Roo Code changelog

Release v3.54.0

Release v3.53.0

Release v3.52.1

Release v3.52.0

Release v3.51.1

Release v3.51.0

Release v3.50.5

[3.50.4] - 2026-02-21

• Feat: Add MiniMax M2.5 model support (#11471 by @love8ko, PR #11458 by @roomote)

Release v3.50.3

Release v3.50.2

Release v3.50.1

Release v3.50.0

Release v3.49.0

Release v3.48.1

[3.48.0]

• Add Anthropic Claude Sonnet 4.6 support across all providers — Anthropic, Bedrock, Vertex, OpenRouter, and Vercel AI Gateway (PR #11509 by @PeterDaveHello)
• Add lock toggle to pin API config across all modes in a workspace (PR #11295 by @hannesrudolph)
• Fix: Prevent parent task state loss during orchestrator delegation (PR #11281 by @hannesrudolph)
• Fix: Resolve race condition in new_task delegation that loses parent task history (PR #11331 by @daniel-lxs)
• Fix: Serialize taskHistory writes and fix delegation status overwrite race (PR #11335 by @hannesrudolph)
• Fix: Prevent chat history loss during cloud/settings navigation (#11371 by @SannidhyaSah, PR #11372 by @SannidhyaSah)
• Fix: Preserve condensation summary during task resume (#11487 by @SannidhyaSah, PR #11488 by @SannidhyaSah)
• Fix: Resolve chat scroll anchoring and task-switch scroll race conditions (PR #11385 by @hannesrudolph)
• Fix: Preserve pasted images in chatbox during chat activity (PR #11375 by @app/roomote)
• Add disabledTools setting to globally disable native tools (PR #11277 by @daniel-lxs)
• Rename search_and_replace tool to edit and unify edit-family UI (PR #11296 by @hannesrudolph)
• Render nested subtasks as recursive tree in history view (PR #11299 by @hannesrudolph)
• Remove 9 low-usage providers and add retired-provider UX (PR #11297 by @hannesrudolph)
• Remove browser use functionality entirely (PR #11392 by @hannesrudolph)
• Remove built-in skills and built-in skills mechanism (PR #11414 by @hannesrudolph)
• Remove footgun prompting (file-based system prompt override) (PR #11387 by @hannesrudolph)
• Batch consecutive tool calls in chat UI with shared utility (PR #11245 by @hannesrudolph)
• Validate Gemini thinkingLevel against model capabilities and handle empty streams (PR #11303 by @hannesrudolph)
• Add GLM-5 model support to Z.ai provider (PR #11440 by @app/roomote)
• Fix: Prevent double notification sound playback (PR #11283 by @hannesrudolph)
• Fix: Prevent false unsaved changes prompt with OpenAI Compatible headers (#8230 by @hannesrudolph, PR #11334 by @daniel-lxs)
• Fix: Cancel backend auto-approval timeout when auto-approve is toggled off mid-countdown (PR #11439 by @SannidhyaSah)
• Fix: Add follow_up param validation in AskFollowupQuestionTool (PR #11484 by @rossdonald)
• Fix: Prevent webview postMessage crashes and make dispose idempotent (PR #11313 by @0xMink)
• Fix: Avoid zsh process-substitution false positives in assignments (PR #11365 by @hannesrudolph)
• Fix: Harden command auto-approval against inline JS false positives (PR #11382 by @hannesrudolph)
• Fix: Make tab close best-effort in DiffViewProvider.open (PR #11363 by @0xMink)
• Fix: Canonicalize core.worktree comparison to prevent Windows path mismatch failures (PR #11346 by @0xMink)
• Fix: Make removeClineFromStack() delegation-aware to prevent orphaned parent tasks (PR #11302 by @app/roomote)
• Fix task resumption in the API module (PR #11369 by @cte)
• Make defaultTemperature required in getModelParams to prevent silent temperature overrides (PR #11218 by @app/roomote)
• Remove noisy console.warn logs from NativeToolCallParser (PR #11264 by @daniel-lxs)
• Consolidate getState calls in resolveWebviewView (PR #11320 by @0xMink)
• Clean up repo-facing mode rules (PR #11410 by @hannesrudolph)
• Implement ModelMessage storage layer with AI SDK response messages (PR #11409 by @daniel-lxs)
• Extract translation and merge resolver modes into reusable skills (PR #11215 by @app/roomote)
• Add blog section with initial posts to roocode.com (PR #11127 by @app/roomote)
• Replace Roomote Control with Linear Integration in cloud features grid (PR #11280 by @app/roomote)
• Add IPC query handlers for commands, modes, and models (PR #11279 by @cte)
• Add stdin stream mode for the CLI (PR #11476 by @cte)
• Make CLI auto-approve by default with require-approval opt-in (PR #11424 by @cte)
• Update CLI default model from Opus 4.5 to Opus 4.6 (PR #11273 by @app/roomote)
• Add linux-arm64 support for the Roo CLI (PR #11314 by @cte)
• CLI release: v0.0.51 (PR #11274 by @cte)
• CLI release: v0.0.52 (PR #11324 by @cte)
• CLI release: v0.0.53 (PR #11425 by @cte)
• CLI release: v0.0.54 (PR #11477 by @cte)

