Harnesses / hx-medgauge-driver

MedGauge reference driver (containerised tool loop) Active

Vendor-neutral driver shipped with every task package. Shell commands run inside the task container. Speaks a chat-completions-compatible interface and a messages-style interface with the same six tools. Each run writes a run record (model, interface, steps and step limit, tool-call counts, tokens, timing, temperature, system-prompt hash, whether inputs were modified, grade). The second-environment reproduction (env_delta) has not been run yet.

Id and version

[email protected]

Base runtime

medgauge-driver 0.1.0

Model interface

Chat-completions-compatible API · Messages-style API

Model slots

executor: any model reachable through the interface

Tools

[email protected][email protected][email protected][email protected][email protected][email protected]

Limits

max_steps=100 max_output_tokens=65536 temperature=0.2 timeout_sec=3600

Sandbox

One container per task · network disabled · readonly inputs

Autonomy

A3

Capabilities targeted

B.3B.4X.2X.4

Reproduction

run_agent.py --task <task-dir> --model <model> --max-steps 100 --temperature 0.2

Adapter check

reference solution full score: yes (2026-09-16) · env_delta: not yet measured

Maintainer

MedGauge

Licence

Apache-2.0

Versions

• 0.1.0 Active

Results with this harness