apps/docs/docs/update-notes/v3.48.0.mdx
This release adds API configuration locking, Claude Sonnet 4.6 support across major providers, and a recursive subtask history tree, alongside significant removals of browser automation, nine low-usage providers, built-in skills, and file-based system prompt overrides.
Claude Sonnet 4.6 (claude-sonnet-4-6) is now available across Anthropic, Amazon Bedrock, Google Vertex, OpenRouter, and Vercel AI Gateway with a 1M token context window and 64K max output tokens. (thanks PeterDaveHello!) (#11509)
A new lock icon in the API config selector lets you pin your active provider and model across all mode switches in the current workspace. When locked, switching modes no longer swaps out your API configuration — the same provider and model stay active regardless of which mode you switch to. Unlock at any time to restore normal per-mode config behavior. The lock state is stored per-workspace and never affects other projects. (#11295)
The History view now renders the complete nested subtask hierarchy as an expandable tree. Each level of nesting is indented independently and can be expanded or collapsed, making it easy to navigate deep orchestrator task chains and drill into any branch at arbitrary depth. Previously only flat or single-level subtask groupings were visible. (#11299)
list_files, read_file, and file-edit operations now collapse into a single grouped block instead of stacking as separate identical headers. Batch "Save All" / "Deny All" buttons are included for file-edit groups, reducing chat clutter during multi-file tasks. (#11245)edit Tool: The search_and_replace tool has been renamed to edit with a flatter parameter model (file_path, old_string, new_string, replace_all). One replacement per call with uniqueness enforced by default. search_and_replace is retained as a backward-compatible alias. (#11296)disabledTools Setting: Administrators can globally disable specific native tools via the new disabledTools extension or organization setting. Disabled tools are stripped from API requests and blocked at execution time, targeting enterprise deployments that require restricted tool access. (#11277)browser_action tool has been removed along with all related settings, the browser session panel, and the puppeteer-core dependencies. Migrate to Playwright MCP or another MCP-based browser tool. (#11392).roo/system-prompt-{mode} file override mechanism has been removed along with the in-chat warning banner and the "Advanced: Override System Prompt" disclosure in Mode settings. Migrate to custom instructions or mode-level prompt customization. (#11387)gemini-3-pro-preview) caused by stale reasoning effort values being forwarded from other models. (#11303)ask_followup_question Validation: Fixes a crash when AI models omit the follow_up parameter in ask_followup_question calls, replacing a cryptic TypeError with an actionable retry prompt. (thanks rossdonald!) (#11484)defaultTemperature being silently ignored when models returned undefined from getModelParams, which could cause unexpected temperature values for some providers. (#11218)echo "task" | roo. (#11476)-a / --require-approval flag to opt in to manual approval prompts. The legacy --yes and --dangerously-skip-permissions flags have been removed. (#11424)roo-cli-linux-arm64.tar.gz binary now ships with each release for AWS Graviton, Raspberry Pi, and other ARM64 Linux environments. (#11314)