[3.47.3] - 2026-02-06

• Remove "Enable URL context" and "Enable Grounding with Google search" checkboxes that are no longer needed (PR #11253 by @roomote)
• Revert refactor that appended environment details into existing blocks, restoring original behavior (PR #11256 by @mrubens)
• Revert removal of stripAppendedEnvironmentDetails and helpers, restoring necessary utility functions (PR #11255 by @mrubens)

[3.47.2] - 2026-02-05

• Add support for .agents/skills directory (PR #11181 by @roomote)
• Fix: Restore Gemini thought signature round-tripping after AI SDK migration (PR #11237 by @hannesrudolph)
• Fix: Capture and round-trip thinking signature for Bedrock Claude (PR #11238 by @hannesrudolph)

Release v3.47.1

[3.47.0] - 2026-02-05

• Add Claude Opus 4.6 support across all providers (#11223 by @hannesrudolph, PR #11224 by @hannesrudolph and @PeterDaveHello)
• Add GPT-5.3-Codex model to OpenAI - ChatGPT provider (PR #11225 by @roomote)
• Migrate Gemini and Vertex providers to AI SDK for improved reliability and consistency (PR #11180 by @daniel-lxs)
• Improve Skills and Slash Commands settings UI with multi-mode support (PR #11157 by @brunobergher)
• Add support for AGENTS.local.md personal override files (PR #11183 by @roomote)
• Add Kimi K2.5 model to Fireworks provider (PR #11177 by @daniel-lxs)
• Improve CLI dev experience and Roo provider API key support (PR #11203 by @cte)
• Fix: Preserve reasoning parts in AI SDK message conversion (#11199 by @hannesrudolph, PR #11217 by @hannesrudolph)
• Refactor: Append environment details into existing blocks for cleaner context (#11200 by @hannesrudolph, PR #11198 by @hannesrudolph)
• Fix: Resolve race condition causing provider switch during CLI mode changes (PR #11205 by @cte)
• Roo Code CLI v0.0.50 (PR #11204 by @cte)
• Chore: Remove dead toolFormat code from getEnvironmentDetails (#11206 by @hannesrudolph, PR #11207 by @roomote)
• Refactor: Simplify docs-extractor mode to focus on raw fact extraction (PR #11129 by @hannesrudolph)
• Revert then re-land AI SDK reasoning fix (PR #11216 by @mrubens, PR #11196 by @hannesrudolph)

[3.46.2] - 2026-02-03

• Fix: Queue messages during command execution instead of losing them (PR #11140 by @mrubens)
• Fix: Transform tool blocks to text before condensing to prevent context corruption (PR #10975 by @daniel-lxs)
• Fix: Add image content support to MCP tool responses (PR #10874 by @roomote)
• Fix: Remove deprecated text-embedding-004 and migrate code index to gemini-embedding-001 (PR #11038 by @roomote)
• Feat: Use custom Base URL for OpenRouter model list fetch (#11150 by @sebastianlang84, PR #11154 by @roomote)
• Feat: Migrate Mistral provider to AI SDK (PR #11089 by @daniel-lxs)
• Feat: Migrate SambaNova provider to AI SDK (PR #11153 by @roomote)
• Feat: Migrate xAI provider to AI SDK (PR #11158 by @roomote)
• Chore: Remove Feature Request from issue template options (PR #11141 by @roomote)
• Fix: IPC improvements for task cancellation and queued message handling (PR #11162 by @cte)

