@mastra/livekit 0.3.0 → 0.3.1-alpha.1

package/dist/workflow-generator-B67QrY8L.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"workflow-generator-B67QrY8L.cjs","names":["ReadableStream"],"sources":["../src/constants.ts","../src/metadata.ts","../src/workflow-generator.ts"],"sourcesContent":["/** Default LiveKit agent name used for explicit dispatch when none is configured. */\nexport const DEFAULT_LIVEKIT_AGENT_NAME = 'mastra-voice';\n","/**\n * Session metadata passed from the Mastra server to the LiveKit agent worker through\n * LiveKit's job dispatch metadata (a plain string, so this is JSON-serialized).\n */\nexport interface LiveKitSessionMetadata {\n  /** Mastra agent to run, by registered key or agent id. */\n  agentId?: string;\n  /** Memory thread id. Defaults to the LiveKit room name when omitted. */\n  threadId?: string;\n  /** Memory resource id (typically the end user id). */\n  resourceId?: string;\n  /** Plain-object entries restored into a RequestContext for agent execution. */\n  requestContext?: Record<string, unknown>;\n}\n\nexport function parseSessionMetadata(raw: string | undefined | null): LiveKitSessionMetadata {\n  if (!raw) return {};\n  try {\n    const parsed = JSON.parse(raw);\n    if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {\n      return parsed as LiveKitSessionMetadata;\n    }\n  } catch {\n    // Dispatch metadata is user-controlled and may not be JSON; treat as absent.\n  }\n  return {};\n}\n\nexport function serializeSessionMetadata(metadata: LiveKitSessionMetadata): string {\n  return JSON.stringify(metadata);\n}\n","import { ReadableStream } from 'node:stream/web';\nimport type { WritableStream } from 'node:stream/web';\nimport type { TracingContext } from '@mastra/core/observability';\nimport type { RequestContext } from '@mastra/core/request-context';\nimport type { Workflow } from '@mastra/core/workflows';\nimport type {\n  VoiceReplyGenerator,\n  VoiceToolCall,\n  VoiceTurnCompleteContext,\n  VoiceTurnCompleteHook,\n  VoiceTurnContext,\n} from './bridge';\n\nexport interface WorkflowReplyGeneratorOptions {\n  /** The Mastra workflow that generates each turn's reply. Runs once per turn (no suspend/resume). */\n  workflow: Workflow;\n  /**\n   * Maps a turn into the workflow's `inputData`. Required — input schemas are caller-defined.\n   * A common shape passes the full transcript so the workflow is stateless between turns, e.g.\n   * `ctx => ({ history: chatContextToMessages(ctx.chatCtx) })`.\n   */\n  workflowInput: (ctx: VoiceTurnContext) => unknown | Promise<unknown>;\n  /**\n   * Only stream text from this step (by id). Defaults to every step that writes text to its\n   * `writer`. Set when multiple steps write and only one produces the spoken reply.\n   */\n  replyStep?: string;\n  /**\n   * Fallback when the workflow streams no text via `writer`: derive the spoken reply from the\n   * final run result. Without this, a non-streaming workflow stays silent. Streaming via the\n   * step `writer` is preferred — it gives the caller low time-to-first-token.\n   */\n  resultText?: (result: unknown) => string | undefined | void;\n  /**\n   * Speak a short phrase while a tool call runs. Fires only for tool calls the reply step surfaces\n   * to its `writer` — use {@link pipeAgentReplyToWriter} (or pipe the agent's `fullStream`) so\n   * `tool-call` chunks reach the stream. See {@link MastraVoiceAgentOptions.toolFeedback}.\n   */\n  toolFeedback?: (toolCall: VoiceToolCall) => string | undefined | void;\n  /**\n   * Fired off the audio path after the reply streams, fire-and-forget. Carries the produced reply\n   * text and any tool calls the workflow surfaced. See {@link MastraVoiceAgentOptions.onTurnComplete}.\n   */\n  onTurnComplete?: VoiceTurnCompleteHook;\n}\n\n/**\n * Unwraps the text carried by a `workflow-step-output` chunk's `payload.output`. A step that\n * pipes `agent.stream().textStream` into its `writer` produces raw strings; a step built from an\n * agent (`createStep(agent)`) produces full `text-delta` chunks. Returns the text for both, or\n * `undefined` for any other shape.