@voltro/ai 0.70.0 → 0.71.0

package/dist/inferenceQueue-TwTEBG3Q.js

@@ -0,0 +1,864 @@
+import { c as e, o as t, r as n } from "./provider-C_Sms_DA.js";
+import { NoObjectGeneratedError as r, Output as i, experimental_transcribe as a, generateObject as o, generateText as s, jsonSchema as c, stepCountIs as l, tool as u } from "ai";
+import { Context as d, Effect as f, JSONSchema as p, Metric as m, MetricBoundaries as ee, Runtime as te, Schema as h, Stream as ne } from "effect";
+import { and as g, bigint as re, eq as _, generateId as v, gte as ie, id as y, integer as b, isNull as ae, json as x, queryFor as S, registerRetention as C, retentionTtlMsFromEnv as w, table as T, text as E, timestamp as D } from "@voltro/database";
+import { Ajv as oe } from "ajv";
+import { Ajv2019 as se } from "ajv/dist/2019.js";
+import { Ajv2020 as ce } from "ajv/dist/2020.js";
+import le from "ajv-formats";
+import { createHash as ue } from "node:crypto";
+//#region src/tools.ts
+var de = (e, t) => e(t), fe = Symbol.for("@voltro/ai/Tool"), pe = (e) => ({
+	...e,
+	[fe]: !0
+}), me = (e) => typeof e == "object" && !!e && e[fe] === !0, he = (e) => {
+	let t = e.default;
+	if (typeof t != "function") return 0;
+	let n = 0;
+	for (let r of Object.values(e)) me(r) && r.execute === void 0 && r.handler === void 0 && (r.handler = t, n += 1);
+	return n;
+}, ge = (e, t, n) => {
+	let r = t ?? te.defaultRuntime;
+	return te.runPromise(r)(e, n === void 0 ? {} : { signal: n });
+}, _e = (e, t, n) => n ? e.pipe(f.timeoutFail({
+	duration: `${n} millis`,
+	onTimeout: () => /* @__PURE__ */ Error(`tool '${t}' timed out`)
+})) : e, ve = (e, t, n) => {
+	if (e.execute) return ge(_e(de(e.execute, t), e.name, e.timeoutMs), n.runtime, n.abortSignal);
+	if (e.handler) {
+		let r = e.handler, i = r(t, n);
+		return f.isEffect(i) ? ge(_e(i, e.name, e.timeoutMs), n.runtime, n.abortSignal) : i;
+	}
+	return Promise.reject(/* @__PURE__ */ Error(`tool '${e.name}' has neither execute nor handler`));
+}, ye = (e) => {
+	let t = p.make(e);
+	return t && t.type === "object" ? t : {
+		type: "object",
+		properties: {},
+		additionalProperties: !1
+	};
+}, be = (e) => {
+	let t = e.inputJsonSchema;
+	return t === void 0 ? ye(e.input) : t.type === "object" ? t : {
+		type: "object",
+		properties: {},
+		additionalProperties: !0
+	};
+}, O = (e, t = { store: void 0 }) => {
+	let n = {};
+	for (let [r, i] of Object.entries(e)) n[i.name || r] = u({
+		description: i.description,
+		inputSchema: c(be(i)),
+		execute: async (e, n) => {
+			let r = n?.abortSignal ?? t.abortSignal;
+			r?.throwIfAborted();
+			let a = {
+				...t,
+				...n?.toolCallId === void 0 ? {} : { invocationId: n.toolCallId },
+				...r === void 0 ? {} : { abortSignal: r }
+			}, o = await f.runPromise(h.decodeUnknown(i.input)(e));
+			r?.throwIfAborted();
+			let s = await ve(i, o, a);
+			return i.output ? f.runPromise(h.decodeUnknown(i.output)(s)) : s;
+		}
+	});
+	return n;
+}, k = class extends h.TaggedError()("AiBudgetExceeded", {
+	tenantId: h.NullOr(h.String),
+	limitUsd: h.Number,
+	spentUsd: h.Number,
+	attemptedUsd: h.Number
+}) {}, A = "_voltro_ai_budget", j = "_voltro_ai_budget_reservations", M = T(A, {
+	id: y({ prefix: "aib" }),
+	bucket: E().maxLength(191),
+	tenantId: E().nullable(),
+	reservedMicroUsd: re().default("0").narrowedFrom("integer"),
+	updatedAt: D().default("now")
+}).unique("byAiBudgetBucket", ["bucket"]), xe = T(j, {
+	id: y({ prefix: "air" }),
+	reservationId: E().maxLength(191),
+	bucket: E().maxLength(191),
+	tenantId: E().nullable(),
+	amountMicroUsd: re(),
+	releasedAt: D().nullable(),
+	updatedAt: D().default("now")
+}).unique("byAiBudgetReservation", ["reservationId"]), N = class extends Error {}, Se = (e, t, n) => {
+	if (!Number.isFinite(e) || e < 0) throw RangeError(`${t} must be finite and non-negative`);
+	let r = e * 1e6, i = n === "up" ? Math.ceil(r) : Math.floor(r);
+	if (!Number.isSafeInteger(i)) throw RangeError(`${t} exceeds the safe micro-USD range`);
+	return BigInt(i);
+}, P = (e) => {
+	if (typeof e == "number" && !Number.isSafeInteger(e)) throw Error("AI budget counter is not an exact integer");
