@voltro/ai 0.63.0 → 0.64.0

package/dist/{inferenceQueue-CS1VouiW.js → inferenceQueue-BS3PVSkk.js}

@@ -1,6 +1,6 @@
-import { a as e, r as t, s as n } from "./provider-BOgeCRkg.js";
+import { c as e, o as t, r as n } from "./provider-RA0getTL.js";
 import { Output as r, experimental_transcribe as i, generateObject as a, generateText as o, jsonSchema as s, stepCountIs as c, tool as l } from "ai";
-import { Context as u, Effect as d, JSONSchema as ee, Metric as f, MetricBoundaries as te, Runtime as ne, Schema as p, Stream as m } from "effect";
+import { Context as u, Effect as d, JSONSchema as ee, Metric as f, MetricBoundaries as te, Runtime as p, Schema as m, Stream as ne } from "effect";
 import { and as h, eq as g, generateId as _, gte as re, id as v, integer as y, json as b, queryFor as x, registerRetention as S, retentionTtlMsFromEnv as C, table as w, text as T, timestamp as E } from "@voltro/database";
 import { createHash as ie } from "node:crypto";
 //#region src/tools.ts
@@ -14,8 +14,8 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 	for (let r of Object.values(e)) ce(r) && r.execute === void 0 && r.handler === void 0 && (r.handler = t, n += 1);
 	return n;
 }, ue = (e, t) => {
-	let n = t ?? ne.defaultRuntime;
-	return ne.runPromise(n)(e);
+	let n = t ?? p.defaultRuntime;
+	return p.runPromise(n)(e);
 }, de = (e, t, n) => n ? e.pipe(d.timeoutFail({
 	duration: `${n} millis`,
 	onTimeout: () => /* @__PURE__ */ Error(`tool '${t}' timed out`)
@@ -40,18 +40,18 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 		properties: {},
 		additionalProperties: !0
 	};
-}, he = (e, t = { store: void 0 }) => {
+}, D = (e, t = { store: void 0 }) => {
 	let n = {};
 	for (let [r, i] of Object.entries(e)) n[i.name || r] = l({
 		description: i.description,
 		inputSchema: s(me(i)),
 		execute: async (e) => {
-			let n = await d.runPromise(p.decodeUnknown(i.input)(e)), r = await fe(i, n, t);
-			return i.output ? d.runPromise(p.decodeUnknown(i.output)(r)) : r;
+			let n = await d.runPromise(m.decodeUnknown(i.input)(e)), r = await fe(i, n, t);
+			return i.output ? d.runPromise(m.decodeUnknown(i.output)(r)) : r;
 		}
 	});
 	return n;
-}, D = {
+}, he = {
 	"claude-opus-4-8": {
 		inputPer1M: 15,
 		outputPer1M: 75
@@ -100,7 +100,7 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 	let t = O;
 	return O = { ...e }, t;
 }, k = () => ({
-	...D,
+	...he,
 	...O
 }), A = (e) => {
 	let t = k();
@@ -207,11 +207,11 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 		...t.promptDigest === void 0 ? [] : [g("promptDigest", t.promptDigest)]
 	], r = x(F), i = n.length === 0 ? r : r.where(n.length === 1 ? n[0] : h(...n));
 	return (await e.query(i.descriptor)).reduce((e, t) => e + (Number(t.costMicroUsd) || 0), 0) / 1e6;
-}), I = class extends p.TaggedError()("AiBudgetExceeded", {
-	tenantId: p.NullOr(p.String),
-	limitUsd: p.Number,
-	spentUsd: p.Number,
-	attemptedUsd: p.Number
+}), I = class extends m.TaggedError()("AiBudgetExceeded", {
+	tenantId: m.NullOr(m.String),
+	limitUsd: m.Number,
+	spentUsd: m.Number,
+	attemptedUsd: m.Number
 }) {}, L = "_voltro_ai_budget", R = w(L, {
 	id: v({ prefix: "aib" }),
 	bucket: T().maxLength(191),
@@ -297,41 +297,52 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 	} }));
 }), Oe = (e, t) => {
 	let n = Date.now();
-	return t.pipe(m.tap((t) => t._tag === "done" ? U(f.increment(B), e, "ok").pipe(d.zipRight(U(f.update(V, (Date.now() - n) / 1e3), e, "ok")), d.zipRight(De(e, t.usage ?? {}))) : t._tag === "error" ? U(f.increment(B), e, "error").pipe(d.zipRight(U(f.increment(Ce), e, "error")), d.zipRight(U(f.update(V, (Date.now() - n) / 1e3), e, "error"))) : d.void), m.withSpan("voltro.ai.call", { attributes: {
+	return t.pipe(ne.tap((t) => t._tag === "done" ? U(f.increment(B), e, "ok").pipe(d.zipRight(U(f.update(V, (Date.now() - n) / 1e3), e, "ok")), d.zipRight(De(e, t.usage ?? {}))) : t._tag === "error" ? U(f.increment(B), e, "error").pipe(d.zipRight(U(f.increment(Ce), e, "error")), d.zipRight(U(f.update(V, (Date.now() - n) / 1e3), e, "error"))) : d.void), ne.withSpan("voltro.ai.call", { attributes: {
 		"ai.provider": e.provider,
 		"ai.model": e.model,
 		"ai.operation": e.operation
 	} }));
