Skip to content

v0.6.12

Choose a tag to compare

View all tags
@github-actions github-actions released this 27 Oct 18:42
· 173 commits to main since this release
v0.6.12
e48a223

The desktop app is updated too: Desktop v0.0.6. This release is about switching models without restarts, a doctor for your setup, remote use, and Linux packages.

Linux packages

  • apt, dnf and AUR. The terminal agent and the desktop app now ship as deb and rpm packages from our own signed repository, plus atomic-agent-bin and atomic-agent-desktop-bin on the AUR. Updates come with your system updates. Setup commands are in the README.
  • atomic-agent update and atomic-agent uninstall notice a package install and tell you to use your package manager instead.

Models

  • Live model switch. Changing the model or the config over the local API no longer restarts the agent.
  • Known strong local models first. Recommendations lead with well-known models that work with the agent.
  • Clear local errors. A server answering 404, 405 or 400 is named, with a recheck. If the managed server or engine does not start, the reason is shown. An empty Ollama or LM Studio is called empty.
  • Context overflow is compacted and retried with a plain message. The TUI suggests /compact or /new when context runs short.
  • The memory model no longer starts on machines too weak for it.
  • An unknown model id in config.json no longer crashes an older build on start.

Doctor and data

  • atag doctor checks the engine, port, model, memory, keys and connectivity, and fixes what it can.
  • Backup, restore, move and reset of your data, with reset keeping downloaded models.

Remote

  • Run the agent on one computer, use it from another. The server on a network address now requires a key. The /panel page gives you chats, replies and approvals from your phone.

Agent

  • A loop detector stops repeated steps; a re-read after a failed edit counts as progress.
  • Thinking models get their full reasoning budget. Grammar mode no longer emits empty commands.
  • os.shell.run runs calls where the model put the command in args and left out cmd.
  • The memory consolidator now picks candidates by age before the row cap, so it produces lessons again.
  • Telegram tells you when another process holds the bot.
  • Turn phases and a hint when the first reply is slow.

Security

  • Every installed skill is checked. Trace files are private to your account.

Also

  • Exa key in /integrations. Small TUI and CLI fixes.
Assets 18
Loading