+	let t = String(e);
+	if (!/^(0|[1-9][0-9]*)$/.test(t)) throw Error("AI budget counter is invalid; refusing to assume zero");
+	return BigInt(t);
+}, Ce = (e) => {
+	if (typeof e != "string" || e.length === 0 || e.length > 191) throw RangeError("AI budget reservationId must contain 1–191 characters");
+	return e;
+}, we = () => {
+	let e = w(process.env.VOLTRO_AI_USAGE_TTL_HOURS, 8760);
+	for (let t of [A, j]) C({
+		source: "plugin",
+		table: t,
+		timeColumn: "updatedAt",
+		ttlMs: e
+	});
+}, Te = async (e, t) => {
+	for (let n = 0;; n++) try {
+		return await e.transactional(t);
+	} catch (e) {
+		if (!(e instanceof N) || n >= 31) throw e;
+		await new Promise((e) => setTimeout(e, n + 1));
+	}
+}, Ee = (e) => f.tryPromise({
+	try: e,
+	catch: (e) => e
+}).pipe(f.catchAll((e) => e instanceof k ? f.fail(e) : f.die(e))), De = (e, t) => f.gen(function* () {
+	let n = Se(t.limitUsd, "limitUsd", "down"), r = Se(t.addUsd ?? 0, "addUsd", "up"), i = t.tenantId ?? null;
+	if (t.since !== void 0 && !Number.isFinite(t.since.getTime())) throw RangeError("AI budget since must be a valid Date");
+	let a = `${i ?? "global"}::${t.since === void 0 ? "all" : t.since.getTime()}`;
+	if (a.length > 191) throw RangeError("AI budget tenant/window key exceeds 191 characters");
+	let o = Ce(t.reservationId ?? v({
+		kind: "typeid",
+		prefix: "air"
+	}, "_voltro_ai_budget_reservations"));
+	we();
+	let s = async (e) => {
+		let [t] = await e.query(S(M).where(_("bucket", a)).descriptor);
+		if (t && t.tenantId !== i) throw Error("AI budget bucket belongs to a different tenant");
+		return t;
+	}, c = (e) => {
+		if (e + r > n) throw new k({
+			tenantId: i,
+			limitUsd: t.limitUsd,
+			spentUsd: Number(e) / 1e6,
+			attemptedUsd: t.addUsd ?? 0
+		});
+	};
+	return r === 0n ? yield* Ee(async () => {
+		let t = await s(e);
+		return c(t ? P(t.reservedMicroUsd) : 0n), null;
+	}) : yield* Ee(() => Te(e, async (e) => {
+		let t = v({
+			kind: "typeid",
+			prefix: "air"
+		}, j), n = /* @__PURE__ */ new Date(), l = await e.insertIgnore(j, {
+			id: t,
+			reservationId: o,
+			bucket: a,
+			tenantId: i,
+			amountMicroUsd: String(r),
+			releasedAt: null,
+			updatedAt: n
+		}, { conflictColumns: ["reservationId"] });
+		if (l.id !== t) {
+			if (l.tenantId !== i || l.bucket !== a || P(l.amountMicroUsd) !== r) throw Error("AI budget reservationId is already bound to another tenant, amount or window");
+			if (l.releasedAt !== null) throw Error("AI budget reservation was released; new spend requires a new reservationId");
+			return {
+				reservationId: o,
+				tenantId: i,
+				amountMicroUsd: String(r),
+				status: "replayed"
+			};
+		}
+		let u = await s(e), d = u ? P(u.reservedMicroUsd) : 0n;
+		if (c(d), u) {
+			if (await e.updateMany("_voltro_ai_budget", {
+				reservedMicroUsd: String(d + r),
+				updatedAt: n
+			}, { where: g(_("id", u.id), _("reservedMicroUsd", u.reservedMicroUsd)) }) !== 1) throw new N("AI budget counter changed; discard this receipt and retry");
+		} else {
+			let t = v({
+				kind: "typeid",
+				prefix: "aib"
+			}, A);
+			if ((await e.insertIgnore("_voltro_ai_budget", {
+				id: t,
+				bucket: a,
+				tenantId: i,
+				reservedMicroUsd: String(r),
+				updatedAt: n
+			}, { conflictColumns: ["bucket"] })).id !== t) throw new N("AI budget counter was created concurrently; discard this receipt and retry");
+		}
+		return {
+			reservationId: o,
+			tenantId: i,
+			amountMicroUsd: String(r),
+			status: "reserved"
+		};
+	}));
+}), Oe = (e, t) => f.gen(function* () {
+	return Ce(t.reservationId), yield* f.promise(() => Te(e, async (e) => {
+		let [n] = await e.query(S(xe).where(g(_("reservationId", t.reservationId), t.tenantId === null ? ae("tenantId") : _("tenantId", t.tenantId))).descriptor);
+		if (!n) return "not-found";
+		if (n.releasedAt !== null) return "already-released";
+		let [r] = await e.query(S(M).where(_("bucket", n.bucket)).descriptor);
+		if (!r || r.tenantId !== t.tenantId) throw Error("AI budget counter missing or mismatched; refusing unproved credit");
+		let i = P(r.reservedMicroUsd), a = P(n.amountMicroUsd);
