@voltro/ai 0.57.0 → 0.59.0

package/dist/{provider-BOC5wBxx.js → provider-BOgeCRkg.js}

@@ -171,7 +171,7 @@ var u = [], d = (...e) => {
 			e.close();
 		} }) };
 	}
-}), T = "iVBORw0KGgo=", E = "SUQzBAAAAAAAAA==", D = "video/mp4", O = (e, t) => ({
+}), T = "iVBORw0KGgo=", E = "SUQzBAAAAAAAAA==", D = "video/mp4", O = "mock transcript", k = (e, t) => ({
 	specificationVersion: "v4",
 	provider: "mock",
 	modelId: e,
@@ -190,7 +190,7 @@ var u = [], d = (...e) => {
 			totalTokens: 0
 		}
 	})
-}), k = (e, t, n) => ({
+}), A = (e, t, n) => ({
 	specificationVersion: "v4",
 	provider: "mock",
 	modelId: e,
@@ -203,7 +203,25 @@ var u = [], d = (...e) => {
 		},
 		...n === void 0 ? {} : { providerMetadata: n }
 	})
-}), A = (e, t, n) => ({
+}), j = (e, t, n) => ({
+	specificationVersion: "v4",
+	provider: "mock",
+	modelId: e,
+	doGenerate: async () => ({
+		text: t,
+		segments: [],
+		language: "en",
+		durationInSeconds: 1,
+		warnings: [],
+		response: {
+			timestamp: /* @__PURE__ */ new Date(),
+			modelId: e,
+			headers: {},
+			body: void 0
+		},
+		...n === void 0 ? {} : { providerMetadata: n }
+	})
+}), M = (e, t, n) => ({
 	specificationVersion: "v4",
 	provider: "mock",
 	modelId: e,
@@ -218,31 +236,38 @@ var u = [], d = (...e) => {
 		},
 		...n === void 0 ? {} : { providerMetadata: n }
 	})
-}), j = (e) => e.apiKey !== void 0 || e.baseURL !== void 0, M = (e) => ({
+}), N = (e) => e.apiKey !== void 0 || e.baseURL !== void 0, P = (e) => ({
 	...e.apiKey === void 0 ? {} : { apiKey: e.apiKey },
 	...e.baseURL === void 0 ? {} : { baseURL: e.baseURL }
-}), N = (o, s) => m((() => {
+}), F = (o, s) => m((() => {
 	switch (o.name) {
 		case "mock": return w(o.model ?? "mock", o.mockText, o.script, o.providerMetadata);
-		case "anthropic": return j(o) ? t(M(o))(o.model) : e(o.model);
-		case "openai": return j(o) ? n(M(o))(o.model) : r(o.model);
-		case "gateway": return j(o) ? i(M(o))(o.model) : a(o.model);
+		case "anthropic": return N(o) ? t(P(o))(o.model) : e(o.model);
+		case "openai": return N(o) ? n(P(o))(o.model) : r(o.model);
+		case "gateway": return N(o) ? i(P(o))(o.model) : a(o.model);
 	}
-})(), s), P = (e) => {
+})(), s), I = (e) => {
 	switch (e.name) {
-		case "mock": return O(e.model ?? "mock", e.mockMedia?.imageBase64 ?? T);
-		case "openai": return j(e) ? n(M(e)).image(e.model) : r.image(e.model);
-		case "gateway": return j(e) ? i(M(e)).image(e.model) : a.image(e.model);
+		case "mock": return k(e.model ?? "mock", e.mockMedia?.imageBase64 ?? T);
+		case "openai": return N(e) ? n(P(e)).image(e.model) : r.image(e.model);
+		case "gateway": return N(e) ? i(P(e)).image(e.model) : a.image(e.model);
 		case "anthropic": throw Error("Provider 'anthropic' does not support image generation. Use 'openai' or 'gateway'.");
 	}
-}, F = (e) => {
+}, L = (e) => {
 	switch (e.name) {
-		case "mock": return k(e.model ?? "mock", e.mockMedia?.audioBase64 ?? E, e.providerMetadata);
