Skip to content

v0.6.9

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Oct 02:16
· 7 commits to main since this release
v0.6.9
6d28652

The desktop app is updated too: Desktop v0.0.3. From this release on, the terminal agent and the desktop app share one setup.

One setup for the terminal and the desktop

  • The desktop app now uses the terminal agent's folder, ~/.atomic-agent. Chats, memory, skills, keys and settings are the same in both.
  • Both can run at once. Quitting one no longer stops a local model the other is using.
  • Scheduled tasks run once, in whichever app was started first. Telegram works the same way.
  • Settings saved in one app no longer overwrite a change the other made a moment earlier. A save that would have lost it is refused and retried.
  • New: atomic-agent import atomic-desktop --source <dir> brings chats, scheduled tasks and memory from an older desktop folder. The desktop app runs it for you on first start.

Faster context compaction

  • Compaction takes seconds. The agent keeps a running summary in the background, so it is ready when the context fills up, and a slow model no longer starts a chain of retries.
  • A paused task is summarised from its own records if the summary step fails.

Clearer failures

  • A failed turn says what went wrong in plain words, with the raw error on a Details line, and offers [try again].
  • Common provider HTTP statuses are put into words while the agent waits.
  • The Fallback pane says which failure caused a switch to the fallback model.
  • Fusion flags a worker that has run more than twice as long as expected.

Tools

  • New os.fs.move moves and renames files without the shell.
  • os.fs.grep keeps the matches it found when ripgrep times out, and os.fs.glob stops a walk that takes too long.
  • The Fusion planner can read the web (os.web.fetch, os.web.search). Reading a folder or a device such as /dev/null as a file now says what to do instead.
  • Files the agent attaches to a reply reach HTTP clients that ask for them, not only Telegram and Discord.

Local models

  • Atomic Chat Core 0.16.0. An existing install keeps working and switches over on its next idle start. Memory search by meaning keeps working on Core 0.13.6.
  • An embedding-only OpenAI-compatible provider no longer needs a chat model.
  • The reasoning budget is capped at what llama.cpp grammars accept.
Assets 10
Loading