+		if (a > i) throw Error("AI budget receipt exceeds its counter; refusing unproved credit");
+		let o = /* @__PURE__ */ new Date();
+		if (await e.updateMany("_voltro_ai_budget_reservations", {
+			releasedAt: o,
+			updatedAt: o
+		}, { where: g(_("id", n.id), ae("releasedAt")) }) !== 1) throw new N("AI budget release raced; retry");
+		if (await e.updateMany("_voltro_ai_budget", {
+			reservedMicroUsd: String(i - a),
+			updatedAt: o
+		}, { where: g(_("id", r.id), _("reservedMicroUsd", r.reservedMicroUsd)) }) !== 1) throw new N("AI budget counter changed; discard this release and retry");
+		return "released";
+	}));
+}), ke = {
+	"claude-opus-4-8": {
+		inputPer1M: 15,
+		outputPer1M: 75
+	},
+	"claude-opus-4-6": {
+		inputPer1M: 15,
+		outputPer1M: 75
+	},
+	"claude-sonnet-4-6": {
+		inputPer1M: 3,
+		outputPer1M: 15
+	},
+	"claude-haiku-4-5": {
+		inputPer1M: 1,
+		outputPer1M: 5
+	},
+	"claude-haiku-4-5-20251001": {
+		inputPer1M: 1,
+		outputPer1M: 5
+	},
+	"gpt-5.5": {
+		inputPer1M: 5,
+		outputPer1M: 30
+	},
+	"gpt-4o": {
+		inputPer1M: 2.5,
+		outputPer1M: 10
+	},
+	"gpt-4o-mini": {
+		inputPer1M: .15,
+		outputPer1M: .6
+	},
+	"gemini-2.5-pro": {
+		inputPer1M: 1.25,
+		outputPer1M: 10
+	},
+	"gemini-2.5-flash": {
+		inputPer1M: .3,
+		outputPer1M: 2.5
+	},
+	"gemini-2.0-flash": {
+		inputPer1M: .1,
+		outputPer1M: .4
+	}
+}, F = {}, Ae = (e) => {
+	let t = F;
+	return F = { ...e }, t;
+}, je = () => ({
+	...ke,
+	...F
+}), Me = (e) => {
+	let t = je();
+	if (t[e] !== void 0) return t[e];
+	let n = e.indexOf("/");
+	if (n > 0) {
+		let r = e.slice(n + 1);
+		if (t[r] !== void 0) return t[r];
+	}
+}, Ne = (e) => typeof e.unit == "string", Pe = (e) => Ne(e) ? {
+	unit: e.unit,
+	input: e.input ?? 0,
+	output: e.output ?? 0
+} : {
+	unit: "tokens",
+	input: e.inputTokens ?? 0,
+	output: e.outputTokens ?? 0
+}, I = (e, t) => {
+	let { unit: n, input: r, output: i } = Pe(e);
+	if (t.actualCostUsd !== void 0) {
+		if (!Number.isFinite(t.actualCostUsd) || t.actualCostUsd < 0) throw RangeError("estimateCostUsd: actualCostUsd must be a finite, non-negative USD amount.");
+		let e = r + i, a = e > 0 ? r / e : 0;
+		return {
+			inputTokens: r,
+			outputTokens: i,
+			inputCostUsd: t.actualCostUsd * a,
+			outputCostUsd: t.actualCostUsd * (1 - a),
+			totalCostUsd: t.actualCostUsd,
+			costSource: "gateway",
+			unit: n
+		};
+	}
+	if (n !== "tokens" && t.price === void 0) return {
+		inputTokens: r,
+		outputTokens: i,
+		inputCostUsd: 0,
+		outputCostUsd: 0,
+		totalCostUsd: 0,
+		costSource: "unpriced",
+		unit: n
+	};
+	let a = t.price ?? Me(t.model) ?? {
+		inputPer1M: 0,
+		outputPer1M: 0
+	}, o = r / 1e6 * a.inputPer1M, s = i / 1e6 * a.outputPer1M;
+	return {
+		inputTokens: r,
+		outputTokens: i,
+		inputCostUsd: o,
+		outputCostUsd: s,
+		totalCostUsd: o + s,
+		costSource: "estimated",
+		unit: n
+	};
+}, L = "_voltro_ai_usage", Fe = T(L, {
+	id: y({ prefix: "aiu" }),
+	tenantId: E().nullable(),
+	provider: E(),
+	model: E(),
+	operation: E(),
+	agent: E().nullable(),
+	inputTokens: b().default(0),
+	outputTokens: b().default(0),
+	costMicroUsd: b().default(0),
+	costSource: E().default("estimated"),
+	unit: E().default("tokens"),
+	promptId: E().nullable(),
+	promptDigest: E().nullable(),
+	promptRevision: b().nullable(),
+	calledAt: D().default("now")
+}).index(["tenantId"]).index(["promptDigest"]), Ie = (e) => {
+	let t = e?.gateway?.cost, n = typeof t == "string" ? t.trim() : void 0, r = typeof t == "number" ? t : n !== void 0 && /^(?:\d+(?:\.\d*)?|\.\d+)(?:e[+-]?\d+)?$/i.test(n) ? Number(n) : NaN;
+	return Number.isFinite(r) && r >= 0 ? r : void 0;
+}, Le = (e, t) => (C({
+	source: "plugin",
+	table: L,
+	timeColumn: "calledAt",
+	ttlMs: w(process.env.VOLTRO_AI_USAGE_TTL_HOURS, 8760)
+}), f.promise(async () => {
+	let n = I(t.usage, {
+		model: t.model,
+		...t.price === void 0 ? {} : { price: t.price },
+		...t.actualCostUsd === void 0 ? {} : { actualCostUsd: t.actualCostUsd }
+	});
+	return await e.insert(L, {
+		tenantId: t.tenantId ?? null,
+		provider: t.provider,
+		model: t.model,
+		operation: t.operation,
