@capybara-hq/kernel 0.0.0-stage → 0.1.0-rc.1

package/dist/runtime/usage-report.js

@@ -0,0 +1,551 @@
+import { DelayedError, Job, Queue, UnrecoverableError, Worker } from "bullmq";
+import { sql } from "kysely";
+import { enabledExtensionsQuery } from "../extensions/enabled.js";
+import { deepFreeze, errMsg } from "../extensions/hooks.js";
+import { withJobContext } from "../log-context.js";
+import { bullConnection } from "./run-queue.js";
+/** Queue of `usage.report` deliveries (no `:` in the name). */
+export const USAGE_REPORT_QUEUE = "cap-kernel-usage-report";
+/** One job per `usage` event (no `:`: conversation ids are UUIDs, seq is a number). */
+export const usageReportJobId = (conversationId, seq) => `usage-${conversationId}-${seq}`;
+/** First attempt plus 5 retries (F11). */
+export const USAGE_REPORT_ATTEMPTS = 6;
+/** Retry backoff: 2 s, doubling (2, 4, 8, 16, 32 s). */
+export const USAGE_REPORT_BACKOFF_MS = 2000;
+/** Failed jobs (given up) are kept for inspection, at most {@link KEEP_FAILED_MAX} of them. */
+const KEEP_FAILED_S = 7 * 86_400;
+/** Upper bound on kept failed jobs (I-2): a long outage of a billing endpoint cannot grow Redis. */
+const KEEP_FAILED_MAX = 100_000;
+/** Warnings about a plugin that cannot take deliveries are logged at most once per this window. */
+const WARN_WINDOW_MS = 60_000;
+/**
+ * Upper bound on kept finished jobs (Redis memory). At the default retention it only bites above
+ * ~1.1 usage events per second sustained for a day; dedupe needs the jobs only for the sweeper overlap.
+ */
+const KEEP_COMPLETED_MAX = 100_000;
+/** Maintenance job of the sweeper (no `:`), its scheduler id and period. */
+export const USAGE_SWEEP_JOB = "usage-report-sweep";
+export const USAGE_SWEEP_SCHEDULER_ID = "usage-report-sweep-every";
+export const USAGE_SWEEP_EVERY_MS = 60_000;
+/** The sweep leaves the newest events to the fast path (and to transactions still committing). */
+export const USAGE_SWEEP_LAG_MS = 30_000;
+/**
+ * Each sweep re-scans this far behind the watermark: events whose transaction committed after a
+ * previous sweep passed their `created_at` are still found. Far inside the dedupe window
+ * (`hooks.usageReportKeepHours` ≥ 1 h), so re-enqueued events dedupe on the kept job.
+ */
+export const USAGE_SWEEP_OVERLAP_MS = 10 * 60_000;
+const SWEEP_PAGE = 500;
+const JOB = "usage-report";
+/**
+ * Plugins that subscribe to `usage.report` according to deployment state: an approved permission
+ * snapshot that declares the `usage.report` hook, for a plugin the deployment still knows: an enabled
+ * install, or a config-declared plugin with a replica status row that is not `unloaded`. With `orgId`, only plugins
+ * the organization enabled (org config rows and required bootstrap plugins).
+ */
+export async function queryUsageSubscribers(db, orgId) {
+    let q = db
+        .selectFrom("extension_permission_snapshots as s")
+        .select("s.plugin")
+        .where(sql `s.permissions -> 'permissions' -> 'hooks' ? 'usage.report'`)
+        .where((eb) => eb.or([
+        eb.exists(eb
+            .selectFrom("extension_installs as i")
+            .select(sql `1`.as("one"))
+            .whereRef("i.name", "=", "s.plugin")
+            .where("i.desired_state", "=", "enabled")),
+        // config-declared plugins still declared (and not unloaded) on some replica; installs are
+        // covered by desired_state above (their replica rows linger as `unloaded` when disabled)
+        eb.exists(eb
+            .selectFrom("extension_replica_status as r")
+            .select(sql `1`.as("one"))
+            .whereRef("r.plugin", "=", "s.plugin")
+            .where("r.source", "=", "config")
+            .where("r.state", "<>", "unloaded")),
+    ]));
+    if (orgId !== undefined) {
+        q = q.where("s.plugin", "in", enabledExtensionsQuery(db, orgId));
+    }
+    return new Set((await q.execute()).map((r) => r.plugin));
+}
+/**
+ * Of `plugins`, those that are **definitely** unservable right now (I-2, N-3): live replicas report
+ * them (a `extension_replica_status` row with a heartbeat newer than `staleSeconds`), no live row is
+ * `loaded` or still serves an older version (`serving_version`, e.g. an upgrade pending approval), and
+ * the rows are not all `unloaded` (that is the state right after a re-enable, before the replicas
+ * reconcile). A plugin with no live row at all is unknown too (slow to load, or a host without a
+ * status store). Unknown plugins are never listed.
