v0.2.2

Sixteen PRs on top of v0.2.1, mostly polish for the first-run and local-model experience, plus honest CLI behavior. Nothing here changes config or breaks existing setups.

Highlights

A calmer first run

A fresh install used to greet new users with a red ○ down for a llama-server nobody had chosen to run. The default localModels.url ships with every install, so the very first health probe painted a failure on the one screen where a new user has the least context to judge it. Now the local health indicator only appears once local is actually your route, seeded from config at startup and latched on by a healthy probe. Somebody who runs llama-server on the default URL without touching config still gets the indicator, and keeps it when their server later dies. (#136)

The CLI answers instead of hanging

atomic-agent run --help, tui --help, and --version now do what you expect. run --help and tui --help print usage and exit, --version prints the version, and a bare help works as an alias. The repl scaffold no longer hijacks the terminal when you only asked it for help, and it stays out of the top-level --help listing until it is real. Exit codes are honest too: a stalled run fails instead of reporting success, an invalid --cwd is caught up front, and a skill lookup that finds nothing exits 1. (#143, #147, #145)

Local health that answers fast

When llama-server is unreachable, the agent used to hang before admitting it. Now /health answers in a single probe, llama-down failures come back quickly with a message that tells you what to do, and the whole path is faster and clearer end to end. (#140, #141)

Everything else

Added

• run --help, tui --help, --version, and a help alias for the CLI. (#143)
• Cloud providers can pass vendor-specific fields in the request body, so provider-only options no longer need a code change. (#149)
• Esc from an idle Manage panel returns you to the Run screen, and the hotkey hint strip stops repeating itself. (#148)

Fixed

• Fresh installs no longer report a llama-server that was never meant to run; the local health badge waits until local is actually the route. (#136)
• /health answers in one probe and admits when llama is down instead of hanging. (#140)
• Llama-down failures answer fast and say what to do next. (#141)
• Keyless local providers now pass the startup gate instead of being treated as unconfigured. (#123)
• The first keystroke after entering a panel no longer leaks into the chat buffer. (#139)
• Honest CLI exit codes: stalled runs fail, an invalid --cwd is rejected up front, and a lookup miss exits 1. (#145)
• config --help examples are generated from the live schema, so they cannot rot, with a test pinning it. (#146)
• The repl scaffold is hidden from --help, and repl --help answers instead of taking over the terminal. (#147)
• Non-interactive stdin gets one clear sentence instead of a stack trace. (#142)
• Model download progress feels alive, and a download already running is no longer offered a second time. (#137)
• The empty cloud-providers hint is bold instead of muted grey, so it reads as guidance rather than a disabled state. (#134)
• The install download meter collapses to a single blue progress line instead of scrolling. (#133)
• Model pricing in analytics is resolved from the provider catalog. (#138)

Install / Update

macOS / Linux:

curl -fsSL https://atomicagent.io/install | sh

Windows (PowerShell):

irm https://atomicagent.io/install.ps1 | iex

Atomic Agent updates itself in place. After an update the TUI prompts you to restart.

Full Changelog: v0.2.1...v0.2.2

Built by @ValeryB and @sosidudku1, reviewed by @Ooooze.