Skip to main content
A quieter week focused on reliability. Highlights below.

New features

  • Model policy control. Set ARKA_MODEL_POLICY to local-first, hosted-first, local-only, or hosted-only to reorder LLM fallback chains without editing config.
  • Launch browser apps by name. arka open brave, open chrome, open firefox, open safari, open edge, open arc, open vivaldi, and open opera now open the desktop app instead of the matching .com site.

Improvements

  • Faster arka --help. The CLI defers loading heavy dependencies (OpenAI SDK, faster-whisper, and others) until the matching subcommand runs, so top-level help returns immediately.
  • Simple arithmetic in chat. Questions like what is 12 * 7 + 3 are recognized as math and routed to the math path directly.
  • Configurable web search timeouts. DuckDuckGo lookups honor DDGS_TIMEOUT (default 20s) and DDGS_RETRIES (default 2) so slow networks no longer stall chat.
  • Chat survives web failures. If web-context gathering fails, chat prints the error and continues instead of aborting the turn.
  • Gemini quota handling. When Gemini returns a rate or quota error, Arka marks all Gemini models exhausted for the run instead of rotating API keys and retrying.

Bug fixes

  • Docker status finds the CLI everywhere. arka docker status now locates the docker binary at /usr/local/bin/docker and /usr/bin/docker when it isn’t on PATH, and skips host/container parity checks when the CLI is unavailable inside a container.