From a domain plan to authorized computer work

Phase 6 organizes domain policy and capacity. The existing computer-work pipeline executes separately admitted tasks through an operator-configured OpenClaw Responses gateway. An operator can also use ChatGPT Work to perform a scoped task and export evidence for independent review. This demo does not provide a remote ChatGPT Work dispatch API.

Choose a work surface

Surface	How to use it	Evidence to retain
OpenClaw with native Codex Computer Use	Commission the actual gateway, desktop plugin, OS identity, app access and task policy	Exact runtime/plugin versions, live desktop readiness result, admitted task digest, session receipt and artifacts
OpenClaw managed browser / approved integrations	Use isolated profiles and the tools appropriate to the job	Allowed origins, provider configuration, exact input/output and observed application state
ChatGPT Work Computer Use	An authorized operator performs the task with Work's app and OS permissions	Original task, permitted actions, final artifacts and a review of actual effects
Custom OpenAI API worker	Implement and commission an isolated tool-execution loop behind an approved endpoint	Runtime enforcement, tool observations, limits, recovery tests and artifact integrity

Official guidance checked 2026-10-07: OpenClaw Codex Computer Use, OpenClaw Responses endpoint, OpenClaw sandboxing, ChatGPT Work Computer Use, and OpenAI API computer use.

The current OpenClaw desktop plugin distinguishes installed tools from a successful live desktop check. For turns requiring desktop access, deliberately configure and test plugins.entries.codex.config.computerUse.strictReadiness: true; its default is false. Use /codex computer-use status in an OpenClaw chat surface to check the actual desktop. The gateway Responses endpoint is separately enabled and authenticated. OpenClaw sandboxing is not enabled by default, and a new conversation is not an isolation boundary.

Work supports scoped graphical tasks on supported macOS/Windows setups. App approvals and OS permissions are separate controls; a permitted origin or application does not authorize every consequential action inside it. Prefer structured integrations when they suit the task. For a custom API implementation, current guidance supports code execution in a controlled runtime and structured computer actions. Your application must execute those calls, preserve the session and enforce policy; the model alone does none of that enforcement.

A complete first handoff

Use the public cross-domain brief task to test the boundary without private data or writes. It requests a sourced comparison of the five Phase 6 profiles. It authorizes no login, messaging, trading, payments, publishing or infrastructure changes.

npm run build:orchestrator
node demo/One-Box/computer-work/run.cjs inspect \
  demo/Phase-6-Scaling-Multi-Domain-Expansion/examples/cross-domain-task.json

Inspection produces a normalized task and digest; it is not admission or dispatch. Choose a commissioned profile in protected operator configuration, review the exact task digest and admit it to a specific job. Follow the canonical commissioning procedure. The example profile name must correspond to a profile you commission; a label is not proof of capability.

Retain the original task, provider/session metadata, candidate artifacts, checksums and independently checked findings. Current repository receipts support bounded UTF-8 JSON, CSV, Markdown and plain text artifacts. Screenshots, Office files and other binary evidence need a separately commissioned artifact store; do not relabel unsupported binaries as text. Provider success and matching hashes do not establish task correctness or authenticate a worker.

Use the evidence reviewer for local integrity and acceptance findings. Its exports are unsigned review records, not settlement authorization. Marketplace execution still requires the existing identity, stake, validation, dispute and finalized settlement lifecycle.

Domain-specific acceptance examples

Domain	Appropriate scoped first job	Independent acceptance
Finance	Reconcile synthetic ledger entries and explain discrepancies	Row counts, exact arithmetic, source references; no trading or funds movement
Health	Compare public service documentation or synthetic administrative records	Source accuracy, privacy rules and qualified review; no diagnosis or treatment decisions
Logistics	Prepare a shipment exception report from synthetic records	Identifier reconciliation, timing calculations and operator review before any rerouting
Climate	Check public dataset provenance and prepare an analysis brief	Reproducible calculations, units, uncertainty and traceable inputs
Education	Produce an accessible learning draft using permitted materials	Factual review, source rights, accessibility and instructor acceptance

The same interfaces can support research, software QA, document/spreadsheet preparation, design tools and many operational workflows when properly commissioned. Success is measured per task. Expanding domains means adding evidence-backed capability and review capacity, not inferring competence from a generic desktop connection.

Before admitting a consequential task

Define exact origins/apps, least-privilege accounts, data rights, output locations, action and spending limits, stop controls and independent review. Treat web pages and documents as untrusted inputs. Test wrong accounts, redirects, malicious instructions, downloads, app crashes, timeouts and duplicate requests. A timeout can follow a completed external action; preserve journals and reconcile before authorizing a new attempt. Keep approval for external messages, publication, funds or destructive changes explicit and scoped to the task.