+		agent: t.agent ?? null,
+		inputTokens: n.inputTokens,
+		outputTokens: n.outputTokens,
+		costMicroUsd: Math.round(n.totalCostUsd * 1e6),
+		costSource: n.costSource,
+		unit: n.unit,
+		promptId: t.prompt?.promptId ?? null,
+		promptDigest: t.prompt?.digest ?? null,
+		promptRevision: t.prompt?.revision ?? null,
+		calledAt: t.calledAt ?? /* @__PURE__ */ new Date()
+	}), n;
+})), Re = (e, t = {}) => f.promise(async () => {
+	let n = [
+		...t.tenantId === void 0 ? [] : [_("tenantId", t.tenantId)],
+		...t.since === void 0 ? [] : [ie("calledAt", t.since)],
+		...t.promptDigest === void 0 ? [] : [_("promptDigest", t.promptDigest)]
+	], r = S(Fe), i = n.length === 0 ? r : r.where(n.length === 1 ? n[0] : g(...n));
+	return (await e.query(i.descriptor)).reduce((e, t) => e + (Number(t.costMicroUsd) || 0), 0) / 1e6;
+}), ze = ee.exponential({
+	start: .001,
+	factor: 2,
+	count: 16
+}), R = m.counter("voltro_ai_calls_total", { description: "Total AI provider calls (generate + stream)." }), Be = m.counter("voltro_ai_call_errors_total", { description: "AI provider calls that errored." }), z = m.histogram("voltro_ai_call_duration_seconds", ze, "AI provider call duration in seconds."), Ve = m.counter("voltro_ai_input_tokens_total", {
+	description: "Total prompt (input) tokens across AI calls.",
+	incremental: !0
+}), He = m.counter("voltro_ai_output_tokens_total", {
+	description: "Total completion (output) tokens across AI calls.",
+	incremental: !0
+}), Ue = m.counter("voltro_ai_cost_microusd_total", {
+	description: "Total estimated AI spend in micro-USD (USD × 1e6).",
+	incremental: !0
+}), B = (e, t) => ({
+	provider: e.name,
+	model: e.model ?? e.name,
+	operation: t
+}), V = (e, t, n) => e.pipe(f.tagMetrics("provider", t.provider), f.tagMetrics("model", t.model), f.tagMetrics("operation", t.operation), f.tagMetrics("status", n)), We = (e, t) => {
+	let n = t.inputTokens ?? 0, r = t.outputTokens ?? 0, i = I({
+		inputTokens: n,
+		outputTokens: r
+	}, { model: e.model }), a = Math.round(i.totalCostUsd * 1e6);
+	return f.all([
+		V(m.incrementBy(Ve, n), e, "ok"),
+		V(m.incrementBy(He, r), e, "ok"),
+		V(m.incrementBy(Ue, a), e, "ok")
+	], { discard: !0 });
+}, H = (e, t, n) => f.gen(function* () {
+	let r = yield* f.clockWith((e) => e.currentTimeMillis);
+	return yield* n.pipe(f.tap((n) => We(e, t(n))), f.tap(() => V(m.increment(R), e, "ok").pipe(f.zipRight(f.clockWith((e) => e.currentTimeMillis).pipe(f.flatMap((t) => V(m.update(z, (t - r) / 1e3), e, "ok")))))), f.tapError(() => V(m.increment(R), e, "error").pipe(f.zipRight(V(m.increment(Be), e, "error")), f.zipRight(f.clockWith((e) => e.currentTimeMillis).pipe(f.flatMap((t) => V(m.update(z, (t - r) / 1e3), e, "error")))))), f.withSpan("voltro.ai.call", { attributes: {
+		"ai.provider": e.provider,
+		"ai.model": e.model,
+		"ai.operation": e.operation
+	} }));
+}), Ge = (e, t) => {
+	let n = Date.now();
+	return t.pipe(ne.tap((t) => t._tag === "done" ? V(m.increment(R), e, "ok").pipe(f.zipRight(V(m.update(z, (Date.now() - n) / 1e3), e, "ok")), f.zipRight(We(e, t.usage ?? {}))) : t._tag === "error" ? V(m.increment(R), e, "error").pipe(f.zipRight(V(m.increment(Be), e, "error")), f.zipRight(V(m.update(z, (Date.now() - n) / 1e3), e, "error"))) : f.void), ne.withSpan("voltro.ai.call", { attributes: {
+		"ai.provider": e.provider,
+		"ai.model": e.model,
+		"ai.operation": e.operation
+	} }));
+}, Ke = (e) => {
+	let t = structuredClone(e), n = {
+		strict: !0,
+		strictTypes: !1,
+		strictTuples: !1,
+		strictRequired: !1,
+		allowUnionTypes: !0
+	}, r = t.$schema, i = r === "https://json-schema.org/draft/2020-12/schema" ? new ce(n) : r === "https://json-schema.org/draft/2019-09/schema" ? new se(n) : new oe(n);
+	le(i);
+	let a = i.compile(t);
+	if ("$async" in a && a.$async === !0) throw Error("asynchronous output schemas are not supported");
+	return c(t, { validate: (e) => a(e) ? {
+		success: !0,
+		value: e
+	} : {
+		success: !1,
+		error: /* @__PURE__ */ Error("model output does not satisfy the declared JSON Schema")
+	} });
+}, U = h.Number.pipe(h.int(), h.nonNegative(), h.lessThanOrEqualTo(2 ** 53 - 1)), qe = h.Struct({
