@voltro/ai 0.74.0 → 0.75.0

package/dist/workflow.d.ts

@@ -1,5 +1,6 @@
 import { DataStore } from '@voltro/database';
 import { Effect } from 'effect';
+import { PluginMediaPersistMetadata } from '@voltro/protocol';
 import { Runtime } from 'effect';
 import { Schema } from 'effect';
 import { StepRetryPolicy } from '@voltro/workflow';
@@ -244,16 +245,23 @@ declare type GatewayModel = `${GatewayCreator}/${string}`;
  *
  * The queue row is the job's identity, allocated before the first paid call.
  * A replay finds the row; a dispatcher that dies mid-job leaves a lease that
- * expires — and the reclaiming dispatcher does NOT submit again: the provider
- * has no lookup, so the outcome is reported to the run as `unknown`
- * (`AiOffloadError.outcome`) rather than paid for twice. Artifacts are
- * persisted through the storage plugin's `mediaPersist` capability under
- * `<row>#<attempt>#<index>`, and the run receives URLs and ref ids.
+ * expires — and the reclaiming dispatcher does NOT submit again. With a
+ * provider registered through `registerMediaTaskProvider` the row also
+ * carries the provider's task handle, so the reclaim polls the same task and
+ * the run receives its result. A built-in generator has no lookup — and a
+ * task job whose handle never reached the row cannot be found either — so
+ * those are reported to the run as `unknown` (`AiOffloadError.outcome`)
+ * rather than paid for twice. Artifacts are persisted through the storage
+ * plugin's `mediaPersist` capability under `<row>#<attempt>#<index>`, and
+ * the run receives URLs and ref ids.
  */
 export declare const mediaStep: (options: MediaStepOptions) => Effect.Effect<MediaStepResult, unknown>;
 
-export declare interface MediaStepOptions extends Omit<AiStepBaseOptions, 'system' | 'promptTemplate' | 'promptVars' | 'offload' | 'tools' | 'maxSteps' | 'toolContext' | 'fallbacks' | 'maxTokens'> {
+export declare interface MediaStepOptions extends Omit<AiStepBaseOptions, 'system' | 'promptTemplate' | 'promptVars' | 'offload' | 'tools' | 'maxSteps' | 'toolContext' | 'fallbacks' | 'maxTokens' | 'provider'> {
     readonly modality: 'image' | 'video' | 'speech';
+    /** A built-in provider, or the name of one registered with
+     *  `registerMediaTaskProvider` — the dispatcher routes by the name. */
+    readonly provider?: ProviderConfig | RegisteredMediaTaskConfig;
     /** Provider parameters — `n`, `size`, `aspectRatio`, `seed` for an image;
      *  `durationSeconds`, `resolution`, `fps` for a video; `voice`, `format` for
      *  speech. Passed through to the generator, journaled with the job. */
@@ -263,6 +271,9 @@ export declare interface MediaStepOptions extends Omit<AiStepBaseOptions, 'syste
     readonly inputs?: Record<string, unknown>;
     /** Required: a media job is always an external job with a durable identity. */
     readonly store: DataStore;
+    /** Who the stored artifacts belong to, how they are served, and the tags they
+     *  are found by — handed to the storage plugin's `mediaPersist` as-is. */
+    readonly persist?: PluginMediaPersistMetadata;
 }
 
 export declare interface MediaStepResult {
@@ -490,6 +501,15 @@ declare type ProviderConfig = {
  */
 declare type ProviderFallbacks = ReadonlyArray<ProviderConfig>;
 
+/**
+ * The provider config `mediaStep` accepts for a registered task provider —
+ * named, not enumerated, exactly like `RegisteredSpeechConfig`.
