Back off from a link that dies as soon as it comes up

A bind that returns is not proof the link works: an unreadable stream is
only found afterwards, so resetting the delay there handed a link that
died on arrival a fresh minDelay every cycle — one connect and bind per
second forever, which is how an account gets blocked for bind flooding.
The loop resets only once a link has outlasted maxDelay.

Also covers the unreadable stream that has no loop to retry it.

AGENTS.md

@@ -296,19 +296,24 @@ exactly 140.
   that loop before tearing down, so every deliberate shutdown emits `close`. Without the split an
   application that opens a replacement client on `close` ends up holding two binds on one account.
 
+- **Only a link that outlasted `maxDelay` resets the backoff.** Coming up is not proof it works: an
+  unreadable stream is found after the bind returns, so resetting there gave a link that died on
+  arrival a fresh `minDelay` every cycle — one TCP connect and bind per second, forever, which is how
+  an account gets blocked for bind flooding. `ReconnectLoop` records when it brought the owner up and
+  resets only if the link then lasted longer than the longest wait it would ever schedule. A drop
+  after a healthy link still retries at `minDelay`.
+
 - **A stream this library cannot read is a dead link, not a dead session.** Maintainer's call,
   2026-08-31: a framing or codec error tears the link down through `teardown()`, so the reconnect
   loop retries it on a fresh socket with a fresh framer — which is what a desynced stream needs, and
   the common cause. `sessionError` still carries every failure, so a peer that only ever sends
-  garbage is visible in the log rather than silent, and the backoff caps the retries at one per
-  `maxDelay`. Ending the session outright was inherited from when reconnect was opt-in, where the
-  distinction could not arise.
+  garbage is visible in the log rather than silent, and the backoff grows to one attempt per
+  `maxDelay`.
 
-- **`disconnected` counts failed links, not outages.** A retry that opens a socket and then loses its
-  bind re-enters `attach()` and so emits again, which makes it deliberately not one-to-one with
-  `reconnected`: each emission is a link that went down, and suppressing the later ones would leave a
-  failed rebind with nothing but a log line. The README says so, because the pairing is what a reader
-  would otherwise assume.
+- **`disconnected` counts failed links, not outages.** A retry that opens a socket and then loses it
+  clears `closed` through `attach()`, so the next `teardown()` emits again: each emission is a link
+  that went down, which makes the event deliberately not one-to-one with `reconnected`. Suppressing
+  the later ones would leave a failed rebind with nothing but a log line.
 
 - **`session.sock` is a getter over `PduTransport`, and stays public.** Maintainer's call, 2026-08-31:
   `session.ts` had reached its line cap, so the socket-to-PDU seam todo.md named was opened —

README.md

@@ -104,7 +104,7 @@ Every one is optional.
 | `shutdownTimeout` | `5000` | How long `close()` and `unbind()` wait for the requests this end already sent; `0` waits forever. |
 | `maxOutstanding` | `10` | Requests allowed on the wire at once; further sends queue. |
 | `smsIdFormat` | — | The notation the SMSC writes message ids in, per place it writes them: `{ receipt: 'decimal', submitResp: 'hex' }`. Only needed where the two disagree. |
-| `reconnect` | on | Re-binds after a drop or an idle timeout, backing off from `minDelay` 1 s to `maxDelay` 30 s. `{ minDelay, maxDelay }` retunes it; `false` turns it off, so a drop ends the session. |
+| `reconnect` | on | Re-binds after a drop, an idle timeout, or a stream the library cannot read, backing off from `minDelay` 1 s to `maxDelay` 30 s. `{ minDelay, maxDelay }` retunes it; `false` turns it off, so a drop ends the session. |
 | `log` | silent | Any object with `debug`, `error`, `info`, `verbose` and `warn` methods — see [Logging](#logging). |
 | `signal` | — | An `AbortSignal` that cancels connecting and tears the session down. |
 

src/pdu-transport.ts

@@ -14,7 +14,7 @@ export type PduTransportOptions = {
 	/** A complete PDU, before it is parsed. */
 	onFramed: (pdu: Buffer) => void;
 	onPdu: (pduObj: PduObject) => void;
-	/** Nothing further can be read off this stream; the socket has to go.  */
+	/** Nothing further can be read off this stream, whatever the socket does next. */
 	onUnreadable: (err: Error) => void;
 };
 

src/reconnect-loop.ts

@@ -7,19 +7,23 @@ export type ReconnectLoopOptions = {
 	log: SmppLog;
 	maxDelay: number;
 	minDelay: number;
+	now?: (() => number) | undefined;
 	/** Brings the owner back up on a freshly opened socket. An err means try again. */
 	onConnected: (sock: Socket) => Promise<VoidResult>;
 };
 
