Atlarix CLI 15.0.1: Sub-Agents in the Terminal, Coloured Code and a Plan You Can Read

Sub-agents are how an agent explores in parallel: one reads the tests while another traces a function. In 15.0.0 every sub-agent the CLI started failed before it ran and then showed as running forever. Two separate bugs were behind it, and a third turned up while fixing them: outside Atlarix's own repository the CLI could not find the prompts it ships with, so every turn quietly ran without its mode's instructions and every sub-agent failed to start. All three are fixed, with tests that run the CLI from a folder that is not ours.

While Atlarix works, the space under the prompt box now lists what is running: the main agent and what it is doing, then one row per sub-agent with its current step, how long it has run and its tokens. Finished ones drop off, and background commands share one row. Press ↓ and enter on a sub-agent to watch it live, its tool calls and what it is writing; ← and → step between sub-agents and esc backs out.

Answers and edit diffs are now syntax highlighted, and inline code stands out in lavender rather than a dim grey chip. The interesting part is how languages are handled. Atlarix does not ship a list of grammars. The first time a language shows up, its grammar is found by name: one already on your machine from another terminal app built on the same toolkit is reused, otherwise the language's own grammar project is fetched once. Only the languages you read ever land on disk, and offline the code simply stays plain.

Plan mode used to end with a popup that asked you to approve a plan you could not see. Now the plan stays pinned above the prompt box while it waits to be built. You read it, type what to change underneath, and watch it redraw as Atlarix revises it. Build it, or /build, starts the work; esc hides the plan and /plan brings it back.

The first message of a session used to freeze the terminal for about a second before it appeared. The cause was token counting, which rebuilt its tokenizer every time it counted, dozens of times on a first turn. It is built once now. The desktop app got faster from the same fix.

• The context meter shows how full the context is as a share and as tokens of the model's window, instead of a running total that could pass the window and look like a contradiction.
• A message you send while Atlarix works is drawn where the agent read it, not under the whole reply.
• /compact typed mid-answer waits for the answer to finish, and messages typed after it still go first.
• Pasting the same text again expands the pasted-text chip so you can edit it.
• The approval dialog's diff is coloured like the transcript's.
• The terminal tab pulses while Atlarix works, so a busy session is visible from another tab.
• Long session names are shortened, the prompt box sits a step above the page, and checklists read on one line.

If you already have the CLI it updates itself. Otherwise: npm install -g atlarix, brew install amariahak/atlarix/atlarix, or the one-line installer on atlarix.dev/cli. The CLI shares the desktop app's sessions, models, keys and MCP servers, and runs on your own API keys, local models or Atlarix Core.