+ */
+declare interface RegisteredMediaTaskConfig {
+    readonly name: string;
+    readonly model?: string;
+}
+
 /** What `.render(vars)` produces — a stamp plus the text to send. */
 declare interface RenderedPrompt extends PromptStamp {
     readonly prompt: string;

package/dist/workflow.js

@@ -1,5 +1,5 @@
 import { r as e } from "./provider-Cts1BRhY.js";
-import { D as t, E as n, O as r, R as i, U as a, Z as o, b as s, m as c, o as l, p as u, s as d, t as f, w as p, y as m } from "./inferenceQueue-CXEVU6mI.js";
+import { D as t, O as n, T as r, V as i, b as a, k as o, m as s, o as c, p as l, q as u, s as d, t as f, tt as p, y as m } from "./inferenceQueue-CdWTzRp7.js";
 import { Effect as h, Schema as g } from "effect";
 import { eq as _ } from "@voltro/database";
 import { CurrentWorkflowExecutionId as v, awaitSignalSuspending as y, step as b, suspendForBudget as x } from "@voltro/workflow";
@@ -24,14 +24,14 @@ var S = {
 		...t === void 0 ? {} : { system: t }
 	} : {
 		...i,
-		promptDigest: c(e),
+		promptDigest: s(e),
 		promptChars: e.length,
 		...t === void 0 ? {} : {
-			systemDigest: c(t),
+			systemDigest: s(t),
 			systemChars: t.length
 		}
 	};
-}, E = (e) => u(e.prompt) ? {
+}, E = (e) => l(e.prompt) ? {
 	text: e.prompt.prompt,
 	system: e.system ?? e.prompt.system,
 	rendered: e.prompt
@@ -40,22 +40,22 @@ var S = {
 	system: e.system,
 	rendered: void 0
 }, D = (e) => ({
-	provider: String(e.provider ?? "unknown"),
+	provider: String(e.name),
 	model: String(e.model ?? "unknown")
-}), O = (e, t, n, r, o) => {
-	let s = e.store;
-	if (s === void 0) return h.void;
-	let { provider: c, model: l } = D(t), u = i(r.providerMetadata);
-	return a(s, {
-		provider: c,
-		model: l,
+}), O = (e, t, n, r, a) => {
+	let o = e.store;
+	if (o === void 0) return h.void;
+	let { provider: s, model: c } = D(t), l = i(r.providerMetadata);
+	return u(o, {
+		provider: s,
+		model: c,
 		operation: n,
 		agent: e.name,
 		usage: r.usage,
 		...e.tenantId === void 0 ? {} : { tenantId: e.tenantId },
-		...u === void 0 ? {} : { actualCostUsd: u },
-		...o === void 0 ? {} : { prompt: o }
-	}).pipe(h.asVoid, h.catchAllCause(() => h.void));
+		...l === void 0 ? {} : { actualCostUsd: l },
+		...a === void 0 ? {} : { prompt: a }
+	});
 }, k = (e, t) => {
 	let n = e.store;
 	return n === void 0 || t === void 0 ? h.void : m(n, {
@@ -80,11 +80,11 @@ var S = {
 }, P = (e, t, n, r, i, a, o) => h.gen(function* () {
 	let s = e.store;
 	if (s === void 0) return yield* h.die(new N(e.name));
-	let c = yield* v;
-	if (c === void 0) return yield* h.die(/* @__PURE__ */ Error(`aiStep('${e.name}'): offload: true is only valid inside a workflow body — there is no execution to suspend here.`));
-	let { provider: u, model: p } = D(t), m = d(c, e.name), g = yield* h.promise(() => s.query({
+	let l = yield* v;
+	if (l === void 0) return yield* h.die(/* @__PURE__ */ Error(`aiStep('${e.name}'): offload: true is only valid inside a workflow body — there is no execution to suspend here.`));
+	let { provider: u, model: p } = D(t), m = d(l, e.name), g = yield* h.promise(() => s.query({
 		table: "_voltro_workflow_runs",
-		predicate: _("executionId", c),
+		predicate: _("executionId", l),
 		order: [],
 		take: 1,
 		skip: void 0,
@@ -93,10 +93,10 @@ var S = {
