dsh-compat-probe

When a harness points at a local llama-server, the first real call can fail even though the server was started and the endpoint looks normal. The docs say one thing, the build does another. dsh-compat-probe sends five small fixture requests to an OpenAI-compatible endpoint and returns a protocol card with concrete booleans, so the model can see what the server actually supports before a larger task breaks.

The card covers the details that often escape documentation: whether the endpoint responds, whether GET /models succeeds, whether stream: true comes back as Server-Sent Events or is silently converted to one JSON completion, and whether the server accepts max_tokens, max_completion_tokens, both, or neither. Failures are returned as fields, not exceptions, so unreachable, DNS, timeout, HTTP error, and status details land in one stable shape.

It is designed for operators pairing DeepSeek Harness with dsh-llamacpp or another OpenAI-compatible server. It does not download weights, start processes, benchmark GPUs, test load, or speak Ollama. Its default-deny loopback guard also stops it from aiming at remote hosts unless the profile explicitly opts in, making it a safe local compatibility check for model agents.