+ */
+export async function queryUnservable(db, plugins, staleSeconds) {
+    if (plugins.length === 0)
+        return new Set();
+    const rows = await db
+        .selectFrom("extension_replica_status")
+        .select("plugin")
+        .where("plugin", "in", [...plugins])
+        .where("heartbeat_at", ">", sql `now() - make_interval(secs => ${staleSeconds})`)
+        .groupBy("plugin")
+        .having(sql `bool_or(state = 'loaded' OR serving_version IS NOT NULL)`, "=", false)
+        .having(sql `bool_and(state = 'unloaded')`, "=", false)
+        .execute();
+    return new Set(rows.map((r) => r.plugin));
+}
+/** `" dropped: x (why), y (why)"`, or empty. */
+const droppedNote = (dropped) => dropped.size === 0
+    ? ""
+    : ` dropped: ${[...dropped].map(([p, why]) => `${p} (${why})`).join(", ")}`;
+/** A target this replica cannot deliver to: not loaded here, or loaded without a handler. */
+class Undeliverable extends Error {
+    plugin;
+    why;
+    constructor(plugin, why) {
+        super(why === "no_handler"
+            ? `plugin ${plugin} is loaded on this replica but registered no usage.report handler`
+            : `plugin ${plugin} has no usage.report handler on this replica`);
+        this.plugin = plugin;
+        this.why = why;
+        this.name = "Undeliverable";
+    }
+}
+/**
+ * The next `pending` set (m-2 fence): plugins that failed now, narrowed to what the stored job data
+ * still lists. A plugin another (stalled-over) execution already delivered never comes back.
+ */
+export function narrowPending(stored, failed) {
+    return stored === undefined ? [...failed] : failed.filter((p) => stored.includes(p));
+}
+/**
+ * Reliable `usage.report` delivery (F11, P-17, R13). Every `usage` event written to the event log (chat
+ * runs and full-run eval replays) is delivered at least once to every subscribing plugin.
+ * - **Fast path:** `model-call` enqueues job `usage-<conversationId>-<seq>` right after writing the event
+ *   (never awaited; an enqueue failure only logs). Skipped without I/O when this replica has no
+ *   `usage.report` handler (M-2); the sweeper then offers the event.
+ * - **Sweeper (R13):** a maintenance job (once across replicas, every minute) re-enqueues the `usage`
+ *   events in `[watermark - 10 min, now - 30 s]` with the same job ids, so a lost fast-path enqueue is
+ *   caught up and duplicates dedupe on the kept job (`hooks.usageReportKeepHours`).
+ * - **Worker** (worker role, `hooks.usageReportConcurrency` per replica): reads the payload from the log
+ *   at execution time (the `usage` event, the conversation with `kind`, the run). On the first attempt
+ *   the targets are the deployment's subscribers the organization enabled (I-2: not this replica's
+ *   handlers), written into `pending`; plugins run concurrently (`allSettled`); failed ones stay in
+ *   `pending` and are retried (5 retries, 2 s doubling). A plugin not loaded on this replica (or
+ *   between generations) is a failure, so another replica or a later attempt delivers it. On the
+ *   last attempt it is dropped only when it is definitely unservable ({@link queryUnservable});
+ *   otherwise the job fails as usual. A plugin loaded here that registered no handler is dropped at
+ *   once (I-2). A drop (N-3) is never silent: after delivering to the other targets the job ends
+ *   failed with an `UnrecoverableError` whose failedReason names the dropped plugins, plus a
+ *   throttled warning. Plugins the organization disabled by delivery time are dropped silently (not
+ *   targets any more). The final failure logs an error (no audit); failed jobs are kept 7 days, at
+ *   most 100 000. Each call runs through the instance's `inflight.run`, limited by `hooks.timeoutMs`;
+ *   when this replica stops, in-flight calls are aborted and the job goes back to the queue without
+ *   using up an attempt.
+ * Handlers must be idempotent by `eventId` (`<conversationId>/<seq>`). Nothing is written to the event
+ * log (never model-visible). Compaction calls log usage too (`purpose: "compaction"`) and are reported,
+ * as are conversation title calls (`purpose: "title"`, no run: `runId` is "").