[3.46.1] - 2026-01-30

• Fix: Sanitize tool_use_id in tool_result blocks to match API history, preventing message format errors (PR #11131 by @daniel-lxs)
• Add: Mode dropdown to change skill mode dynamically, allowing more flexible skill configuration (PR #11102 by @SannidhyaSah)
• Add: Import settings option in the initial welcome screen for easier onboarding (#10992 by @emeraldcheshire, PR #10994 by @roomote)
• Chore: Treat extension .env as optional to simplify development setup (PR #11116 by @hannesrudolph)

[3.46.0] - 2026-01-27

• Parallel tool calls enabled by default for improved performance (PR #11031 by @daniel-lxs)
• Codex-inspired read_file refactor introduces indentation mode for extracting complete semantic code blocks without mid-function truncation, ideal when targeting specific lines from search results or errors (#10239 by @pwilkin, PR #10981 by @hannesrudolph)
• Lossless terminal output with new read_command_output tool allows retrieving full command output from truncated executions with pagination and regex filtering (#10941 by @hannesrudolph, PR #10944 by @hannesrudolph)
• New skill system replaces fetch_instructions with a dedicated skill tool and built-in skills for create-mcp-server and create-mode, with configurable skill locations and mandatory skill checks (#11062 by @hannesrudolph, PR #11084 by @hannesrudolph)
• Skills management UI added to settings panel for managing workspace and global skills (#10513 by @SannidhyaSah, PR #10844 by @SannidhyaSah)
• AI SDK provider migrations: Moonshot (PR #11063 by @daniel-lxs), DeepSeek (PR #11079 by @daniel-lxs), Cerebras (PR #11086 by @daniel-lxs), Groq (PR #11088 by @daniel-lxs), and Fireworks (PR #11118 by @daniel-lxs) now use the AI SDK for better streaming and tool support
• Add OpenAI-compatible base provider infrastructure for AI SDK migrations (PR #11063 by @daniel-lxs)
• Add AI SDK dependencies and message conversion utilities (PR #11047 by @daniel-lxs)
• React Compiler integration added to webview-ui for automatic memoization and performance improvements (#9916 by @In-line, PR #9565 by @In-line)
• Fix: Include reserved output tokens in task header percentage calculation (PR #11034 by @app/roomote)
• Fix: Calculate header percentage based on available input space (PR #11054 by @app/roomote)
• Fix: Prevent time-travel bug in parallel tool calling (PR #11046 by @daniel-lxs)
• Docs: Clarify read_command_output search param should be omitted when not filtering (PR #11056 by @hannesrudolph)
• Add pnpm serve command for code-server development (PR #10964 by @mrubens)
• Update Next.js to latest version (PR #11108 by @cte)
• Replace bespoke navigation menu with shadcn navigation menu on website (PR #11117 by @app/roomote)
• Add Linear integration marketing page to website (PR #11028 by @app/roomote)

[3.45.0] - 2026-01-27

• Smart Code Folding: Context condensation now intelligently preserves a lightweight map of files you worked on—function signatures, class declarations, and type definitions—so Roo can continue referencing them accurately after condensing. Files are prioritized by most recent access, with a ~50k character budget ensuring your latest work is always preserved. (Idea by @shariqriazz, PR #10942 by @hannesrudolph)

[3.44.2] - 2026-01-27

• Re-enable parallel tool calling with new_task isolation safeguards (PR #11006 by @mrubens)
• Fix worktree indexing by using relative paths in isPathInIgnoredDirectory (PR #11009 by @daniel-lxs)
• Fix local model validation error for Ollama models (PR #10893 by @roomote)
• Fix duplicate tool_call emission from Responses API providers (PR #11008 by @daniel-lxs)

[3.44.1] - 2026-01-27

• Fix LiteLLM tool ID validation errors for Bedrock proxy (PR #10990 by @daniel-lxs)
• Add temperature=0.9 and top_p=0.95 to zai-glm-4.7 model for better generation quality (PR #10945 by @sebastiand-cerebras)
• Add quality checks to marketing site deployment workflows (PR #10959 by @mp-roocode)