\n */\nexport function unwrapStepText(output: unknown): string | undefined {\n  if (typeof output === 'string') return output;\n  if (output && typeof output === 'object' && (output as { type?: unknown }).type === 'text-delta') {\n    const text = (output as { payload?: { text?: unknown } }).payload?.text;\n    return typeof text === 'string' ? text : undefined;\n  }\n  return undefined;\n}\n\n/**\n * Unwraps a tool call carried by a `workflow-step-output` chunk's `payload.output`. A step that\n * pipes the agent's `fullStream` (rather than just `textStream`) into its `writer` surfaces\n * `tool-call` chunks; this returns the {@link VoiceToolCall} for those, or `undefined` otherwise.\n */\nexport function unwrapStepToolCall(output: unknown): VoiceToolCall | undefined {\n  if (!output || typeof output !== 'object' || (output as { type?: unknown }).type !== 'tool-call') return undefined;\n  const payload = (output as { payload?: { toolCallId?: unknown; toolName?: unknown; args?: unknown } }).payload;\n  if (payload && typeof payload.toolCallId === 'string' && typeof payload.toolName === 'string') {\n    return { toolCallId: payload.toolCallId, toolName: payload.toolName, args: payload.args };\n  }\n  return undefined;\n}\n\n/** The minimal shape of a Mastra agent stream consumed by {@link pipeAgentReplyToWriter}. */\nexport interface AgentReplyStreamLike {\n  fullStream: AsyncIterable<unknown>;\n}\n\n/**\n * Streams a Mastra agent's reply into a workflow step's `writer` for the LiveKit workflow\n * entrypoint — the recommended way to drive a turn's reply from a step.\n *\n * It forwards the agent's text deltas (so text-to-speech starts before the full reply is ready)\n * AND its `tool-call` chunks (so {@link WorkflowReplyGeneratorOptions.toolFeedback} fires and\n * {@link WorkflowReplyGeneratorOptions.onTurnComplete}'s `result.toolCalls` is populated). This is\n * the difference from piping only `agent.stream().textStream`, which silently drops tool calls.\n * Other chunk types (reasoning, tool results, lifecycle) are not forwarded, keeping the spoken\n * stream clean. Returns the accumulated reply text for the step to return.\n *\n * Pass the step's `abortSignal` to `agent.stream(...)` so barge-in stops generation promptly.\n *\n * ```ts\n * const generateResponse = createStep({\n *   // ...\n *   execute: async ({ inputData, mastra, writer, abortSignal }) => {\n *     const stream = await mastra.getAgent('callCenter').stream(inputData.turn, { abortSignal });\n *     const reply = await pipeAgentReplyToWriter(stream, writer);\n *     return { reply };\n *   },\n * });\n * ```\n */\nexport function pipeAgentReplyToWriter(\n  agentStream: AgentReplyStreamLike,\n  writer: WritableStream<unknown>,\n): Promise<string> {\n  let text = '';\n  // Re-emit only the chunks the voice path cares about, then pipe through the step writer — which\n  // reuses pipeTo's proven backpressure + close handling rather than driving the writer by hand.\n  const forwarded = new ReadableStream<unknown>({\n    start: async controller => {\n      for await (const chunk of agentStream.fullStream) {\n        const type = (chunk as { type?: unknown })?.type;\n        if (type === 'text-delta') {\n          const delta = (chunk as { payload?: { text?: unknown } }).payload?.text;\n          if (typeof delta === 'string' && delta) {\n            text += delta;\n            controller.enqueue(chunk);\n          }\n        } else if (type === 'tool-call') {\n          controller.enqueue(chunk);\n        }\n      }\n      controller.close();\n    },\n  });\n  return forwarded.pipeTo(writer).then(() => text);\n}\n\n/**\n * A {@link VoiceReplyGenerator} backed by a Mastra workflow. Per turn it starts a fresh run to\n * completion (LiveKit owns the turn boundary, so there is no suspend/resume and no conversation\n * state carried between turns) and streams the text its steps write to their `writer`.