OCM Operator

Operating Model

Use OCM as the control plane for OpenClaw. Treat an OCM environment as the isolation boundary: it owns OPENCLAW_HOME, state, config, workspace, gateway port, runtime/launcher binding, and service policy.

Prefer OCM commands over raw OpenClaw commands whenever the task touches:

• env creation, cloning, migration, export/import, snapshots, or cleanup
• runtime install/build/verify/update/remove
• upgrades or release testing
• background gateway services
• logs, status, TUI/dashboard, plugin checks, model checks

Run OpenClaw inside an env with:

ocm @<env> -- <openclaw args>

Use ocm env exec <env> -- <command> only when you need a non-OpenClaw command inside the env, such as env, node, or shell tooling.

First Steps

1. Inspect current state before changing it:

ocm env list
ocm service status
ocm runtime list

2. Read help for the exact surface before using unfamiliar flags:

ocm help start
ocm help env
ocm help runtime
ocm help service
ocm help logs
ocm help upgrade

3. For detailed command patterns, read references/command-cookbook.md.

4. For destructive operations, service boundaries, or real user state, read references/safety-and-state.md before acting.

5. For this machine's common paths and release-test defaults, read references/local-paths.md.

Decision Guide

• Start a normal runtime-backed env: use ocm start <env> --runtime <runtime> or ocm start <env> --version <version>.
• Start a fresh user env: use ocm start <env> --runtime <runtime> first; use --onboard only when testing interactive onboarding specifically.
• Use source-backed OpenClaw development: use ocm dev <env> --repo <repo>. Do not use this as proof of release packaging.
• Test release behavior from local OpenClaw source: use ocm runtime build-local <runtime> --repo <repo> --force, then test envs against that runtime.
• Test existing-user upgrade behavior: clone the existing env, upgrade the clone, and never mutate the source env directly.
• Run OpenClaw commands in an env: use ocm @<env> -- ....
• Read logs: use ocm logs <env>.
• Manage background gateway state: use ocm service status/start/stop/restart.
• Clean up: use ocm env destroy <env> --yes; this removes env snapshots, OCM service policy, and OCM-created dev worktrees.

Safety Rules

• Treat important existing user envs as durable state. Clone first for testing unless the user explicitly asks to mutate them.
• Clones retain durable auth and settings. Treat them as secret-bearing, keep services stopped by default, and require explicit authorization before provider, channel, webhook, browser, or other external calls.
• Use a credential-free fresh env for untrusted plugins or authentication tests. Never export, publish, or attach an unsanitized clone.
• Remove envs with ocm env destroy.
• Manage OCM env gateway lifecycle with ocm service.
• Remember OCM injects OPENCLAW_SERVICE_REPAIR_POLICY=external for env-run OpenClaw processes, so openclaw doctor --fix should repair normal env state but skip OpenClaw-owned service lifecycle repair.
• Do not treat ocm dev, pnpm openclaw, or a source worktree as release validation. For release behavior, build and use a package runtime.
• Clean up temporary envs, services, worktrees, tarballs, and retained simulation envs.
• If a command could change real user data, run read-only inspection or clone first, then report the exact command you plan to run if there is ambiguity.

Verification Standard

Finish by exercising the relevant command path. Prefer these checks:

ocm env list
ocm service status <env>
ocm logs <env> --tail 100
ocm @<env> -- --version
ocm @<env> -- status
ocm @<env> -- doctor --fix

Add plugin/model/TUI checks when the task touches plugins, model auth/catalog, onboarding, or gateway connectivity:

ocm @<env> -- plugins list
ocm @<env> -- plugins update --all --dry-run
ocm @<env> -- models list
ocm @<env> -- tui

Reporting

Keep reports quiet and useful:

• lead with pass/fail/blocker status
• include the exact command for failures
• include concise evidence for passes
• separate OCM issues from OpenClaw runtime/package issues
• say which relevant checks were skipped
• include cleanup performed

Use this failure shape:

Failure:
- Command: `<exact command>`
- Expected: <expected state>
- Actual: <observed state/output>
- Impact: <why it matters>
- Likely owner: OCM | OpenClaw | test setup | external dependency
- Paste to fixer: <short actionable summary>