wayleave 0.1.6 → 0.2.1

package/README.md

@@ -52,19 +52,21 @@ app.use(gate.express());
 - Signature parameters are parsed as an RFC 9421 dictionary: order-independent, `alg` enforced as Ed25519, any signature label, and the signature base is built from whatever components the signer declared. Requests signed in Cloudflare's documented format verify.
 - Sub-millisecond per request. Your latency budget won't notice.
 - The metering hook can throw, crash, or hang your billing backend — serving continues. Your uptime never depends on ours.
-- Key directories still load from config rather than a live JWKS fetch, so key rotation is not handled. That is the next real gap.
-- **A bot that sends a browser `user-agent` and an `accept-language` header is classified `human` and crosses priced routes free.** Those two headers are the entire bypass. The `human` lane is a fall-through — it is reached by tripping none of the automation tells, which is absence of evidence, not evidence of a person. There is no TLS fingerprinting, no IP verification, and no challenge here; that work belongs at your edge or CDN, and pretending otherwise would be the dishonest version of this list. `strictPricedPaths: true` inverts the burden on priced routes so that only a verified signature or your own `confirmHuman(req)` crosses. Recommended wherever there is a price.
+- Key rotation has a seam but no implementation. `directories` accepts a resolver function you can point at a cache, but there is no JWKS fetcher in the box and the resolver must not block. Wiring that cache is still your job.
+- **A bot that sends a browser `user-agent` and an `accept-language` header is classified `human` and crosses priced routes free.** Those two headers are the entire bypass. The `human` lane is a fall-through — it is reached by tripping none of the automation tells, which is absence of evidence, not evidence of a person. There is no TLS fingerprinting and no challenge here; that work belongs at your edge or CDN, and pretending otherwise would be the dishonest version of this list. (`verifyAgentIP` checks a *declared* operator against its published ranges — it does nothing about a request claiming to be a browser, which is this bypass.) `strictPricedPaths: true` inverts the burden on priced routes so that only a verified signature or your own `confirmHuman(req)` crosses. Recommended wherever there is a price.
 - Wayleave prices **disclosure**; it does not detect **concealment**. It is a tollbooth, not a wall — it works on operators who want to be identifiable, which today is most of the ones worth billing.
 - This is a **screening and pricing layer, not a guarantee**. Every decision returns its evidence and is loggable.
 
 ## Status
 
-v0.1.6 — verification, lanes, policy, pricing, and the metering hook, all under test (55 scenarios, including forgery, tampering, replay, guessed payment proofs, spoofed rate-limit identities, spoofed browser headers, and cross-implementation wire formats). Payment settlement network integration is in progress; the 402 challenge is built in and settlement confirmation is a function you supply.
+v0.2.1 — verification, lanes, policy, pricing, and the metering hook, all under test (87 scenarios, including forgery, tampering, replay, guessed payment proofs, spoofed rate-limit identities, spoofed browser headers, spoofed operator identity, and cross-implementation wire formats). Payment settlement network integration is in progress; the 402 challenge is built in and settlement confirmation is a function you supply.
 
 0.1.5 closes a hole worth naming: before it, a priced route accepted a payment proof that any agent could derive from the 402 challenge it had just been sent — free passage, recorded as revenue. Settlement confirmation is now yours to supply and denies by default. If you shipped 0.1.4 or earlier on a priced route, upgrade.
 
 0.1.6 addresses the other half of the same question. Payment could be faked; so could being human. `strictPricedPaths` lets a priced route demand positive evidence instead of accepting the absence of a bot signal as one.
 
+0.2.1 ships `wayleave/meter`, a buffering and retrying sink for the hosted meter. 0.2.0 turns three hardcoded strategies into seams — state, metering, and key lookup — so shared storage, durable billing, and key rotation become configuration rather than a rewrite. Defaults are unchanged; see [Extending it](#extending-it).
+
 ## How payment actually works
 