\n *\n * A workflow's own stream emits structured step events, not token deltas — text only surfaces\n * when a step pipes it into the injected `writer`, arriving as `workflow-step-output` chunks. The\n * simplest correct reply step calls {@link pipeAgentReplyToWriter}, which forwards both text and\n * tool calls (and passes the step's `abortSignal` through `agent.stream` so barge-in stops\n * generation promptly). Piping only `agent.stream(...).textStream.pipeTo(writer)` works for text\n * but silently drops tool calls, so {@link WorkflowReplyGeneratorOptions.toolFeedback} and\n * {@link WorkflowReplyGeneratorOptions.onTurnComplete}'s `result.toolCalls` stay empty.\n */\nexport function createWorkflowReplyGenerator(options: WorkflowReplyGeneratorOptions): VoiceReplyGenerator {\n  const { workflow, workflowInput, replyStep, resultText, toolFeedback, onTurnComplete } = options;\n  return async ctx => {\n    const inputData = await workflowInput(ctx);\n    const run = await workflow.createRun();\n\n    const streamArgs: { inputData: unknown; tracingContext?: TracingContext; requestContext?: RequestContext } = {\n      inputData,\n    };\n    if (ctx.tracingContext) streamArgs.tracingContext = ctx.tracingContext;\n    // Forward the per-session request context so workflow steps see it, mirroring the agent path.\n    if (ctx.requestContext) streamArgs.requestContext = ctx.requestContext;\n    const output = run.stream(streamArgs);\n\n    let cancelled = false;\n    // Accumulated as the turn streams so the post-turn hook can see what was actually produced.\n    let replyText = '';\n    const toolCalls: VoiceToolCall[] = [];\n\n    // Fire-and-forget after the reply has streamed: off the audio path and not awaited, so it\n    // never delays the next turn. Errors are logged, not thrown. Mirrors createAgentReplyGenerator.\n    const emitTurnComplete = (interrupted: boolean) => {\n      if (!onTurnComplete) return;\n      const completeCtx: VoiceTurnCompleteContext = { ...ctx, result: { text: replyText, toolCalls, interrupted } };\n      Promise.resolve()\n        .then(() => onTurnComplete(completeCtx))\n        .catch(error => {\n          console.warn('@mastra/livekit: onTurnComplete hook threw', error);\n        });\n    };\n\n    return new ReadableStream<string>({\n      start: async controller => {\n        let streamedAny = false;\n        try {\n          for await (const chunk of output.fullStream) {\n            if (cancelled) break;\n            if (chunk.type !== 'workflow-step-output') continue;\n            const payload = chunk.payload as { output?: unknown; stepName?: unknown };\n            if (replyStep && payload.stepName !== replyStep) continue;\n            const text = unwrapStepText(payload.output);\n            if (text) {\n              streamedAny = true;\n              replyText += text;\n              controller.enqueue(text);\n              continue;\n            }\n            // A tool call only surfaces when the step pipes the agent's fullStream; when it does,\n            // mirror the agent path — record it and speak any toolFeedback filler.\n            const toolCall = unwrapStepToolCall(payload.output);\n            if (toolCall) {\n              toolCalls.push(toolCall);\n              if (toolFeedback) {\n                const filler = toolFeedback(toolCall);\n                if (filler) controller.enqueue(filler.endsWith(' ') ? filler : `${filler} `);\n              }\n            }\n          }\n          if (!cancelled && !streamedAny && resultText) {\n            const finalText = resultText(await output.result);\n            if (finalText) {\n              replyText += finalText;\n              controller.enqueue(finalText);\n            }\n          }\n          if (!cancelled) controller.close();\n          // Success, or a clean barge-in break out of the loop: the turn is done either way.\n          emitTurnComplete(cancelled);\n        } catch (error) {\n          // Barge-in cancels the run; that's not a failure — the turn still completed\n          // (interrupted), so the hook still fires for memory reconciliation.