openhands-coding-agent

name	openhands-coding-agent
description	Run OpenHands headless CLI/SDK missions from BioKernel swarms for autonomous software work.
keywords	["openhands","coding-agent","cli","sdk","headless"]
metadata	{"upstream_repo":"https://github.com/OpenHands/OpenHands","version":"2026-04-02"}
allowed-tools	["run_shell_command","python3","read_file"]

OpenHands Coding Agent Skill

OpenHands (formerly OpenDevin) is a 70k+ star open-source coding agent stack that exposes the same engine through a Python SDK, a CLI, a local GUI, a hosted cloud workspace, and a self-hosted enterprise edition.¹ The CLI ships with a headless mode so you can script missions, capture JSONL traces, and hand progress back to other agents.²

When to Trigger This Skill

• You need a persistent IDE-grade coding agent that can run shells, editors, and browsers against your repo while other biomedical agents handle research/safety steps.
• You want to re-use OpenHands’ plan/act/react loop instead of building a sandbox from scratch.
• You must export JSONL traces for downstream reviewers or for ingestion into platform/compliance/ dashboards.

Prerequisites

1. Install OpenHands (CLI bundles a uv launcher):

   uv tool install openhands --python 3.12
   # or pip install openhands
   

2. Provide an LLM key via env (Claude, GPT-4.1, Gemini, Qwen, etc.). You can also pass them via --env KEY=value when using the runner below:

   export OPENAI_API_KEY=sk-...
   

3. Ensure your repo mount has git + docker access if you want containerized sandboxes; otherwise use the lightweight local mode.

Workflow

1. Define the mission text that OpenHands should solve (e.g., “Triage KRAS G12D tox notebook failures”).
2. Run headless CLI so the agent streams work to stdout and (optionally) JSONL. Use the helper below or invoke the CLI directly:

   # direct
   openhands --headless -t "${MISSION}" --json > traces/${RUN_ID}.jsonl
   
   # helper script (handles env vars and tee'ing output)
   python Skills/Agentic_AI/OpenHands_Coding_Agent/openhands_runner.py \
     --mission "${MISSION}" \
     --json-out Skills/Agentic_AI/OpenHands_Coding_Agent/runs/${RUN_ID}.jsonl \
     --env OPENAI_API_KEY=sk-xxx CLAUDE_API_KEY=sk-ant-yyy
   

3. Optional task files: place a multi-step YAML/Markdown brief in openhands_task.md and run openhands --headless -f openhands_task.md.
4. Synthesize results: parse the JSONL (each event has role, tool, content) and feed summaries back into Multi_Agent_Systems/orchestrator.py as part of the Reviewer/Safety checkpoints.
5. Escalate/stop: send SIGINT (Ctrl+C) to halt the agent; it checkpoints the shell/IDE history so you can resume the same run ID later.

Swarm Handoff Pattern

Kick-off from orchestrator: have the Overmind delegate coding subtasks by writing MISSION.txt, then invoke OpenHands via subprocess.run (or manually) using the command above.

Return artifacts: drop patch files, terminal transcripts, and the JSON trace under Skills/Agentic_AI/OpenHands_Coding_Agent/runs/<timestamp>/ so other agents (Reviewer, SafetyOfficer) can inspect diffs.

Assets

File	Purpose
openhands_runner.py	Wrapper that shells out to the OpenHands CLI with --headless, --json, optional task files, env injection, and JSONL teeing.

Safety hooks: tail the JSONL stream to automatically look for PHI/PII before allowing patches to merge.

References

1. GitHub – OpenHands/OpenHands (README outlines SDK, CLI, GUI, Cloud, Enterprise surfaces). https://github.com/OpenHands/OpenHands
2. GitHub – OpenHands CLI headless usage (openhands --headless -t ... --json). https://github.com/OpenHands/OpenHands?tab=readme-ov-file#openhands-cli