Operations agent CLI

areno agent

Run a local operations assistant for AReno training and serving tasks. The agent uses an OpenAI-compatible chat model, inspects the current checkout, reads command help, runs diagnostics, and can produce or execute AReno commands for the current machine.

Configure the agent

Store the endpoint, model, and API key once:

areno agent --set \
  --base-url http://127.0.0.1:8000/v1 \
  --model deepseek-v4-flash \
  --api-key "$OPENAI_API_KEY"

The config is stored under ~/.areno. After this, normal agent runs do not need --base-url, --model, or --api-key on the command line.

Run an agent task

Pass the requested job as one natural-language argument:

areno agent "Give me a complete command to run the math demo with n-samples=8, fitting the current GPU and using as much GPU memory as practical."

The agent can inspect GPUs, read example files, run areno check and areno train --help, ask follow-up questions through the terminal when a required value is missing, and stream command output while it works.

From a source checkout, use the repository-local wrapper when AReno is not installed:

./agent.sh "Give me a complete command to run the math demo with n-samples=8, fitting the current GPU and using as much GPU memory as practical."

Refresh agent knowledge

The built-in operations knowledge tells the model how to reason about AReno train and serve commands, GPU memory, smoke checks, ModelScope defaults, and common recovery steps. Refresh the local copy when CLI behavior or examples change:

areno agent --refresh-knowledge

When to use it

Use areno agent when you want help choosing runnable train or serve parameters for the current machine, especially when GPU memory, tensor parallelism, dataset loaders, reward functions, or agentic rollout settings are unclear. For deterministic scripts and CI, prefer explicit areno train or areno serve commands after the agent has helped you settle on parameters.