Skip to main content

External Policy Conversation Key, Feedback, and Verifier Hints

This page describes caller-visible and admin-configured extensions to External Routing Policy for conversation pinning, post-completion feedback, and offline verifier metadata. Operator detail lives in the repository runbook docs/EXTERNAL_POLICY_CONTEXT.md.

Conversation key​

When a model group uses strategy: external, the policy request includes context.conversationKey. Callers can keep multi-turn pins stable by sending a deployment-defined conversation id in request metadata:

{
  "model": "adaptive",
  "messages": [{"role": "user", "content": "Continue the previous task."}],
  "metadata": {
    "conversation_id": "app-thread-123"
  }
}

Accepted metadata keys: conversation_id, conversationId, session_id, sessionId, and metrum_conversation_id. Values must be 1–128 characters of letters, numbers, and ._:-.

If no conversation id is present, the router hashes the first user-turn text locally and still emits a pseudonymous key. The policy service does not receive that raw text unless the operator separately enables include_request.

Keys are scoped by caller identity and model group. They never include bearer tokens or token hashes. Treat them as routing pin identifiers, not authorization secrets.

Completion feedback​

Operators may enable an authenticated post-completion callback on the policy service. Callers do not configure this; it is deployment-owned. The callback carries only request ID, selected target, status, stored usage/cost/latency, and related safe scalars. Prompts, tools, and credentials are never included.

Verifier hints​

When the operator enables external_policy.verifier_hints, callers may attach bounded offline-evaluation metadata:

{
  "model": "adaptive",
  "messages": [{"role": "user", "content": "What is 2+2?"}],
  "metadata": {
    "metrum_verifier_hint": {
      "kind": "exact",
      "version": "v1",
      "spec": {"expected": "4"}
    }
  }
}

Supported kinds are none, exact, regex, and json_schema. Hints are metadata only and are never executed by the router. Executable kinds are rejected. Size and version limits are configured by the operator. Hint ground truth is trusted policy/evaluation metadata; do not place secrets in verifier specs.