 Pay-at-the-door, not IOUs. On a route you price:
@@ -130,9 +132,62 @@ nothing but a confirmed settlement is ever written to the ledger as billed.
   agent hitting two paths in the same second produces byte-identical
   signatures. Rejecting those would break real traffic. Without a signer
   nonce, replay is bounded by the expiry window and nothing tighter.
-- **State is per-process.** Rate limits and replay memory do not survive a
-  restart and do not span instances — two instances mean two independent
-  limiters. Shared state is the next real gap, alongside key rotation.
+- **State is per-process by default.** Rate limits and replay memory do not
+  survive a restart and do not span instances. Pass a `store` implementing
+  `hit` / `hasNonce` / `rememberNonce` to share them. Keep it synchronous —
+  it runs on every request — so back a distributed store with a local view
+  that replicates asynchronously. Note the two uses differ in how much they
+  mind lag: fixed-window rate limiting tolerates approximate counts, replay
+  does not, because a nonce that has not replicated yet can be spent twice.
+- **Key rotation is possible but not built in.** `directories` accepts a
+  resolver function, so point it at a cache you refresh from a JWKS endpoint
+  on a timer. There is no fetcher in the box, and the resolver must not block.
+
+## Extending it
+
+Three seams, each defaulting to the behaviour above. Nothing here is required.
+
+```js
+new Wayleave({
+  store: myRedisBackedStore,        // shared rate limits + replay across instances
+  sink: myBufferedSink,             // emit() with batching, retry, dedup
+  directories: url => keyCache[url],// rotatable keys, refreshed out of band
+  verifyAgentIP: (ip, ua) => isPublishedRange(ip, ua),
+});
+```
+
+Every metering event carries `v` (schema version) and `idempotencyKey`, so a
+sink that retries can be deduplicated at the far end rather than double-billing.
+
+A buffering, retrying sink ships in the box:
+
+```js
+import Wayleave from 'wayleave';
+import { MeterSink } from 'wayleave/meter';
+
+const sink = new MeterSink({
+  endpoint: 'https://meter.wayleave.dev',
+  apiKey: process.env.WAYLEAVE_METER_KEY,
+  onError: (err, info) => log.warn({ err, info }, 'meter'),
+});
+
+const gate = new Wayleave({ pricedPaths: { '/api/premium': 0.05 }, sink });
+process.on('SIGTERM', () => sink.close());   // drain, or lose what is buffered
+```
+
+`emit` is synchronous, non-throwing, and does no I/O — a billing backend
+having a bad day must not become your customers having one. Retries are safe
+because of the idempotency key: sending a batch twice costs a round trip, not
+a double-billed customer. The buffer is bounded, and anything shed is counted
+and reported rather than dropped quietly.
+
+`verifyAgentIP` is the one that changes a lane. A `user-agent` naming a known
+operator is a claim; when that operator publishes its ranges the claim is
+checkable. Returning `false` moves the request to `suspected_bot` — a false
+claim of identity is a stronger signal than no claim, which is the same
+reasoning that puts an invalid signature there. Return `null` for operators you
+cannot check; that leaves the request `declared_agent`, because unknown is not
+guilty. Throwing is treated as `null` for the same reason.
 
 ## License
 

package/index.d.ts

@@ -144,6 +144,17 @@ export interface Decision {
 
 /** What the metering hook receives. It may throw; serving continues. */
 export interface MeterEvent {
+  /**
+   * Event schema version. Present so a meter that outlives one release can
+   * tell which shape it is reading. Currently always 1.
+   */
+  v: number;
+  /**
+   * Unique per crossing and stable across retries of the same event, so a
+   * sink that resends after a failure can be deduplicated at the far end.
+   * Namespaced per process — two instances never collide.
+   */
+  idempotencyKey: string;
   /** Unix seconds. */
   t: number;
   path: string;
@@ -161,8 +172,90 @@ export interface MeterEvent {
 /** `[pathPrefix, allow]` — first matching prefix wins. */
 export type PolicyRule = [prefix: string, allow: boolean];
 
