See the Roo Code Router guide for model selection details.

Task Sync​

Introducing our new cloud connectivity feature that lets you monitor long-running tasks from your phone - no more waiting at your desk! (#7799, #7805, #7850):

Important: Roo Code remains completely free and open source. Task Sync is an optional supplementary service that connects your local VS Code to the cloud - all processing still happens in your VS Code instance.

Task Sync interface showing real-time task monitoring from browser

Task Sync (FREE for All Users):

  • Monitor from Anywhere: Check on long-running tasks from your phone while away from your desk
  • Real-time Updates: Live streaming of your local task messages and progress
  • Task History: Your tasks are saved to the cloud for later reference
  • Cloud Visibility: View your VS Code tasks from any browser

Documentation: See Task Sync.

GPT-5-Codex lands in OpenAI Native​

  • Work with repository-scale context: Keep multi-file specs and long reviews in a single thread thanks to a 400k token window.
  • Reuse prompts faster and include visuals: Prompt caching and image support help you iterate on UI fixes without re-uploading context.
  • Let the model adapt its effort: GPT-5-Codex automatically balances quick responses for simple questions with deeper reasoning on complex builds.

This gives teams a higher-capacity OpenAI option without extra configuration.#8260:

Documentation: See OpenAI Provider Guide for capabilities and setup guidance.

QOL Improvements​

  • Cloud Account Management: Added account switcher and "Create Team Account" option for logged-in users, making it easier to manage multiple accounts (#8291)

  • Chat Icon Sizing: Chat interface icons now maintain consistent size in limited space (#8343)

  • Social Media Sharing: Improved OpenGraph images for better preview cards when sharing Roo Code links on Slack, Twitter, and Facebook (#8285)

  • Auto-approve keyboard shortcut: Toggle approvals with Cmd/Ctrl+Alt+A from anywhere in the editor so you can stay in the flow while reviewing changes (via #8214)

  • Click-to-Edit Chat Messages: Click directly on any message text to edit it, with ESC to cancel and improved padding consistency (#7790)

  • Enhanced Reasoning Display: The AI's thinking process now shows a persistent timer and displays reasoning content in clean italic text (#7752)

  • Easier-to-scan reasoning transcripts: Added clear line breaks before reasoning headers inside the UI so long thoughts are easier to skim (via #7868)

  • Manual Auth URL Input: Users in containerized environments can now paste authentication redirect URLs manually when automatic redirection fails (#7805)

  • Active Mode Centering: The mode selector dropdown now automatically centers the active mode when opened (#7883)

  • Preserve First Message: The first message containing slash commands or initial context is now preserved during conversation condensing instead of being replaced with a summary (#7910)

  • Checkpoint Initialization Notifications: You'll now receive clear notifications when checkpoint initialization fails, particularly with nested Git repositories (#7766)

  • Translation coverage auditing: The translation checker now validates package.nls locales by default to catch missing strings before release (via #8255)

  • Smaller and more subtle auto-approve UI (thanks brunobergher!) (#7894)

  • Add padding to the cloudview for better visual spacing (#7954)

  • Temporarily pause auto-approve with a global “Enabled” switch so you can pause without losing your per-toggle selections; the UI clearly distinguishes global vs. per-toggle states (#8024)

  • Slash commands management moved to Settings > Slash Commands with a gear shortcut from the chat popover; the popover is now selection-only for quick use (#7988)

  • Keyboard shortcut to quickly “Add to Context” (macOS: Cmd+Y; Windows/Linux: Ctrl+Y) (#7908)

  • Add Image button moved inside the text area (top-right) and Enhance Prompt positioned above Send to reduce pointer travel and clarify layout (#7989)

  • Composer action buttons now appear only when there's text input, reducing visual clutter and accidental clicks (#7987)

  • New Task button icon switched from "+" to a compose/pencil icon to better match its purpose and align with codicon standards (#7942)

  • Redesigned Message Feed: Enjoy a cleaner, more readable chat interface with improved visual hierarchy that helps you focus on what matters (#7985)