\n          if (cancelled) {\n            emitTurnComplete(true);\n            return;\n          }\n          controller.error(error);\n        }\n      },\n      cancel: () => {\n        cancelled = true;\n        // Barge-in fires this synchronously; swallow any rejection from the run cancellation so a\n        // failed cancel can't surface as an unhandled promise rejection.\n        void Promise.resolve(run.cancel()).catch(error => {\n          console.warn('@mastra/livekit: failed to cancel the workflow run on barge-in', error);\n        });\n      },\n    });\n  };\n}\n"],"mappings":";;;AACA,MAAa,6BAA6B;;;ACc1C,SAAgB,qBAAqB,KAAwD;CAC3F,IAAI,CAAC,KAAK,OAAO,CAAC;CAClB,IAAI;EACF,MAAM,SAAS,KAAK,MAAM,GAAG;EAC7B,IAAI,UAAU,OAAO,WAAW,YAAY,CAAC,MAAM,QAAQ,MAAM,GAC/D,OAAO;CAEX,QAAQ,CAER;CACA,OAAO,CAAC;AACV;AAEA,SAAgB,yBAAyB,UAA0C;CACjF,OAAO,KAAK,UAAU,QAAQ;AAChC;;;;;;;;;ACsBA,SAAgB,eAAe,QAAqC;CAClE,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,UAAU,OAAO,WAAW,YAAa,OAA8B,SAAS,cAAc;EAChG,MAAM,OAAQ,OAA4C,SAAS;EACnE,OAAO,OAAO,SAAS,WAAW,OAAO,KAAA;CAC3C;AAEF;;;;;;AAOA,SAAgB,mBAAmB,QAA4C;CAC7E,IAAI,CAAC,UAAU,OAAO,WAAW,YAAa,OAA8B,SAAS,aAAa,OAAO,KAAA;CACzG,MAAM,UAAW,OAAsF;CACvG,IAAI,WAAW,OAAO,QAAQ,eAAe,YAAY,OAAO,QAAQ,aAAa,UACnF,OAAO;EAAE,YAAY,QAAQ;EAAY,UAAU,QAAQ;EAAU,MAAM,QAAQ;CAAK;AAG5F;;;;;;;;;;;;;;;;;;;;;;;;;AA+BA,SAAgB,uBACd,aACA,QACiB;CACjB,IAAI,OAAO;CAoBX,OAAO,IAjBeA,WAAAA,eAAwB,EAC5C,OAAO,OAAM,eAAc;EACzB,WAAW,MAAM,SAAS,YAAY,YAAY;GAChD,MAAM,OAAQ,OAA8B;GAC5C,IAAI,SAAS,cAAc;IACzB,MAAM,QAAS,MAA2C,SAAS;IACnE,IAAI,OAAO,UAAU,YAAY,OAAO;KACtC,QAAQ;KACR,WAAW,QAAQ,KAAK;IAC1B;GACF,OAAO,IAAI,SAAS,aAClB,WAAW,QAAQ,KAAK;EAE5B;EACA,WAAW,MAAM;CACnB,EACF,CACe,CAAC,CAAC,OAAO,MAAM,CAAC,CAAC,WAAW,IAAI;AACjD;;;;;;;;;;;;;;AAeA,SAAgB,6BAA6B,SAA6D;CACxG,MAAM,EAAE,UAAU,eAAe,WAAW,YAAY,cAAc,mBAAmB;CACzF,OAAO,OAAM,QAAO;EAClB,MAAM,YAAY,MAAM,cAAc,GAAG;EACzC,MAAM,MAAM,MAAM,SAAS,UAAU;EAErC,MAAM,aAAuG,EAC3G,UACF;EACA,IAAI,IAAI,gBAAgB,WAAW,iBAAiB,IAAI;EAExD,IAAI,IAAI,gBAAgB,WAAW,iBAAiB,IAAI;EACxD,MAAM,SAAS,IAAI,OAAO,UAAU;EAEpC,IAAI,YAAY;EAEhB,IAAI,YAAY;EAChB,MAAM,YAA6B,CAAC;EAIpC,MAAM,oBAAoB,gBAAyB;GACjD,IAAI,CAAC,gBAAgB;GACrB,MAAM,cAAwC;IAAE,GAAG;IAAK,QAAQ;KAAE,MAAM;KAAW;KAAW;IAAY;GAAE;GAC5G,QAAQ,QAAQ,CAAC,CACd,WAAW,eAAe,WAAW,CAAC,CAAC,CACvC,OAAM,UAAS;IACd,QAAQ,KAAK,8CAA8C,KAAK;GAClE,CAAC;EACL;EAEA,OAAO,IAAIA,WAAAA,eAAuB;GAChC,OAAO,OAAM,eAAc;IACzB,IAAI,cAAc;IAClB,IAAI;KACF,WAAW,MAAM,SAAS,OAAO,YAAY;MAC3C,IAAI,WAAW;MACf,IAAI,MAAM,SAAS,wBAAwB;MAC3C,MAAM,UAAU,MAAM;MACtB,IAAI,aAAa,QAAQ,aAAa,WAAW;MACjD,MAAM,OAAO,eAAe,QAAQ,MAAM;MAC1C,IAAI,MAAM;OACR,cAAc;OACd,aAAa;OACb,WAAW,QAAQ,IAAI;OACvB;MACF;MAGA,MAAM,WAAW,mBAAmB,QAAQ,MAAM;MAClD,IAAI,UAAU;OACZ,UAAU,KAAK,QAAQ;OACvB,IAAI,cAAc;QAChB,MAAM,SAAS,aAAa,QAAQ;QACpC,IAAI,QAAQ,WAAW,QAAQ,OAAO,SAAS,GAAG,IAAI,SAAS,GAAG,OAAO,EAAE;OAC7E;MACF;KACF;KACA,IAAI,CAAC,aAAa,CAAC,eAAe,YAAY;MAC5C,MAAM,YAAY,WAAW,MAAM,OAAO,MAAM;MAChD,IAAI,WAAW;OACb,aAAa;OACb,WAAW,QAAQ,SAAS;MAC9B;KACF;KACA,IAAI,CAAC,WAAW,WAAW,MAAM;KAEjC,iBAAiB,SAAS;IAC5B,SAAS,OAAO;KAGd,IAAI,WAAW;MACb,iBAAiB,IAAI;MACrB;KACF;KACA,WAAW,MAAM,KAAK;IACxB;GACF;GACA,cAAc;IACZ,YAAY;IAGZ,QAAa,QAAQ,IAAI,OAAO,CAAC,CAAC,CAAC,OAAM,UAAS;KAChD,QAAQ,KAAK,kEAAkE,KAAK;IACtF,CAAC;GACH;EACF,CAAC;CACH;AACF"}

package/dist/workflow-generator-BtfClQcM.js

@@ -0,0 +1,180 @@
+import { ReadableStream } from "stream/web";
+//#region src/constants.ts
+/** Default LiveKit agent name used for explicit dispatch when none is configured. */
+const DEFAULT_LIVEKIT_AGENT_NAME = "mastra-voice";
+//#endregion
+//#region src/metadata.ts
+function parseSessionMetadata(raw) {
+	if (!raw) return {};
+	try {
+		const parsed = JSON.parse(raw);
+		if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) return parsed;
+	} catch {}
+	return {};
+}
+function serializeSessionMetadata(metadata) {
+	return JSON.stringify(metadata);
+}
+//#endregion
+//#region src/workflow-generator.ts
+/**
+* Unwraps the text carried by a `workflow-step-output` chunk's `payload.output`. A step that
+* pipes `agent.stream().textStream` into its `writer` produces raw strings; a step built from an
+* agent (`createStep(agent)`) produces full `text-delta` chunks. Returns the text for both, or
+* `undefined` for any other shape.
