@capybara-hq/kernel 0.1.0-rc.8 → 0.1.0-rc.9

package/dist/devices/device-gateway.d.ts

@@ -0,0 +1,128 @@
+import { type ExtensionContext } from "@capybara-hq/plugin-sdk";
+import type { WebSocket } from "@fastify/websocket";
+import type { Redis } from "ioredis";
+import { KernelConfigService } from "../infra/config/kernel-config.service.js";
+import { LockService } from "../infra/lock/lock.service.js";
+import type { KernelLogger } from "../log.js";
+import { type MachineIdentity, MachinesService } from "./machines.js";
+/** WebSocket close codes the gateway uses (the program reads 4002 and 4003). */
+export declare const DEVICE_CLOSE: {
+    /** The kernel is shutting down: reconnect (another gateway replica takes over). */
+    readonly shutdown: 1001;
+    /** A frame larger than `environments.device.maxFrameBytes` (sent by `ws` itself). */
+    readonly tooBig: 1009;
+    /** The kernel failed (e.g. the database): reconnect. */
+    readonly internal: 1011;
+    /** The lease was lost (renewal kept failing or the key was taken): reconnect. */
+    readonly leaseLost: 4001;
+    /** A newer connection of the environment took over: the program exits. */
+    readonly replaced: 4002;
+    /** The machine's credential was revoked: the program exits. */
+    readonly revoked: 4003;
+    /** No valid `hello` first, or a frame that is not protocol v1 from a device. */
+    readonly protocol: 4400;
+    /** The lease could not be claimed in time: reconnect. */
+    readonly claimTimeout: 4409;
+};
+/** How long the program has to send `hello` after the upgrade. */
+export declare const HELLO_TIMEOUT_MS = 10000;
+/** How often a connection is pinged; two pings without a pong close it. */
+export declare const PING_INTERVAL_MS = 20000;
+/** How long a new connection retries the lease after asking the holder to let go. */
+export declare const CLAIM_TIMEOUT_MS = 5000;
+/** The pause between two tries at the lease while claiming. */
+export declare const CLAIM_RETRY_MS = 250;
+type State = "hello" | "claiming" | "holding" | "closed";
+/**
+ * One machine connection on this replica. Its state (the socket, the lease handle, the requests
+ * it forwarded) is the socket's own and lives exactly as long as the socket.
+ */
+declare class DeviceConnection {
+    private readonly gw;
+    private readonly socket;
+    readonly identity: MachineIdentity;
+    private readonly credential;
+    readonly token: `${string}-${string}-${string}-${string}-${string}`;
+    state: State;
+    private lease;
+    private readonly unsubscribes;
+    private helloTimer;
+    private pingTimer;
+    private missedPongs;
+    private lastTouch;
+    private lastAuth;
+    /** The drains (and the close after a revocation) run one after another. */
+    private chain;
+    /** Resolves once the lease is handed back (or was never held). */
+    private released;
+    constructor(gw: DeviceGateway, socket: WebSocket, identity: MachineIdentity, credential: string);
+    get environmentId(): string;
+    start(): void;
+    private onFrame;
+    /**
+     * Stores one result and announces it (the relay's wire, relay.ts), only for a request taken for
+     * this environment (ruling R20); anything else is dropped.
+     */
+    private store;
+    /**
+     * Claims the environment's lease; true when this connection holds it. The newest connection
+     * wins (ruling R21): when another connection holds the lease, a takeover is published (the
+     * holder yields to any takeover) and the lease retried every {@link CLAIM_RETRY_MS} for
+     * {@link CLAIM_TIMEOUT_MS}, then 4409 (the program reconnects). A claimer never yields.
