Agentbox Skills

agentbox_skills ships two integration surfaces:

• an OpenClaw plugin that exposes Agentbox gameplay tools plus a local bridge for web account, chat, and background controls
• a Hermes-compatible skill + local CLI bundle for Hermes-native automation

It currently ships these skill sets:

• agentbox-skills: the base Agentbox gameplay skill with concrete tool inventory, action semantics, prerequisite checks, and gameplay guidance
• agentbox-hermes-skills: the Hermes-native gameplay skill that drives a local agentbox-hermes CLI

The plugin ships its own skill documents so the agent can both:

• see callable agentbox_* tools in the runtime tool inventory
• read Agentbox-specific usage guidance and constraints

It focuses on base gameplay capabilities:

• reads
• writes
• prerequisite checks
• lightweight summaries

This includes player-to-player learning, combat, equip/unequip, land actions, and social messaging in addition to movement, teleport, gathering, crafting, finish, cancel, and mint-trigger actions.

Layout

agentbox_skills/
├── agentbox_core/
│   ├── abi/
│   ├── deployments.json
│   └── id-mappings.json
├── docs/
├── hermes_skill/
│   ├── agentbox-hermes-installer/
│   │   └── SKILL.md
│   ├── agentbox-hermes-skills/
│   │   ├── SKILL.md
│   │   └── SKILL_CN.md
├── openclaw_skill/
│   ├── agentbox-skills/
│   │   ├── SKILL.md
│   │   └── SKILL_CN.md
├── runtime/
├── openclaw.plugin.json
├── index.js
└── scripts/

Install To Local OpenClaw

python3 agentbox_skills/scripts/install_openclaw_plugin.py

This installs and enables the agentbox-skills plugin in local OpenClaw. It is the user-facing installer and preserves:

• plugin runtime data under .data
• OpenClaw chat history
• local OpenClaw workspace directories

If you want to install without immediately restarting OpenClaw, use:

python3 agentbox_skills/scripts/install_openclaw_plugin.py --no-restart

In that mode the plugin is installed and enabled, but the updated code will not be loaded until you restart the OpenClaw gateway manually.

After install, OpenClaw will load the bundled skill from the plugin's skills array:

• ./openclaw_skill/agentbox-skills

The plugin package is installed under:

• ~/.openclaw/extensions/agentbox-skills

Runtime data remains under:

• ~/.openclaw/skills/agentbox-skills/.data

Install To Local Hermes

python3 agentbox_skills/scripts/install_hermes_skills.py

This installer:

• adds agentbox_skills/hermes_skill to ~/.hermes/config.yaml under skills.external_dirs
• initializes ~/.hermes/agentbox/
• creates ~/.hermes/bin/agentbox-hermes as the Hermes CLI entrypoint
• installs and starts the Agentbox Hermes local bridge as a macOS LaunchAgent

Hermes runtime state is stored under:

• ~/.hermes/agentbox/active_signer.json
• ~/.hermes/agentbox/active_role.json
• ~/.hermes/agentbox/background_runner_state.json
• ~/.hermes/agentbox/bridge.json

If you only want the skill mount and state directory, but do not want the CLI symlink:

python3 agentbox_skills/scripts/install_hermes_skills.py --no-bin-link

If you want to skip the auto-managed local bridge during development:

python3 agentbox_skills/scripts/install_hermes_skills.py --skip-bridge-service

Hermes CLI

The Hermes integration uses a local CLI:

• preferred command: agentbox-hermes
• fallback path: ~/.hermes/bin/agentbox-hermes

Examples:

agentbox-hermes signer read
agentbox-hermes registration confirm --profile-mode auto_generate
agentbox-hermes roles list-owned
agentbox-hermes roles select-active --role-wallet 0x...
agentbox-hermes read role-snapshot --source chain
agentbox-hermes check craft --recipe-id 2
agentbox-hermes action finish

The CLI writes signer and active-role state to ~/.hermes/agentbox/, not to OpenClaw.

Hermes bridge management commands:

agentbox-hermes bridge status
agentbox-hermes bridge token
agentbox-hermes bridge restart
agentbox-hermes bridge uninstall-service

agentbox-hermes bridge start remains available as a foreground debugging command. Normal installs should use the LaunchAgent installed by install_hermes_skills.py.

Background Runs

Agentbox background automation is split into two jobs:

• planner: maintains Operation Manager plans and strategy
• executor: executes safe next actions from Operation Manager

OpenClaw uses OPENCLAW_PLANNER_PROMPT.md and OPENCLAW_EXECUTOR_PROMPT.md. Hermes uses HERMES_PLANNER_PROMPT.md and HERMES_EXECUTOR_PROMPT.md.

Long-running gameplay state is stored in the Operation Manager files under each agent data directory.

Inspect Final Model Prompt

When OpenClaw cacheTrace is enabled, you can inspect the latest fully-expanded model input for a session with:

python3 agentbox_skills/scripts/show_openclaw_prompt_trace.py agentbox-background-executor

This script resolves the session by session key, session id, or trailing session label, then prints:

• PROMPT BEFORE: the new prompt added for that turn, often the cron text
• FINAL SYSTEM PROMPT: the fully-expanded system prompt after OpenClaw injects workspace files, skills, memory, and runtime guidance
• FINAL MESSAGES: the message history sent with that turn
• ASSISTANT OUTPUT: the assistant messages recorded after that run

To get the raw structured result instead of a text report:

python3 agentbox_skills/scripts/show_openclaw_prompt_trace.py agentbox-background-executor --json

To write the report directly to a file:

python3 agentbox_skills/scripts/show_openclaw_prompt_trace.py \
  agentbox-background-executor \
  --output /tmp/agentbox-background-executor-trace.txt