Classification

aegis.classification labels each request so later nodes can make decisions without re-scanning the text. It runs a small, ordered set of regex rules over the latest user message and writes the first match to state.labels["classification"].

guardrails:
  classify:
    pack: aegis.classification

pipeline:
  ingress: [classify]

Labels

Rules are checked in this order; the first match wins.

Label	Matches
pii	email addresses, US-style phone numbers
financial	16-digit card-number patterns
secret	api_key=…, password: …, token=… and similar
medical	diagnosis, prescription, patient, HIPAA
legal	attorney-client, privileged, confidential
public	anything else

The node never blocks — it only labels. The label is recorded in the run's events.

Acting on the label

Put the classifier first, then a guard that reads the label:

from aegis_core.pipeline.state import RunState
from aegis_core.pipeline.verdict import Verdict


class NoSecretsGuard:
    name = "no_secrets"
    streaming = "none"

    async def scan(self, state: RunState) -> Verdict:
        if state.labels.get("classification") == "secret":
            return Verdict.block("credentials must not be sent to a model")
        return Verdict.allow()

pipeline:
  ingress: [classify, no_secrets]

The rules are intentionally cheap and conservative; for richer detection, combine with PII masking or LLM Guard, or write a node with your own classifier.