chatlunalab/agent-config-admin

Use this skill for any request that inspects or changes the agent's own admin state. That includes skill enablement, sub-agent permissions, tool routing, MCP config, and syncing sandbox-authored skills or sub-agents back to local storage.

Core protocol

• Do not guess config, paths, or effective permissions.
• Use the agentcli tool for commands that start with agentcli.
• Treat agentcli as a dedicated tool that accepts the same CLI syntax.
• Do not mix agentcli work with normal shell commands in the same tool call.
• Use agentcli --help or agentcli <command> --help when you need the exact command shape.
• Before create, update, remove, or config work on skills, sub-agents, tools, or MCP, run the full activation sweep.
• After the activation sweep, inspect the exact target with agentcli show ....
• Before any config change, use agentcli preview ....
• Repeated agentcli preview ... commands in the same session append to the same pending preview until agentcli apply last or agentcli cancel pending.
• Many named preview commands accept multiple targets in one call, including skills, sub-agents, tools, MCP tools, and removable MCP servers.
• Tool authority can be set with agentcli preview set tool <name...> authority <0-5>.
• Only commit a pending preview with agentcli apply last.
• If the user changes direction, discard the pending preview with agentcli cancel pending.

Creation rules for skills and sub-agents

• agentcli cannot create a skill folder or a sub-agent markdown file by itself.
• When the task creates or edits a skill, you must load skill-creator first and follow that skill while authoring the files.
• When the task creates or edits a sub-agent markdown file, you must load sub-agent-creator first and follow that skill while authoring the file.
• Use agentcli to inspect paths, preview config, and sync sandbox-authored files. Use normal file tools or normal shell commands only for the actual file creation or editing work.
• If the current computer backend is not local, the sandbox may not already contain the target skills or sub-agents directory. Create the missing directory first before writing files.
• After writing skill or sub-agent files in a non-local sandbox, run agentcli sync, agentcli sync skills, or agentcli sync subagents so the files are staged back to the local instance paths.
• If a sync preview shows overwrites, list the local files that will be replaced and wait for explicit confirmation before agentcli apply last.

Path rules

• The local skill and sub-agent paths shown by agentcli belong to the current instance, not to your own computer environment.
• When the backend is local, write directly to those local instance paths.
• When the backend is not local, write into the sandbox paths shown in the prompt or returned by agentcli show overview, then sync those files back to the local instance paths.
• Never invent a home directory or machine-specific path that was not provided by ChatLuna.

Important rules

• The activation sweep is mandatory for create, update, remove, or config work on skills, sub-agents, tools, or MCP, even if the user named only one area.
• Prefer exact selectors like builtin:plan for sub-agents when possible.
• If a preview shows the wrong target, do not apply it.
• When changing permissions, always verify the result with agentcli show subagent <selector> effective or another matching show command.
• When the task authors new skill or sub-agent files, verify both the file result and the admin result before saying the task is done.
• If the agentcli tool is unavailable, say so instead of inventing results.

Activation sweep

agentcli show skills
agentcli show subagents
agentcli show tools
agentcli show mcp servers
agentcli show mcp tools

Common commands

agentcli --help
agentcli show overview
agentcli show skills
agentcli show skill coding-agent
agentcli show subagents
agentcli show subagent builtin:plan effective
agentcli show tools
agentcli show tool bash
agentcli show mcp servers
agentcli show mcp server filesystem
agentcli show mcp tools
agentcli show pending
agentcli sync
agentcli sync skills
agentcli sync subagents

agentcli preview enable skill coding-agent onboard delight
agentcli preview enable skill coding-agent
agentcli preview disable tool bash file_edit file_write --main
agentcli preview disable tool bash --main && agentcli preview disable tool file_edit file_write --main
agentcli preview disable subagent builtin:plan
agentcli preview set subagent builtin:plan builtin:general tools allow file_read glob grep
agentcli preview set subagent builtin:plan tools allow file_read glob grep
agentcli preview set tool bash grep authority 3
agentcli preview set tool bash grep subagents allow builtin:general
agentcli preview set mcp tool filesystem_read filesystem_write enabled false
agentcli preview remove mcp server filesystem browser
agentcli preview save mcp server filesystem json '{"command":"npx","args":["-y","@modelcontextprotocol/server-filesystem","."]}'
agentcli preview remove mcp server filesystem
agentcli apply last
agentcli cancel pending

Workflow

1. Run the activation sweep for create, update, remove, or config work.
2. Run agentcli show overview and inspect the exact target with agentcli show ....
3. If the task creates or edits a skill or sub-agent file, load skill-creator or sub-agent-creator first.
4. If the backend is not local, make sure the sandbox skills or sub-agents directory exists before writing files.
5. Author the skill or sub-agent files in the correct local or sandbox directory.
6. For config changes, create one focused preview with agentcli preview ... and append more preview calls if the task needs multiple staged changes.
7. For sandbox-authored files, run agentcli sync to create a preview for the local write-back.
8. Check the preview carefully. If it overwrites local files, get the user's confirmation before applying it.
9. Apply with agentcli apply last.
10. Read back the affected config, effective permissions, or synced local files.

Command chaining

• agentcli command lines may contain &, &&, |, |&, ||, and ;.
• Treat | and |& as command separators for multiple agentcli calls on one line. They do not provide stdin piping between agentcli tool commands.
• Named preview commands accept multiple targets when the syntax uses <name...> or <selector...>.
• agentcli preview enable tool and agentcli preview disable tool accept multiple tool names before --main.
• agentcli preview set tool <name...> authority <0-5> updates required Koishi authority for one or more tools.