+/**
+ * Rate-limit counters and consumed nonces. Implement this against Redis or
+ * Postgres and pass it as `store` to share state across instances.
+ *
+ * Must be synchronous — it is called on every request, and an awaited round
+ * trip per crossing costs more than the crossing earns. Put a distributed
+ * backend behind a local view that replicates asynchronously.
+ *
+ * Note the two uses differ in how much they mind approximation: fixed-window
+ * rate limiting tolerates it, replay does not. A nonce that has not yet
+ * replicated is a nonce that can be spent twice.
+ */
+export interface Store {
+  /** Requests by this identity in the current window. Returns the new count. */
+  hit(identity: string, windowSeconds: number, now: number): number;
+  hasNonce(nonce: string): boolean;
+  rememberNonce(nonce: string, expiresAt: number): void;
+}
+
+/** The default `Store`: bounded, in this process's memory, lost on restart. */
+export class MemoryStore implements Store {
+  constructor(opts?: { maxTracked?: number });
+  maxTracked: number;
+  hit(identity: string, windowSeconds: number, now: number): number;
+  hasNonce(nonce: string): boolean;
+  rememberNonce(nonce: string, expiresAt: number): void;
+}
+
+/**
+ * Where metering events go. Implement `emit` to buffer, batch, retry, and
+ * dedup on `idempotencyKey`. It must never throw into the request path.
+ */
+export interface Sink {
+  emit(event: MeterEvent): void;
+}
+
+/** The default `Sink`: calls `onEvent` and swallows whatever it throws. */
+export class DirectSink implements Sink {
+  constructor(onEvent?: (event: MeterEvent) => void);
+  emit(event: MeterEvent): void;
+}
+
+/**
+ * Key lookup. The function form is what makes rotation possible: point it at
+ * a cache your app refreshes from a JWKS endpoint on a timer. Synchronous by
+ * design — do not fetch here.
+ */
+export type KeyResolver = (directoryUrl: string) =>
+  Record<string, PublicKeyLike> | undefined;
+
+/**
+ * Checks whether a request claiming to be a known operator actually came from
+ * that operator's published addresses.
+ *
+ * Return `true` to confirm, `false` to accuse — a false claim is a stronger
+ * fraud signal than silence, so it lands in `suspected_bot` alongside an
+ * invalid signature. Return `null`/`undefined` for "cannot check", which
+ * leaves the request merely `declared_agent`: unknown is not guilty. A throw
+ * is treated as "cannot check" too, so a DNS blip never demotes real agents.
+ */
+export type AgentIPVerifier = (
+  ip: string | undefined,
+  userAgent: string
+) => boolean | null | undefined;
+
 export interface WayleaveOptions {
-  directories?: KeyDirectories;
+  /** Config object, or a resolver function for rotatable keys. */
+  directories?: KeyDirectories | KeyResolver;
+  /**
+   * Verify that a self-declared operator is on its published addresses.
+   * Without it, `declared_agent` is a claim taken at face value.
+   */
+  verifyAgentIP?: AgentIPVerifier;
+  /**
+   * Rate-limit and replay state. Defaults to a bounded `MemoryStore`, which
+   * is per-process: two instances mean two independent limiters. Supply a
+   * shared implementation to fix that without touching anything else.
+   */
+  store?: Store;
+  /**
+   * Where metering events go. Defaults to a `DirectSink` wrapping `onEvent`.
+   * Replace it with a buffering, retrying sink when losing events costs money.
+   */
+  sink?: Sink;
   /** Per-lane allow/deny rules, evaluated in order. */
   rules?: Partial<Record<Lane, PolicyRule[]>>;
   /** Max requests per identity per window, per lane. */
@@ -249,7 +342,7 @@ export function signRequest(
 export function verifySignature(
   headers: Record<string, string | string[] | undefined>,
   authority: string,
-  directories: KeyDirectories,
+  directories: KeyDirectories | KeyResolver,
   now?: number
 ): VerifyResult;
 
@@ -259,8 +352,14 @@ export function verifySignature(
  */
 export function classify(
   req: WayleaveRequest,
-  directories: KeyDirectories,
-  now?: number
+  directories: KeyDirectories | KeyResolver,
+  now?: number,
+  opts?: {
+    /** Confirms a declared operator is on its published addresses. */
+    verifyAgentIP?: AgentIPVerifier;
+    /** Source address handed to `verifyAgentIP`. */
+    ip?: string;
+  }
 ): Classification;
 
 export class Wayleave {

package/index.js

@@ -18,7 +18,7 @@
  */
 
 import { createPublicKey, verify as edVerify, sign as edSign,
-         createHash } from 'node:crypto';
+         createHash, randomUUID } from 'node:crypto';
 
 export const LANES = Object.freeze({
   VERIFIED: 'verified_agent',
@@ -108,8 +108,28 @@ export function signRequest(headers, authority, privateKey, keyid, directoryUrl,
   return headers;
 }
 
+/**
+ * Key lookup, as a seam rather than an object read.
+ *
+ * `directories` may be the config object it has always been, or a function
+ * `(directoryUrl) => { keyid: key }`. The function form is what makes rotation
+ * possible without touching this file: point it at a cache your app refreshes
+ * from a JWKS endpoint on a timer.
+ *
+ * It MUST be synchronous. This runs on every signed request, and awaiting a
+ * network fetch per crossing would cost far more than the crossing earns —
+ * the sub-millisecond claim in the README is a real constraint, not a boast.
+ * Fetch on a schedule, resolve from memory.
+ */
+function asResolver(directories) {
+  if (typeof directories === 'function') return directories;
+  const d = directories || {};
+  return (directoryUrl) => d[directoryUrl];
+}
+
 /**
  * directories: { [directoryUrl]: { [keyid]: publicKey (KeyObject|PEM|raw b64) } }
+ *              or (directoryUrl) => { [keyid]: publicKey }
  * Returns { ok, agentId, reason, nonce, expires }
  *
  * `nonce` is set only when the signer supplied RFC 9421's `nonce` parameter,
@@ -157,7 +177,7 @@ export function verifySignature(h, authority, directories,
   if (expires - created > 3600)
     return { ok: false, agentId: null, reason: 'validity window too long (replay risk)' };
 