  • Responsive Auto-Approve: The auto-approve dropdown now adapts to different window sizes with smart 1-2 column layouts, and tooltips show all enabled actions without truncation (#8032)

  • Network Resilience: Telemetry data now automatically retries on network failures, ensuring analytics and diagnostics aren't lost during connectivity issues (#7597)

  • Code blocks wrap by default: Code blocks now wrap text by default, and the snippet toolbar no longer includes language or wrap toggles, keeping snippets readable across locales (via #8194; #8208)

  • Reset logs you out of Roo Code Cloud: Reset now also logs you out of Roo Code Cloud for a truly clean slate; the reset still completes even if logout fails. (#8312)

  • Measure engagement with upsell messages: Telemetry for DismissibleUpsell tracks dismiss vs click (upsellId only) to tune messaging and reduce noise; no user-facing behavior change. (#8309)

  • Welcome screen Roo provider experiment: Surfaces the Roo provider on the welcome screen (localized, with click telemetry) to speed up setup. (#8317)

  • Cleaner prompts (no “thinking” tags): Removes <thinking> tags for cleaner output, fewer tokens, and better model compatibility; preserves the plain‑language rule to confirm tool success. (#8319)

  • Simpler Requesty model refresh: Removes an unnecessary warning when refreshing the Requesty models list so you can fetch new policies without leaving the screen. (#7710)

  • Bedrock 1M Context Checkbox: Adds checkbox to enable 1M context window support for Claude Sonnet 4 and 4.5 on Bedrock, providing more flexibility for large-scale projects (#8384)

  • Include reasoning messages in cloud task histories so you can review complete thinking after runs (#8401)

  • Identify Cloud tasks in the extension bridge to improve diagnostics, logging, and future UI behavior (#8539)

  • Add the parent task ID in telemetry to improve traceability (#8532)

  • Clarify the zh‑TW “Run command” label to match the tooltip and reduce confusion (thanks PeterDaveHello!) (#8631)

  • Image generation model selection update: defaults to Gemini 2.5 Flash Image; adds OpenAI GPT‑5 Image and GPT‑5 Image Mini; clearer settings dropdown (thanks chrarnoldus!) (#8698)

Bug Fixes​

  • AWS Bedrock Claude Sonnet 4.5: Corrected model identifier for Claude Sonnet 4.5 on AWS Bedrock (thanks sunhyung!) (#8371)

  • OpenRouter Claude Sonnet 4.5: Corrected model ID format to align with naming and prevent selection/API mismatches (#8373)

  • GPT-5 LiteLLM Compatibility: Fixed GPT-5 models failing with LiteLLM provider by using correct max_completion_tokens parameter (thanks lx1054331851!) (#6980)

  • Fixed "No tool used" Errors: Resolved the situation where LLMs would sometimes not make tool calls in their response which improves Roo's overall flow. (#8292)

  • Context Condensing: Fixed an issue where the initial task request was being lost during context condensing, causing Roo to try to re-answer the original task ask when resuming after condensing (#8298)

  • Roo provider stays signed in: Roo provider tokens refresh automatically and the local evals app binds to port 3446 for predictable scripts (via #8224)

  • Checkpoint text stays on one line: Prevented multi-line wrapping in languages such as Chinese, Korean, Japanese, and Russian so the checkpoint UI stays compact (via #8207; reported in #8206)

  • Ollama respects Modelfile num_ctx: Roo now defers to your Modelfile’s context window to avoid GPU OOMs while still allowing explicit overrides when needed (via #7798; reported in #7797)

  • Groq Context Window: Fixed incorrect display of cached tokens in context window (#7839)

  • Chat Message Operations: Resolved duplication issues when editing messages and "Couldn't find timestamp" errors when deleting (#7793)

  • UI Overlap: Fixed CodeBlock button z-index to prevent overlap with popovers and configuration panels (thanks A0nameless0man!) (#7783)

  • Fix message queue re-queue loop in Task.ask() that could degrade performance (#7823)

  • Restrict @-mention parsing to line-start or whitespace boundaries to prevent false triggers (#7876)