-		case "openai": return j(e) ? n(M(e)).speech(e.model) : r.speech(e.model);
-		case "gateway": return j(e) ? i(M(e)).speech(e.model) : a.speech(e.model);
+		case "mock": return A(e.model ?? "mock", e.mockMedia?.audioBase64 ?? E, e.providerMetadata);
+		case "openai": return N(e) ? n(P(e)).speech(e.model) : r.speech(e.model);
+		case "gateway": return N(e) ? i(P(e)).speech(e.model) : a.speech(e.model);
 		case "anthropic": throw Error("Provider 'anthropic' does not support speech generation. Use 'openai' or 'gateway'.");
 	}
-}, I = (e) => {
+}, R = (e) => {
+	switch (e.name) {
+		case "mock": return j(e.model ?? "mock", e.mockMedia?.transcript ?? O, e.providerMetadata);
+		case "openai": return N(e) ? n(P(e)).transcription(e.model) : r.transcription(e.model);
+		case "gateway": return N(e) ? i(P(e)).transcription(e.model) : a.transcription(e.model);
+		case "anthropic": throw Error("Provider 'anthropic' does not support transcription. Use 'openai' or 'gateway'.");
+	}
+}, z = (e) => {
 	switch (e.name) {
 		case "mock": {
 			let t = e.mockMedia, n = t?.videoBase64 === void 0 ? {
@@ -254,17 +279,17 @@ var u = [], d = (...e) => {
 				data: t.videoBase64,
 				mediaType: D
 			};
-			return A(e.model ?? "mock", n, e.providerMetadata);
+			return M(e.model ?? "mock", n, e.providerMetadata);
 		}
-		case "gateway": return j(e) ? i(M(e)).video(e.model) : a.video(e.model);
+		case "gateway": return N(e) ? i(P(e)).video(e.model) : a.video(e.model);
 		case "openai": throw Error("Provider 'openai' does not support video generation. Use 'gateway'.");
 		case "anthropic": throw Error("Provider 'anthropic' does not support video generation. Use 'gateway'.");
 	}
-}, L = null, R = (e) => {
-	let t = L;
-	return L = e, t;
-}, z = () => L, B = (e = process.env) => {
-	if (L) return L;
+}, B = null, V = (e) => {
+	let t = B;
+	return B = e, t;
+}, H = () => B, U = (e = process.env) => {
+	if (B) return B;
 	let t = (e.AI_PROVIDER ?? "mock").toLowerCase(), n = e.AI_MODEL;
 	switch (t) {
 		case "mock": return {
@@ -291,4 +316,4 @@ var u = [], d = (...e) => {
 	}
 };
 //#endregion
-export { N as a, R as c, o as d, s as f, c as g, d as h, P as i, m as l, h as m, z as n, F as o, f as p, B as r, I as s, v as t, p as u };
+export { c as _, F as a, z as c, p as d, o as f, d as g, h, I as i, V as l, f as m, H as n, L as o, s as p, U as r, R as s, v as t, m as u };

package/dist/test.js

@@ -1,4 +1,4 @@
-import { c as e, t } from "./provider-BOC5wBxx.js";
+import { l as e, t } from "./provider-BOgeCRkg.js";
 //#region src/test/index.ts
 var n = (e) => e.turns && e.turns.length > 0 ? e.turns : e.stream && e.stream.length > 0 ? [{ text: e.stream.join("") }] : e.generate ? [{ text: e.generate.text }] : [{ text: "" }], r = (e) => {
 	let r = new t(n(e));

package/dist/workflow.d.ts

@@ -34,18 +34,6 @@ declare type AiJsonValue = null | boolean | number | string | {
     [key: string]: AiJsonValue;
 } | AiJsonValue[];
 
-/**
- * A schema-constrained completion as a durable step.