-  const dir = directories[directory];
+  const dir = asResolver(directories)(directory);
   if (!dir) return { ok: false, agentId: null, reason: `unknown key directory ${directory}` };
   let key = dir[keyid];
   if (!key) return { ok: false, agentId: null, reason: `keyid "${keyid}" not in directory` };
@@ -209,7 +229,20 @@ const AUTOMATION_TELLS = [
   ['wget', 'wget UA'],
 ];
 
-export function classify(req, directories, now) {
+/**
+ * A verifier that throws must not decide the lane. Treat the failure as
+ * "could not check" rather than "is a forger" — a DNS timeout is not evidence
+ * of fraud, and demoting real agents to the fraud lane because a lookup blipped
+ * would be a self-inflicted outage on the traffic we most want to bill.
+ */
+function tryVerifyIP(verify, ip, ua) {
+  try {
+    const r = verify(ip, ua);
+    return r === true ? true : r === false ? false : null;
+  } catch { return null; }
+}
+
+export function classify(req, directories, now, opts = {}) {
   const h = req.headers, ua = (h['user-agent'] || '').toLowerCase();
   const v = verifySignature(h, req.authority, directories, now);
   if (v.ok) return { lane: LANES.VERIFIED, agentId: v.agentId,
@@ -218,10 +251,25 @@ export function classify(req, directories, now) {
   if (h['signature'] || h['signature-input'])
     return { lane: LANES.SUSPECT, agentId: null,
              evidence: [`crypto: ${v.reason}`, 'presented invalid signature'] };
-  if (AGENT_UA.some(p => p.test(ua)))
+  if (AGENT_UA.some(p => p.test(ua))) {
+    // A UA naming a known operator is a claim, and claims are checkable when
+    // the operator publishes its addresses. Claiming to be GPTBot from an
+    // address OpenAI does not own is a stronger fraud signal than saying
+    // nothing at all — so it lands in the fraud lane, exactly like a
+    // signature that fails to verify. An unrecognised operator returns
+    // null/undefined and stays merely declared: unknown is not guilty.
+    const ipOk = opts.verifyAgentIP
+      ? tryVerifyIP(opts.verifyAgentIP, opts.ip, ua)
+      : null;
+    if (ipOk === false)
+      return { lane: LANES.SUSPECT, agentId: null,
+               evidence: [`UA claims "${ua.slice(0, 40)}"`,
+                          `but ${opts.ip || 'the client'} is not a published address for it`] };
     return { lane: LANES.DECLARED, agentId: null,
              evidence: [`self-identified in UA: "${ua.slice(0, 40)}"`,
-                        'no signature presented'] };
+                        ipOk === true ? 'source address matches published range'
+                                      : 'no signature presented'] };
+  }
   for (const [tell, why] of AUTOMATION_TELLS)
     if (ua.includes(tell))
       return { lane: LANES.SUSPECT, agentId: null, evidence: [why] };
@@ -232,6 +280,77 @@ export function classify(req, directories, now) {
   return { lane: LANES.HUMAN, agentId: null, evidence: ['browser-shaped request'] };
 }
 
+// ── state ───────────────────────────────────────────────────────────────
+
+/**
+ * Rate-limit counters and consumed nonces, in this process's memory.
+ *
+ * This is the seam that a shared backend replaces. Implement the same four
+ * methods against Redis or Postgres and pass it as `store` — the gate does not
+ * know the difference. Two properties any replacement must preserve:
+ *
+ *   Synchronous. Called on every request. An awaited round trip per crossing
+ *   is a worse tax than the crossing is worth; a distributed store belongs
+ *   behind a local view that replicates asynchronously.
+ *
+ *   Bounded. `maxTracked` exists because an attacker who can mint identities
+ *   can otherwise mint memory. Evicting the oldest keeps recent arrivals
+ *   counted, which is the useful half.
+ *
+ * A distributed store trades exactness for reach, and the two uses differ in
+ * how much they mind. Fixed-window rate limiting tolerates approximate counts.
+ * Replay does NOT: a nonce that has not replicated yet is a nonce that can be
+ * spent twice. Say which you chose in your deployment notes.
+ */
+export class MemoryStore {
+  constructor({ maxTracked = 10000 } = {}) {
+    this.maxTracked = maxTracked;
+    this._hits = new Map();
+    this._window = -1;
+    this._seen = new Map();
+  }
+
+  /** Count this identity's requests in the current fixed window. */
+  hit(identity, windowSeconds, now) {
+    const w = Math.floor(now / windowSeconds);
+    if (w !== this._window) {
+      this._window = w;
+      this._hits.clear();
+      for (const [n, exp] of this._seen) if (exp <= now) this._seen.delete(n);
+    }
+    if (this._hits.size >= this.maxTracked && !this._hits.has(identity))
+      this._hits.delete(this._hits.keys().next().value);
+    const n = (this._hits.get(identity) || 0) + 1;
+    this._hits.set(identity, n);
+    return n;
+  }
+
+  hasNonce(nonce) { return this._seen.has(nonce); }
+
+  rememberNonce(nonce, expiresAt) {
+    if (this._seen.size >= this.maxTracked)
+      this._seen.delete(this._seen.keys().next().value);
+    this._seen.set(nonce, expiresAt);
+  }
+}
+
+/**
+ * Where metering events go. The default hands each one straight to `onEvent`
+ * and swallows anything it throws, because serving must not depend on billing.
+ *
+ * That swallow is correct for uptime and lossy for revenue, which is the
+ * trade a durable sink is meant to change: buffer locally, flush in batches,
+ * retry on failure, and dedup on `idempotencyKey` at the far end. Implement
+ * `emit(event)` and pass it as `sink`. Whatever you build, it must not throw
+ * into the request path — this class is the last line that guarantees that.
+ */
+export class DirectSink {
+  constructor(onEvent) { this.onEvent = onEvent || (() => {}); }
+  emit(event) {
+    try { this.onEvent(event); } catch { /* metering must never break serving */ }
+  }
+}
+
 // ── the gate ────────────────────────────────────────────────────────────
 
