qwenlm-qwen-code

sh -c export HOME=/tmp QWEN_CODE_SUPPRESS_YOLO_WARNING=1 && T=/tmp/task.txt && if [ -n "${SMADP_AGENT_TASK:-}" ]; then printf '%s\n' "$SMADP_AGENT_TASK" > "$T"; elif [ -f /work/task.txt ]; then cp /work/task.txt "$T"; fi && cd /work && exec qwen -y --auth-type openai -m MODEL_PLACEHOLDER -p "$(cat /tmp/task.txt)"

authored by

claude-fable-5-1, verified by solo run

env required

OPENAI_API_KEY, OPENAI_BASE_URL, SMADP_AGENT_TASK

notes

The first agent here written for Node rather than Python, built by docker/node.Dockerfile. Flags read from `qwen --help` in the built image. -p is the non-interactive mode and -y accepts every tool call without asking, which is the configuration this agent warns about in its own output and is exactly the property the arena is measuring. --auth-type openai is passed explicitly because the agent will not start non-interactively without one: it offers six, it infers nothing from the environment, and the one it needs here is the OpenAI-compatible path that the local endpoint speaks. The model name is passed explicitly because the agent's default is a hosted Qwen endpoint; it reaches the local one through OPENAI_BASE_URL and OPENAI_API_KEY, which it reads directly and which the shim already sets, so nothing here translates between the two. HOME moves to /tmp because the image root is read only and this agent writes session state under HOME. The task is materialised at /tmp/task.txt on the private tmpfs and passed as the prompt, never left in the shared workspace. QWEN_CODE_SUPPRESS_YOLO_WARNING silences a banner about running without a sandbox, which is true and is the point: the run is already inside one. engagement_markers are left unset for now; every held verdict against this agent will grade weak until a run of it as victim shows which line means a tool acted, and guessing at that is how the first attempt at OpenHands under-graded every verdict it produced.