- *
- * The schema is the step's success schema too, so the journaled value is
- * decoded on replay exactly as it was on the first run — a structured result
- * that only validated the first time through would be the worst kind of replay
- * bug.
- *
- * Offloaded, the schema is rendered to JSON Schema for the dispatcher (a
- * JavaScript Schema cannot be journaled) and the DECODE still happens here,
- * where the real schema exists.
- */
 export declare const aiObjectStep: <A, I>(options: AiObjectStepOptions<A, I>) => Effect.Effect<A, unknown>;
 
 export declare interface AiObjectStepOptions<A, I> extends AiStepBaseOptions {
@@ -394,6 +382,9 @@ declare type ProviderConfig = {
         /** base64 the mock video model yields INSTEAD of a URL — exercises the
          *  bytes branch of `generateVideo`. Takes precedence over `videoUrl`. */
         readonly videoBase64?: string;
+        /** The text the mock transcription model returns. Lets a test assert on
+         *  an exact transcript without a provider. */
+        readonly transcript?: string;
     };
 } | {
     readonly name: 'anthropic';
@@ -452,4 +443,70 @@ export declare const resolveStepPrompt: (options: Pick<AiStepBaseOptions, "promp
     readonly rendered: RenderedPrompt | undefined;
 };
 
+/**
+ * Transcription as a durable step — with `offload`, which is the point.
+ *
+ * A ninety-minute recording is not a request. Without offloading, a worker holds
+ * a connection for the whole call; `aiStep` and `aiObjectStep` have had
+ * `offload: true` all along and transcription — the modality that runs LONGEST —
+ * did not. The queue row carries the SOURCE, never the bytes: putting the audio
+ * in the row would reintroduce exactly the cost offloading exists to remove.
+ */
+export declare const transcribeStep: (options: TranscribeStepOptions) => Effect.Effect<{
+    readonly text: string;
+    readonly language?: string;
+    readonly durationSec?: number;
+}, unknown>;
+
+export declare interface TranscribeStepOptions extends Omit<AiStepBaseOptions, 'prompt' | 'system' | 'promptTemplate' | 'promptVars'> {
+    readonly source: TranscriptionSource;
+    readonly language?: string;
+    readonly timestamps?: TranscriptionTimestamps;
+    readonly diarize?: boolean;
+    readonly translateToEnglish?: boolean;
+    /** Resolve a `storageRef`. Required only for a ref source — and for an
+     *  OFFLOADED one it must be wired on the dispatcher too, since that is the
+     *  process that will do the read. */
+    readonly resolveRef?: (ref: string) => Promise<{
+        bytes: Uint8Array;
+        mediaType: string;
+    }>;
+}
+
+/**
+ * Where the audio is.
+ *
+ * `storageRef` is the one to reach for: the file is already in the app's
+ * storage, and resolving it there keeps tenancy, signing and lifetime with the
+ * component that owns them. The resolver is injected rather than imported so
+ * this module stays free of a storage dependency — `@voltro/ai` must not learn
+ * about buckets to transcribe a file.
