Kill Switch

The kill switch lets you disable any registered agent instantly. Once an agent is blocked, the proxy rejects every subsequent request from that agent with HTTP 403 — without forwarding anything to the LLM upstream. No redeployment, no config change, no restart required.

How it works

Every inbound request to the proxy carries an X-Agent-ID header (or the agent ID is inferred from the API key). Before forwarding the request, interceptor.rs queries the agent registry:

inbound request
      |
      v
interceptor.rs: SELECT status FROM agents WHERE agent_id = $1
      |
      +-- status = "active"  --> forward to upstream
      |
      +-- status = "blocked" --> return HTTP 403 immediately
      |
      +-- DB error           --> forward (fail-open)

The check happens before any LLM call is made. A blocked agent consumes zero LLM tokens.

Blocking an agent

curl -X PUT http://localhost:4001/api/v1/agents/billing-agent \
  -H "Authorization: Bearer $GOVRIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"status": "blocked"}'

Response:

{
  "agent_id": "billing-agent",
  "status": "blocked",
  "updated_at": "2026-03-05T10:30:00.000Z"
}

The change takes effect on the next request from that agent. There is no propagation delay beyond the time it takes the next request to hit the proxy.

Re-enabling an agent

curl -X PUT http://localhost:4001/api/v1/agents/billing-agent \
  -H "Authorization: Bearer $GOVRIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"status": "active"}'

HTTP 403 response format

When a blocked agent makes a request, the proxy returns:

HTTP/1.1 403 Forbidden
Content-Type: application/json

{
  "error": "agent_blocked",
  "message": "Agent 'billing-agent' is currently blocked. Contact your administrator.",
  "agent_id": "billing-agent"
}

Your agent code should treat a 403 with error: "agent_blocked" as a signal to stop retrying. Standard 403 retry logic will loop indefinitely — add a check for this error code.

Fail-open behavior

If the agent registry database is unreachable (connection pool exhausted, PostgreSQL down), the kill switch fails open: the request is forwarded to the upstream as if the agent were active.

A tracing::warn log line is emitted whenever the kill switch falls back to fail-open:

WARN govrix_scout_proxy::interceptor: agent status check failed, failing open agent_id="billing-agent" error="connection pool timed out"

Implementation