-}, G = class extends p.TaggedError()("AiError", {
-	reason: p.Literal("generation", "decode"),
-	message: p.String
+}, G = class extends m.TaggedError()("AiError", {
+	reason: m.Literal("generation", "decode"),
+	message: m.String
 }) {}, K = (e, t, n) => {
 	let r = [e, ...t ?? []], i = (e) => n(r[e]).pipe(d.catchTag("AiError", (t) => t.reason === "generation" && e + 1 < r.length ? i(e + 1) : d.fail(t)));
 	return i(0);
-}, ke = (n) => K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateText"), (e) => e.usage, d.tryPromise({
-	try: async () => {
-		let r = e(t), i = await o({
-			model: r,
-			prompt: n.prompt,
-			...n.system === void 0 ? {} : { system: n.system },
-			...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
-			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
-		});
-		return {
-			text: i.text,
-			usage: {
-				inputTokens: i.usage.inputTokens,
-				outputTokens: i.usage.outputTokens,
-				totalTokens: i.usage.totalTokens
-			},
-			providerMetadata: i.providerMetadata
-		};
-	},
-	catch: (e) => new G({
-		reason: "generation",
-		message: String(e?.message ?? e)
-	})
-}))), q = (e, t) => d.gen(function* () {
+}, ke = (e) => d.gen(function* () {
+	let r = e.tools === void 0 ? void 0 : yield* d.runtime(), i = e.tools === void 0 || r === void 0 ? void 0 : D(e.tools, {
+		store: void 0,
+		...e.toolContext,
+		runtime: r
+	});
+	return yield* K(e.provider ?? n(), e.fallbacks, (n) => W(H(n, "generateText"), (e) => e.usage, d.tryPromise({
+		try: async () => {
+			let r = t(n), a = await o({
+				model: r,
+				prompt: e.prompt,
+				...i === void 0 ? {} : {
+					tools: i,
+					stopWhen: c(e.maxSteps ?? 8)
+				},
+				...e.system === void 0 ? {} : { system: e.system },
+				...e.maxTokens === void 0 ? {} : { maxOutputTokens: e.maxTokens },
+				...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+			}), s = i !== void 0 && a.totalUsage !== void 0 ? a.totalUsage : a.usage;
+			return {
+				text: a.text,
+				usage: {
+					inputTokens: s.inputTokens,
+					outputTokens: s.outputTokens,
+					totalTokens: s.totalTokens
+				},
+				providerMetadata: a.providerMetadata
+			};
+		},
+		catch: (e) => new G({
+			reason: "generation",
+			message: String(e?.message ?? e)
+		})
+	})));
+}), q = (e, t) => d.gen(function* () {
 	let n = ee.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)";
@@ -341,15 +352,15 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 		}));
 	}
 	return n;
-}), Ae = (e, t) => q(e, t), je = (n) => K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateObject"), (e) => e.usage, d.gen(function* () {
-	let r = e(t), i = yield* d.tryPromise({
+}), Ae = (e, t) => q(e, t), je = (e) => K(e.provider ?? n(), e.fallbacks, (n) => W(H(n, "generateObject"), (e) => e.usage, d.gen(function* () {
+	let r = t(n), i = yield* d.tryPromise({
 		try: () => a({
 			model: r,
-			schema: s(n.jsonSchema),
-			prompt: n.prompt,
-			...n.system === void 0 ? {} : { system: n.system },
-			...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
-			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
+			schema: s(e.jsonSchema),
+			prompt: e.prompt,
+			...e.system === void 0 ? {} : { system: e.system },
+			...e.maxTokens === void 0 ? {} : { maxOutputTokens: e.maxTokens },
+			...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
 		}),
 		catch: (e) => new G({
 			reason: "generation",
@@ -374,7 +385,7 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 		...e.fallbacks === void 0 ? {} : { fallbacks: e.fallbacks },
 		...e.maxTokens === void 0 ? {} : { maxTokens: e.maxTokens },
 		...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
-	}), n = yield* p.decodeUnknown(e.schema)(t.object).pipe(d.mapError((e) => new G({
+	}), n = yield* m.decodeUnknown(e.schema)(t.object).pipe(d.mapError((e) => new G({
 		reason: "decode",
 		message: `model output failed schema decode: ${String(e)}`
 	})));
@@ -382,23 +393,23 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 		...t,
 		object: n
 	};
-}), Ne = (n) => d.gen(function* () {
-	let i = yield* d.runtime(), a = yield* q(n.schema, "generateObjectWithTools"), l = {
+}), Ne = (e) => d.gen(function* () {
+	let i = yield* d.runtime(), a = yield* q(e.schema, "generateObjectWithTools"), l = {