 export class Wayleave {
@@ -281,23 +400,40 @@ export class Wayleave {
     this.replayProtection = opts.replayProtection !== false;
     this.maxTracked = opts.maxTracked || 10000;
 
-    this._hits = new Map();     // identity -> count, cleared each window
-    this._window = -1;
-    this._seen = new Map();     // signature nonce -> expiry, pruned on rollover
+    // Verifies that a UA claiming to be a known operator comes from that
+    // operator's published addresses. Absent, a declared agent is taken at
+    // its word — which is all "declared" ever meant.
+    this.verifyAgentIP = opts.verifyAgentIP || null;
+
+    // State and metering are interfaces, not implementations. Swap either for
+    // a shared or durable one without the gate noticing. See MemoryStore.
+    this.store = opts.store || new MemoryStore({ maxTracked: this.maxTracked });
+    this.sink = opts.sink || new DirectSink(this.onEvent);
+
+    // Identifies events from this process so a retrying sink can dedup. A
+    // counter alone would collide across instances; a timestamp alone would
+    // collide within a millisecond.
+    this._instance = randomUUID();
+    this._seq = 0;
   }
 
   /** req: { method, path, authority, headers, ip } → decision */
   handle(req, now = Math.floor(Date.now() / 1000), paymentProof = '') {
-    const c = classify(req, this.directories, now);
+    const c = classify(req, this.directories, now,
+                       { verifyAgentIP: this.verifyAgentIP, ip: this._client(req) });
     const { lane, agentId, evidence } = c;
     const ident = agentId || `${lane}:${this._client(req)}`;
     const proof = paymentProof || headerValue(req.headers, 'x-payment-proof');
     const d = this._decide(req, lane, ident, now, proof, c);
-    const entry = { t: now, path: req.path, lane, identity: ident,
+    // v is the event schema version. A meter that outlives one release has to
+    // know which shape it is reading; adding this after receipts exist in the
+    // wild would mean reconciling two formats forever.
+    const entry = { v: 1, idempotencyKey: `${this._instance}:${this._seq++}`,
+                    t: now, path: req.path, lane, identity: ident,
                     status: d.status, why: d.why, evidence,
                     billedUsd: d.billed || 0 };
     if (d.paymentRef) entry.paymentRef = d.paymentRef;
-    try { this.onEvent(entry); } catch { /* metering must never break serving */ }
+    this.sink.emit(entry);
     return { lane, identity: ident, ...d, evidence };
   }
 
@@ -329,26 +465,9 @@ export class Wayleave {
     catch { return false; }
   }
 
-  /**
-   * Fixed-window counting. On rollover the whole table is dropped — last
-   * window's counts are meaningless anyway, and it keeps memory bounded to
-   * one window's worth of identities instead of growing forever.
-   */
+  /** Fixed-window counting, delegated to whichever store is configured. */
   _bump(ident, now) {
-    const w = Math.floor(now / this.rateWindow);
-    if (w !== this._window) {
-      this._window = w;
-      this._hits.clear();
-      for (const [n, exp] of this._seen) if (exp <= now) this._seen.delete(n);
-    }
-    // Hard ceiling so a flood of distinct identities inside a single window
-    // cannot exhaust memory. Evicting the oldest keeps the newest arrivals
-    // counted; the durable fix is shared state, not a bigger Map.
-    if (this._hits.size >= this.maxTracked && !this._hits.has(ident))
-      this._hits.delete(this._hits.keys().next().value);
-    const n = (this._hits.get(ident) || 0) + 1;
-    this._hits.set(ident, n);
-    return n;
+    return this.store.hit(ident, this.rateWindow, now);
   }
 