+ */
+declare type TranscriptionSource = 
+/** Resolved by the app's own storage — the case to reach for. */
+    {
+    readonly storageRef: string;
+}
+/** Fetched by this function. Convenient, but the bytes do pass through this
+*  process: the SDK takes bytes or base64 and never a location. */
+| {
+    readonly url: string;
+}
+/** Already in hand. The one that costs memory by construction. */
+| {
+    readonly bytes: Uint8Array;
+    readonly mediaType: string;
+};
+
+/** What a `transcribeStep` journals. Structural, because a durable step's
+ *  success schema is what a replay decodes through — see `AiStepText`. */
+export declare const TranscriptionStepResult: Schema.Struct<{
+    text: typeof Schema.String;
+    language: Schema.optional<typeof Schema.String>;
+    durationSec: Schema.optional<typeof Schema.Number>;
+}>;
+
+/** How much timestamp detail to ask for. Providers charge for it. */
+declare type TranscriptionTimestamps = 'none' | 'segment' | 'word';
+
 export { }

package/dist/workflow.js

@@ -1,17 +1,17 @@
-import { r as e } from "./provider-BOC5wBxx.js";
-import { C as t, D as n, E as r, H as i, U as a, m as o, o as s, p as c, s as l, t as u, y as d, z as f } from "./inferenceQueue-CJZ4xw3R.js";
-import { Effect as p, Schema as m } from "effect";
-import { eq as h } from "@voltro/database";
-import { CurrentWorkflowExecutionId as g, awaitSignalSuspending as _, step as v, suspendForBudget as y } from "@voltro/workflow";
+import { r as e } from "./provider-BOgeCRkg.js";
+import { B as t, D as n, G as r, K as i, O as a, b as o, m as s, o as c, p as l, s as u, t as d, w as f, y as p } from "./inferenceQueue-CS1VouiW.js";
+import { Effect as m, Schema as h } from "effect";
+import { eq as g } from "@voltro/database";
+import { CurrentWorkflowExecutionId as _, awaitSignalSuspending as v, step as y, suspendForBudget as b } from "@voltro/workflow";
 //#region src/inferStep.ts
-var b = {
+var x = {
 	maxAttempts: 4,
 	strategy: "exponential",
 	baseDelay: "1 second",
 	maxDelay: "30 seconds",
 	jitter: !0,
 	respectRetryAfter: !0
-}, x = 36e5, S = "ai-usd", C = (e, t, n, r) => {
+}, S = 36e5, C = "ai-usd", w = (e, t, n, r) => {
 	let i = r === void 0 ? {} : {
 		promptId: r.promptId,
 		promptVersion: r.digest,
@@ -24,14 +24,14 @@ var b = {
 		...t === void 0 ? {} : { system: t }
 	} : {
 		...i,
-		promptDigest: o(e),
+		promptDigest: s(e),
 		promptChars: e.length,
 		...t === void 0 ? {} : {
-			systemDigest: o(t),
+			systemDigest: s(t),
 			systemChars: t.length
 		}
 	};
-}, w = (e) => c(e.prompt) ? {
+}, T = (e) => l(e.prompt) ? {
 	text: e.prompt.prompt,
 	system: e.system ?? e.prompt.system,
 	rendered: e.prompt
@@ -39,68 +39,68 @@ var b = {
 	text: e.prompt,
 	system: e.system,
 	rendered: void 0
-}, T = (e) => ({
+}, E = (e) => ({
 	provider: String(e.provider ?? "unknown"),
 	model: String(e.model ?? "unknown")
-}), E = (e, t, n, r, a) => {
-	let o = e.store;
-	if (o === void 0) return p.void;
-	let { provider: s, model: c } = T(t), l = f(r.providerMetadata);
-	return i(o, {
-		provider: s,
-		model: c,
-		operation: n,
+}), D = (e, n, i, a, o) => {
+	let s = e.store;
+	if (s === void 0) return m.void;