[3.44.0] - 2026-01-26

• Add worktree selector and creation UX (PR #10940 by @brunobergher, thanks Cline!)
• Improve subtask visibility and navigation in history and chat views (PR #10864 by @brunobergher)
• Add wildcard support for MCP alwaysAllow configuration (PR #10948 by @app/roomote)
• Fix: Prevent nested condensing from including previously-condensed content (PR #10985 by @hannesrudolph)
• Fix: VS Code LM token counting returns 0 outside requests, breaking context condensing (#10968 by @srulyt, PR #10983 by @daniel-lxs)
• Fix: Record truncation event when condensation fails but truncation succeeds (PR #10984 by @hannesrudolph)
• Replace hyphen encoding with fuzzy matching for MCP tool names (PR #10775 by @daniel-lxs)
• Remove MCP SERVERS section from system prompt for cleaner prompts (PR #10895 by @daniel-lxs)
• new_task tool creates checkpoint the same way write_to_file does (PR #10982 by @daniel-lxs)
• Update Fireworks provider with new models (#10674 by @hannesrudolph, PR #10679 by @ThanhNguyxn)
• Fix: Truncate AWS Bedrock toolUseId to 64 characters (PR #10902 by @daniel-lxs)
• Fix: Restore opaque background to settings section headers (PR #10951 by @app/roomote)
• Fix: Remove unsupported Fireworks model tool fields (PR #10937 by @app/roomote)
• Update and improve zh-TW Traditional Chinese locale and docs (PR #10953 by @PeterDaveHello)
• Chore: Remove POWER_STEERING experiment remnants (PR #10980 by @hannesrudolph)

[3.43.0] - 2026-01-23

• Intelligent Context Condensation v2: New context condensation system that intelligently summarizes conversation history when approaching context limits, preserving important information while reducing token usage (PR #10873 by @hannesrudolph)
• Improved context condensation with environment details, accurate token counts, and lazy evaluation for better performance (PR #10920 by @hannesrudolph)
• Move condense prompt editor to Context Management tab for better discoverability and organization (PR #10909 by @hannesrudolph)
• Update Z.AI models with new variants and pricing (#10859 by @ErdemGKSL, PR #10860 by @ErdemGKSL)
• Add pnpm install:vsix:nightly command for easier nightly build installation (PR #10912 by @hannesrudolph)
• Fix: Convert orphaned tool_results to text blocks after condensing to prevent API errors (PR #10927 by @daniel-lxs)
• Fix: Auto-migrate v1 condensing prompt and handle invalid providers on import (PR #10931 by @hannesrudolph)
• Fix: Use json-stream-stringify for pretty-printing MCP config files to prevent memory issues with large configs (#9862 by @Michaelzag, PR #9864 by @Michaelzag)
• Fix: Correct Gemini 3 pricing for Flash and Pro models (#10432 by @rossdonald, PR #10487 by @roomote)
• Fix: Skip thoughtSignature blocks during markdown export for cleaner output (#10199 by @rossdonald, PR #10932 by @rossdonald)
• Fix: Duplicate model display for OpenAI Codex provider (PR #10930 by @roomote)
• Remove diffEnabled and fuzzyMatchThreshold settings as they are no longer needed (#10648 by @hannesrudolph, PR #10298 by @hannesrudolph)
• Remove MULTI_FILE_APPLY_DIFF experiment (PR #10925 by @hannesrudolph)
• Remove POWER_STEERING experimental feature (PR #10926 by @hannesrudolph)
• Remove legacy XML tool calling code (getToolDescription) for cleaner codebase (PR #10929 by @hannesrudolph)

[3.42.0] - 2026-01-22

• Added UI to track your ChatGPT usage limits in the OpenAI Codex provider (PR #10813 by @hannesrudolph)
• Removed deprecated Claude Code provider (PR #10883 by @daniel-lxs)
• Streamlined codebase by removing legacy XML tool calling functionality (#10848 by @hannesrudolph, PR #10841 by @hannesrudolph)
• Standardize model selectors across all providers: Improved consistency of model selection UI (#10650 by @hannesrudolph, PR #10294 by @hannesrudolph)
• Enable prompt caching for Cerebras zai-glm-4.7 model (#10601 by @jahanson, PR #10670 by @app/roomote)
• Add Kimi K2 thinking model to VertexAI provider (#9268 by @diwakar-s-maurya, PR #9269 by @app/roomote)
• Warn users when too many MCP tools are enabled (PR #10772 by @app/roomote)
• Migrate context condensing prompt to customSupportPrompts (PR #10881 by @hannesrudolph)
• Unify export path logic and default to Downloads folder (PR #10882 by @hannesrudolph)
• Performance improvements for webview state synchronization (PR #10842 by @hannesrudolph)
• Fix: Handle mode selector empty state on workspace switch (#10660 by @hannesrudolph, PR #9674 by @app/roomote)
• Fix: Resolve race condition in context condensing prompt input (PR #10876 by @hannesrudolph)
• Fix: Prevent double emission of text/reasoning in OpenAI native and codex handlers (PR #10888 by @hannesrudolph)
• Fix: Prevent task abortion when resuming via IPC/bridge (PR #10892 by @cte)
• Fix: Enforce file restrictions for all editing tools (PR #10896 by @app/roomote)
• Fix: Remove custom condensing model option (PR #10901 by @hannesrudolph)
• Unify user content tags to <user_message> for consistent prompt formatting (#10658 by @hannesrudolph, PR #10723 by @app/roomote)
• Clarify linked SKILL.md file handling in prompts (PR #10907 by @hannesrudolph)
• Fix: Padding on Roo Code Cloud teaser (PR #10889 by @app/roomote)