+*/
+function unwrapStepText(output) {
+	if (typeof output === "string") return output;
+	if (output && typeof output === "object" && output.type === "text-delta") {
+		const text = output.payload?.text;
+		return typeof text === "string" ? text : void 0;
+	}
+}
+/**
+* Unwraps a tool call carried by a `workflow-step-output` chunk's `payload.output`. A step that
+* pipes the agent's `fullStream` (rather than just `textStream`) into its `writer` surfaces
+* `tool-call` chunks; this returns the {@link VoiceToolCall} for those, or `undefined` otherwise.
+*/
+function unwrapStepToolCall(output) {
+	if (!output || typeof output !== "object" || output.type !== "tool-call") return void 0;
+	const payload = output.payload;
+	if (payload && typeof payload.toolCallId === "string" && typeof payload.toolName === "string") return {
+		toolCallId: payload.toolCallId,
+		toolName: payload.toolName,
+		args: payload.args
+	};
+}
+/**
+* Streams a Mastra agent's reply into a workflow step's `writer` for the LiveKit workflow
+* entrypoint — the recommended way to drive a turn's reply from a step.
+*
+* It forwards the agent's text deltas (so text-to-speech starts before the full reply is ready)
+* AND its `tool-call` chunks (so {@link WorkflowReplyGeneratorOptions.toolFeedback} fires and
+* {@link WorkflowReplyGeneratorOptions.onTurnComplete}'s `result.toolCalls` is populated). This is
+* the difference from piping only `agent.stream().textStream`, which silently drops tool calls.
+* Other chunk types (reasoning, tool results, lifecycle) are not forwarded, keeping the spoken
+* stream clean. Returns the accumulated reply text for the step to return.
+*
+* Pass the step's `abortSignal` to `agent.stream(...)` so barge-in stops generation promptly.
+*
+* ```ts
+* const generateResponse = createStep({
+*   // ...
+*   execute: async ({ inputData, mastra, writer, abortSignal }) => {
+*     const stream = await mastra.getAgent('callCenter').stream(inputData.turn, { abortSignal });
+*     const reply = await pipeAgentReplyToWriter(stream, writer);
+*     return { reply };
+*   },
+* });
+* ```
+*/
+function pipeAgentReplyToWriter(agentStream, writer) {
+	let text = "";
+	return new ReadableStream({ start: async (controller) => {
+		for await (const chunk of agentStream.fullStream) {
+			const type = chunk?.type;
+			if (type === "text-delta") {
+				const delta = chunk.payload?.text;
+				if (typeof delta === "string" && delta) {
+					text += delta;
+					controller.enqueue(chunk);
+				}
+			} else if (type === "tool-call") controller.enqueue(chunk);
+		}
+		controller.close();
+	} }).pipeTo(writer).then(() => text);
+}
+/**
+* A {@link VoiceReplyGenerator} backed by a Mastra workflow. Per turn it starts a fresh run to
+* completion (LiveKit owns the turn boundary, so there is no suspend/resume and no conversation
+* state carried between turns) and streams the text its steps write to their `writer`.
+*
+* A workflow's own stream emits structured step events, not token deltas — text only surfaces
+* when a step pipes it into the injected `writer`, arriving as `workflow-step-output` chunks. The
+* simplest correct reply step calls {@link pipeAgentReplyToWriter}, which forwards both text and
+* tool calls (and passes the step's `abortSignal` through `agent.stream` so barge-in stops
+* generation promptly). Piping only `agent.stream(...).textStream.pipeTo(writer)` works for text
+* but silently drops tool calls, so {@link WorkflowReplyGeneratorOptions.toolFeedback} and
+* {@link WorkflowReplyGeneratorOptions.onTurnComplete}'s `result.toolCalls` stay empty.