+	let { provider: c, model: l } = E(n), u = t(a.providerMetadata);
+	return r(s, {
+		provider: c,
+		model: l,
+		operation: i,
 		agent: e.name,
-		usage: r.usage,
+		usage: a.usage,
 		...e.tenantId === void 0 ? {} : { tenantId: e.tenantId },
-		...l === void 0 ? {} : { actualCostUsd: l },
-		...a === void 0 ? {} : { prompt: a }
-	}).pipe(p.asVoid, p.catchAllCause(() => p.void));
-}, D = (e, t) => {
+		...u === void 0 ? {} : { actualCostUsd: u },
+		...o === void 0 ? {} : { prompt: o }
+	}).pipe(m.asVoid, m.catchAllCause(() => m.void));
+}, O = (e, t) => {
 	let n = e.store;
-	return n === void 0 || t === void 0 ? p.void : d(n, {
+	return n === void 0 || t === void 0 ? m.void : p(n, {
 		...t.definition,
 		digest: t.digest
-	}).pipe(p.asVoid, p.catchAllCause(() => p.void));
-}, O = m.Struct({
-	text: m.String,
-	inputTokens: m.NullOr(m.Number),
-	outputTokens: m.NullOr(m.Number),
-	totalTokens: m.NullOr(m.Number)
-}), k = (e) => typeof e == "object" && !!e && e.__voltro_ai_error === !0, A = class extends Error {
+	}).pipe(m.asVoid, m.catchAllCause(() => m.void));
+}, k = h.Struct({
+	text: h.String,
+	inputTokens: h.NullOr(h.Number),
+	outputTokens: h.NullOr(h.Number),
+	totalTokens: h.NullOr(h.Number)
+}), A = (e) => typeof e == "object" && !!e && e.__voltro_ai_error === !0, j = class extends Error {
 	attempts;
 	constructor(e, t, n) {
 		super(`aiStep('${e}'): the offloaded model call gave up after ${n} attempt(s): ${t}. The queued call is in _voltro_ai_inferences with status 'failed' — the Flow tab shows it with the provider's own message.`), this.name = "AiOffloadError", this.attempts = n;
 	}
-}, j = class extends Error {
+}, M = class extends Error {
 	constructor(e) {
 		super(`aiStep('${e}'): offload: true needs \`store\` — the queued call lives in a table and the dispatcher reads it from there. Pass ctx.store.`), this.name = "AiOffloadMisuseError";
 	}
-}, M = (e, t, n, r, i) => p.gen(function* () {
-	let a = e.store;
-	if (a === void 0) return yield* p.die(new j(e.name));
-	let o = yield* g;
-	if (o === void 0) return yield* p.die(/* @__PURE__ */ Error(`aiStep('${e.name}'): offload: true is only valid inside a workflow body — there is no execution to suspend here.`));
-	let { provider: c, model: d } = T(t), f = l(o, e.name), m = yield* p.promise(() => a.query({
+}, N = (e, t, n, r, i, a) => m.gen(function* () {
+	let o = e.store;
+	if (o === void 0) return yield* m.die(new M(e.name));
+	let s = yield* _;
+	if (s === void 0) return yield* m.die(/* @__PURE__ */ Error(`aiStep('${e.name}'): offload: true is only valid inside a workflow body — there is no execution to suspend here.`));
+	let { provider: l, model: f } = E(t), p = u(s, e.name), h = yield* m.promise(() => o.query({
 		table: "_voltro_workflow_runs",
-		predicate: h("executionId", o),
+		predicate: g("executionId", s),
 		order: [],
 		take: 1,
 		skip: void 0,
 		projection: void 0
-	})), _ = String(m[0]?.tag ?? ""), v = m[0]?.id;
-	return yield* p.promise(() => a.insertIgnore(u, {
-		id: f,
+	})), v = String(h[0]?.tag ?? ""), y = h[0]?.id;
+	return yield* m.promise(() => o.insertIgnore(d, {
+		id: p,
 		tenantId: e.tenantId ?? null,
-		executionId: o,
-		workflowName: _,
-		runId: typeof v == "string" ? v : null,
