Skip to main content

​
New features

  • bolt arena command — Race the same task across multiple models in parallel, each in its own isolated git worktree so contenders can freely edit files without stepping on each other. A judge model then ranks the anonymized results using both the final answer and the actual git diff, and the winner’s output goes to stdout. Losing worktrees are kept by default so you can inspect or merge any branch, or pass --cleanup to drop them. Duplicates in --models are allowed, so you can race a model against itself. Run bolt arena --models anthropic/claude-sonnet-4,openai/gpt-5 "add retry logic to the fetch helper", or pass --format json for machine-readable output. See Providers and models.
  • 11 new UI languages — The desktop app and TUI now ship translations for Hindi, Dutch, Indonesian, Vietnamese, Italian, Urdu, Punjabi, Azerbaijani, Finnish, Swedish, and Thai, bringing the total to 28 supported languages. Switch languages in Settings → Appearance → Language. See Desktop features and Install.
  • Right-to-left layout support — Navigation, editors, dialogs, controls, and text rendering now mirror correctly for right-to-left languages like Arabic, Hebrew, and Urdu, so the entire interface reads naturally in those locales.
  • Native desktop translations — OS-level menus, updater dialogs, the recovery flow, WSL onboarding, and native file pickers are now localized alongside the in-app UI, so the whole desktop experience follows your language setting.
  • Expanded localization coverage — Menus, dialogs, prompts, help content, error messages, and accessibility labels are now translated in every supported language, closing gaps where strings previously fell back to English.
  • Poolside provider — Bolt now connects to Poolside through its OpenAI-compatible API. Run /connect, choose Poolside, and paste your API key from platform.poolside.ai. Self-hosted Poolside deployments can also be configured as a custom provider. See Providers and models.
  • Locale-aware pluralization — Counts, statuses, files, and session list items now pluralize correctly per language, so timeline and status text reads naturally across all supported locales.
  • Richer Markdown rendering — Assistant messages get improved Markdown formatting, smoother streaming updates, better syntax highlighting, and support for mathematical content, making long-form and technical replies easier to read.

​
Updates

  • GPT-5 reasoning effort matches model version — Reasoning effort tiers (including none and xhigh) are now offered only for the GPT-5 variants that support them, so --variant no longer surfaces options a given model would reject. See Controlling reasoning effort.
  • Diff annotation polish — The session UI diff view tightens up comment indicators, gutter behavior, and light/dark syntax colors for better readability.
  • Refined existing translations — Existing languages get clearer terminology and expanded UI labels for a more consistent experience across the app.

​
Bug fixes

  • Session action timing — Session actions now record and report their timing more accurately, so the timeline reflects when work actually started and finished.
  • Clearer remote workspace errors — Failures when connecting to or running commands in a remote workspace now surface more descriptive error messages instead of generic ones.
  • Workspace proxy no longer leaks working directory — Requests routed through the workspace proxy no longer include working-directory details, keeping local paths off the wire.
  • Map fills in empty regions — The activity map now stays visible in regions with no recorded activity instead of collapsing to a blank area.
  • Locale detection fixed — Bolt now picks up your system locale reliably at startup, so the interface opens in the right language on more platforms.
  • German timeline translation fixed — Timeline entries in German now render with the correct strings instead of falling back to English or misformatted text.
  • Toggle file selections off — Clicking a file selection a second time now clears it, so you can undo a selection without picking a different one first.
  • Cleaner draft comment interactions — Starting or dismissing a draft comment now clears the underlying selection consistently and no longer competes with hover state.
  • Animated reasoning header restored — The reasoning header in the TUI shows an animated spinner again while thinking, and its gradient dims when expanded so the collapsed and open states are easier to tell apart.
  • Background process and test tools restored — bash_background, process_output, list_processes, kill_process, and test_run are available to agents again after being dropped in an earlier merge. Agents can start long-running commands like dev servers and watchers, read their output, and run the project’s tests with structured pass/fail results. See Built-in tools.