Agent Integration & Skills

CLI does not replace the agent. It gives the agent a stable command entrypoint.

HENGSHI CLI works for direct human use as well as agent integration. This page focuses on the second case: how to make CLI execution more stable and reviewable when you integrate it into OpenClaw or another agent runtime.

The recommended mental model looks like this:

Layer	Purpose
Human	Provide intent, review results, confirm risk boundaries
Agent	Understand the request, break down steps, decide what capability to call
Skills	Encode domain knowledge into reusable standard workflows
HENGSHI CLI	Execute reads, previews, creates, updates, and authorization actions

Why agents should go through CLI

• Stable structured output: json / yaml works well for machine read/write paths
• Clear command boundaries: safer than letting the agent assemble private APIs
• Preview support: risky writes can go through --dry-run first
• Reusable everywhere: local terminals, CI, automation jobs, and agent runtimes share the same execution surface
• Auditable execution: the chain is easier to log, replay, and review

What official skills are for

HENGSHI CLI is delivered together with official bundled skills, including:

• hbi-core
• hbi-data
• hbi-dashboard
• hbi-permission
• hbi-workflow

The main value of skills is to encode:

1. What to read before writing
2. Which actions must go through --dry-run
3. How results should be structured for humans or downstream systems

Recommended integration steps

1. Ensure the runtime is provisioned with the CLI

See Installation & Upgrade for the access path and post-install checks. For team or customer-facing agent runtimes, it is usually better to provision CLI through the HENGSHI delivery channel instead of asking each runtime owner to discover a download path separately.

2. Pin the instance and authentication

export HBI_API_URL="https://<your-everest-instance>"

See Authentication & Connection for the available auth paths. For agent runtimes, it is usually better to pin HBI_CLIENT_ID / HBI_CLIENT_SECRET or HBI_TOKEN explicitly instead of depending on ad-hoc credential input.

3. Provision the bundled official skills

Official skills are now delivered together with the CLI package, so a separate pull from an external distribution surface is no longer the recommended path. If you are preparing runtimes for OpenClaw, Claude Code, GitHub Copilot, or similar agents, specify the target agents during delivery so the bundled skills land in the correct global skill directories from the start.

4. Constrain the agent’s default execution strategy

These rules should usually live in the agent prompt or runtime policy:

1. Read state before writing
2. Use --dry-run whenever possible
3. Prefer --output json or --output yaml
4. Do not assemble private HENGSHI APIs manually
5. For risky grants, deletes, or config changes, return the plan to a human before execution

A typical agent execution pattern

For example, if an agent is asked to “create a sales dashboard and grant access to a teammate”, the recommended order is usually:

1. Read app and dataset context
2. Confirm whether the target dashboard already exists
3. Create the dashboard
4. Add charts or other elements
5. Preview the authorization with --dry-run
6. Execute the real grant after human confirmation

This is exactly the kind of sequencing that skills are good at capturing.

Recommended practice

If you plan to make CLI a formal part of an agent runtime, these practices usually help the most:

1. Pin the CLI version and the bundled-skill delivery path across environments
2. Prefer json or yaml output by default to reduce downstream parsing cost
3. Keep --dry-run in front of risky grants, deletes, and system configuration changes
4. Log HBI_API_URL, the auth path, and the key command results for later review

What agents should return

When an agent uses CLI, ask it to summarize results in at least three buckets:

• What it did: which reads, previews, and writes were executed
• What it saw: what the important commands returned
• What is still missing: which steps did not continue because of permissions, missing context, or human confirmation

That gives teams something far more useful than a raw wall of terminal output.