@siere.ai/core

Core building blocks for Siere AI agent detection, content transformation, analytics, health checks, and licensing.

npm install @siere.ai/core

See Siere documentation for integration guidance.

Agent response caching

Transformed Markdown responses default to Cache-Control: private, no-store and Vary: User-Agent, Accept. Custom Vary values are preserved and normalized with those required fields. A custom Cache-Control header cannot weaken this default unless unsafePublicAgentCaching: true is set explicitly.

That advanced escape hatch is only for deployments whose shared-cache key has been verified to separate every agent representation by both User-Agent and Accept; the internal Hermes transformation cache remains available without it.

Detection contract

detectAgent() always returns a typed result. A User-Agent match is classified as unverified; it is not proof of the crawler's identity. The canonical classification is purpose (search-index, user-retrieval, model-training, other, or unknown). The legacy type projection remains available for the 0.x compatibility cycle. Every result also exposes the registry version and evidence level used for the classification.

import { detectAgent } from "@siere.ai/core";

const detection = detectAgent(request.headers.get("user-agent") ?? "");
if (detection.matchedAgent) {
  console.log(detection.detectionMethod, detection.verificationStatus);
}

Delivery decision contract

decideDeliveryAction() is the framework-neutral authority for whether a request passes through to the origin or receives Markdown. It returns a typed action, reason, strategy, purpose, evidence level, detection method, and verification status so adapters can expose the decision without retaining raw request headers.

import { decideDeliveryAction, detectAgent } from "@siere.ai/core";

const detection = detectAgent(request.headers.get("user-agent") ?? "");
const decision = decideDeliveryAction({
  strategy: "observe",
  detection,
  acceptHeader: request.headers.get("accept"),
});

The strategies are:

• observe: always pass through while recording the classified request.
• negotiate-markdown: serve Markdown only when Accept negotiates it.
• recognized-agent-markdown: serve recognized User-Agent or negotiated requests; this remains the omitted-config default for existing integrations.
• verified-agent-markdown: reserved for a future verified-identity provider; unverified requests currently pass through.

Missing, stale, unknown, transform-error, fidelity-drifted, and fidelity-unverifiable content always fails open to the origin. Adapter response headers and /__aeo/health expose the active strategy, and onDeliveryDecision can receive the typed decision for structured logging.