AANA Guardrail Skill Bundled Helper Variant

Provides a local Python helper to send minimal review payloads to a trusted AANA HTTP bridge for policy guidance inside the skill package.

Use this variant only when the user or administrator wants a bundled, inspectable helper inside the skill package.

This package includes a small Python helper at bin/aana_guardrail_check.py. The helper does not implement AANA policy itself. It only sends a minimal review payload to a separately running AANA HTTP bridge on localhost.

Safety Boundary

• Do not run the helper unless the user explicitly approves using the configured AANA bridge for this task.
• Do not send payloads to remote hosts. The bundled helper rejects non-localhost URLs.
• Do not include API keys, bearer tokens, passwords, full payment numbers, unnecessary account records, or unrelated private messages.
• Prefer redacted summaries over raw sensitive content.
• Treat AANA recommendations as advisory unless the bridge and policy are trusted and auditable.
• If the helper or bridge is unavailable, use manual review.

Review Payload

Use a minimal JSON payload:

json

{
  "adapter_id": "support_reply",
  "request_summary": "draft a refund support reply",
  "candidate_summary": "reply would promise refund eligibility",
  "evidence_summary": ["refund eligibility is unknown"],
  "allowed_actions": ["accept", "revise", "ask", "defer", "refuse"]
}

Helper Behavior

The helper:

• reads one JSON payload file,
• validates that it is a JSON object,
• blocks obvious secret-like keys,
• sends the payload to a localhost AANA bridge,
• prints the JSON result,
• exits nonzero if the result recommends anything other than accept.

The helper does not:

• install dependencies,
• execute other scripts,
• infer checker paths,
• contact remote hosts,
• create event files,
• delete user files.

Decision Rule

• If recommended_action is accept, proceed.
• If recommended_action is revise, revise before acting.
• If recommended_action is ask, ask the user for missing information.
• If recommended_action is defer, route to a stronger tool or human review.
• If recommended_action is refuse, pause, explain the reason, and ask for review when the decision affects important work.

If no trusted bridge is configured, do not use this helper.