+*/
+function createWorkflowReplyGenerator(options) {
+	const { workflow, workflowInput, replyStep, resultText, toolFeedback, onTurnComplete } = options;
+	return async (ctx) => {
+		const inputData = await workflowInput(ctx);
+		const run = await workflow.createRun();
+		const streamArgs = { inputData };
+		if (ctx.tracingContext) streamArgs.tracingContext = ctx.tracingContext;
+		if (ctx.requestContext) streamArgs.requestContext = ctx.requestContext;
+		const output = run.stream(streamArgs);
+		let cancelled = false;
+		let replyText = "";
+		const toolCalls = [];
+		const emitTurnComplete = (interrupted) => {
+			if (!onTurnComplete) return;
+			const completeCtx = {
+				...ctx,
+				result: {
+					text: replyText,
+					toolCalls,
+					interrupted
+				}
+			};
+			Promise.resolve().then(() => onTurnComplete(completeCtx)).catch((error) => {
+				console.warn("@mastra/livekit: onTurnComplete hook threw", error);
+			});
+		};
+		return new ReadableStream({
+			start: async (controller) => {
+				let streamedAny = false;
+				try {
+					for await (const chunk of output.fullStream) {
+						if (cancelled) break;
+						if (chunk.type !== "workflow-step-output") continue;
+						const payload = chunk.payload;
+						if (replyStep && payload.stepName !== replyStep) continue;
+						const text = unwrapStepText(payload.output);
+						if (text) {
+							streamedAny = true;
+							replyText += text;
+							controller.enqueue(text);
+							continue;
+						}
+						const toolCall = unwrapStepToolCall(payload.output);
+						if (toolCall) {
+							toolCalls.push(toolCall);
+							if (toolFeedback) {
+								const filler = toolFeedback(toolCall);
+								if (filler) controller.enqueue(filler.endsWith(" ") ? filler : `${filler} `);
+							}
+						}
+					}
+					if (!cancelled && !streamedAny && resultText) {
+						const finalText = resultText(await output.result);
+						if (finalText) {
+							replyText += finalText;
+							controller.enqueue(finalText);
+						}
+					}
+					if (!cancelled) controller.close();
+					emitTurnComplete(cancelled);
+				} catch (error) {
+					if (cancelled) {
+						emitTurnComplete(true);
+						return;
+					}
+					controller.error(error);
+				}
+			},
+			cancel: () => {
+				cancelled = true;
+				Promise.resolve(run.cancel()).catch((error) => {
+					console.warn("@mastra/livekit: failed to cancel the workflow run on barge-in", error);
+				});
+			}
+		});
+	};
+}
+//#endregion
+export { DEFAULT_LIVEKIT_AGENT_NAME as a, serializeSessionMetadata as i, pipeAgentReplyToWriter as n, parseSessionMetadata as r, createWorkflowReplyGenerator as t };
+
+//# sourceMappingURL=workflow-generator-BtfClQcM.js.map

package/dist/workflow-generator-BtfClQcM.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"workflow-generator-BtfClQcM.js","names":[],"sources":["../src/constants.ts","../src/metadata.ts","../src/workflow-generator.ts"],"sourcesContent":["/** Default LiveKit agent name used for explicit dispatch when none is configured. */\nexport const DEFAULT_LIVEKIT_AGENT_NAME = 'mastra-voice';\n","/**\n * Session metadata passed from the Mastra server to the LiveKit agent worker through\n * LiveKit's job dispatch metadata (a plain string, so this is JSON-serialized).\n */\nexport interface LiveKitSessionMetadata {\n  /** Mastra agent to run, by registered key or agent id. */\n  agentId?: string;\n  /** Memory thread id. Defaults to the LiveKit room name when omitted. */\n  threadId?: string;\n  /** Memory resource id (typically the end user id). */\n  resourceId?: string;\n  /** Plain-object entries restored into a RequestContext for agent execution. */\n  requestContext?: Record<string, unknown>;\n}\n\nexport function parseSessionMetadata(raw: string | undefined | null): LiveKitSessionMetadata {\n  if (!raw) return {};\n  try {\n    const parsed = JSON.parse(raw);\n    if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {\n      return parsed as LiveKitSessionMetadata;\n    }\n  } catch {\n    // Dispatch metadata is user-controlled and may not be JSON; treat as absent.\n  }\n  return {};\n}\n\nexport function serializeSessionMetadata(metadata: LiveKitSessionMetadata): string {\n  return JSON.stringify(metadata);\n}\n","import { ReadableStream } from 'node:stream/web';\nimport type { WritableStream } from 'node:stream/web';\nimport type { TracingContext } from '@mastra/core/observability';\nimport type { RequestContext } from '@mastra/core/request-context';\nimport type { Workflow } from '@mastra/core/workflows';\nimport type {\n  VoiceReplyGenerator,\n  VoiceToolCall,\n  VoiceTurnCompleteContext,\n  VoiceTurnCompleteHook,\n  VoiceTurnContext,\n} from './bridge';\n\nexport interface WorkflowReplyGeneratorOptions {\n  /** The Mastra workflow that generates each turn's reply. Runs once per turn (no suspend/resume). */\n  workflow: Workflow;\n  /**\n   * Maps a turn into the workflow's `inputData`. Required — input schemas are caller-defined.