  • Make nested git repository warning persistent with path info for better visibility (#7885)

  • Include API key in Ollama /api/tags requests to support authenticated instances (thanks ItsOnlyBinary!) (#7903)

  • Preserve original first message context during conversation condensing for more coherent long chats (#7939)

  • Temperature Parameter: Restored temperature parameter to fix crashes in TabbyApi and ExLlamaV2 backends (thanks drknyt!) (#7594)

  • Ollama Model Info: Fixed max context window display for Ollama models (thanks ItsOnlyBinary!) (#7679)

  • Image Preview Caching: This ensures the chat always displays the newly generated image instead of a cached version (#7860)

  • Restore square brackets in Gemini-generated code so arrays/indexers render correctly (e.g., string[] items = new string[5];) (thanks AtkinsDynamicObjects!) (#7577)

  • Filter out Claude Code built-in tools (ExitPlanMode, BashOutput, KillBash) to avoid erroneous tool calls and distracting warnings (thanks juliettefournier-econ!) (#7818)

  • Fix an issue where the context menu could be obscured while editing messages; the menu stays visible and usable across screen sizes (thanks NaccOll, mini2s!) (#7951)

  • Surface a clear, actionable error instead of cryptic ByteString conversion failures during code indexing when API keys contain invalid/special characters (thanks PavelA85!) (#8008)

  • Correct C# tree-sitter query so C# files are parsed and indexed properly, improving semantic search, navigation, and context quality (thanks mubeen-zulfiqar, vadash, Source-GuyCoder!) (#7813)

  • Keyboard Shortcut: Fixed the command+y shortcut in Nightly builds, restoring quick content addition to context (#8070)

  • Todo list compatibility: Todo lists now display correctly when AI models return checklists with dash prefixes (- [ ] Task), improving compatibility with various models (thanks NaccOll!) (#8055)

  • Vertex AI pricing accuracy: Fixed local cost calculation for Gemini models via Vertex AI to properly apply tiered pricing and cache discounts, ensuring displayed costs match actual Google Cloud billing (thanks ikumi3!) (#8018)

  • Conversation history preservation: Fixed duplicate/overlaid tasks and lost conversation history when canceling during model responses, especially during reasoning phases (#8171)

  • Tool‑use guidance clarity: Clarifies the retry suggestion so the model selects the correct file‑reading step during apply‑diff retries, reducing failed file edits. (#8315)

  • Anthropic Sonnet 4.5 Model ID: Corrects the model ID to the official claude-sonnet-4-5, resolving API errors when using Claude Sonnet 4.5 (#8384)

  • AWS Bedrock: Removed topP parameter to avoid conflicts with extended thinking, improving reliability for streaming and non‑streaming responses (thanks ronyblum!) (#8388)

  • Vertex AI: Fixed Sonnet 4.5 model configuration to prevent 404s and align with the correct default model ID (thanks nickcatal!) (#8391)

  • "Reset and Continue" now truly resets cost tracking; subsequent requests only count new message costs (#6890)

  • Prompts settings: Save button activates on first change and preserves empty strings (#8267)

  • Settings: Remove overeager "unsaved changes" dialog when nothing changed (#8410)

  • Show the Send button when only images are attached (#8423)

  • Fix grey screens caused by long context task sessions (#8696)

  • Editor targeting fix: avoids editing read‑only git diff views; edits the actual file (thanks hassoncs!) (#8676)

Provider Updates​

  • Claude 4.5 Sonnet: Added support for Anthropic's latest Claude 4.5 Sonnet model across all providers (Anthropic, AWS Bedrock, Claude Code, Google Vertex AI, and OpenRouter) with 1M context window support (#8368)

  • New Free Models: Added two more free models to the Roo provider - xai/grok-4-fast and deepseek/deepseek-chat-v3.1 (#8304)

  • Vertex AI Models: Added support for 6 new models including DeepSeek-V3, GPT-OSS, and Qwen models, plus the us-south1 region (thanks ssweens!) (#7727)