+	provider: h.String,
+	model: h.String,
+	usage: h.Struct({
+		inputTokens: h.optional(U),
+		outputTokens: h.optional(U),
+		totalTokens: h.optional(U),
+		completeness: h.Literal("complete", "partial", "unknown")
+	}),
+	cost: h.Union(h.Struct({
+		status: h.Literal("actual"),
+		source: h.Literal("gateway"),
+		usd: h.Number.pipe(h.finite(), h.nonNegative())
+	}), h.Struct({
+		status: h.Literal("unknown"),
+		source: h.Literal("unreported")
+	}))
+}), W = (e) => typeof e == "number" && Number.isSafeInteger(e) && e >= 0 ? e : void 0, G = (e, t, n, r) => {
+	let i = typeof n == "object" && n ? n : {}, a = W("inputTokens" in i ? i.inputTokens : void 0), o = W("outputTokens" in i ? i.outputTokens : void 0), s = W("totalTokens" in i ? i.totalTokens : void 0), c = [
+		a,
+		o,
+		s
+	].filter((e) => e !== void 0).length, l = Ie(r);
+	return {
+		provider: e,
+		model: t,
+		usage: {
+			...a === void 0 ? {} : { inputTokens: a },
+			...o === void 0 ? {} : { outputTokens: o },
+			...s === void 0 ? {} : { totalTokens: s },
+			completeness: c === 3 ? "complete" : c === 0 ? "unknown" : "partial"
+		},
+		cost: l === void 0 ? {
+			status: "unknown",
+			source: "unreported"
+		} : {
+			status: "actual",
+			source: "gateway",
+			usd: l
+		}
+	};
+}, K = class extends h.TaggedError()("AiError", {
+	reason: h.Literal("generation", "decode"),
+	message: h.String,
+	evidence: h.optional(qe)
+}) {}, q = (e, t, n) => {
+	let r = [e, ...t ?? []], i = (e) => n(r[e]).pipe(f.catchTag("AiError", (t) => t.reason === "generation" && e + 1 < r.length ? i(e + 1) : f.fail(t)));
+	return i(0);
+}, Je = (e) => f.gen(function* () {
+	let r = e.messages === void 0 ? e.prompt === void 0 ? void 0 : { prompt: e.prompt } : { messages: [...e.messages] };
+	if (r === void 0 || e.prompt !== void 0 && e.messages !== void 0) return yield* f.fail(new K({
+		reason: "generation",
+		message: "generateText: supply exactly one of prompt or messages"
+	}));
+	let i = e.tools === void 0 ? void 0 : yield* f.runtime(), a = e.tools === void 0 || i === void 0 ? void 0 : O(e.tools, {
+		store: void 0,
+		...e.toolContext,
+		runtime: i
+	});
+	return yield* q(e.provider ?? n(), e.fallbacks, (n) => H(B(n, "generateText"), (e) => e.usage, f.tryPromise({
+		try: async () => {
+			let i = t(n), o = await s({
+				model: i,
+				...r,
+				...a === void 0 ? {} : {
+					tools: a,
+					stopWhen: l(e.maxSteps ?? 8)
+				},
+				...e.system === void 0 ? {} : { system: e.system },
+				...e.maxTokens === void 0 ? {} : { maxOutputTokens: e.maxTokens },
+				...e.temperature === void 0 ? {} : { temperature: e.temperature },
+				...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+			}), c = a !== void 0 && o.totalUsage !== void 0 ? o.totalUsage : o.usage;
+			return {
+				text: o.text,
+				sources: o.sources,
+				usage: {
+					inputTokens: c.inputTokens,
+					outputTokens: c.outputTokens,
+					totalTokens: c.totalTokens
+				},
+				providerMetadata: o.providerMetadata
+			};
+		},
+		catch: (e) => new K({
+			reason: "generation",
+			message: String(e?.message ?? e)
+		})
+	})));
+}), J = (e, t) => f.gen(function* () {
+	let n = p.make(e);
+	if (n?.type !== "object") {
+		let e = n?.anyOf === void 0 ? `type:${JSON.stringify(n?.type ?? null)}` : "anyOf (e.g. Schema.Struct({}) or a union)";
+		return yield* f.fail(new K({
+			reason: "generation",
+			message: `${t}: the output schema must render as a top-level JSON object (type:"object"), but got ${e}. Use a Schema.Struct({ ... }) as the output schema — wrap arrays/scalars in a field (e.g. Schema.Struct({ items: Schema.Array(...) })).`
+		}));
+	}
+	return n;
+}), Ye = (e, t) => J(e, t), Xe = (e) => q(e.provider ?? n(), e.fallbacks, (n) => H(B(n, "generateObject"), (e) => e.usage, f.gen(function* () {
+	let i = yield* f.try({
+		try: () => Ke(e.jsonSchema),
+		catch: () => new K({
+			reason: "decode",
+			message: "generateObjectJson: invalid or unsupported output JSON Schema; no provider call was made"
+		})
+	}), a = t(n), s = yield* f.tryPromise({
+		try: () => o({
+			model: a,
+			schema: i,
