How to report an issue

Good issue reports include enough environment, command, and task context for another person to reproduce the failure.

Include:

• The exact command.

• The relevant stack trace or error message.

• areno check output.

• areno env --json output.

• GPU type and count.

• PyTorch, CUDA, FlashAttention, and AReno versions.

• Whether ARENO_BUILD_EXT=0 was used.

• Dataset loader, reward function, and agent function paths when relevant.

• The smallest batch size or task seed that reproduces the issue.

For private data, replace examples with a minimal synthetic record that keeps the same shape and failure.