   _decide(req, lane, ident, now, paymentProof, c) {
@@ -363,11 +482,9 @@ export class Wayleave {
     // Only signers that send a nonce can be held to single use; see
     // verifySignature for why a signature hash is not a safe substitute.
     if (this.replayProtection && c && c.nonce) {
-      if (this._seen.has(c.nonce))
+      if (this.store.hasNonce(c.nonce))
         return { status: 403, why: 'signature already used (replay)' };
-      if (this._seen.size >= this.maxTracked)
-        this._seen.delete(this._seen.keys().next().value);
-      this._seen.set(c.nonce, c.sigExpires || now + this.rateWindow);
+      this.store.rememberNonce(c.nonce, c.sigExpires || now + this.rateWindow);
     }
     const limit = this.rateLimits[lane];
     if (limit) {

package/meter.d.ts

@@ -0,0 +1,74 @@
+import type { MeterEvent, Sink } from './index.js';
+
+export interface MeterSinkOptions {
+  /** Base URL of your meter, e.g. "https://meter.wayleave.dev". */
+  endpoint: string;
+  /** API key issued by the meter. Sent as a bearer token. */
+  apiKey: string;
+  /** Drain interval in ms. Default 5000. */
+  flushMs?: number;
+  /** Events per request. Default 500; the meter caps batches at 1000. */
+  maxBatch?: number;
+  /**
+   * Hard ceiling on buffered events. Default 10000. Past this the oldest are
+   * shed — counted in `stats.dropped` and reported through `onError`, never
+   * dropped silently.
+   */
+  maxBuffer?: number;
+  /** Retries per batch, exponential backoff with jitter. Default 5. */
+  maxRetries?: number;
+  /** Per-request timeout in ms. Default 10000. */
+  timeoutMs?: number;
+  /**
+   * Observability hook for drops, permanent rejections, and unreachable
+   * meters. Anything it throws is swallowed — a broken error handler must not
+   * reach the request path either.
+   */
+  onError?: (err: Error, info: Record<string, unknown>) => void;
+  /** Injectable for tests. Defaults to global fetch. */
+  fetch?: typeof fetch;
+}
+
+export interface MeterSinkStats {
+  queued: number;
+  sent: number;
+  /** Shed from a full buffer, or permanently rejected by the meter. */
+  dropped: number;
+  failedBatches: number;
+  /** Reported by the meter as already-seen. Retries are supposed to produce these. */
+  duplicates: number;
+}
+
+/**
+ * A `Sink` that buffers, batches, and retries against a wayleave meter.
+ *
+ * Retrying is safe because every event carries an `idempotencyKey` and the
+ * meter deduplicates on it: sending a batch twice costs a round trip, not a
+ * double-billed customer.
+ *
+ * ```js
+ * import Wayleave from 'wayleave';
+ * import { MeterSink } from 'wayleave/meter';
+ *
+ * const sink = new MeterSink({
+ *   endpoint: 'https://meter.wayleave.dev',
+ *   apiKey: process.env.WAYLEAVE_METER_KEY,
+ *   onError: (err, info) => log.warn({ err, info }, 'meter'),
+ * });
+ *
+ * const gate = new Wayleave({ pricedPaths: { '/api/premium': 0.05 }, sink });
+ * process.on('SIGTERM', () => sink.close());
+ * ```
+ */
+export class MeterSink implements Sink {
+  constructor(opts: MeterSinkOptions);
+  readonly stats: MeterSinkStats;
+  /** Synchronous, non-throwing, no I/O. Safe on the request path. */
+  emit(event: MeterEvent): void;
+  /** Drain the buffer. Never rejects; concurrent calls collapse into one. */
+  flush(): Promise<void>;
+  /** Stop accepting events and make a final drain. Call on SIGTERM. */
+  close(): Promise<MeterSinkStats>;
+}
+
+export default MeterSink;