+     */
+    private claim;
+    /** Subscribes for this connection; a subscription that completes after the close is ended. */
+    private subscribe;
+    /** Queues one drain of the request list after the ones already queued. */
+    private drain;
+    private drainOnce;
+    /** The next ping (a chain of timeouts; the lease renewal is LockService's, not this timer's). */
+    private schedulePing;
+    private ping;
+    /** The credential still stands (membership, org and environment are read live). */
+    private reauthenticate;
+    private traffic;
+    /** Closes the socket with `code` and hands the lease back; idempotent. */
+    close(code: number, reason: string): Promise<void>;
+    /** Stops timers and subscriptions and releases the lease (a lost one is abandoned). */
+    shutdown(): Promise<void>;
+}
+/**
+ * The gateway role's half of the device relay (environments spec §13.4): holds the WebSocket
+ * connections of connected machines (`GET /device/connect`, device-routes.ts). One connection per
+ * environment across replicas, under the lease `capk:device-lease:<environmentId>` (the newest
+ * connection wins: it publishes `device.takeover.<environmentId>` and the holder lets go). The
+ * holder drains `capk:device-req:<environmentId>` on every `device.req.<environmentId>` nudge,
+ * once after claiming and on every ping, marking each forwarded request taken for its environment
+ * (`capk:device-taken:<id>`, TTL the reconnect grace plus the request's deadline plus a margin).
+ * A `res` is accepted only for a request taken for the socket's environment: it is stored under
+ * `capk:device-res:<id>` for the rest of that TTL and published on `device.res.<id>`.
+ * `device.revoked.<environmentId>` closes the connection of that machine only (ruling R16). The
+ * sockets are this process's own; everything other replicas need is in Redis.
+ */
+export declare class DeviceGateway {
+    readonly machines: Pick<MachinesService, "authenticate" | "recordHello" | "touch">;
+    readonly locks: Pick<LockService, "hold">;
+    readonly bus: ExtensionContext["bus"];
+    readonly redis: Redis;
+    private readonly config;
+    readonly log: Pick<KernelLogger, "info" | "warn">;
+    /** This replica's open connections (the sockets cannot live anywhere else). */
+    private readonly connections;
+    private stopping;
+    /** How often connections are pinged ({@link PING_INTERVAL_MS}; a test seam). */
+    pingIntervalMs: number;
+    constructor(machines: Pick<MachinesService, "authenticate" | "recordHello" | "touch">, locks: Pick<LockService, "hold">, bus: ExtensionContext["bus"], redis: Redis, config: KernelConfigService, log: Pick<KernelLogger, "info" | "warn">);
+    graceMs(): number;
+    leaseTtlMs(): number;
+    maxFrameBytes(): number;
+    /** Who a `Bearer` machine credential stands for, or null (checked before the upgrade). */
+    authenticate(credential: string): Promise<MachineIdentity | null>;
+    /** Serves one upgraded socket of an authenticated machine until it closes. */
+    accept(socket: WebSocket, identity: MachineIdentity, credential: string): void;
+    /** @internal a connection that ended. */
+    forget(conn: DeviceConnection): void;
+    /** Closes every connection (1001) and hands their leases back. */
+    stop(): Promise<void>;
+}
+export {};

package/dist/devices/device-gateway.js

@@ -0,0 +1,530 @@
+var __decorate = (this && this.__decorate) || function (decorators, target, key, desc) {
+    var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
+    if (typeof Reflect === "object" && typeof Reflect.decorate === "function") r = Reflect.decorate(decorators, target, key, desc);
+    else for (var i = decorators.length - 1; i >= 0; i--) if (d = decorators[i]) r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
+    return c > 3 && r && Object.defineProperty(target, key, r), r;
+};
+var __param = (this && this.__param) || function (paramIndex, decorator) {
+    return function (target, key) { decorator(target, key, paramIndex); }
