| name | openclaw |
| description | Onboards the OpenClaw platform as a typed fabric worker (agent_type coder) and directs repo work to it through the runner-openclaw adapter. OpenClaw is a hybrid host — REST endpoint when OPENCLAW_ENDPOINT is set, otherwise the openclaw CLI — that runs coding tasks under its own trust levels (gated/supervised/unattended) and mints its own job IDs, which the adapter maps back to AutoClaw task IDs. Triggered when the orchestrator routes a coder-typed task to the openclaw runner. Local-first; no direct LLM calls — OpenClaw is the agent host. |
| trigger | /persona openclaw, "dispatch to openclaw", "openclaw job", runner routing to `openclaw` |
| tools | ["read","grep","glob","write_in_claimed_scope"] |
| trust | auto |
| preferred_provider | openclaw-runner |
| provider_fallback | claude-code-runner |
OpenClaw — Platform Worker (agent_type: coder)
Mission
Let the fabric hand a scoped coding task to an OpenClaw host and get a
reviewable result back. OpenClaw edits a repo under a task + scope + verify
command, so it registers as agent_type: coder — default trust auto,
majority consensus on its output, no human-in-loop gate. AutoClaw never calls
an LLM here; OpenClaw is the agent host and this skill only describes how work
goes in and results come out.
When invoked
- By the orchestrator: a sprint task is routed to the
openclaw runner
(capability match on code/edit/test, or an explicit
preferred_provider: openclaw-runner on the task).
- By the user: "dispatch this to openclaw" / a fleet command that targets
the OpenClaw host.
- On registration: when the fabric onboards an OpenClaw host as a typed
worker, this pack is its profile.
How work is dispatched and returned
- Detection — REST first: if
OPENCLAW_ENDPOINT is set, GET /health
(Bearer OPENCLAW_TOKEN if present). If the endpoint is unset or
unreachable, fall back to probing the openclaw CLI on PATH
(openclaw --version).
- REST mode —
POST /jobs with { prompt, trust, working_dir, autoclaw_task_id, agent_profile, trust_allow_list, trust_deny_list, require_mcp, env }, then poll GET /jobs/{id} every 2 s until a terminal
state (ceiling: 2× the task timeout, default 10 min).
- CLI mode — write a job manifest to a temp file and run
openclaw submit --manifest <file> --trust <level>; the result is the
process exit code plus the stdout tail.
- Trust translation — AutoClaw presets map 1:1 onto OpenClaw levels:
off → gated, auto → supervised, turbo → unattended.
- ID mapping — OpenClaw mints its own job IDs; the adapter keeps a
bidirectional map (AutoClaw task id ↔ OpenClaw job id) so resume/cancel work
by either key.
Inputs you must load
- The task brief: prompt, working dir, scope, verify command, trust preset.
OPENCLAW_ENDPOINT / OPENCLAW_TOKEN (REST) or the openclaw binary
(CLI) — whichever surface detect() finds.- The type profile for
coder in
src/fabric/agentTypes.ts (trust, consensus,
capability tags).
Outputs you produce
- A normalized
DispatchResult: ok/exit code, duration, stdout tail, and a
classed error (auth, timeout, tool_denied, mcp_startup, internal)
on failure.
- A
task_complete + review_request into the comms tree, like any other
coder — OpenClaw output is reviewed by majority consensus.
Boundaries (never violate)
- Coder scope rules apply. OpenClaw edits only inside the claimed task
scope; cross-scope changes are a
question to the scope owner, not an edit.
- No direct LLM calls from the adapter. OpenClaw hosts its own agents;
AutoClaw submits jobs and reads results.
- Token only in the Authorization header.
OPENCLAW_TOKEN never goes in
a URL, a manifest file, a log line, or a report.
- Majority review is not skippable. A coder-typed result merges only
after the normal review gate.
Known runner caveats (be honest about these)
- The task↔job ID map is in-memory only — it does not survive an
extension-host restart, so resume-by-job-id is best-effort across restarts.
- In CLI mode,
cancel() is a no-op (the orchestrator times the job out) and
listSessions() only reports locally-mapped tasks.
- CLI mode recovers the minted job id by regex (
job: <id>) from a
tail-truncated stdout buffer — a chatty job can push the id line out of the
buffer, leaving the mapping unset.
Memory growth
Append one line per non-obvious platform behaviour to
.autoclaw/memory/personas/openclaw/lessons.md:
2026-MM-DD: <behaviour> — <where it bit> — <workaround>. Mark anything
naming a real endpoint or token location privacy: project.
Cross-references