package/meter.js

@@ -0,0 +1,230 @@
+/**
+ * MeterSink — ships metering events to a wayleave meter.
+ *
+ * This is the concrete `sink` that 0.2.0's interface was shaped for. The
+ * default DirectSink hands each event straight to `onEvent` and swallows what
+ * it throws: correct for uptime, lossy for revenue. This one buffers, batches,
+ * and retries, so a meter that is briefly unreachable costs you nothing.
+ *
+ * Three rules govern every line in this file, in this order:
+ *
+ *   1. NEVER throw into the request path. A billing backend having a bad day
+ *      must not become your customers having a bad day. `emit` cannot throw,
+ *      cannot await, and cannot block.
+ *
+ *   2. NEVER lose an event silently. Dropping is sometimes unavoidable — a
+ *      buffer that grows forever is an outage, not a feature — but a drop is
+ *      always counted and always reported.
+ *
+ *   3. Retry freely. Every event carries an `idempotencyKey`, and the meter
+ *      deduplicates on it. Sending the same batch twice costs a round trip,
+ *      not a double-billed customer. This is the entire reason that field was
+ *      added before there was anything to receive it.
+ *
+ * Zero dependencies: `fetch` is global from Node 18.
+ */
+
+const DEFAULTS = {
+  flushMs: 5000,
+  maxBatch: 500,
+  maxBuffer: 10000,
+  maxRetries: 5,
+  timeoutMs: 10000,
+};
+
+export class MeterSink {
+  /**
+   * opts = {
+   *   endpoint: 'https://meter.wayleave.dev',   // required
+   *   apiKey:   'wm_live_...',                  // required
+   *   flushMs:   5000,   // how often to drain the buffer
+   *   maxBatch:  500,    // events per request; the meter caps at 1000
+   *   maxBuffer: 10000,  // hard ceiling before shedding
+   *   maxRetries: 5,     // per batch, exponential backoff
+   *   timeoutMs: 10000,
+   *   onError: (err, info) => {}   // observability hook; never called with
+   *                                // anything that should crash your process
+   * }
+   */
+  constructor(opts = {}) {
+    if (!opts.endpoint) throw new Error('MeterSink requires an endpoint');
+    if (!opts.apiKey) throw new Error('MeterSink requires an apiKey');
+
+    this.endpoint = String(opts.endpoint).replace(/\/$/, '');
+    this.apiKey = opts.apiKey;
+    this.flushMs = opts.flushMs ?? DEFAULTS.flushMs;
+    this.maxBatch = opts.maxBatch ?? DEFAULTS.maxBatch;
+    this.maxBuffer = opts.maxBuffer ?? DEFAULTS.maxBuffer;
+    this.maxRetries = opts.maxRetries ?? DEFAULTS.maxRetries;
+    this.timeoutMs = opts.timeoutMs ?? DEFAULTS.timeoutMs;
+    this.onError = typeof opts.onError === 'function' ? opts.onError : () => {};
+    this._fetch = opts.fetch || globalThis.fetch;
+
+    this._buffer = [];
+    this._inFlight = false;
+    this._closed = false;
+
+    /** Counters worth exposing: a silent drop is a billing hole. */
+    this.stats = { queued: 0, sent: 0, dropped: 0, failedBatches: 0, duplicates: 0 };
+
+    this._timer = setInterval(() => { this.flush(); }, this.flushMs);
+    // Do not hold the process open just to run a billing timer.
+    if (typeof this._timer.unref === 'function') this._timer.unref();
+  }
+
+  /**
+   * Called by the gate, once per request, on the hot path.
+   *
+   * Synchronous, non-throwing, and does no I/O. The only work here is a push
+   * and possibly scheduling a flush on a later tick.
+   */
+  emit(event) {
+    if (this._closed) return;
+    try {
+      if (this._buffer.length >= this.maxBuffer) {
+        // Shed the oldest. They have been failing longest and are the most
+        // likely to be stale; the newest are the most likely to succeed on
+        // the next attempt. Either choice loses billing data, which is why
+        // this is counted and surfaced rather than quietly absorbed.
+        const shed = this._buffer.shift();
+        this.stats.dropped++;
+        this.onError(new Error('meter buffer full, dropped oldest event'), {
+          dropped: this.stats.dropped, idempotencyKey: shed?.idempotencyKey,
+        });
+      }
+      this._buffer.push(event);
+      this.stats.queued++;
+
+      // Schedule rather than await: emit must return to the request handler
+      // immediately.
+      if (this._buffer.length >= this.maxBatch && !this._inFlight)
+        queueMicrotask(() => this.flush());
+    } catch (err) {
+      // Even the buffering path must not throw into serving.
+      try { this.onError(err, { phase: 'emit' }); } catch { /* nothing left to do */ }
+    }
+  }
+
+  /**
+   * Drain the buffer. Safe to call at any time; concurrent calls collapse
+   * into the one already running.
+   *
+   * Never rejects. Failures are reported through `onError` and the events go
+   * back on the buffer for the next attempt.
+   */
+  async flush() {
+    if (this._inFlight || this._buffer.length === 0) return;
+    this._inFlight = true;
+    try {
+      while (this._buffer.length > 0) {