+ * Single-step eval replays and judge requests log no usage: not reported (M7).
+ */
+export class UsageReporter {
+    d;
+    queue;
+    queueConn;
+    worker;
+    workerConn;
+    closed = false;
+    stopping = new AbortController();
+    constructor(d) {
+        this.d = d;
+    }
+    jobOpts(conversationId, seq) {
+        return {
+            jobId: usageReportJobId(conversationId, seq),
+            attempts: this.d.retry?.attempts ?? USAGE_REPORT_ATTEMPTS,
+            backoff: { type: "exponential", delay: this.d.retry?.backoffMs ?? USAGE_REPORT_BACKOFF_MS },
+            removeOnComplete: {
+                age: this.d.config().hooks.usageReportKeepHours * 3600,
+                count: KEEP_COMPLETED_MAX,
+            },
+            removeOnFail: { age: KEEP_FAILED_S, count: KEEP_FAILED_MAX },
+        };
+    }
+    /**
+     * Enqueues the delivery of one `usage` event. Returns without a job when no plugin subscribes.
+     * Never throws: an enqueue failure (e.g. a Redis hiccup) only logs a warning; the sweeper catches up.
+     */
+    async enqueue(conversationId, seq) {
+        if (this.closed)
+            return;
+        // M-2: no usage.report handler on this replica → no I/O at all. A plugin loaded only on other
+        // replicas still gets the event: the sweeper re-offers it within about a minute and a half.
+        if (!this.d.hooks.has("usage.report"))
+            return;
+        try {
+            if ((await this.d.subscribers()).size === 0)
+                return;
+            await this.getQueue().add(JOB, { conversationId, seq }, this.jobOpts(conversationId, seq));
+        }
+        catch (err) {
+            this.d.log.warn("cannot enqueue usage report", { conversationId, seq, err });
+        }
+    }
+    /**
+     * One sweep (R13): re-enqueues the `usage` events created in `[watermark - overlap, now - lag]` in
+     * pages, then advances the watermark to `now - lag`. With no subscriber it only advances. The first
+     * sweep starts at `now - lag - overlap`. Returns the number of events offered to the queue.
+     */
+    async sweep(opts = {}) {
+        // bounds and keyset as full-precision timestamp text (microseconds), computed with the DB clock
+        // unless a test fixes "now" (m-a, m-c)
+        const to = opts.now !== undefined
+            ? new Date(opts.now - USAGE_SWEEP_LAG_MS).toISOString()
+            : (await sql `SELECT (now() - make_interval(secs => ${USAGE_SWEEP_LAG_MS / 1000}))::text AS to`.execute(this.d.db)).rows[0]?.to;
+        if (to === undefined)
+            return 0;
+        const row = await this.d.db
+            .selectFrom("usage_report_sweep")
+            .select(sql `watermark::text`.as("watermark"))
+            .where("id", "=", 1)
+            .executeTakeFirst();
+        const watermark = row?.watermark ?? to;
+        let offered = 0;
+        if ((await this.d.subscribers()).size > 0) {
+            let after;
+            for (;;) {
+                let q = this.d.db
+                    .selectFrom("conversation_events")
+                    .select(["conversation_id", "seq", sql `created_at::text`.as("at")])
+                    .where("type", "=", "usage")
+                    .where(sql `created_at >= ${watermark}::timestamptz - make_interval(secs => ${USAGE_SWEEP_OVERLAP_MS / 1000})`)
+                    .where(sql `created_at <= ${to}::timestamptz`);
+                if (after) {
+                    q = q.where(sql `(created_at, conversation_id, seq) > (${after.at}::timestamptz, ${after.conv}, ${after.seq}::bigint)`);
+                }
+                const page = await q
+                    .orderBy("created_at")
+                    .orderBy("conversation_id")
+                    .orderBy("seq")
+                    .limit(SWEEP_PAGE)
+                    .execute();
+                if (page.length > 0) {
+                    await this.getQueue().addBulk(page.map((r) => ({
+                        name: JOB,
+                        data: { conversationId: r.conversation_id, seq: Number(r.seq) },
+                        opts: this.jobOpts(r.conversation_id, Number(r.seq)),
+                    })));
+                    offered += page.length;
+                }
+                const last = page.at(-1);
+                if (!last || page.length < SWEEP_PAGE)
+                    break;
+                after = { at: last.at, conv: last.conversation_id, seq: String(last.seq) };
+            }
+        }
+        await this.d.db
+            .insertInto("usage_report_sweep")
+            .values({ id: 1, watermark: sql `${to}::timestamptz` })
+            .onConflict((oc) => oc.column("id").doUpdateSet({
+            watermark: sql `GREATEST(usage_report_sweep.watermark, excluded.watermark)`,
+            updated_at: sql `now()`,
+        }))
+            .execute();
+        return offered;
+    }
+    /** Starts the worker (worker role only). Job lock and stalled detection follow the run queue. */
+    startWorker() {
+        if (this.worker || this.closed)
+            return;
+        const { jobLockMs, stalledIntervalMs } = this.d.config().runtime;
+        // the queue exists before any job runs: in-flight jobs fence and requeue through it, even on a
+        // replica that never enqueued (m-b)
+        this.getQueue();