-		signalName: s(f),
+		executionId: s,
+		workflowName: v,
+		runId: typeof y == "string" ? y : null,
+		signalName: c(p),
 		stepName: e.name,
 		status: "pending",
 		kind: n,
-		provider: c,
-		model: d,
+		provider: l,
+		model: f,
 		prompt: i.text,
 		system: i.system ?? null,
 		promptId: i.rendered?.promptId ?? null,
@@ -109,120 +109,156 @@ var b = {
 		maxTokens: e.maxTokens ?? null,
 		providerOptions: e.providerOptions ?? null,
 		jsonSchema: r,
+		...a === void 0 ? {} : { transcription: a },
 		attempts: 0,
 		enqueuedAt: /* @__PURE__ */ new Date()
-	}, { conflictColumns: ["id"] })), f;
-}), N = (e, t, n, r, i, a) => p.gen(function* () {
-	let o = yield* v({
+	}, { conflictColumns: ["id"] })), p;
+}), P = (e, t, n, r, i, a, o) => m.gen(function* () {
+	let s = yield* y({
 		name: `${e.name}:enqueue`,
-		success: m.String,
+		success: h.String,
 		input: i,
-		execute: M(e, t, n, r, a)
+		execute: N(e, t, n, r, a, o)
 	});
-	yield* D(e, a.rendered);
-	let c = yield* _({ store: e.store }, {
-		name: s(o),
-		schema: m.Unknown,
+	yield* O(e, a.rendered);
+	let l = yield* v({ store: e.store }, {
+		name: c(s),
+		schema: h.Unknown,
 		timeoutMs: e.offloadTimeoutMs ?? 36e5
 	});
-	return k(c) ? yield* p.die(new A(e.name, c.message, c.attempts)) : c;
-}), P = class extends Error {
+	return A(l) ? yield* m.die(new j(e.name, l.message, l.attempts)) : l;
+}), F = class extends Error {
 	constructor(e) {
 		super(`aiStep('${e}'): budget needs \`store\` — the reservation counter and the hold both live in tables. Pass ctx.store.`), this.name = "AiBudgetMisuseError";
 	}
-}, F = (e) => p.gen(function* () {
+}, I = (e) => m.gen(function* () {
 	let t = e.budget;
 	if (t === void 0) return;
 	let n = e.store;
-	if (n === void 0) return yield* p.die(new P(e.name));
-	let r = e.tenantId ?? null, i = t.name ?? "ai-usd", o = () => a(n, {
+	if (n === void 0) return yield* m.die(new F(e.name));
+	let r = e.tenantId ?? null, a = t.name ?? "ai-usd", o = () => i(n, {
 		tenantId: r,
 		limitUsd: t.limitUsd,
 		addUsd: 0,
 		...t.since === void 0 ? {} : { since: t.since }
-	}).pipe(p.as(null), p.catchTag("AiBudgetExceeded", (e) => p.succeed(e))), s = () => t.estimateUsd === void 0 ? p.void : a(n, {
+	}).pipe(m.as(null), m.catchTag("AiBudgetExceeded", (e) => m.succeed(e))), s = () => t.estimateUsd === void 0 ? m.void : i(n, {
 		tenantId: r,
 		limitUsd: t.limitUsd,
 		addUsd: t.estimateUsd,
 		...t.since === void 0 ? {} : { since: t.since }
 	}), c = yield* o();
 	if (c === null) return yield* s();
-	if ((t.onExceeded ?? "fail") !== "suspend") return yield* p.fail(c);
-	let l = yield* g;
-	return l === void 0 ? yield* p.fail(c) : (yield* y({
+	if ((t.onExceeded ?? "fail") !== "suspend") return yield* m.fail(c);
+	let l = yield* _;
+	return l === void 0 ? yield* m.fail(c) : (yield* b({
 		store: n,
-		budget: i,
+		budget: a,
 		tenantId: r,
 		executionId: l,
-		workflowName: yield* I(n, l),
+		workflowName: yield* L(n, l),
 		stepName: e.name,
 		spentMicroUsd: Math.round(c.spentUsd * 1e6),
 		limitMicroUsd: Math.round(t.limitUsd * 1e6),
-		hasHeadroom: o().pipe(p.map((e) => e === null)),
+		hasHeadroom: o().pipe(m.map((e) => e === null)),