+			prompt: e.prompt,
+			...e.system === void 0 ? {} : { system: e.system },
+			...e.maxTokens === void 0 ? {} : { maxOutputTokens: e.maxTokens },
+			...e.temperature === void 0 ? {} : { temperature: e.temperature },
+			...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+		}),
+		catch: (e) => r.isInstance(e) ? new K({
+			reason: "decode",
+			message: "model output failed JSON Schema validation",
+			evidence: G(n.name, a.modelId, e.usage)
+		}) : new K({
+			reason: "generation",
+			message: "structured provider call failed",
+			evidence: G(n.name, a.modelId, void 0)
+		})
+	});
+	return {
+		object: s.object,
+		evidence: G(n.name, a.modelId, s.usage, s.providerMetadata),
+		usage: {
+			inputTokens: s.usage.inputTokens,
+			outputTokens: s.usage.outputTokens,
+			totalTokens: s.usage.totalTokens
+		},
+		providerMetadata: s.providerMetadata
+	};
+}))), Ze = (e) => f.gen(function* () {
+	let t = yield* Xe({
+		jsonSchema: yield* J(e.schema, "generateObject"),
+		prompt: e.prompt,
+		...e.system === void 0 ? {} : { system: e.system },
+		...e.provider === void 0 ? {} : { provider: e.provider },
+		...e.fallbacks === void 0 ? {} : { fallbacks: e.fallbacks },
+		...e.maxTokens === void 0 ? {} : { maxTokens: e.maxTokens },
+		...e.temperature === void 0 ? {} : { temperature: e.temperature },
+		...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+	}), n = yield* h.decodeUnknown(e.schema)(t.object).pipe(f.mapError(() => new K({
+		reason: "decode",
+		message: "model output failed schema decode",
+		evidence: t.evidence
+	})));
+	return {
+		...t,
+		object: n
+	};
+}), Qe = (e) => f.gen(function* () {
+	let a = yield* f.runtime(), o = yield* J(e.schema, "generateObjectWithTools"), u = {
+		store: void 0,
+		...e.toolContext,
+		runtime: a
+	}, d = O(e.tools, u);
+	return yield* q(e.provider ?? n(), e.fallbacks, (n) => H(B(n, "generateObjectWithTools"), (e) => e.usage, f.gen(function* () {
+		let a = t(n), u = yield* f.tryPromise({
+			try: () => s({
+				model: a,
+				prompt: e.prompt,
+				tools: d,
+				stopWhen: l(e.maxSteps ?? 8),
+				output: i.object({ schema: c(o) }),
+				...e.system === void 0 ? {} : { system: e.system },
+				...e.maxTokens === void 0 ? {} : { maxOutputTokens: e.maxTokens },
+				...e.temperature === void 0 ? {} : { temperature: e.temperature },
+				...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+			}),
+			catch: (e) => r.isInstance(e) ? new K({
+				reason: "decode",
+				message: "model output failed structured decode",
+				evidence: G(n.name, a.modelId, e.usage)
+			}) : new K({
+				reason: "generation",
+				message: "structured provider call failed",
+				evidence: G(n.name, a.modelId, void 0)
+			})
+		}), p = u.totalUsage, m = G(n.name, a.modelId, p, u.steps.length === 1 ? u.providerMetadata : void 0), ee = yield* f.try({
+			try: () => u.output,
+			catch: () => new K({
+				reason: "decode",
+				message: "model output failed structured decode",
+				evidence: m
+			})
+		});
+		return {
+			object: yield* h.decodeUnknown(e.schema)(ee).pipe(f.mapError(() => new K({
+				reason: "decode",
+				message: "model output failed schema decode",
+				evidence: m
+			}))),
+			evidence: m,
+			usage: {
+				inputTokens: p.inputTokens,
+				outputTokens: p.outputTokens,
+				totalTokens: p.totalTokens
+			},
+			providerMetadata: u.providerMetadata
+		};
+	})));
+}), $e = class extends d.Tag("@voltro/AiService")() {}, et = {
+	generateText: (e) => Je(e),
+	generateObject: (e) => Ze(e)
+}, tt = async (e, t) => {
+	if ("bytes" in e) return {
+		audio: e.bytes,
+		mediaType: e.mediaType
+	};
+	if ("url" in e) {
+		let t = await fetch(e.url);
+		if (!t.ok) throw Error(`transcribe: GET ${e.url} answered ${t.status}.`);
+		let n = t.headers.get("content-type") ?? void 0;
+		return {
+			audio: new Uint8Array(await t.arrayBuffer()),
+			...n === void 0 ? {} : { mediaType: n }
+		};
+	}
+	if (t === void 0) throw Error("transcribe: `source.storageRef` needs a `resolveRef` to turn the reference into bytes.\n  → pass `resolveRef: (ref) => storage.read(ref)`. It is injected rather than imported so @voltro/ai does not depend on storage; the app owns tenancy and signing for that read.");