+        this.workerConn = bullConnection(this.d.redisUrl, this.d.log);
+        this.worker = new Worker(USAGE_REPORT_QUEUE, (j, t) => withJobContext(USAGE_REPORT_QUEUE, j, () => this.process(j, t)), {
+            connection: this.workerConn,
+            concurrency: this.d.config().hooks.usageReportConcurrency,
+            lockDuration: jobLockMs,
+            stalledInterval: stalledIntervalMs,
+            maxStalledCount: 3,
+        });
+        this.worker.on("error", (err) => this.d.log.error("usage report worker error", { err }));
+        this.worker.on("failed", (j, err) => {
+            if (!j || err instanceof DelayedError || err.name === "DelayedError")
+                return;
+            // a definite drop (N-3) was already warned about (throttled); the failed job is the trace
+            if (err instanceof UnrecoverableError || err.name === "UnrecoverableError")
+                return;
+            const final = j.attemptsMade >= (j.opts.attempts ?? 1);
+            const fields = {
+                jobId: j.id,
+                conversationId: j.data.conversationId,
+                seq: j.data.seq,
+                pending: j.data.pending,
+                attempt: j.attemptsMade,
+                error: err.message,
+            };
+            if (final)
+                this.d.log.error("usage report given up after retries", fields);
+            else
+                this.d.log.warn("usage report attempt failed; retrying", fields);
+        });
+    }
+    /** Stops the worker (in-flight calls are aborted, their jobs requeued) and the queue. */
+    async stop() {
+        this.closed = true;
+        this.stopping.abort();
+        const w = this.worker;
+        this.worker = undefined;
+        // the worker first: its in-flight jobs still use the queue to fence and requeue
+        await w?.close().catch(() => { });
+        const q = this.queue;
+        this.queue = undefined;
+        await q?.close().catch(() => { });
+        for (const c of [this.workerConn, this.queueConn]) {
+            if (c)
+                await c.quit().catch(() => c.disconnect());
+        }
+        this.workerConn = undefined;
+        this.queueConn = undefined;
+    }
+    getQueue() {
+        if (!this.queue) {
+            if (this.closed)
+                throw new Error("usage report queue is closed");
+            this.queueConn = bullConnection(this.d.redisUrl, this.d.log);
+            this.queue = new Queue(USAGE_REPORT_QUEUE, { connection: this.queueConn });
+            this.queue.on("error", (err) => this.d.log.error("usage report queue error", { err }));
+        }
+        return this.queue;
+    }
+    /** The payload of one usage event, read from the log; null when the event no longer exists. */
+    async payloadOf(conversationId, seq) {
+        const row = await this.d.db
+            .selectFrom("conversation_events as e")
+            .innerJoin("conversations as c", "c.id", "e.conversation_id")
+            .leftJoin("runs as r", "r.id", "e.run_id")
+            // a sub-agent's run: its dispatch says where it was dispatched from (spec §11)
+            .leftJoin("subagent_dispatches as sd", "sd.child_run_id", "e.run_id")
+            .select([
+            "e.payload",
+            "e.run_id",
+            "e.created_at",
+            "e.type",
+            "c.org_id",
+            "c.agent_id",
+            "c.account_id",
+            "c.kind",
+            "r.trigger_account_id",
+            "sd.parent_conversation_id",
+            "sd.parent_run_id",
+            "sd.parent_call_id",
+        ])
+            .where("e.conversation_id", "=", conversationId)
+            .where("e.seq", "=", String(seq))
+            .executeTakeFirst();
+        if (row?.type !== "usage")
+            return null;
+        const p = row.payload;
+        return {
+            eventId: `${conversationId}/${seq}`,
+            purpose: p.purpose ?? "step",
+            orgId: row.org_id,
+            accountId: row.trigger_account_id ?? row.account_id,
+            agentId: row.agent_id,
+            conversationId,
+            // a sub-agent's calls report "chat" (spec §11: the SDK union is not widened)
+            conversationKind: row.kind === "eval" ? "eval" : "chat",
+            runId: row.run_id ?? "",
+            ...(row.parent_conversation_id !== null &&
+                row.parent_run_id !== null &&
+                row.parent_call_id !== null && {
+                parent: {
+                    conversationId: row.parent_conversation_id,
+                    runId: row.parent_run_id,
+                    callId: row.parent_call_id,
+                },
+            }),
+            provider: p.provider,
+            model: p.model,
+            usage: {
+                inputTokens: p.inputTokens,
+                outputTokens: p.outputTokens,
+                cacheReadTokens: p.cacheReadTokens,
+                cacheWriteTokens: p.cacheWriteTokens,
+            },
+            occurredAt: new Date(row.created_at).toISOString(),
+        };
+    }
+    /**
+     * Writes the next `pending` set and the accumulated drops, fenced against a concurrent
+     * (stalled-over) execution (m-2 fence): only plugins the stored data still lists as pending or
+     * dropped can stay, so a plugin another execution delivered never comes back.
