hol-guard

HOL Guard

HOL Guard protects local AI harnesses before tools run. Use this skill when the user wants AI antivirus behavior, local approval review, Codex protection, Claude Code protection, MCP safety checks, skill/package verification, or release gates from hol-guard and plugin-scanner.

Hard Rules

• Never read .env files.
• Never bypass Guard approvals.
• Do not mark a workspace protected until a Guard command proves status.
• Prefer reversible Guard commands over direct harness config edits.
• Do not mutate user-level harness config unless the hol-guard command owns that mutation.
• Treat scanner failures as real until inspected.
• Preserve existing user changes and inspect git status --short before edits in a repo.

Install Check

Check both CLIs independently:

command -v hol-guard
command -v plugin-scanner

If hol-guard is missing and the user asked for runtime setup, prefer:

pipx install hol-guard

If plugin-scanner is missing and the user asked for scanning, install the separate scanner distribution:

pipx install plugin-scanner

Do not assume the hol-guard distribution provides the plugin-scanner command. If pipx is unavailable, explain that isolated CLI installation is recommended rather than silently changing the user's Python environment.

After runtime installation:

hol-guard status
hol-guard detect --json

Protect A Local Harness

Use this flow for Codex, Claude Code, Copilot CLI, Cursor, Gemini, Hermes, OpenClaw, OpenCode, or Antigravity.

hol-guard bootstrap
hol-guard install <harness>
hol-guard run <harness> --dry-run
hol-guard run <harness>
hol-guard status

Harness names:

• codex
• claude-code
• copilot
• cursor
• gemini
• hermes
• openclaw
• opencode
• antigravity

Harness aliases:

• claude maps to claude-code
• gemini-cli maps to gemini
• open-code maps to opencode
• open-claw maps to openclaw
• copilot-cli maps to copilot

Use harness-specific bootstrap when available:

hol-guard hermes bootstrap

Claude Code

Use this when the workspace has .claude/settings.local.json, .claude/agents, Claude hooks, .mcp.json, or Claude-managed tool approval surfaces.

hol-guard install claude-code
hol-guard run claude-code --dry-run
hol-guard run claude-code
hol-guard doctor claude-code --json

Claude Code is a first-class Guard target. Prefer Guard-owned Claude hooks over direct manual edits to Claude config.

Codex

Use this when the workspace has Codex config, .codex/hooks.json, Codex MCP servers, or Codex App/CLI tool flows.

hol-guard install codex
hol-guard run codex --dry-run
hol-guard run codex
hol-guard doctor codex --json

Codex supports Guard-owned PreToolUse Bash hooks and same-chat MCP elicitation where available.

Other Harnesses

Use the same Guard flow for:

• Copilot CLI: hol-guard install copilot
• Cursor: hol-guard install cursor
• Gemini CLI: hol-guard install gemini
• Hermes: hol-guard hermes bootstrap
• OpenClaw: hol-guard install openclaw
• OpenCode: hol-guard install opencode
• Antigravity: hol-guard install antigravity

Approval Work

If Guard blocks or queues work:

hol-guard approvals
hol-guard approvals open
hol-guard receipts
hol-guard diff <harness>

For terminal-only resolution:

hol-guard approvals approve <request-id>
hol-guard approvals deny <request-id>

Only approve after reading the risk reason and understanding the requested scope.

Evidence Work

Use evidence commands when user needs proof, audit trail, or handoff artifacts:

hol-guard receipts
hol-guard inventory
hol-guard abom --format json
hol-guard events
hol-guard explain <artifact-id>

For cloud sync, keep it optional and user-directed:

hol-guard connect
hol-guard connect status
hol-guard connect repair
hol-guard sync

Scan A Plugin Or Skill Package

Use scanner mode for Codex plugins, Claude Code project surfaces, .agents marketplaces, skills, MCP server configs, and release gates.

plugin-scanner lint .
plugin-scanner verify .

If scanning a specific package:

plugin-scanner lint <path>
plugin-scanner verify <path>

If the target is a Codex marketplace root with .agents/plugins/marketplace.json, scan the repo root so local plugin entries can be discovered.

Scanner target guidance:

• Codex plugin: scan the repo root or plugin folder containing .codex-plugin/plugin.json.
• Codex marketplace: scan the repo root containing .agents/plugins/marketplace.json.
• Claude Code project: scan the workspace root containing .claude/, .mcp.json, hooks, or agent folders.
• MCP server package: scan the package root containing server config and package metadata.
• Skill package: scan the folder containing SKILL.md.
• Mixed agent workspace: scan the repo root so local plugin, skill, MCP, and harness config surfaces are discovered together.

Common Debug Commands

hol-guard doctor
hol-guard doctor <harness> --json
hol-guard detect --json
hol-guard settings show
hol-guard explain install-connect
plugin-scanner verify . --json

Response Pattern

When using Guard, report:

• What command ran.
• What Guard found.
• What remains blocked or risky.
• What proof exists.
• Exact next command if user must act.

Do not claim protection, approval, or release readiness without command output proving it.