+	let { bytes: n, mediaType: r } = await t(e.storageRef);
+	return {
+		audio: n,
+		mediaType: r
+	};
+}, nt = (e) => {
+	if (!Array.isArray(e) || e.length === 0) return;
+	let t = [];
+	for (let n of e) {
+		let e = typeof n.startSecond == "number" ? n.startSecond : typeof n.start == "number" ? n.start : void 0, r = typeof n.endSecond == "number" ? n.endSecond : typeof n.end == "number" ? n.end : void 0, i = typeof n.text == "string" ? n.text : void 0;
+		if (e === void 0 || r === void 0 || i === void 0) continue;
+		let a = typeof n.speaker == "string" ? n.speaker : void 0;
+		t.push({
+			start: e,
+			end: r,
+			text: i,
+			...a === void 0 ? {} : { speaker: a }
+		});
+	}
+	return t.length > 0 ? t : void 0;
+}, rt = (t) => {
+	let r = t.provider ?? n();
+	return H(B(r, "transcribe"), () => ({}), f.tryPromise({
+		try: async () => {
+			let { audio: n, mediaType: i } = await tt(t.source, t.resolveRef), o = await a({
+				model: e(r),
+				audio: n,
+				...i === void 0 ? {} : { mediaType: i },
+				...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions }
+			}), s = t.timestamps === "none" ? void 0 : nt(o.segments);
+			return {
+				text: o.text,
+				...o.language === void 0 ? {} : { language: o.language },
+				...o.durationInSeconds === void 0 ? {} : { durationSec: o.durationInSeconds },
+				...s === void 0 ? {} : { segments: s },
+				...o.providerMetadata === void 0 ? {} : { providerMetadata: o.providerMetadata }
+			};
+		},
+		catch: (e) => new K({
+			reason: "generation",
+			message: String(e?.message ?? e)
+		})
+	}));
+}, it = (e) => `sha256:${ue("sha256").update(e).digest("hex").slice(0, 16)}`, at = class extends h.TaggedError()("PromptRenderError", {
+	promptId: h.String,
+	missing: h.Array(h.String),
+	message: h.String
+}) {}, ot = /\{\{\s*([A-Za-z0-9_.]+)\s*\}\}/g, st = (e) => {
+	let t = /* @__PURE__ */ new Set();
+	for (let n of e.matchAll(ot)) {
+		let e = n[1];
+		e !== void 0 && t.add(e);
+	}
+	return [...t];
+}, ct = (e, t) => it(`${e.length}:${e}${(t ?? "").length}:${t ?? ""}`), lt = (e) => {
+	let t = ct(e.template, e.system), n = st(e.template);
+	return {
+		...e,
+		digest: t,
+		variables: n,
+		render: (r = {}) => f.suspend(() => {
+			let i = n.filter((e) => r[e] === void 0);
+			if (i.length > 0) return f.fail(new at({
+				promptId: e.id,
+				missing: i,
+				message: `prompt '${e.id}' needs ${i.map((e) => `{{${e}}}`).join(", ")}, which ${i.length === 1 ? "was" : "were"} not supplied.`
+			}));
+			let a = e.template.replace(ot, (e, t) => String(r[t]));
+			return f.succeed({
+				promptId: e.id,
+				digest: t,
+				...e.label === void 0 ? {} : { label: e.label },
+				prompt: a,
+				...e.system === void 0 ? {} : { system: e.system },
+				definition: e
+			});
+		})
+	};
+}, ut = (e) => typeof e == "object" && !!e && typeof e.prompt == "string" && typeof e.digest == "string" && typeof e.promptId == "string", Y = "_voltro_prompts", X = T(Y, {
+	id: y({ prefix: "prm" }),
+	promptId: E().maxLength(191),
+	digest: E().maxLength(191),
+	revision: b().default(1),
+	label: E().nullable(),
+	template: E().nullable(),
+	system: E().nullable(),
+	templateChars: b().default(0),
+	firstSeenAt: D().default("now"),
+	lastUsedAt: D().default("now")
+}).unique("byPromptsVersion", ["promptId", "digest"]).index(["promptId"]), dt = 8760, ft = () => {
+	C({
+		source: "framework",
+		table: Y,
+		timeColumn: "lastUsedAt",
+		ttlMs: w(process.env.VOLTRO_AI_PROMPTS_TTL_HOURS, dt),
+		envVar: "VOLTRO_AI_PROMPTS_TTL_HOURS",
+		label: "ai prompt versions"
+	});
+}, Z = (e) => ({
+	promptId: String(e.promptId),
+	digest: String(e.digest),
+	revision: Number(e.revision) || 1,
+	label: e.label ?? null,
+	template: e.template ?? null,
+	system: e.system ?? null,
+	firstSeenAt: e.firstSeenAt ?? null,
+	lastUsedAt: e.lastUsedAt ?? null
+}), pt = (e, t) => f.gen(function* () {
+	ft();
+	let n = t.digest ?? lt(t).digest;
+	return yield* f.promise(async () => {
+		let r = (await e.query(S(X).where(g(_("promptId", t.id), _("digest", n))).descriptor))[0];