+};
+import { randomUUID } from "node:crypto";
+import { DEVICE_PROTOCOL_VERSION, parseDeviceFrame, } from "@capybara-hq/plugin-sdk";
+import { Inject, Injectable } from "@nestjs/common";
+import { KernelConfigService } from "../infra/config/kernel-config.service.js";
+import { LockService } from "../infra/lock/lock.service.js";
+import { KERNEL_BUS, KERNEL_LOGGER, REDIS } from "../infra/tokens.js";
+import { DEVICE_TAKEN_PREFIX, deviceLeaseKey, deviceReqKey, deviceReqTopic, deviceResKey, deviceResTopic, deviceRevokedTopic, deviceTakenKey, deviceTakeoverTopic, } from "./keys.js";
+import { MachinesService } from "./machines.js";
+import { LIST_TTL_MARGIN_MS } from "./relay.js";
+/** WebSocket close codes the gateway uses (the program reads 4002 and 4003). */
+export const DEVICE_CLOSE = {
+    /** The kernel is shutting down: reconnect (another gateway replica takes over). */
+    shutdown: 1001,
+    /** A frame larger than `environments.device.maxFrameBytes` (sent by `ws` itself). */
+    tooBig: 1009,
+    /** The kernel failed (e.g. the database): reconnect. */
+    internal: 1011,
+    /** The lease was lost (renewal kept failing or the key was taken): reconnect. */
+    leaseLost: 4001,
+    /** A newer connection of the environment took over: the program exits. */
+    replaced: 4002,
+    /** The machine's credential was revoked: the program exits. */
+    revoked: 4003,
+    /** No valid `hello` first, or a frame that is not protocol v1 from a device. */
+    protocol: 4400,
+    /** The lease could not be claimed in time: reconnect. */
+    claimTimeout: 4409,
+};
+/** How long the program has to send `hello` after the upgrade. */
+export const HELLO_TIMEOUT_MS = 10_000;
+/** How often a connection is pinged; two pings without a pong close it. */
+export const PING_INTERVAL_MS = 20_000;
+/** How long a new connection retries the lease after asking the holder to let go. */
+export const CLAIM_TIMEOUT_MS = 5_000;
+/** The pause between two tries at the lease while claiming. */
+export const CLAIM_RETRY_MS = 250;
+/** Traffic marks the machine seen at most this often per connection (the row has its own gate). */
+const TOUCH_EVERY_MS = 30_000;
+/** How often a held connection re-checks its credential (membership, org, environment). */
+const REAUTH_EVERY_MS = 60_000;
+/**
+ * Puts a frame the holder popped but could not send back at the head of the list, keeping the
+ * list's TTL at least `ARGV[2]` ms (the list never exists without one).
+ */
+const PUT_BACK = `
+redis.call('LPUSH', KEYS[1], ARGV[1])
+if redis.call('PTTL', KEYS[1]) < tonumber(ARGV[2]) then
+  redis.call('PEXPIRE', KEYS[1], ARGV[2])
+end
+return 1
+`;
+/**
+ * Pops the next frame of the request list (`KEYS[1]`) and, when it is a `req`, marks it taken in
+ * the same step: `<ARGV[1]><id>` → the environment id `ARGV[2]`, for the reconnect grace
+ * `ARGV[3]` ms plus the request's deadline plus the margin `ARGV[4]` ms (ruling R20). Returns the
+ * frame, or false when the list is empty. A frame that is not JSON is returned as is (the caller
+ * drops it). The taken key cannot be declared in KEYS (its id is known only after the pop): like
+ * the rest of the kernel, this assumes a single-node Valkey / Redis, not a cluster.
+ */
+const TAKE = `
+local raw = redis.call('LPOP', KEYS[1])
+if not raw then return false end
+local ok, f = pcall(cjson.decode, raw)
+if ok and type(f) == 'table' and f.type == 'req' and type(f.id) == 'string'
+  and type(f.deadlineMs) == 'number' and f.deadlineMs > 0 then
+  local ttl = tonumber(ARGV[3]) + math.ceil(f.deadlineMs) + tonumber(ARGV[4])
+  redis.call('SET', ARGV[1] .. f.id, ARGV[2], 'PX', ttl)
+end
+return raw
+`;
+/**
+ * Accepts a result (ruling R20): only when its request is marked taken (`KEYS[1]`) for the
+ * environment `ARGV[1]`. Stores the frame `ARGV[2]` under `KEYS[2]` for what is left of the taken
+ * mark's TTL (the caller's whole window, also after a reconnect) and clears the mark. Returns 1
+ * when stored, 0 otherwise.