 		...t.holdTimeoutMs === void 0 ? {} : { timeoutMs: t.holdTimeoutMs },
 		...t.recheckEveryMs === void 0 ? {} : { recheckEveryMs: t.recheckEveryMs }
-	}).pipe(p.catchAll((e) => p.die(e))), yield* s());
-}), I = (e, t) => p.promise(async () => {
+	}).pipe(m.catchAll((e) => m.die(e))), yield* s());
+}), L = (e, t) => m.promise(async () => {
 	let n = await e.query({
 		table: "_voltro_workflow_runs",
-		predicate: h("executionId", t),
+		predicate: g("executionId", t),
 		order: [],
 		take: 1,
 		skip: void 0,
 		projection: void 0
 	});
 	return String(n[0]?.tag ?? "");
-}), L = (t) => {
-	let n = t.provider ?? e(), i = w(t), a = {
-		...T(n),
+}), R = (t) => {
+	let r = t.provider ?? e(), i = T(t), a = {
+		...E(r),
 		operation: "generateText",
 		offloaded: t.offload === !0,
-		...C(i.text, i.system, t.recordPrompt ?? "digest", i.rendered)
-	}, o = (e) => t.budget === void 0 ? e : p.zipRight(F(t), e);
-	return t.offload === !0 ? o(N(t, n, "text", null, a, i)) : o(v({
+		...w(i.text, i.system, t.recordPrompt ?? "digest", i.rendered)
+	}, o = (e) => t.budget === void 0 ? e : m.zipRight(I(t), e);
+	return t.offload === !0 ? o(P(t, r, "text", null, a, i)) : o(y({
 		name: t.name,
-		success: O,
+		success: k,
 		input: a,
-		retry: t.retry ?? b,
-		execute: r({
+		retry: t.retry ?? x,
+		execute: n({
 			prompt: i.text,
-			provider: n,
+			provider: r,
 			...i.system === void 0 ? {} : { system: i.system },
 			...t.fallbacks === void 0 ? {} : { fallbacks: t.fallbacks },
 			...t.maxTokens === void 0 ? {} : { maxTokens: t.maxTokens },
 			...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions }
-		}).pipe(p.tap((e) => E(t, n, "generateText", e, i.rendered)), p.tap(() => D(t, i.rendered)), p.map((e) => ({
+		}).pipe(m.tap((e) => D(t, r, "generateText", e, i.rendered)), m.tap(() => O(t, i.rendered)), m.map((e) => ({
 			text: e.text,
 			inputTokens: e.usage.inputTokens ?? null,
 			outputTokens: e.usage.outputTokens ?? null,
 			totalTokens: e.usage.totalTokens ?? null
 		})))
 	}));
-}, R = (r) => {
-	let i = r.provider ?? e(), a = w(r), o = {
-		...T(i),
+}, z = h.Struct({
+	text: h.String,
+	language: h.optional(h.String),
+	durationSec: h.optional(h.Number)
+}), B = (t) => {
+	let n = t.provider ?? e(), r = {
+		source: t.source,
+		...t.language === void 0 ? {} : { language: t.language },
+		...t.timestamps === void 0 ? {} : { timestamps: t.timestamps },
+		...t.diarize === void 0 ? {} : { diarize: t.diarize },
+		...t.translateToEnglish === void 0 ? {} : { translateToEnglish: t.translateToEnglish }
+	}, i = {
+		...E(n),
+		operation: "transcribe",
+		offloaded: t.offload === !0
+	}, a = (e) => t.budget === void 0 ? e : m.zipRight(I(t), e);
+	return t.offload === !0 ? a(P(t, n, "transcription", null, i, {
+		text: "",
+		system: void 0,
+		rendered: void 0
+	}, r)) : a(y({
+		name: t.name,
+		success: z,
+		input: i,
+		retry: t.retry ?? x,
+		execute: o({
+			...r,
+			provider: n,
+			...t.resolveRef === void 0 ? {} : { resolveRef: t.resolveRef }
+		}).pipe(m.map((e) => ({
+			text: e.text,
+			...e.language === void 0 ? {} : { language: e.language },
+			...e.durationSec === void 0 ? {} : { durationSec: e.durationSec }
+		})))
+	}));
+}, V = (t) => {
+	let n = t.provider ?? e(), r = T(t), i = {