\n   * A common shape passes the full transcript so the workflow is stateless between turns, e.g.\n   * `ctx => ({ history: chatContextToMessages(ctx.chatCtx) })`.\n   */\n  workflowInput: (ctx: VoiceTurnContext) => unknown | Promise<unknown>;\n  /**\n   * Only stream text from this step (by id). Defaults to every step that writes text to its\n   * `writer`. Set when multiple steps write and only one produces the spoken reply.\n   */\n  replyStep?: string;\n  /**\n   * Fallback when the workflow streams no text via `writer`: derive the spoken reply from the\n   * final run result. Without this, a non-streaming workflow stays silent. Streaming via the\n   * step `writer` is preferred — it gives the caller low time-to-first-token.\n   */\n  resultText?: (result: unknown) => string | undefined | void;\n  /**\n   * Speak a short phrase while a tool call runs. Fires only for tool calls the reply step surfaces\n   * to its `writer` — use {@link pipeAgentReplyToWriter} (or pipe the agent's `fullStream`) so\n   * `tool-call` chunks reach the stream. See {@link MastraVoiceAgentOptions.toolFeedback}.\n   */\n  toolFeedback?: (toolCall: VoiceToolCall) => string | undefined | void;\n  /**\n   * Fired off the audio path after the reply streams, fire-and-forget. Carries the produced reply\n   * text and any tool calls the workflow surfaced. See {@link MastraVoiceAgentOptions.onTurnComplete}.\n   */\n  onTurnComplete?: VoiceTurnCompleteHook;\n}\n\n/**\n * Unwraps the text carried by a `workflow-step-output` chunk's `payload.output`. A step that\n * pipes `agent.stream().textStream` into its `writer` produces raw strings; a step built from an\n * agent (`createStep(agent)`) produces full `text-delta` chunks. Returns the text for both, or\n * `undefined` for any other shape.\n */\nexport function unwrapStepText(output: unknown): string | undefined {\n  if (typeof output === 'string') return output;\n  if (output && typeof output === 'object' && (output as { type?: unknown }).type === 'text-delta') {\n    const text = (output as { payload?: { text?: unknown } }).payload?.text;\n    return typeof text === 'string' ? text : undefined;\n  }\n  return undefined;\n}\n\n/**\n * Unwraps a tool call carried by a `workflow-step-output` chunk's `payload.output`. A step that\n * pipes the agent's `fullStream` (rather than just `textStream`) into its `writer` surfaces\n * `tool-call` chunks; this returns the {@link VoiceToolCall} for those, or `undefined` otherwise.\n */\nexport function unwrapStepToolCall(output: unknown): VoiceToolCall | undefined {\n  if (!output || typeof output !== 'object' || (output as { type?: unknown }).type !== 'tool-call') return undefined;\n  const payload = (output as { payload?: { toolCallId?: unknown; toolName?: unknown; args?: unknown } }).payload;\n  if (payload && typeof payload.toolCallId === 'string' && typeof payload.toolName === 'string') {\n    return { toolCallId: payload.toolCallId, toolName: payload.toolName, args: payload.args };\n  }\n  return undefined;\n}\n\n/** The minimal shape of a Mastra agent stream consumed by {@link pipeAgentReplyToWriter}. */\nexport interface AgentReplyStreamLike {\n  fullStream: AsyncIterable<unknown>;\n}\n\n/**\n * Streams a Mastra agent's reply into a workflow step's `writer` for the LiveKit workflow\n * entrypoint — the recommended way to drive a turn's reply from a step.\n *\n * It forwards the agent's text deltas (so text-to-speech starts before the full reply is ready)\n * AND its `tool-call` chunks (so {@link WorkflowReplyGeneratorOptions.toolFeedback} fires and\n * {@link WorkflowReplyGeneratorOptions.onTurnComplete}'s `result.toolCalls` is populated). This is\n * the difference from piping only `agent.stream().textStream`, which silently drops tool calls.\n * Other chunk types (reasoning, tool results, lifecycle) are not forwarded, keeping the spoken\n * stream clean. Returns the accumulated reply text for the step to return.\n *\n * Pass the step's `abortSignal` to `agent.stream(...)` so barge-in stops generation promptly.\n *\n * ```ts\n * const generateResponse = createStep({\n *   // ...\n *   execute: async ({ inputData, mastra, writer, abortSignal }) => {\n *     const stream = await mastra.getAgent('callCenter').stream(inputData.turn, { abortSignal });\n *     const reply = await pipeAgentReplyToWriter(stream, writer);\n *     return { reply };\n *   },\n * });\n * ```\n */\nexport function pipeAgentReplyToWriter(\n  agentStream: AgentReplyStreamLike,\n  writer: WritableStream<unknown>,\n): Promise<string> {\n  let text = '';\n  // Re-emit only the chunks the voice path cares about, then pipe through the step writer — which\n  // reuses pipeTo's proven backpressure + close handling rather than driving the writer by hand.