+ */
+const ACCEPT = `
+if redis.call('GET', KEYS[1]) ~= ARGV[1] then return 0 end
+local ttl = redis.call('PTTL', KEYS[1])
+if ttl <= 0 then return 0 end
+redis.call('SET', KEYS[2], ARGV[2], 'PX', ttl)
+redis.call('DEL', KEYS[1])
+return 1
+`;
+const errMsg = (e) => (e instanceof Error ? e.message : String(e));
+/**
+ * One machine connection on this replica. Its state (the socket, the lease handle, the requests
+ * it forwarded) is the socket's own and lives exactly as long as the socket.
+ */
+class DeviceConnection {
+    gw;
+    socket;
+    identity;
+    credential;
+    token = randomUUID();
+    state = "hello";
+    lease;
+    unsubscribes = [];
+    helloTimer;
+    pingTimer;
+    missedPongs = 0;
+    lastTouch = 0;
+    lastAuth = Date.now();
+    /** The drains (and the close after a revocation) run one after another. */
+    chain = Promise.resolve();
+    /** Resolves once the lease is handed back (or was never held). */
+    released;
+    constructor(gw, socket, identity, credential) {
+        this.gw = gw;
+        this.socket = socket;
+        this.identity = identity;
+        this.credential = credential;
+    }
+    get environmentId() {
+        return this.identity.environmentId;
+    }
+    start() {
+        const { socket } = this;
+        this.helloTimer = setTimeout(() => this.close(DEVICE_CLOSE.protocol, "hello expected"), HELLO_TIMEOUT_MS);
+        this.schedulePing();
+        socket.on("pong", () => {
+            this.missedPongs = 0;
+            this.traffic();
+        });
+        socket.on("message", (data, isBinary) => {
+            if (this.state === "closed")
+                return;
+            this.traffic();
+            if (isBinary) {
+                this.close(DEVICE_CLOSE.protocol, "frames are JSON text");
+                return;
+            }
+            const text = Array.isArray(data)
+                ? Buffer.concat(data).toString("utf8")
+                : Buffer.isBuffer(data)
+                    ? data.toString("utf8")
+                    : Buffer.from(data).toString("utf8");
+            void this.onFrame(text);
+        });
+        socket.on("close", () => {
+            void this.shutdown();
+        });
+    }
+    async onFrame(text) {
+        let frame;
+        try {
+            frame = parseDeviceFrame(text, "device");
+        }
+        catch (e) {
+            this.gw.log.warn("device sent an invalid frame", {
+                environmentId: this.environmentId,
+                error: errMsg(e),
+            });
+            this.close(DEVICE_CLOSE.protocol, "invalid frame");
+            return;
+        }
+        if (frame.type === "hello") {
+            if (this.state !== "hello") {
+                this.close(DEVICE_CLOSE.protocol, "hello sent twice");
+                return;
+            }
+            if (frame.protocol !== DEVICE_PROTOCOL_VERSION) {
+                this.close(DEVICE_CLOSE.protocol, "unsupported protocol");
+                return;
+            }
+            clearTimeout(this.helloTimer);
+            this.state = "claiming";
+            if (!(await this.claim()))
+                return;
+            try {
+                await this.gw.machines.recordHello(this.identity.machineId, frame);
+            }
+            catch (e) {
+                this.gw.log.warn("could not record a device hello", {
+                    environmentId: this.environmentId,
+                    error: errMsg(e),
+                });
+            }
+            return;
+        }
+        if (frame.type !== "res" || this.state === "hello") {
+            this.close(DEVICE_CLOSE.protocol, "hello expected");
+            return;
+        }
+        // results are stored in any state after hello: the program sends those it kept while
+        // disconnected right after its hello, before the lease is claimed
+        await this.store(frame);
+    }
+    /**
+     * Stores one result and announces it (the relay's wire, relay.ts), only for a request taken for
+     * this environment (ruling R20); anything else is dropped.
