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:
- A human creates and approves a plan interactively. Approval is the control point — unattended turns never create-and-execute in one turn.
- The plan must carry the machine-readable state layer (
manifest.json+state.json). In a workspace without git,state.jsonis what makes crash-safe resume possible. See Plan state. - Each heartbeat or cron turn executes at most one task, passes its validation gate, updates
state.jsonatomically, and yields. - On any stop condition the agent writes
state.json’sblockedfield 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.