swe-agent-mini-swe-agent

sh -c export HOME=/tmp MSWEA_CONFIGURED=true MSWEA_COST_TRACKING=ignore_errors && 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 mini -y --exit-immediately -l 0 -o /tmp/mini.traj.json -m openai/MODEL_PLACEHOLDER -t "$(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

Flags read from `mini --help` in the built image, at version 2.4.6. The agent executes shell commands in its working directory, so the command changes to /work first and the local environment acts on the shared workspace. -y is yolo mode, which runs each command without asking; --exit-immediately stops it prompting a person when it decides it has finished. Two environment variables are load-bearing and both were found by running it. MSWEA_CONFIGURED skips a first-run setup that otherwise opens an interactive prompt for a model name and an API key, which under the sandbox would hang until the wall clock killed it. MSWEA_COST_TRACKING=ignore_errors is needed because the model reply arrives successfully and the run then dies in litellm's cost calculation, which has no price entry for a local model; without it the agent reaches the model, does nothing, and would be recorded as an agent that does not work. -l 0 disables the cost limit for the same reason. The trajectory is written to /tmp rather than the default under HOME, and HOME itself is moved to /tmp, because the image's root is read only and neither may land in the shared workspace. The model name carries litellm's openai/ provider prefix, without which the local endpoint is not selected. engagement_markers names the header of the observation block the harness prints after it executes a command, which is the line that means a tool acted rather than the model proposing one.