+     */
+    async store(res) {
+        try {
+            const stored = await this.gw.redis.eval(ACCEPT, 2, deviceTakenKey(res.id), deviceResKey(res.id), this.environmentId, JSON.stringify(res));
+            if (Number(stored) !== 1) {
+                this.gw.log.warn("dropping a device result for a request not taken for its environment", {
+                    environmentId: this.environmentId,
+                });
+                return;
+            }
+            await this.gw.bus.publish(deviceResTopic(res.id), res);
+        }
+        catch (e) {
+            this.gw.log.warn("could not store a device result", {
+                environmentId: this.environmentId,
+                error: errMsg(e),
+            });
+        }
+    }
+    /**
+     * Claims the environment's lease; true when this connection holds it. The newest connection
+     * wins (ruling R21): when another connection holds the lease, a takeover is published (the
+     * holder yields to any takeover) and the lease retried every {@link CLAIM_RETRY_MS} for
+     * {@link CLAIM_TIMEOUT_MS}, then 4409 (the program reconnects). A claimer never yields.
+     */
+    async claim() {
+        const env = this.environmentId;
+        try {
+            // subscribed before claiming, so a takeover right after the claim is not missed
+            await this.subscribe(deviceTakeoverTopic(env), (p) => {
+                const token = p?.token;
+                if (this.state === "holding" && token !== this.token) {
+                    this.close(DEVICE_CLOSE.replaced, "replaced by a newer connection");
+                }
+            });
+            if (this.state !== "claiming")
+                return false;
+            const ttl = this.gw.leaseTtlMs();
+            let lease = await this.gw.locks.hold(deviceLeaseKey(env), this.token, ttl);
+            if (!lease) {
+                await this.gw.bus.publish(deviceTakeoverTopic(env), { token: this.token });
+                const until = Date.now() + CLAIM_TIMEOUT_MS;
+                while (!lease && this.state === "claiming" && Date.now() < until) {
+                    await new Promise((r) => setTimeout(r, CLAIM_RETRY_MS));
+                    if (this.state !== "claiming")
+                        break;
+                    lease = await this.gw.locks.hold(deviceLeaseKey(env), this.token, ttl);
+                }
+            }
+            if (this.state !== "claiming") {
+                // closed while claiming (replaced, revoked, shutdown, the machine left)
+                await lease?.release();
+                return false;
+            }
+            if (!lease) {
+                this.close(DEVICE_CLOSE.claimTimeout, "the environment is held by another connection");
+                return false;
+            }
+            this.lease = lease;
+            this.state = "holding";
+            const onLost = () => this.close(DEVICE_CLOSE.leaseLost, "lease lost");
+            if (lease.lost.aborted) {
+                onLost();
+                return false;
+            }
+            lease.lost.addEventListener("abort", onLost, { once: true });
+            await this.subscribe(deviceReqTopic(env), () => this.drain());
+            await this.subscribe(deviceRevokedTopic(env), (p) => {
+                const id = p?.machineId;
+                if (id !== this.identity.machineId)
+                    return;
+                // one more drain first: a stop frame whose nudge was lost still goes out before the close
+                this.drain();
+                this.chain = this.chain.then(() => this.close(DEVICE_CLOSE.revoked, "the machine was disconnected"));
+            });
+            if (this.state !== "holding")
+                return false;
+            this.gw.log.info("device connected", {
+                environmentId: env,
+                machineId: this.identity.machineId,
+            });
+            // what was queued while nobody held the machine
+            this.drain();
+            return true;
+        }
+        catch (e) {
+            this.gw.log.warn("could not claim a device connection", {
+                environmentId: env,
+                error: errMsg(e),
+            });
+            this.close(DEVICE_CLOSE.internal, "internal error");
+            return false;
+        }
+    }
+    /** Subscribes for this connection; a subscription that completes after the close is ended. */
+    async subscribe(topic, handler) {
+        const unsubscribe = await this.gw.bus.subscribe(topic, handler);
+        if (this.state === "closed")
+            await unsubscribe().catch(() => { });
