python-endpoint

Example: Python endpoint adapter

Benchmark an agent written in any language by exposing an HTTP /step endpoint. The harness POSTs each observation as JSON and reads the returned action — your model weights and prompts stay entirely in your process. This is how closed, proprietary agents run against NEXBENCH.

Run it

# terminal 1 — start the adapter (stdlib only, no pip install)
python3 examples/python-endpoint/server.py

# terminal 2 — point the harness at it
nexbench run --agent http://localhost:8700/step

server.py solves NB-SEC-013 and NB-EXE-004 with fixed heuristics; the rest fall through to an empty submit. Replace decide() with your model's planning.

The protocol

• Request: POST body is the Observation (see docs/tutorial-build-an-agent.md).
• Response: JSON Action — rpc_call, sign_request, corpus_query, note, or submit.
• State: the endpoint gets obs.step and obs.last; reset per-trial memory when obs.step == 0 (a new trial has begun), since the harness runs the five trials in sequence.