 	return yield* h.promise(() => s.insertIgnore(f, {
 		id: m,
 		tenantId: e.tenantId ?? null,
-		executionId: c,
+		executionId: l,
 		workflowName: y,
 		runId: typeof b == "string" ? b : null,
-		signalName: l(m),
+		signalName: c(m),
 		stepName: e.name,
 		status: "pending",
 		kind: n,
@@ -117,7 +117,7 @@ var S = {
 		enqueuedAt: /* @__PURE__ */ new Date()
 	}, { conflictColumns: ["id"] })), m;
 }), F = (e, t, n, r, i, a, o, s) => h.gen(function* () {
-	let c = yield* b({
+	let l = yield* b({
 		name: `${e.name}:enqueue`,
 		success: g.String,
 		input: i,
@@ -125,7 +125,7 @@ var S = {
 	});
 	yield* k(e, a.rendered);
 	let u = yield* y({ store: e.store }, {
-		name: l(c),
+		name: c(l),
 		schema: g.Unknown,
 		timeoutMs: e.offloadTimeoutMs ?? 36e5
 	});
@@ -139,12 +139,12 @@ var S = {
 	if (t === void 0) return;
 	let n = e.store;
 	if (n === void 0) return yield* h.die(new I(e.name));
-	let r = e.tenantId ?? null, i = t.name ?? "ai-usd", a = yield* v, s = () => o(n, {
+	let r = e.tenantId ?? null, i = t.name ?? "ai-usd", a = yield* v, o = () => p(n, {
 		tenantId: r,
 		limitUsd: t.limitUsd,
 		addUsd: 0,
 		...t.since === void 0 ? {} : { since: t.since }
-	}).pipe(h.as(null), h.catchTag("AiBudgetExceeded", (e) => h.succeed(e))), c = () => t.estimateUsd === void 0 ? h.void : o(n, {
+	}).pipe(h.as(null), h.catchTag("AiBudgetExceeded", (e) => h.succeed(e))), s = () => t.estimateUsd === void 0 ? h.void : p(n, {
 		tenantId: r,
 		limitUsd: t.limitUsd,
 		addUsd: t.estimateUsd,
@@ -154,20 +154,20 @@ var S = {
 			e.name
 		])) },
 		...t.since === void 0 ? {} : { since: t.since }
-	}).pipe(h.asVoid), l = yield* s();
-	return l === null ? yield* c() : (t.onExceeded ?? "fail") !== "suspend" || a === void 0 ? yield* h.fail(l) : (yield* x({
+	}).pipe(h.asVoid), c = yield* o();
+	return c === null ? yield* s() : (t.onExceeded ?? "fail") !== "suspend" || a === void 0 ? yield* h.fail(c) : (yield* x({
 		store: n,
 		budget: i,
 		tenantId: r,
 		executionId: a,
 		workflowName: yield* R(n, a),
 		stepName: e.name,
-		spentMicroUsd: Math.round(l.spentUsd * 1e6),
+		spentMicroUsd: Math.round(c.spentUsd * 1e6),
 		limitMicroUsd: Math.round(t.limitUsd * 1e6),
-		hasHeadroom: s().pipe(h.map((e) => e === null)),
+		hasHeadroom: o().pipe(h.map((e) => e === null)),
 		...t.holdTimeoutMs === void 0 ? {} : { timeoutMs: t.holdTimeoutMs },
 		...t.recheckEveryMs === void 0 ? {} : { recheckEveryMs: t.recheckEveryMs }
-	}).pipe(h.catchAll((e) => h.die(e))), yield* c());
+	}).pipe(h.catchAll((e) => h.die(e))), yield* s());
 }), R = (e, t) => h.promise(async () => {
 	let n = await e.query({
 		table: "_voltro_workflow_runs",
@@ -178,29 +178,29 @@ var S = {
 		projection: void 0
 	});
 	return String(n[0]?.tag ?? "");
-}), z = (n) => {
-	let r = n.provider ?? e(), i = E(n), a = {
+}), z = (t) => {
+	let r = t.provider ?? e(), i = E(t), a = {
 		...D(r),
 		operation: "generateText",
-		offloaded: n.offload === !0,
-		...T(i.text, i.system, n.recordPrompt ?? "digest", i.rendered)
-	}, o = (e) => n.budget === void 0 ? e : h.zipRight(L(n), e);
-	return n.offload === !0 ? o(F(n, r, "text", null, a, i)) : o(b({
-		name: n.name,
+		offloaded: t.offload === !0,
+		...T(i.text, i.system, t.recordPrompt ?? "digest", i.rendered)
+	}, o = (e) => t.budget === void 0 ? e : h.zipRight(L(t), e);
