AGENT SKILL

yard-control-plane

Audit and change the user's agent setup across Claude Code, Codex, and Cursor. Use when asked what is loaded, what is eating the context window, why a hook or MCP server is not working, or to enable/disable a skill, plugin, or MCP server.

License 许可
MIT
所属仓库
yannelli/yard
最近更新
2026-08-17
Stars
★ 0
Forks
0

01 前言 Front Matter

name
yard-control-plane
description
Audit and change the user's agent setup across Claude Code, Codex, and Cursor. Use when asked what is loaded, what is eating the context window, why a hook or MCP server is not working, or to enable/disable a skill, plugin, or MCP server.

03 SKILL.md 正文

Yard

Yard reads the user's real configuration for all three clients and can change it. It is not limited to this repository.

When to reach for it

  • "why is my context window so full" → env_context
  • "what skills/plugins/MCP servers do I have" → env_inventory
  • "my hook isn't firing" / "this MCP server won't start" → env_doctor
  • "turn off X" → env_set_skill, env_set_plugin, env_set_mcp

Reading

env_inventory returns counts plus a bounded list. Filter with client and kind rather than pulling everything — there are often several hundred skills.

env_context prices what enters the context window every turn: skill listing lines, MCP tool schemas, subagent descriptions, slash commands, and memory files. Numbers are estimates and are labelled as such.

env_doctor finds hooks pointing at deleted scripts, MCP servers that fail to start, skills with missing or truncated frontmatter, plugins enabled but not installed, and duplicate skill names.

probe: true on either one starts the user's configured stdio MCP servers to read their real tool lists. That executes their commands, so only pass it when the user asked for real numbers, and say that is what you are doing.

Changing

Every mutating tool takes dryRun. Use it first when the change is not trivially reversible, show the user what would change, then apply.

Writes are atomic and backed up to ~/.yard/backups, and each result names the file that changed.

Skill visibility is the main lever for context, and only Claude Code has all four values:

  • on — name and description in the listing
  • name-only — name only, description dropped
  • user-invocable-only — invisible to the model, still available from /
  • off — gone

user-invocable-only is usually the right suggestion for a skill the user triggers by hand: it costs nothing per turn and stays one keystroke away. Prefer it over off unless they want it gone.

If a name exists in more than one client the call fails and names the candidates. Pass client.

Codex owns ~/.codex/config.toml, so MCP changes there are refused with the codex mcp command to run instead — rewriting that file would destroy the user's comments and ordering.

Authoring

catalog_search and catalog_get browse this marketplace's own plugins. plugin_adapt fills in the files a Claude-only skill or plugin is missing for Codex and Cursor; it only writes what is absent.

One thing worth knowing when editing plugins: plugin hooks are a single PascalCase hooks/hooks.json that all three clients read. The camelCase version: 1 form belongs to the user's own .cursor/hooks.json and is not a plugin file. Cursor also reads plugin MCP servers from .cursor-plugin/plugin.json, not from .mcp.json.

04 同仓库更多 Skills