hol-guard

Install and use HOL Guard as a local pre-execution safety boundary for supported AI coding-agent harnesses, with explicit approval, receipt, and verification flows.

SKILL.md

GitHub

HOL Guard

Use HOL Guard when a user wants a local enforcement boundary around a supported AI coding-agent harness before consequential tool or command execution.

Use this skill when

  • Setting up HOL Guard for a detected local coding-agent harness
  • Running an agent through HOL Guard rather than directly
  • Reviewing Guard approvals, receipts, diagnostics, or protection state
  • Verifying whether a supported harness is actually protected before work continues

Do not use this skill when

  • The user has not authorized changes to their local agent harness
  • The target is an application/server authorization problem rather than a supported local harness
  • The requested workflow would require bypassing a Guard denial, approval, or review gate

Instructions

  1. Check whether HOL Guard is installed without reading secrets using a cross-platform direct CLI probe:

    hol-guard --version
    

    If it is missing and the user requested runtime setup, prefer the isolated CLI install:

    pipx install hol-guard
    
  2. Inspect Guard state and detect supported harnesses:

    hol-guard status
    hol-guard detect --json
    

    Use the exact harness identifier returned by hol-guard detect --json; do not maintain a guessed harness list.

  3. For an authorized detected harness, use Guard-owned setup and validation:

    hol-guard bootstrap
    hol-guard install <harness>
    hol-guard run <harness> --dry-run
    hol-guard doctor <harness> --json
    
  4. Run the harness through HOL Guard only after the dry run and diagnostics support it:

    hol-guard run <harness>
    
  5. If Guard blocks or requires review, stop the protected action and inspect the decision instead of retrying unprotected:

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

    Use the pending request ID returned by hol-guard approvals; do not invent an ID. Approval or denial must remain an explicit user decision after the risk reason and requested scope are understood.

  6. Verify the resulting protection state with Guard-owned evidence:

    hol-guard status
    hol-guard doctor <harness> --json
    hol-guard receipts
    

Safety boundaries

  • Never read .env files as part of this workflow.
  • Never bypass Guard approvals or convert deny, review-required, timeout, malformed, unavailable, or error states into an allow.
  • Preserve host permissions, repository tests, validation, and required human review; HOL Guard does not replace them.
  • Do not claim application/server authorization, downstream correctness, or protection for an unsupported harness.
  • Do not mark a workspace protected until HOL Guard command output proves the state.

Result

Report the commands run, the harness Guard detected, the protection or review state, the evidence available, and the exact next action if the user must intervene.

Attribution

Adapted for this catalog from the maintained HOL Guard Agent Skill in hashgraph-online/hol-guard-plugin. HOL Guard is licensed under Apache-2.0. The bundled LICENSE file preserves the upstream license text.