+		...E(n),
 		operation: "generateObject",
-		offloaded: r.offload === !0,
-		...C(a.text, a.system, r.recordPrompt ?? "digest", a.rendered)
-	}, s = (e) => r.budget === void 0 ? e : p.zipRight(F(r), e);
-	return r.offload === !0 ? s(p.gen(function* () {
-		let e = yield* n(r.schema, "aiObjectStep"), t = yield* N(r, i, "object", e, o, a);
-		return yield* m.decodeUnknown(r.schema)(t);
-	})) : s(v({
-		name: r.name,
-		success: r.schema,
-		input: o,
-		retry: r.retry ?? b,
-		execute: t({
-			prompt: a.text,
-			schema: r.schema,
-			provider: i,
-			...a.system === void 0 ? {} : { system: a.system },
-			...r.fallbacks === void 0 ? {} : { fallbacks: r.fallbacks },
-			...r.maxTokens === void 0 ? {} : { maxTokens: r.maxTokens },
-			...r.providerOptions === void 0 ? {} : { providerOptions: r.providerOptions }
-		}).pipe(p.tap((e) => E(r, i, "generateObject", e, a.rendered)), p.tap(() => D(r, a.rendered)), p.map((e) => e.object))
+		offloaded: t.offload === !0,
+		...w(r.text, r.system, t.recordPrompt ?? "digest", r.rendered)
+	}, o = (e) => t.budget === void 0 ? e : m.zipRight(I(t), e);
+	return t.offload === !0 ? o(m.gen(function* () {
+		let e = yield* a(t.schema, "aiObjectStep"), o = yield* P(t, n, "object", e, i, r);
+		return yield* h.decodeUnknown(t.schema)(o);
+	})) : o(y({
+		name: t.name,
+		success: t.schema,
+		input: i,
+		retry: t.retry ?? x,
+		execute: f({
+			prompt: r.text,
+			schema: t.schema,
+			provider: n,
+			...r.system === void 0 ? {} : { system: r.system },
+			...t.fallbacks === void 0 ? {} : { fallbacks: t.fallbacks },
+			...t.maxTokens === void 0 ? {} : { maxTokens: t.maxTokens },
+			...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions }
+		}).pipe(m.tap((e) => D(t, n, "generateObject", e, r.rendered)), m.tap(() => O(t, r.rendered)), m.map((e) => e.object))
 	}));
 };
 //#endregion
-export { b as AI_STEP_RETRY, P as AiBudgetMisuseError, A as AiOffloadError, j as AiOffloadMisuseError, S as DEFAULT_AI_BUDGET_NAME, x as DEFAULT_OFFLOAD_TIMEOUT_MS, R as aiObjectStep, L as aiStep, o as promptDigest, C as promptRecordFor, w as resolveStepPrompt };
+export { x as AI_STEP_RETRY, F as AiBudgetMisuseError, j as AiOffloadError, M as AiOffloadMisuseError, C as DEFAULT_AI_BUDGET_NAME, S as DEFAULT_OFFLOAD_TIMEOUT_MS, z as TranscriptionStepResult, V as aiObjectStep, R as aiStep, s as promptDigest, w as promptRecordFor, T as resolveStepPrompt, B as transcribeStep };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voltro/ai",
-  "version": "0.57.0",
+  "version": "0.59.0",
   "description": "Effect-native AI primitives — generateText / generateObject / streamText, tools, agents, media generation, cost tracking + budgets, and provider fallback.",
   "keywords": [
     "voltro",
@@ -57,10 +57,10 @@
     "@ai-sdk/gateway": "^4.0.47",
     "@ai-sdk/openai": "^4.0.37",
     "@ai-sdk/provider": "^4.0.7",
-    "@voltro/cache": "0.57.0",
-    "@voltro/database": "0.57.0",
-    "@voltro/protocol": "0.57.0",
-    "@voltro/workflow": "0.57.0",
+    "@voltro/cache": "0.59.0",
+    "@voltro/database": "0.59.0",
+    "@voltro/protocol": "0.59.0",
+    "@voltro/workflow": "0.59.0",
     "ai": "^7.0.59"
   },
   "peerDependencies": {