[3.41.3] - 2026-01-18

• Fix: Thinking block word-breaking to prevent horizontal scroll in the chat UI (PR #10806 by @roomote)
• Add Claude-like CLI flags and authentication fixes for the Roo Code CLI (PR #10797 by @cte)
• Improve CLI authentication by using a redirect instead of a fetch (PR #10799 by @cte)
• Fix: Roo Code Router fixes for the CLI (PR #10789 by @cte)
• Release CLI v0.0.48 with latest improvements (PR #10800 by @cte)
• Release CLI v0.0.47 (PR #10798 by @cte)
• Revert E2E tests enablement to address stability issues (PR #10794 by @cte)

[3.41.2] - 2026-01-16

• Add button to open markdown in VSCode preview for easier reading of formatted content (PR #10773 by @brunobergher)
• Fix: Reset invalid model selection when using OpenAI Codex provider (PR #10777 by @hannesrudolph)
• Fix: Add openai-codex to providers that don't require an API key (PR #10786 by @roomote)
• Fix: Detect Gemini models with space-separated names for proper thought signature injection in LiteLLM (PR #10787 by @daniel-lxs)

[3.41.1] - 2026-01-16

• Feat: Aggregate subtask costs in parent task (#5376 by @hannesrudolph, PR #10757 by @taltas)
• Fix: Prevent duplicate tool_use IDs causing API 400 errors (PR #10760 by @daniel-lxs)
• Fix: Handle missing tool identity in OpenAI Native streams (PR #10719 by @hannesrudolph)
• Fix: Truncate call_id to 64 chars for OpenAI Responses API (PR #10763 by @daniel-lxs)
• Fix: Gemini thought signature validation errors (PR #10694 by @daniel-lxs)
• Fix: Filter out empty text blocks from user messages for Gemini compatibility (PR #10728 by @daniel-lxs)
• Fix: Flatten top-level anyOf/oneOf/allOf in MCP tool schemas (PR #10726 by @daniel-lxs)
• Fix: Filter Ollama models without native tool support (PR #10735 by @daniel-lxs)
• Feat: Add settings tab titles to search index (PR #10761 by @roomote)
• Feat: Clarify Slack and Linear are Cloud Team only features (PR #10748 by @roomote)

[3.41.0] - 2026-01-15

• Add OpenAI - ChatGPT Plus/Pro Provider that gives subscription-based access to Codex models without per-token costs (PR #10736 by @hannesrudolph)
• Add gpt-5.2-codex model to openai-native provider, providing access to the latest GPT model with enhanced coding capabilities (PR #10731 by @hannesrudolph)
• Fix: Clear terminal output buffers to prevent memory leaks that could cause gray screens and performance degradation (#10666, PR #7666 by @hannesrudolph)
• Fix: Inject dummy thought signatures on ALL tool calls for Gemini models, resolving issues with Gemini tool call handling through LiteLLM (PR #10743 by @daniel-lxs)
• Enable E2E tests with 39 passing tests, improving test coverage and reliability (PR #10720 by @ArchimedesCrypto)
• Add alwaysAllow config for MCP time server tools in E2E tests (PR #10733 by @ArchimedesCrypto)

[3.40.1] - 2026-01-13

• Fix: Add allowedFunctionNames support for Gemini to prevent mode switch errors (#10711 by @hannesrudolph, PR #10708 by @hannesrudolph)