+	return t.offload === !0 ? o(F(t, r, "text", null, a, i)) : o(b({
+		name: t.name,
 		success: A,
 		input: a,
-		retry: n.retry ?? S,
-		execute: t({
+		retry: t.retry ?? S,
+		execute: n({
 			prompt: i.text,
 			provider: r,
 			...i.system === void 0 ? {} : { system: i.system },
-			...n.fallbacks === void 0 ? {} : { fallbacks: n.fallbacks },
-			...n.maxTokens === void 0 ? {} : { maxTokens: n.maxTokens },
-			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions },
-			...n.tools === void 0 ? {} : { tools: n.tools },
-			...n.maxSteps === void 0 ? {} : { maxSteps: n.maxSteps },
-			...n.toolContext === void 0 ? {} : { toolContext: n.toolContext }
-		}).pipe(h.tap((e) => O(n, r, "generateText", e, i.rendered)), h.tap(() => k(n, i.rendered)), h.map((e) => ({
+			...t.fallbacks === void 0 ? {} : { fallbacks: t.fallbacks },
+			...t.maxTokens === void 0 ? {} : { maxTokens: t.maxTokens },
+			...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions },
+			...t.tools === void 0 ? {} : { tools: t.tools },
+			...t.maxSteps === void 0 ? {} : { maxSteps: t.maxSteps },
+			...t.toolContext === void 0 ? {} : { toolContext: t.toolContext }
+		}).pipe(h.tap((e) => O(t, r, "generateText", e, i.rendered)), h.tap(() => k(t, i.rendered)), h.map((e) => ({
 			text: e.text,
 			inputTokens: e.usage.inputTokens ?? null,
 			outputTokens: e.usage.outputTokens ?? null,
@@ -222,17 +222,17 @@ var S = {
 		...D(n),
 		operation: "transcribe",
 		offloaded: t.offload === !0
-	}, a = (e) => t.budget === void 0 ? e : h.zipRight(L(t), e);
-	return t.offload === !0 ? a(F(t, n, "transcription", null, i, {
+	}, o = (e) => t.budget === void 0 ? e : h.zipRight(L(t), e);
+	return t.offload === !0 ? o(F(t, n, "transcription", null, i, {
 		text: "",
 		system: void 0,
 		rendered: void 0
-	}, r)) : a(b({
+	}, r)) : o(b({
 		name: t.name,
 		success: B,
 		input: i,
 		retry: t.retry ?? S,
-		execute: s({
+		execute: a({
 			...r,
 			provider: n,
 			...t.resolveRef === void 0 ? {} : { resolveRef: t.resolveRef }
@@ -249,7 +249,8 @@ var S = {
 	}), i = {
 		modality: t.modality,
 		params: t.params ?? {},
-		inputs: t.inputs ?? {}
+		inputs: t.inputs ?? {},
+		...t.persist === void 0 ? {} : { persist: t.persist }
 	}, a = {
 		...D(n),
 		operation: `generate${t.modality[0].toUpperCase()}${t.modality.slice(1)}`,
@@ -261,42 +262,42 @@ var S = {
 		...t,
 		offload: !0
 	}, n, "media", null, a, r, void 0, i));
-}, U = (t) => {
-	let i = t.provider ?? e(), a = E(t), o = {
+}, U = (n) => {
+	let i = n.provider ?? e(), a = E(n), s = {
 		...D(i),
 		operation: "generateObject",
-		offloaded: t.offload === !0,
-		...T(a.text, a.system, t.recordPrompt ?? "digest", a.rendered)
-	}, s = (e) => t.budget === void 0 ? e : h.zipRight(L(t), e);
-	return t.offload === !0 ? s(h.gen(function* () {
-		let e = yield* r(t.schema, "aiObjectStep"), n = yield* F(t, i, "object", e, o, a);
-		return yield* g.decodeUnknown(t.schema)(n);
-	})) : s(b({
-		name: t.name,
-		success: t.schema,
-		input: o,
-		retry: t.retry ?? S,
-		execute: (t.tools === void 0 ? p({
+		offloaded: n.offload === !0,
+		...T(a.text, a.system, n.recordPrompt ?? "digest", a.rendered)
+	}, c = (e) => n.budget === void 0 ? e : h.zipRight(L(n), e);
+	return n.offload === !0 ? c(h.gen(function* () {
