Roo Code 3.28.6 Release Notes (2025-09-23)

This release adds GPT-5-Codex to OpenAI Native, sharpens localization coverage, and smooths UI workflows across languages.

alt text

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.

Keyboard shortcut for auto-approve: Toggle approvals with Cmd/Ctrl+Alt+A from anywhere in the editor, keeping focus on the code review flow (via #8214)
Cleaner code blocks: Removed the snippet language picker and word-wrap toggle so wrapped code is easier to read and copy across locales (via #8208)
More readable reasoning blocks: Added spacing before section headers inside reasoning transcripts to make long explanations easier to scan (via #7868)
Translation checks cover package settings: The missing translation finder now validates package.nls files for 17 locales to catch untranslated VS Code strings earlier (via #8255)

Bare-metal evals stay 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)