+     */
+    async writePending(j, next, dropped) {
+        const fresh = j.id === undefined ? undefined : await Job.fromId(this.getQueue(), j.id);
+        const data = fresh?.data;
+        const stored = data?.pending ?? j.data.pending;
+        const storedDropped = data?.dropped ?? j.data.dropped ?? {};
+        const allowed = stored === undefined ? undefined : [...stored, ...Object.keys(storedDropped)];
+        const keep = [...dropped].filter(([p]) => allowed === undefined || allowed.includes(p));
+        const out = {
+            conversationId: j.data.conversationId,
+            seq: j.data.seq,
+            pending: narrowPending(allowed, next),
+            ...(keep.length > 0 && { dropped: Object.fromEntries(keep) }),
+        };
+        await j.updateData(out);
+        j.data = out;
+    }
+    async process(j, token) {
+        const { conversationId, seq } = j.data;
+        const payload = await this.payloadOf(conversationId, seq);
+        if (payload === null)
+            return; // conversation deleted meanwhile: nothing to report
+        const subscribers = await this.d.subscribers(payload.orgId);
+        if (j.data.pending === undefined) {
+            // first attempt: the targets come from deployment state, recorded so later attempts (on any
+            // replica) deliver exactly the plugins that have not succeeded yet
+            const targets = [...subscribers];
+            await j.updateData({ conversationId, seq, pending: targets });
+            j.data.pending = targets;
+        }
+        // plugins the organization disabled meanwhile are dropped (not a failure)
+        const targets = (j.data.pending ?? []).filter((p) => subscribers.has(p));
+        if (targets.length === 0)
+            return;
+        const byPlugin = new Map();
+        for (const h of this.d.hooks.handlers("usage.report")) {
+            byPlugin.set(h.plugin, [...(byPlugin.get(h.plugin) ?? []), h.handler]);
+        }
+        const frozen = deepFreeze(structuredClone(payload));
+        const timeoutMs = this.d.config().hooks.timeoutMs;
+        const results = await Promise.allSettled(targets.map(async (plugin) => {
+            const handlers = byPlugin.get(plugin);
+            if (!handlers || handlers.length === 0) {
+                throw new Undeliverable(plugin, this.d.isActive?.(plugin) ? "no_handler" : "not_loaded");
+            }
+            for (const h of handlers)
+                await this.call(h, frozen, timeoutMs);
+        }));
+        const failed = [];
+        const notHere = [];
+        /** Definite drops (N-3): plugin → why; they end the job as failed, with no further retries. */
+        const dropped = new Map();
+        results.forEach((r, i) => {
+            if (r.status === "fulfilled")
+                return;
+            const plugin = targets[i];
+            if (r.reason instanceof Undeliverable) {
+                // I-2: loaded here without a handler: not a subscriber (declared but never registered, or
+                // a downgrade); dropped, not retried
+                if (r.reason.why === "no_handler")
+                    dropped.set(plugin, "registered no usage.report handler");
+                else
+                    notHere.push(plugin);
+                return;
+            }
+            failed.push(plugin);
+            if (!this.stopping.signal.aborted) {
+                this.d.log.warn("usage.report delivery to a plugin failed", {
+                    plugin,
+                    eventId: payload.eventId,
+                    error: errMsg(r.reason),
+                });
+            }
+        });
+        // not loaded here: another replica or a later attempt may serve it (retry). On the last attempt a
+        // target is dropped only when it is definitely unservable; unknown ones fail the job as before.
