Goal: Bring OpenHands (formerly OpenDevin) to feature parity with Claude Code as a skill runtime. Grounds: RFC-AI-0004 Principle 3 — Vendor neutrality.
Why OpenHands: Fully open-source (MIT), backend-agnostic (works with any LLM), oriented toward more-autonomous execution (the "fully agentic" end of the spectrum). Adopters who want the framework's skill discipline applied to an autonomous-execution agent — for batch triage, overnight backlog sweeps, automation pipelines — need an OpenHands runtime.
What "parity" means (recalibrated for OpenHands' autonomy posture):
- Skills under
.claude/skills/<name>/SKILL.md are invokable as OpenHands microagents / instruction sets
- The
tools/* bridges are reachable — OpenHands runs commands in a Docker-sandboxed runtime
- Sandbox parity: OpenHands already ships a Docker-sandboxed execution model — this is a plus, not a gap. Document the alignment with the framework's
bubblewrap baseline (similar threat model, different mechanism)
- HITL primitives: OpenHands is the most autonomous of the named runtimes — this is the tightest fit/risk check. Document which framework skills are appropriate for autonomous execution and which require explicit confirmation pauses; the framework's "draft before send" rule (RFC-AI-0004 Principle 5) must be preserved even in autonomous mode
Suggested approach:
- Map
SKILL.md → OpenHands' .openhands/microagents/ convention (microagents are markdown files OpenHands loads as agent context)
- Document the autonomy-tier mapping: which skills can run end-to-end without confirmation (read-only triage), which require explicit pause points (any write to a public artefact, any outbound message)
- Validate the autonomous-batch use case against the eval suite — autonomous execution + HITL discipline is the interesting design tension this port surfaces
Reference:
Goal: Bring OpenHands (formerly OpenDevin) to feature parity with Claude Code as a skill runtime. Grounds: RFC-AI-0004 Principle 3 — Vendor neutrality.
Why OpenHands: Fully open-source (MIT), backend-agnostic (works with any LLM), oriented toward more-autonomous execution (the "fully agentic" end of the spectrum). Adopters who want the framework's skill discipline applied to an autonomous-execution agent — for batch triage, overnight backlog sweeps, automation pipelines — need an OpenHands runtime.
What "parity" means (recalibrated for OpenHands' autonomy posture):
.claude/skills/<name>/SKILL.mdare invokable as OpenHands microagents / instruction setstools/*bridges are reachable — OpenHands runs commands in a Docker-sandboxed runtimebubblewrapbaseline (similar threat model, different mechanism)Suggested approach:
SKILL.md→ OpenHands'.openhands/microagents/convention (microagents are markdown files OpenHands loads as agent context)Reference:
docs/rfcs/RFC-AI-0004.md(especially Principle 5 — write access and outbound messages require human review).claude/skills/