 		store: void 0,
-		...n.toolContext,
+		...e.toolContext,
 		runtime: i
-	}, u = he(n.tools, l);
-	return yield* K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateObjectWithTools"), (e) => e.usage, d.gen(function* () {
-		let i = e(t), l = yield* d.tryPromise({
+	}, u = D(e.tools, l);
+	return yield* K(e.provider ?? n(), e.fallbacks, (n) => W(H(n, "generateObjectWithTools"), (e) => e.usage, d.gen(function* () {
+		let i = t(n), l = yield* d.tryPromise({
 			try: () => o({
 				model: i,
-				prompt: n.prompt,
+				prompt: e.prompt,
 				tools: u,
-				stopWhen: c(n.maxSteps ?? 8),
+				stopWhen: c(e.maxSteps ?? 8),
 				output: r.object({ schema: s(a) }),
-				...n.system === void 0 ? {} : { system: n.system },
-				...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
-				...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
+				...e.system === void 0 ? {} : { system: e.system },
+				...e.maxTokens === void 0 ? {} : { maxOutputTokens: e.maxTokens },
+				...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
 			}),
 			catch: (e) => new G({
 				reason: "generation",
@@ -406,7 +417,7 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 			})
 		});
 		return {
-			object: yield* p.decodeUnknown(n.schema)(l.output).pipe(d.mapError((e) => new G({
+			object: yield* m.decodeUnknown(e.schema)(l.output).pipe(d.mapError((e) => new G({
 				reason: "decode",
 				message: `model output failed schema decode: ${String(e)}`
 			}))),
@@ -456,16 +467,16 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 		});
 	}
 	return t.length > 0 ? t : void 0;
-}, Re = (e) => {
-	let r = e.provider ?? t();
+}, Re = (t) => {
+	let r = t.provider ?? n();
 	return W(H(r, "transcribe"), () => ({}), d.tryPromise({
 		try: async () => {
-			let { audio: t, mediaType: a } = await Ie(e.source, e.resolveRef), o = await i({
-				model: n(r),
-				audio: t,
+			let { audio: n, mediaType: a } = await Ie(t.source, t.resolveRef), o = await i({
+				model: e(r),
+				audio: n,
 				...a === void 0 ? {} : { mediaType: a },
-				...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
-			}), s = e.timestamps === "none" ? void 0 : Le(o.segments);
+				...t.providerOptions === void 0 ? {} : { providerOptions: t.providerOptions }
+			}), s = t.timestamps === "none" ? void 0 : Le(o.segments);
 			return {
 				text: o.text,
 				...o.language === void 0 ? {} : { language: o.language },
@@ -479,10 +490,10 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 			message: String(e?.message ?? e)
 		})
 	}));
-}, ze = (e) => `sha256:${ie("sha256").update(e).digest("hex").slice(0, 16)}`, Be = class extends p.TaggedError()("PromptRenderError", {
-	promptId: p.String,
-	missing: p.Array(p.String),
-	message: p.String
+}, ze = (e) => `sha256:${ie("sha256").update(e).digest("hex").slice(0, 16)}`, Be = class extends m.TaggedError()("PromptRenderError", {
+	promptId: m.String,
+	missing: m.Array(m.String),
+	message: m.String
 }) {}, Ve = /\{\{\s*([A-Za-z0-9_.]+)\s*\}\}/g, He = (e) => {
 	let t = /* @__PURE__ */ new Set();
 	for (let n of e.matchAll(Ve)) {
@@ -650,4 +661,4 @@ var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 	};
 }, nt = (e) => `voltro.ai.infer:${e}`, rt = (e, t) => `aiinf_${ie("sha256").update(`${e}\u0000${t}`).digest("hex").slice(0, 40)}`;
 //#endregion
-export { Oe as A, _e as B, Fe as C, ke as D, Ne as E, D as F, ve as G, k as H, R as I, he as J, xe as K, ye as L, L as M, P as N, Ae as O, I as P, F as R, Pe as S, je as T, M as U, j as V, A as W, ce as X, se as Y, le as Z, Xe as _, tt as a, Re as b, et as c, Be as d, We as f, Ue as g, Ye as h, Ze as i, W as j, H as k, Ke as l, ze as m, $e as n, nt as o, Ge as p, ge as q, Qe as r, rt as s, Z as t, J as u, Y as v, Me as w, G as x, Je as y, N as z };
+export { Oe as A, _e as B, Fe as C, ke as D, Ne as E, he as F, ve as G, k as H, R as I, D as J, xe as K, ye as L, L as M, P as N, Ae as O, I as P, F as R, Pe as S, je as T, M as U, j as V, A as W, ce as X, se as Y, le as Z, Xe as _, tt as a, Re as b, et as c, Be as d, We as f, Ue as g, Ye as h, Ze as i, W as j, H as k, Ke as l, ze as m, $e as n, nt as o, Ge as p, ge as q, Qe as r, rt as s, Z as t, J as u, Y as v, Me as w, G as x, Je as y, N as z };