+        else
+            this.unsubscribes.push(unsubscribe);
+    }
+    /** Queues one drain of the request list after the ones already queued. */
+    drain() {
+        this.chain = this.chain
+            .then(() => this.drainOnce())
+            .catch((e) => {
+            this.gw.log.warn("device request drain failed", {
+                environmentId: this.environmentId,
+                error: errMsg(e),
+            });
+        });
+    }
+    async drainOnce() {
+        const key = deviceReqKey(this.environmentId);
+        while (this.state === "holding") {
+            // no lease fencing needed: the pop is atomic, so each frame reaches one connection only, and
+            // a holder that lost its lease still talks to the environment's own machine (or puts it back)
+            const raw = (await this.gw.redis.eval(TAKE, 1, key, DEVICE_TAKEN_PREFIX, this.environmentId, String(this.gw.graceMs()), String(LIST_TTL_MARGIN_MS)));
+            if (raw === null)
+                return;
+            let frame;
+            try {
+                frame = parseDeviceFrame(raw, "kernel");
+            }
+            catch (e) {
+                this.gw.log.warn("dropping an invalid device request", {
+                    environmentId: this.environmentId,
+                    error: errMsg(e),
+                });
+                continue;
+            }
+            const ttlMs = this.gw.graceMs() + (frame.type === "req" ? frame.deadlineMs : 0) + LIST_TTL_MARGIN_MS;
+            if (this.state !== "holding" || this.socket.readyState !== this.socket.OPEN) {
+                // closed while popping: back to the head of the list for the next connection
+                await this.gw.redis.eval(PUT_BACK, 1, key, raw, String(ttlMs));
+                return;
+            }
+            if (frame.type === "req" && Buffer.byteLength(raw) > this.gw.maxFrameBytes()) {
+                await this.store({
+                    type: "res",
+                    id: frame.id,
+                    ok: false,
+                    error: {
+                        kind: "error",
+                        message: "The request is larger than environments.device.maxFrameBytes",
+                    },
+                });
+                continue;
+            }
+            this.socket.send(raw);
+        }
+    }
+    /** The next ping (a chain of timeouts; the lease renewal is LockService's, not this timer's). */
+    schedulePing() {
+        this.pingTimer = setTimeout(() => {
+            this.ping();
+            if (this.state !== "closed")
+                this.schedulePing();
+        }, this.gw.pingIntervalMs);
+    }
+    ping() {
+        if (this.state === "closed")
+            return;
+        if (this.missedPongs >= 2) {
+            this.gw.log.info("device connection missed two pongs", { environmentId: this.environmentId });
+            this.socket.terminate();
+            void this.shutdown();
+            return;
+        }
+        this.missedPongs++;
+        this.socket.ping();
+        // a backstop for a lost nudge (the bus delivers at most once)
+        if (this.state === "holding")
+            this.drain();
+        if (this.state === "holding" && Date.now() - this.lastAuth >= REAUTH_EVERY_MS) {
+            this.lastAuth = Date.now();
+            void this.reauthenticate();
+        }
+    }
+    /** The credential still stands (membership, org and environment are read live). */
+    async reauthenticate() {
+        try {
+            const id = await this.gw.machines.authenticate(this.credential);
+            if (id === null || id.machineId !== this.identity.machineId) {
+                this.close(DEVICE_CLOSE.revoked, "the machine was disconnected");
+            }
+        }
+        catch (e) {
+            this.gw.log.warn("could not re-check a device credential", {
+                environmentId: this.environmentId,
+                error: errMsg(e),
+            });
+        }
+    }
+    traffic() {
+        const now = Date.now();
+        if (now - this.lastTouch < TOUCH_EVERY_MS)
+            return;
+        this.lastTouch = now;
+        this.gw.machines.touch(this.identity.machineId).catch((e) => {
+            this.gw.log.warn("could not mark a machine seen", {
+                environmentId: this.environmentId,
+                error: errMsg(e),
+            });
+        });
+    }
+    /** Closes the socket with `code` and hands the lease back; idempotent. */
+    close(code, reason) {
+        if (this.state !== "closed") {
+            try {
+                this.socket.close(code, reason);
+            }
