Harness coverage

A harness is an AI client that loads and runs Agent Skills — Claude Code, Codex, Cursor, and many more. skillsvault draws two capabilities and is careful never to conflate them:

• Distribution governance — install the org's approved skills, and remove banned ones and any a Block policy covers. Works for every agent that discovers SKILL.md folders on disk — 17 today, plus the shared .agents/skills location.
• Runtime veto — a hook that blocks a banned skill at the moment of invocation. Requires the harness to expose a hook that can refuse a tool call. Claude Code today.

Coverage matrix

Detecting and installing

skillsvault detect lists the harnesses found on a machine. Claude Code is labelled from the gate hook actually installed — enforced (hook), partially enforced (hook misses …) or not enforced (no gate hook), the same coverage sync reports under Hook coverage — and every other harness is distribution only. With Claude Code installed and no gate hook, detect also prints the fix.

skillsvault detect
# skillsvault detect — AI agents found on this machine
#
#   • claude_code    /Users/me/.claude/skills  enforced (hook)
#   • cursor         /Users/me/.cursor/skills  distribution only

skillsvault pull --all then installs the approved skills into every detected harness and removes any that are banned or covered by a Block policy — in harnesses without a runtime gate, that removal is the enforcement. Target a single harness with --harness, or a directory with --dir. See the command reference.

Claude Code hook

Runtime veto on Claude Code uses its PreToolUse and UserPromptExpansion hooks. skillsvault install-hook prints the snippet to add to your Claude Code settings — ~/.claude/settings.json to gate every project, or a project's .claude/settings.json (the printed command is the absolute path of your skillsvault binary):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Skill|Agent|Task",
        "hooks": [
          { "type": "command", "command": "skillsvault gate" }
        ]
      }
    ],
    "UserPromptExpansion": [
      {
        "hooks": [
          { "type": "command", "command": "skillsvault gate" }
        ]
      }
    ]
  }
}

PreToolUse covers what the agent invokes: skills through the Skill tool, and subagents through the Agent tool (named Task before Claude Code 2.1.63). UserPromptExpansion covers a person typing /name, which never reaches PreToolUse. Claude Code pipes each payload to skillsvault gate, which evaluates it locally and, on a block, vetoes the invocation and surfaces your policy's reason. The machine needs a synced bundle — login syncs one for you. The full protocol is on The runtime gate.

What each path can do:

Claude Code invokes installed skills by their bare name (deploy, not acme/deploy); the gate maps every form back to your registry — see How an invoked name is resolved.

Hook coverage

On every sync, pull, try and login, the CLI (0.5.2 and later) reports which of these paths a hook running skillsvault gate covers. It reads .claude/settings.json and .claude/settings.local.json in the directory you run it from, and ~/.claude/settings.json — so put the hook in ~/.claude/settings.json, or run sync from the project that holds it. The Gating column on the console's Harnesses page shows the result:

To fix a partial machine, run skillsvault install-hook, paste the block over the old skillsvault entry in the settings file that holds it, and run skillsvault sync. Until then sync and detect print a one-line hint on the machine. A machine on a CLI below the minimum supported version (0.5.1) also shows CLI outdated — upgrade; see Upgrading an existing machine. The report is self-attested, so it catches a stale config, not a determined user.

Related