package/dist/{provider-BOgeCRkg.js → provider-RA0getTL.js}

@@ -288,7 +288,14 @@ var u = [], d = (...e) => {
 }, B = null, V = (e) => {
 	let t = B;
 	return B = e, t;
-}, H = () => B, U = (e = process.env) => {
+}, H = () => B, U = (e) => {
+	switch (e.toLowerCase()) {
+		case "anthropic": return "ANTHROPIC_API_KEY";
+		case "openai": return "OPENAI_API_KEY";
+		case "gateway": return "AI_GATEWAY_API_KEY";
+		default: return;
+	}
+}, W = (e = process.env) => {
 	if (B) return B;
 	let t = (e.AI_PROVIDER ?? "mock").toLowerCase(), n = e.AI_MODEL;
 	switch (t) {
@@ -316,4 +323,4 @@ var u = [], d = (...e) => {
 	}
 };
 //#endregion
-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 };
+export { d as _, I as a, R as c, m as d, p as f, h as g, f as h, U as i, z as l, s as m, H as n, F as o, o as p, W as r, L as s, v as t, V as u, c as v };

package/dist/test.js

@@ -1,32 +1,32 @@
-import { l as e, t } from "./provider-BOgeCRkg.js";
+import { t as e, u as t } from "./provider-RA0getTL.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));
+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 = (t) => {
+	let r = new e(n(t));
 	return {
 		config: {
 			name: "mock",
-			model: e.model ?? "mock",
+			model: t.model ?? "mock",
 			script: r,
-			...e.generate ? { mockText: e.generate.text } : {}
+			...t.generate ? { mockText: t.generate.text } : {}
 		},
 		script: r
 	};
-}, i = (t) => {
-	let { config: n, script: i } = r(t), a = e(n);
+}, i = (e) => {
+	let { config: n, script: i } = r(e), a = t(n);
 	return {
 		calls: i.calls,
 		reset: () => {
-			e(a);
+			t(a);
 		}
 	};
-}, a = (t = {}) => {
-	let { config: n, script: i } = r(t), a = e(n);
+}, a = (e = {}) => {
+	let { config: n, script: i } = r(e), a = t(n);
 	return {
-		...t.generate ? { generate: () => ({ text: t.generate.text }) } : {},
-		...t.generateObject ? { generateObject: () => t.generateObject.object } : {},
+		...e.generate ? { generate: () => ({ text: e.generate.text }) } : {},
+		...e.generateObject ? { generateObject: () => e.generateObject.object } : {},
 		calls: i.calls,
 		reset: () => {
-			e(a);
+			t(a);
 		}
 	};
 }, o = (e, t, n = (t) => /* @__PURE__ */ Error(`throwNTimes: attempt ${t} of ${e}`)) => {

package/dist/workflow.d.ts

@@ -1,5 +1,6 @@
 import { DataStore } from '@voltro/database';
 import { Effect } from 'effect';
+import { Runtime } from 'effect';
 import { Schema } from 'effect';
 import { StepRetryPolicy } from '@voltro/workflow';
 