+            catch {
+                this.socket.terminate();
+            }
+        }
+        return this.shutdown();
+    }
+    /** Stops timers and subscriptions and releases the lease (a lost one is abandoned). */
+    shutdown() {
+        if (this.released)
+            return this.released;
+        this.state = "closed";
+        clearTimeout(this.helloTimer);
+        clearTimeout(this.pingTimer);
+        const lease = this.lease;
+        this.released = (async () => {
+            try {
+                if (lease) {
+                    if (lease.lost.aborted)
+                        lease.abandon();
+                    else
+                        await lease.release();
+                }
+            }
+            catch (e) {
+                this.gw.log.warn("could not release a device lease", {
+                    environmentId: this.environmentId,
+                    error: errMsg(e),
+                });
+            }
+            await Promise.all(this.unsubscribes.splice(0).map((u) => u().catch(() => { })));
+            this.gw.forget(this);
+            if (lease) {
+                this.gw.log.info("device disconnected", {
+                    environmentId: this.environmentId,
+                    machineId: this.identity.machineId,
+                });
+            }
+        })();
+        return this.released;
+    }
+}
+/**
+ * The gateway role's half of the device relay (environments spec §13.4): holds the WebSocket
+ * connections of connected machines (`GET /device/connect`, device-routes.ts). One connection per
+ * environment across replicas, under the lease `capk:device-lease:<environmentId>` (the newest
+ * connection wins: it publishes `device.takeover.<environmentId>` and the holder lets go). The
+ * holder drains `capk:device-req:<environmentId>` on every `device.req.<environmentId>` nudge,
+ * once after claiming and on every ping, marking each forwarded request taken for its environment
+ * (`capk:device-taken:<id>`, TTL the reconnect grace plus the request's deadline plus a margin).
+ * A `res` is accepted only for a request taken for the socket's environment: it is stored under
+ * `capk:device-res:<id>` for the rest of that TTL and published on `device.res.<id>`.
+ * `device.revoked.<environmentId>` closes the connection of that machine only (ruling R16). The
+ * sockets are this process's own; everything other replicas need is in Redis.
+ */
+let DeviceGateway = class DeviceGateway {
+    machines;
+    locks;
+    bus;
+    redis;
+    config;
+    log;
+    /** This replica's open connections (the sockets cannot live anywhere else). */
+    connections = new Set();
+    stopping = false;
+    /** How often connections are pinged ({@link PING_INTERVAL_MS}; a test seam). */
+    pingIntervalMs = PING_INTERVAL_MS;
+    constructor(machines, locks, bus, redis, config, log) {
+        this.machines = machines;
+        this.locks = locks;
+        this.bus = bus;
+        this.redis = redis;
+        this.config = config;
+        this.log = log;
+    }
+    graceMs() {
+        return this.config.get().environments.device.reconnectGraceSeconds * 1000;
+    }
+    leaseTtlMs() {
+        return this.config.get().environments.device.leaseTtlMs;
+    }
+    maxFrameBytes() {
+        return this.config.get().environments.device.maxFrameBytes;
+    }
+    /** Who a `Bearer` machine credential stands for, or null (checked before the upgrade). */
+    authenticate(credential) {
+        return this.machines.authenticate(credential);
+    }
+    /** Serves one upgraded socket of an authenticated machine until it closes. */
+    accept(socket, identity, credential) {
+        if (this.stopping) {
+            socket.close(DEVICE_CLOSE.shutdown, "shutting down");
+            return;
+        }
+        const conn = new DeviceConnection(this, socket, identity, credential);
+        this.connections.add(conn);
+        conn.start();
+    }
+    /** @internal a connection that ended. */
+    forget(conn) {
+        this.connections.delete(conn);
+    }
+    /** Closes every connection (1001) and hands their leases back. */
+    async stop() {
+        this.stopping = true;
+        await Promise.all([...this.connections].map((c) => c.close(DEVICE_CLOSE.shutdown, "shutting down")));
+    }
+};
+DeviceGateway = __decorate([
+    Injectable(),
+    __param(0, Inject(MachinesService)),
+    __param(1, Inject(LockService)),
+    __param(2, Inject(KERNEL_BUS)),
+    __param(3, Inject(REDIS)),
+    __param(4, Inject(KernelConfigService)),
+    __param(5, Inject(KERNEL_LOGGER))
+], DeviceGateway);
+export { DeviceGateway };