+        const final = j.attemptsMade + 1 >= (j.opts.attempts ?? 1);
+        const gone = final && notHere.length > 0 && !this.stopping.signal.aborted
+            ? await this.unservable(notHere)
+            : new Set();
+        for (const plugin of notHere) {
+            if (gone.has(plugin))
+                dropped.set(plugin, "no replica has it loaded");
+            else
+                failed.push(plugin);
+        }
+        // m-2: drops accumulate across attempts (minus plugins delivered now), so a drop on an attempt
+        // that also had a transient failure still fails the job once the rest succeeds
+        const delivered = new Set(targets.filter((_, i) => results[i]?.status === "fulfilled"));
+        const allDropped = new Map(
+        // (a plugin the organization disabled meanwhile is no target any more: not traced)
+        Object.entries(j.data.dropped ?? {}).filter(([p]) => !delivered.has(p) && subscribers.has(p)));
+        for (const [p, why] of dropped)
+            allDropped.set(p, why);
+        if (failed.length > 0) {
+            await this.writePending(j, failed, allDropped);
+            if (this.stopping.signal.aborted) {
+                // shutdown is not the plugin's fault: back to the queue without using up an attempt (m-3)
+                await j.moveToDelayed(Date.now(), token);
+                throw new DelayedError();
+            }
+            throw new Error(`usage.report failed for: ${failed.join(", ")}${droppedNote(dropped)}`);
+        }
+        // m-1: pending = the dropped plugins only, so a manual retry reaches them and never the
+        // plugins already served
+        await this.writePending(j, [...allDropped.keys()], allDropped);
+        const traced = new Map([...allDropped].filter(([p]) => j.data.dropped?.[p] !== undefined));
+        if (traced.size === 0)
+            return;
+        // N-3: a definite drop leaves a durable trace: the job ends failed (kept under the failed-job
+        // cap, visible in queue tooling) with the dropped plugins in failedReason; no more retries
+        for (const [plugin, reason] of dropped) {
+            void this.warnOnce(plugin, "dropped", "usage.report target dropped", {
+                plugin,
+                eventId: payload.eventId,
+                reason,
+            });
+        }
+        throw new UnrecoverableError(`usage.report delivered to the rest;${droppedNote(traced)}`);
+    }
+    /** {@link UsageReporterDeps.unservable}; a lookup failure counts as "unknown" (keep retrying). */
+    async unservable(plugins) {
+        if (!this.d.unservable)
+            return new Set();
+        try {
+            return await this.d.unservable(plugins);
+        }
+        catch (err) {
+            this.d.log.warn("cannot read plugin replica status for usage.report", { err });
+            return new Set();
+        }
+    }
+    /** Logs a warning at most once per plugin and kind per minute (Redis key with a TTL). */
+    async warnOnce(plugin, kind, msg, fields) {
+        let first = true;
+        try {
+            this.getQueue();
+            const key = `capk:usage-warn:${plugin}:${kind}`;
+            first = (await this.queueConn?.set(key, "1", "PX", WARN_WINDOW_MS, "NX")) !== null;
+        }
+        catch {
+            // Redis unavailable: log anyway rather than hiding it
+        }
+        if (first)
+            this.d.log.warn(msg, fields);
+    }
+    /** One handler call with its own signal (timeout or replica stop aborts it); rejects on either. */
+    async call(h, payload, timeoutMs) {
+        const ac = new AbortController();
+        const onStop = () => ac.abort(new Error("replica stopping"));
+        if (this.stopping.signal.aborted)
+            onStop();
+        else
+            this.stopping.signal.addEventListener("abort", onStop, { once: true });
+        let timer;
+        try {
+            await Promise.race([
+                Promise.resolve().then(() => h(payload, ac.signal)),
+                new Promise((_, reject) => {
+                    timer = setTimeout(() => {
+                        ac.abort(new Error(`timed out after ${timeoutMs} ms`));
+                        reject(new Error(`timed out after ${timeoutMs} ms`));
+                    }, timeoutMs);
+                    ac.signal.addEventListener("abort", () => reject(ac.signal.reason), { once: true });
+                }),
+            ]);
+        }
+        finally {
+            clearTimeout(timer);
+            this.stopping.signal.removeEventListener("abort", onStop);
+        }
+    }
+}