  • DeepSeek Pricing: Updated to new unified rates effective September 5, 2025 - $0.56/M input tokens, $1.68/M output tokens (thanks NaccOll!) (#7687)

  • Chutes Provider: Add Qwen3 Next 80B A3B models (#7948)

  • Z.ai Coding Plans: Region-aware options (International/China) with automatic base URL selection and sensible defaults to simplify setup (thanks chrarnoldus!) (#8003)

  • SambaNova models: Added DeepSeek-V3.1 and GPT-OSS 120B to the SambaNova provider (thanks snova-jorgep!) (#8186)

  • Supernova 1M context: Upgrades the default Supernova to roo/code-supernova-1-million (1,000,000‑token context). Existing settings auto‑migrate; expect fewer truncations on long tasks. (#8330)

  • Z.ai GLM‑4.6: Adds GLM‑4.6 model support with a 200k (204,800) context window and availability across both international and mainland API lines (thanks dmarkey!) (#8408)

  • Chutes: Add DeepSeek‑V3.1‑Terminus, DeepSeek‑V3.1‑turbo, DeepSeek‑V3.2‑Exp, GLM‑4.6‑FP8, and Qwen3‑VL‑235B‑A22B‑Thinking to the provider picker (#8467)

  • Claude Code: Added a 1M‑context variant of Claude Sonnet 4.5; select via the [1m] suffix to fit larger repos and logs with fewer truncations (thanks ColbySerpa!) (#8586)

  • Claude Haiku 4.5: Available across Anthropic, AWS Bedrock, and Vertex AI with 200k context, up to 64k output, image input, and prompt caching support (#8673)

  • Gemini: Remove unsupported free “2.5 Flash Image Preview”; paid path unaffected (#8359)

  • Roo Code Cloud: Deprecate “Grok 4 Fast”; show warning if selected and hide when not selected (#8481)

  • Bedrock: Claude Sonnet 4 1M‑context handling and compatibility improvements (#8421)

  • Bedrock: add versioned user agent for per‑version metrics and error tracking; no action needed (thanks ajjuaire!) (#8663)

  • Z AI: only two coding endpoints (International/China) are supported; defaults to International; legacy non‑coding endpoints unsupported (#8693)

Misc Improvements​

  • Website improvements: New pricing page, clearer homepage messaging, and tighter navigation/visuals (including mobile). No changes to the extension UX. (#8303, #8326)
  • Telemetry Tracking: Enhanced analytics to track when users change telemetry settings (#8339)
  • Website Testimonials: Updated website with enhanced testimonials section (#8360)
  • Contributor Recognition: Improved contributor badge workflow with automated cache refreshing (#8083)
  • Internal cleanup: Removes the deprecated pr-reviewer mode. No changes to the extension UX. (#8222)
  • Infrastructure: Set port 3446 for web-evals service in production mode (#8288)
  • Cloud Task Button: Added a new button for opening tasks in Roo Code Cloud with QR codes and shareable URLs (#7572)
  • Telemetry Default: Posthog telemetry is now enabled by default with easy opt-out options in settings (#7909)
  • Roo Code Cloud Announcement: Updated in-app announcements about Roo Code Cloud features, localized in 18 languages (#7914)
  • Merge Resolver Mode: Added GIT_EDITOR environment variable to prevent hanging during non-interactive rebase operations (#7819)
  • Updated dependency esbuild to v0.25.9 for improved build performance (#5455)
  • Updated dependency @changesets/cli to v2.29.6 (#7376)
  • Updated dependency nock to v14.0.10 (#6465)
  • Updated dependency eslint-config-prettier to v10.1.8 (#6464)
  • Updated dependency eslint-plugin-turbo to v2.5.6 (#7764)
  • Updated dependency axios to v1.12.0 for improved reliability and security of HTTP requests (#7963)
  • Code Supernova announcement: Added announcement for the new Code Supernova model with improved authentication flow and landing page redirection (#8197)
  • Privacy policy update: Updated privacy policy to include provisions for optional marketing emails with clear opt-out options (thanks jdilla1277!) (#8180)