Skip to content
← Back to the kit
Agent: OpenClaw Support: full

OpenClaw adapter

OpenClaw is an autonomous agent platform that runs long-lived workspaces with scheduled turns. It natively scans <workspace>/.agents/skills/ for installed skill packs, which means the DWP skill loads without any adapter shim — install it there and OpenClaw picks it up automatically.

Support level

Full — OpenClaw reads AGENTS.md and loads the skill via the AgentSkills standard. All DWP sub-skills (onboard, create, execute, refine, resume, status, verify) are available.

Installation

openclaw skills install deepworkplan

Or manually: clone DailybotHQ/deepworkplan-skill into <workspace>/.agents/skills/deepworkplan/.

Invocation

Plain text — OpenClaw does not use a slash-command prefix:

onboard this workspace with deepworkplan
create a plan for <goal>
resume the open plan

Unattended execution mapping

OpenClaw’s scheduling primitives map directly to DWP’s continuation model:

OpenClaw primitive DWP role
Workspace root Agent workspace archetype: AGENTS.md, .agents/, .dwp/ at the workspace root
<workspace>/.agents/skills/ (native tier-2 scan) Where this pack lives — no adapter shim needed
Heartbeat or cron turn One scheduled continuation turn: wake → DWP Resume Protocol → next atomic task → update state → yield
HEARTBEAT.md checks Add one line: “If .dwp/plans/ has an open plan, resume it for one task.”
Standing orders The plan-approval boundary and the bounded authority of the unattended profile

Operationally:

  1. A human creates and approves a plan interactively. Approval is the control point — unattended turns never create-and-execute in one turn.
  2. The plan must carry the machine-readable state layer (manifest.json + state.json). In a workspace without git, state.json is what makes crash-safe resume possible. See Plan state.
  3. Each heartbeat or cron turn executes at most one task, passes its validation gate, updates state.json atomically, and yields.
  4. On any stop condition the agent writes state.json’s blocked field and reports through the workspace’s normal channel. The next human glance — or heartbeat report — sees exactly what is needed.

The result: a multi-day plan that survives restarts, model changes, and session boundaries, executed overnight by the daemon — with the same gates a human-supervised run would have.

Notes

OpenClaw’s workspace is the canonical example of the agent workspace archetype. The DWP skill pack’s HEARTBEAT.md integration and state.json requirement were designed with this platform in mind.