+		if (r !== void 0) {
+			let t = /* @__PURE__ */ new Date();
+			return await e.updateMany(Y, { lastUsedAt: t }, { where: _("id", String(r.id)) }), Z({
+				...r,
+				lastUsedAt: t
+			});
+		}
+		let i = (await e.query(S(X).where(_("promptId", t.id)).descriptor)).reduce((e, t) => Math.max(e, Number(t.revision) || 0), 0) + 1, a = /* @__PURE__ */ new Date(), o = {
+			id: v({
+				kind: "typeid",
+				prefix: "prm"
+			}, Y),
+			promptId: t.id,
+			digest: n,
+			revision: i,
+			label: t.label ?? null,
+			template: t.template,
+			system: t.system ?? null,
+			templateChars: t.template.length,
+			firstSeenAt: a,
+			lastUsedAt: a
+		};
+		return Z(await e.insertIgnore(Y, o, { conflictColumns: ["promptId", "digest"] }) ?? o);
+	});
+}), mt = (e, t) => f.promise(async () => {
+	let n = (await e.query(S(X).where(_("digest", t)).descriptor))[0];
+	return n === void 0 ? null : Z(n);
+}), ht = (e, t) => f.promise(async () => [...await e.query(S(X).where(_("promptId", t)).orderBy("revision", "desc").descriptor)].sort((e, t) => (Number(t.revision) || 0) - (Number(e.revision) || 0)).map(Z)), gt = "_voltro_ai_inferences", _t = T(gt, {
+	id: y({ prefix: "aiinf" }),
+	tenantId: E().nullable(),
+	executionId: E(),
+	workflowName: E(),
+	runId: E().nullable(),
+	signalName: E(),
+	stepName: E(),
+	status: E().oneOf([
+		"pending",
+		"running",
+		"succeeded",
+		"failed"
+	]).default("pending"),
+	kind: E().oneOf([
+		"text",
+		"object",
+		"transcription"
+	]).default("text"),
+	provider: E(),
+	model: E(),
+	prompt: E(),
+	system: E().nullable(),
+	promptId: E().nullable(),
+	promptDigest: E().nullable(),
+	promptRevision: b().nullable(),
+	maxTokens: b().nullable(),
+	providerOptions: x().nullable(),
+	jsonSchema: x().nullable(),
+	transcription: x().nullable(),
+	result: x().nullable(),
+	errorMessage: E().nullable(),
+	attempts: b().default(0),
+	leaseUntil: D().nullable(),
+	claimedBy: E().nullable(),
+	enqueuedAt: D().default("now"),
+	startedAt: D().nullable(),
+	completedAt: D().nullable(),
+	durationMs: b().nullable()
+}).index(["status", "enqueuedAt"]).index(["status", "leaseUntil"]).index(["executionId"]).index(["tenantId"]), vt = 3e5, yt = 4, Q = (e, t) => {
+	let n = typeof e == "number" ? e : Number(e);
+	return Number.isFinite(n) ? Math.trunc(n) : t;
+}, $ = (e) => e == null ? null : String(e), bt = (e) => ({
+	id: String(e.id),
+	tenantId: $(e.tenantId),
+	executionId: String(e.executionId),
+	workflowName: String(e.workflowName),
+	runId: $(e.runId),
+	signalName: String(e.signalName),
+	stepName: String(e.stepName),
+	kind: e.kind === "object" ? "object" : e.kind === "transcription" ? "transcription" : "text",
+	provider: String(e.provider),
+	model: String(e.model),
+	prompt: String(e.prompt ?? ""),
+	system: $(e.system),
+	promptId: $(e.promptId),
+	promptDigest: $(e.promptDigest),
+	promptRevision: e.promptRevision === null || e.promptRevision === void 0 ? null : Q(e.promptRevision, 0),
+	maxTokens: e.maxTokens === null || e.maxTokens === void 0 ? null : Q(e.maxTokens, 0),
+	providerOptions: e.providerOptions ?? null,
+	jsonSchema: e.jsonSchema ?? null,
+	transcription: e.transcription ?? null,
+	attempts: Q(e.attempts, 0)
+}), xt = (e) => {
+	let t = e.attempts + 1;
+	return t >= e.maxAttempts ? {
+		kind: "give-up",
+		attempts: t
+	} : {
+		kind: "retry",
+		attempts: t,
+		delayMs: e.retryAfterMs !== void 0 && e.retryAfterMs > 0 ? e.retryAfterMs : Math.min(e.baseDelayMs * 2 ** (t - 1), e.maxDelayMs)
+	};
+}, St = (e) => `voltro.ai.infer:${e}`, Ct = (e, t) => `aiinf_${ue("sha256").update(`${e}\u0000${t}`).digest("hex").slice(0, 40)}`;
+//#endregion
+export { pe as $, B as A, je as B, et as C, Je as D, Qe as E, Re as F, j as G, Me as H, Fe as I, xe as J, A as K, I as L, H as M, L as N, Ye as O, ke as P, O as Q, Ie as R, $e as S, Xe as T, Le as U, Pe as V, Ae as W, Oe as X, M as Y, De as Z, ht as _, xt as a, rt as b, bt as c, at as d, me as et, lt as f, ct as g, mt as h, _t as i, Ge as j, qe as k, dt as l, it as m, yt as n, St as o, ut as p, k as q, vt as r, Ct as s, gt as t, he as tt, Y as u, X as v, Ze as w, K as x, pt as y, Ne as z };