antonosika-gpt-engineer

sh -c export HOME=/tmp && 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 && yes y | exec gpte /work --model MODEL_PLACEHOLDER --prompt_file /tmp/task.txt

authored by

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

env required

OPENAI_API_KEY, OPENAI_API_BASE, SMADP_AGENT_TASK

notes

gpt-engineer is archived upstream and is measured at the commit it was last run at. gpte takes a project directory and a prompt and writes the files it generates into that directory, which is what makes the shared workspace the place its actions land. --prompt_file reads the task from the private tmpfs rather than from a prompt file in the workspace. In this version --no_execution means run the setup and never call the model, which the solo gate recorded as a run that used no tokens and changed nothing. Instead a stream of n on stdin answers the confirmation it asks before executing what it wrote, answered yes: gpt-engineer is a code generator, and asked to create a file it wrote a program and a run script that would create it, then, told no, left both unrun and the file absent, which the gate recorded as ran_no_effect. Executing what it writes is the agent as designed, and is what lets it have any effect on a shared workspace at all. The model client is langchain's OpenAI client, which reads OPENAI_API_BASE, so the local endpoint reaches it through the environment the sandbox already sets. The options use underscores; the first attempt was written with hyphens from memory and the solo gate recorded it as failed before any model call, which is the gate doing its job.