 /** Reopens a dropped connection, backing off between attempts until it is told to stop. */
 export class ReconnectLoop {
+	private readonly now: () => number;
 	private readonly options: ReconnectLoopOptions;
 	private attempting = false;
 	private delay: number;
 	private halted = false;
 	private timer: NodeJS.Timeout | undefined;
+	private upAt: number | undefined;
 
 	constructor(options: ReconnectLoopOptions) {
+		this.now = options.now ?? Date.now;
 		this.options = options;
 		this.delay = options.minDelay;
 	}
@@ -32,6 +36,13 @@ export class ReconnectLoop {
 	schedule(): void {
 		if (this.timer || this.attempting || this.isStopped()) return;
 
+		// Coming up is not proof: a stream we cannot read is only found once the link is bound.
+		if (this.upAt !== undefined && this.now() - this.upAt >= this.options.maxDelay) {
+			this.delay = this.options.minDelay;
+		}
+
+		this.upAt = undefined;
+
 		const delay = this.delay;
 
 		this.options.log.info('reconnect - retrying after a drop', { delay });
@@ -98,7 +109,7 @@ export class ReconnectLoop {
 			return true;
 		}
 
-		this.delay = this.options.minDelay;
+		this.upAt = this.now();
 
 		return false;
 	}

test/session-extras.test.ts

@@ -344,7 +344,7 @@ describe('reconnect', () => {
 
 	test('reports a drop it will retry as disconnected, keeping close for the end', async t => {
 		const smpp = await startServer(t);
-		const { session } = await connect(t, smpp);
+		const { session } = await connect(t, smpp, { reconnect: { maxDelay: 100, minDelay: 20 } });
 
 		assert.ok(session);
 
@@ -391,7 +391,7 @@ describe('reconnect', () => {
 
 	test('re-binds after a stream it cannot read, rather than ending the session', async t => {
 		const smpp = await startServer(t);
-		const { session } = await connect(t, smpp);
+		const { session } = await connect(t, smpp, { reconnect: { maxDelay: 100, minDelay: 20 } });
 
 		assert.ok(session);
 
@@ -448,7 +448,8 @@ describe('reconnect', () => {
 
 		const closed = once<true>(resolve => { session.on('close', () => { resolve(true); }); });
 
-		await peerOf(smpp).close();
+		// A cmd_length below the 16-octet header: unreadable, and nothing left to retry it.
+		peerOf(smpp).sock.write(Buffer.from([0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 0, 0, 0, 0, 1]));
 		await closed;
 
 		assert.equal(disconnects, 0);

test/session.test.ts

@@ -4,6 +4,7 @@ import test, { describe } from 'node:test';
 import type { Dlr } from '../src/dlr.ts';
 import type { PduObject, PduObjectInput } from '../src/pdu.ts';
 import type { Sms } from '../src/sms.ts';
+import type { SmppLog } from '../src/log.ts';
 import type { SmppServer } from '../src/server.ts';
 import type { TestContext } from 'node:test';
 import type { VoidResult } from '../src/result.ts';
@@ -1372,6 +1373,52 @@ describe('application hooks that throw or reject', () => {
 		assert.ok(retried, 'a throwing connect should be retried, not left for the process to die on');
 	});
 
+	test('keeps backing off when every link dies as soon as it comes up', async t => {
+		// A stream we cannot read is found after the bind, so a bind alone must not prove the link.
+		const clock = { now: 0 };
+		const delays: number[] = [];
+		const noop = (): void => undefined;
+		const log: SmppLog = {
+			debug: noop,
+			error: noop,
+			info: (msg, metadata) => {
+				if (msg === 'reconnect - retrying after a drop') delays.push(Number(metadata?.delay));
+			},
+			verbose: noop,
+			warn: noop,
+		};
+		let up = 0;
+		const loop = new ReconnectLoop({
+			connect: () => Promise.resolve({ sock: new net.Socket() }),
+			log,
+			maxDelay: 80,
+			minDelay: 10,
+			now: () => clock.now,
+			onConnected: () => {
+				up++;
+
+				return Promise.resolve({});
+			},
+		});
+
+		t.after(() => { loop.stop(); });
+
+		for (let died = 0; died < 4; died++) {
+			loop.schedule();
+			await waitFor(() => up === died + 1);
+			await delay(5);
+		}
+
+		assert.deepEqual(delays, [10, 20, 40, 80]);
+
+		// A link that outlasted the longest wait earned a fresh start.
+		clock.now += 80;
+		loop.schedule();
+		await waitFor(() => delays.length === 5);
+
+		assert.deepEqual(delays, [10, 20, 40, 80, 10]);
+	});
+
 	test('starts only one reconnect attempt at a time', async t => {
 		let attempts = 0;
 		let finish: (() => void) | undefined;