@@ -88,6 +89,18 @@ export declare interface AiStepBaseOptions {
     readonly fallbacks?: ProviderFallbacks;
     readonly maxTokens?: number;
     readonly providerOptions?: AiProviderOptions;
+    /**
+     * Tools the model may call inside this durable step — the whole LLM↔tool
+     * loop runs as ONE step: one journal entry, one retry policy, one usage row
+     * summing every round-trip. Before this, a durable tool loop had to leave
+     * `aiStep` for the inline `generateObjectWithTools`, and lost the journal and
+     * the per-step spend recording on the way out.
+     */
+    readonly tools?: Record<string, AnyTool>;
+    /** Max LLM↔tool round-trips before the model must answer. Default 8. */
+    readonly maxSteps?: number;
+    /** Request context threaded into the tool executors. */
+    readonly toolContext?: ToolContext;
     readonly retry?: StepRetryPolicy;
     readonly recordPrompt?: PromptRecording;
     /**
@@ -183,6 +196,24 @@ export declare interface AiStepResult {
 
 declare type AnthropicModel = 'claude-opus-4-8' | 'claude-sonnet-4-6' | 'claude-haiku-4-5-20251001' | (string & {});
 
+/** A heterogeneous tool — the shape `buildSdkTools` / `runAssistant` consume.
+ *  `defineTool` keeps full input typing on its `execute`/`handler` body, then
+ *  erases to this shape so a `Record<string, AnyTool>` accepts tools of
+ *  differing input/output types (Effect's `Schema` encoded type is invariant,
+ *  so a heterogeneous collection can't preserve per-tool generics). */
+declare interface AnyTool {
+    readonly name: string;
+    readonly description: string;
+    readonly input: Schema.Schema.Any;
+    /** See `ToolDef.inputJsonSchema` — a foreign parameter schema shown to the
+     *  model in place of a rendered `input`. */
+    readonly inputJsonSchema?: Record<string, unknown>;
+    readonly output?: Schema.Schema.Any;
+    readonly execute?: (input: never) => Effect.Effect<unknown, never, unknown>;
+    readonly handler?: ToolHandler<never, unknown>;
+    readonly timeoutMs?: number;
+}
+
 /** The name a hold is scoped to when the caller does not choose one. Operators
  *  lift by this string, so it has to be stable and sayable. */
 export declare const DEFAULT_AI_BUDGET_NAME = "ai-usd";
@@ -443,6 +474,31 @@ export declare const resolveStepPrompt: (options: Pick<AiStepBaseOptions, "promp
     readonly rendered: RenderedPrompt | undefined;
 };
 
+/** The slice of the request context a `*.tool.tsx` handler receives. Kept
+ *  structurally typed here (rather than importing `AppContext` from
+ *  `@voltro/runtime`) so `@voltro/ai` carries no runtime dependency — the
+ *  CLI discovery path passes the real `AppContext`, which is a superset.
+ *
+ *  `runtime` is the caller's ambient Effect `Runtime`. When present, tool
+ *  bodies are run through it (`Runtime.runPromise(runtime)(eff)`) instead of
+ *  bare `Effect.runPromise`, so a body may `yield*` any service the caller's
+ *  runtime carries — `JiraService`, the data store layer, `HttpClient`, … —
+ *  rather than being limited to `R = never`. The agent loop
+ *  (`streamText`/`runAssistant`) captures `yield* Effect.runtime()` and threads
+ *  it here; without a runtime, behaviour is unchanged. */
+declare interface ToolContext {
+    readonly store: unknown;
+    readonly request?: {
+        readonly subject?: unknown;
+    };
+    readonly runtime?: Runtime.Runtime<unknown>;
+}
+
+/** A context-receiving tool body — the `*.tool.tsx` default-export form.
+ *  May return an Effect (with service deps the caller's runtime provides) or
+ *  a Promise; `buildSdkTools` normalises both. */
+declare type ToolHandler<A, O> = (input: A, ctx: ToolContext) => Effect.Effect<O, never, unknown> | Promise<O>;
+
 /**
  * Transcription as a durable step — with `offload`, which is the point.
  *