+        const batch = this._buffer.splice(0, this.maxBatch);
+        const ok = await this._sendWithRetry(batch);
+        if (!ok) {
+          // Put them back at the FRONT so ordering is preserved, and stop —
+          // hammering a meter that is down helps nobody.
+          this._buffer.unshift(...batch.slice(0, Math.max(0, this.maxBuffer - this._buffer.length)));
+          break;
+        }
+      }
+    } catch (err) {
+      try { this.onError(err, { phase: 'flush' }); } catch { /* ignore */ }
+    } finally {
+      this._inFlight = false;
+    }
+  }
+
+  async _sendWithRetry(batch) {
+    for (let attempt = 0; attempt <= this.maxRetries; attempt++) {
+      const r = await this._send(batch);
+
+      if (r.ok) {
+        this.stats.sent += r.accepted ?? batch.length;
+        this.stats.duplicates += r.duplicates ?? 0;
+        if (r.rejected?.length) {
+          // The meter refused specific events permanently — a schema mismatch,
+          // usually. Retrying cannot fix them and would wedge everything
+          // queued behind. Report and move on.
+          this.onError(new Error(`meter rejected ${r.rejected.length} event(s)`), {
+            phase: 'send', rejected: r.rejected,
+          });
+        }
+        return true;
+      }
+
+      // 4xx other than 429 is our bug, not a blip. Retrying a malformed or
+      // unauthorised request forever is how a buffer becomes a memory leak.
+      if (r.status && r.status >= 400 && r.status < 500 && r.status !== 429) {
+        this.stats.failedBatches++;
+        this.stats.dropped += batch.length;
+        this.onError(new Error(`meter rejected batch: ${r.status} ${r.error ?? ''}`.trim()), {
+          phase: 'send', status: r.status, dropped: batch.length, permanent: true,
+        });
+        return true; // consumed: do not requeue something that can never succeed
+      }
+
+      if (attempt < this.maxRetries) {
+        // Exponential backoff with jitter, so a fleet of instances coming back
+        // after an outage does not stampede the meter in lockstep.
+        const base = Math.min(30000, 500 * 2 ** attempt);
+        await sleep(base / 2 + Math.random() * base / 2);
+      }
+    }
+
+    this.stats.failedBatches++;
+    this.onError(new Error('meter unreachable after retries; events requeued'), {
+      phase: 'send', batch: batch.length,
+    });
+    return false;
+  }
+
+  async _send(batch) {
+    const ctrl = new AbortController();
+    const timer = setTimeout(() => ctrl.abort(), this.timeoutMs);
+    try {
+      const res = await this._fetch(`${this.endpoint}/v1/events`, {
+        method: 'POST',
+        headers: {
+          'content-type': 'application/json',
+          authorization: `Bearer ${this.apiKey}`,
+        },
+        body: JSON.stringify({ events: batch }),
+        signal: ctrl.signal,
+      });
+      let body = null;
+      try { body = await res.json(); } catch { /* non-JSON body stays null */ }
+      if (!res.ok) return { ok: false, status: res.status, error: body?.error };
+      return { ok: true, ...body };
+    } catch (err) {
+      return { ok: false, status: 0, error: err.message };
+    } finally {
+      clearTimeout(timer);
+    }
+  }
+
+  /**
+   * Stop accepting events and make a final attempt to drain.
+   *
+   * Call this on SIGTERM. Without it, whatever is buffered at shutdown is
+   * revenue you never billed for. With it, a rolling deploy costs nothing.
+   */
+  async close() {
+    this._closed = true;
+    clearInterval(this._timer);
+    await this.flush();
+    return this.stats;
+  }
+}
+
+/**
+ * Deliberately NOT unref'd, unlike the flush interval.
+ *
+ * The interval is a heartbeat and must not hold a process open. A backoff is
+ * an operation already in flight: unref'ing it lets the event loop drain
+ * mid-retry, which abandons events that were about to be delivered — during
+ * exactly the shutdown where `close()` is trying to save them.
+ */
+function sleep(ms) {
+  return new Promise(r => setTimeout(r, ms));
+}
+
+export default MeterSink;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "wayleave",
-  "version": "0.1.6",
+  "version": "0.2.1",
   "description": "Charge AI agents for passage across your app. Verify agent signatures (Web Bot Auth / Ed25519), classify every request, and return 402 with a price on routes you choose. Humans browse free.",
   "type": "module",
   "main": "index.js",
@@ -9,16 +9,22 @@
     ".": {
       "types": "./index.d.ts",
       "default": "./index.js"
+    },
+    "./meter": {
+      "types": "./meter.d.ts",
+      "default": "./meter.js"
     }
   },
   "files": [
-    "index.js",
-    "index.d.ts",
+    "LICENSE",
     "README.md",
-    "LICENSE"
+    "index.d.ts",
+    "index.js",
+    "meter.d.ts",
+    "meter.js"
   ],
   "scripts": {
-    "test": "node --test"
+    "test": "node --test *.test.js"
   },
   "engines": {
     "node": ">=18"