+		let e = yield* o(n.schema, "aiObjectStep"), t = yield* F(n, i, "object", e, s, a);
+		return yield* g.decodeUnknown(n.schema)(t);
+	})) : c(b({
+		name: n.name,
+		success: n.schema,
+		input: s,
+		retry: n.retry ?? S,
+		execute: (n.tools === void 0 ? r({
 			prompt: a.text,
-			schema: t.schema,
+			schema: n.schema,
 			provider: i,
 			...a.system === void 0 ? {} : { system: a.system },
-			...t.fallbacks === void 0 ? {} : { fallbacks: t.fallbacks },
-			...t.maxTokens === void 0 ? {} : { maxTokens: t.maxTokens },
-			...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions }
-		}) : n({
+			...n.fallbacks === void 0 ? {} : { fallbacks: n.fallbacks },
+			...n.maxTokens === void 0 ? {} : { maxTokens: n.maxTokens },
+			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
+		}) : t({
 			prompt: a.text,
-			schema: t.schema,
+			schema: n.schema,
 			provider: i,
-			tools: t.tools,
-			...t.maxSteps === void 0 ? {} : { maxSteps: t.maxSteps },
-			...t.toolContext === void 0 ? {} : { toolContext: t.toolContext },
+			tools: n.tools,
+			...n.maxSteps === void 0 ? {} : { maxSteps: n.maxSteps },
+			...n.toolContext === void 0 ? {} : { toolContext: n.toolContext },
 			...a.system === void 0 ? {} : { system: a.system },
-			...t.fallbacks === void 0 ? {} : { fallbacks: t.fallbacks },
-			...t.maxTokens === void 0 ? {} : { maxTokens: t.maxTokens },
-			...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions }
-		})).pipe(h.tap((e) => O(t, i, "generateObject", e, a.rendered)), h.tap(() => k(t, a.rendered)), h.map((e) => e.object))
+			...n.fallbacks === void 0 ? {} : { fallbacks: n.fallbacks },
+			...n.maxTokens === void 0 ? {} : { maxTokens: n.maxTokens },
+			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
+		})).pipe(h.tap((e) => O(n, i, "generateObject", e, a.rendered)), h.tap(() => k(n, a.rendered)), h.map((e) => e.object))
 	}));
 };
 //#endregion
-export { S as AI_STEP_RETRY, I as AiBudgetMisuseError, M as AiOffloadError, N as AiOffloadMisuseError, w as DEFAULT_AI_BUDGET_NAME, C as DEFAULT_OFFLOAD_TIMEOUT_MS, B as TranscriptionStepResult, U as aiObjectStep, z as aiStep, H as mediaStep, c as promptDigest, T as promptRecordFor, E as resolveStepPrompt, V as transcribeStep };
+export { S as AI_STEP_RETRY, I as AiBudgetMisuseError, M as AiOffloadError, N as AiOffloadMisuseError, w as DEFAULT_AI_BUDGET_NAME, C as DEFAULT_OFFLOAD_TIMEOUT_MS, B as TranscriptionStepResult, U as aiObjectStep, z as aiStep, H as mediaStep, s as promptDigest, T as promptRecordFor, E as resolveStepPrompt, V as transcribeStep };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voltro/ai",
-  "version": "0.74.0",
+  "version": "0.75.0",
   "description": "Effect-native AI primitives — generateText / generateObject / streamText, tools, agents, media generation, cost tracking + budgets, and provider fallback.",
   "keywords": [
     "voltro",
@@ -55,10 +55,10 @@
   "dependencies": {
     "@ai-sdk/anthropic": "^4.0.49",
     "@ai-sdk/openai": "^4.0.58",
-    "@voltro/cache": "0.74.0",
-    "@voltro/database": "0.74.0",
-    "@voltro/protocol": "0.74.0",
-    "@voltro/workflow": "0.74.0",
+    "@voltro/cache": "0.75.0",
+    "@voltro/database": "0.75.0",
+    "@voltro/protocol": "0.75.0",
+    "@voltro/workflow": "0.75.0",
     "ajv": "^8.20.0",
     "ajv-formats": "^3.0.1"
   },