\n  const forwarded = new ReadableStream<unknown>({\n    start: async controller => {\n      for await (const chunk of agentStream.fullStream) {\n        const type = (chunk as { type?: unknown })?.type;\n        if (type === 'text-delta') {\n          const delta = (chunk as { payload?: { text?: unknown } }).payload?.text;\n          if (typeof delta === 'string' && delta) {\n            text += delta;\n            controller.enqueue(chunk);\n          }\n        } else if (type === 'tool-call') {\n          controller.enqueue(chunk);\n        }\n      }\n      controller.close();\n    },\n  });\n  return forwarded.pipeTo(writer).then(() => text);\n}\n\n/**\n * A {@link VoiceReplyGenerator} backed by a Mastra workflow. Per turn it starts a fresh run to\n * completion (LiveKit owns the turn boundary, so there is no suspend/resume and no conversation\n * state carried between turns) and streams the text its steps write to their `writer`.\n *\n * A workflow's own stream emits structured step events, not token deltas — text only surfaces\n * when a step pipes it into the injected `writer`, arriving as `workflow-step-output` chunks. The\n * simplest correct reply step calls {@link pipeAgentReplyToWriter}, which forwards both text and\n * tool calls (and passes the step's `abortSignal` through `agent.stream` so barge-in stops\n * generation promptly). Piping only `agent.stream(...).textStream.pipeTo(writer)` works for text\n * but silently drops tool calls, so {@link WorkflowReplyGeneratorOptions.toolFeedback} and\n * {@link WorkflowReplyGeneratorOptions.onTurnComplete}'s `result.toolCalls` stay empty.\n */\nexport function createWorkflowReplyGenerator(options: WorkflowReplyGeneratorOptions): VoiceReplyGenerator {\n  const { workflow, workflowInput, replyStep, resultText, toolFeedback, onTurnComplete } = options;\n  return async ctx => {\n    const inputData = await workflowInput(ctx);\n    const run = await workflow.createRun();\n\n    const streamArgs: { inputData: unknown; tracingContext?: TracingContext; requestContext?: RequestContext } = {\n      inputData,\n    };\n    if (ctx.tracingContext) streamArgs.tracingContext = ctx.tracingContext;\n    // Forward the per-session request context so workflow steps see it, mirroring the agent path.\n    if (ctx.requestContext) streamArgs.requestContext = ctx.requestContext;\n    const output = run.stream(streamArgs);\n\n    let cancelled = false;\n    // Accumulated as the turn streams so the post-turn hook can see what was actually produced.\n    let replyText = '';\n    const toolCalls: VoiceToolCall[] = [];\n\n    // Fire-and-forget after the reply has streamed: off the audio path and not awaited, so it\n    // never delays the next turn. Errors are logged, not thrown. Mirrors createAgentReplyGenerator.\n    const emitTurnComplete = (interrupted: boolean) => {\n      if (!onTurnComplete) return;\n      const completeCtx: VoiceTurnCompleteContext = { ...ctx, result: { text: replyText, toolCalls, interrupted } };\n      Promise.resolve()\n        .then(() => onTurnComplete(completeCtx))\n        .catch(error => {\n          console.warn('@mastra/livekit: onTurnComplete hook threw', error);\n        });\n    };\n\n    return new ReadableStream<string>({\n      start: async controller => {\n        let streamedAny = false;\n        try {\n          for await (const chunk of output.fullStream) {\n            if (cancelled) break;\n            if (chunk.type !== 'workflow-step-output') continue;\n            const payload = chunk.payload as { output?: unknown; stepName?: unknown };\n            if (replyStep && payload.stepName !== replyStep) continue;\n            const text = unwrapStepText(payload.output);\n            if (text) {\n              streamedAny = true;\n              replyText += text;\n              controller.enqueue(text);\n              continue;\n            }\n            // A tool call only surfaces when the step pipes the agent's fullStream; when it does,\n            // mirror the agent path — record it and speak any toolFeedback filler.\n            const toolCall = unwrapStepToolCall(payload.output);\n            if (toolCall) {\n              toolCalls.push(toolCall);\n              if (toolFeedback) {\n                const filler = toolFeedback(toolCall);\n                if (filler) controller.enqueue(filler.endsWith(' ') ? filler : `${filler} `);\n              }\n            }\n          }\n          if (!cancelled && !streamedAny && resultText) {\n            const finalText = resultText(await output.result);\n            if (finalText) {\n              replyText += finalText;\n              controller.enqueue(finalText);\n            }\n          }\n          if (!cancelled) controller.close();\n          // Success, or a clean barge-in break out of the loop: the turn is done either way.\n          emitTurnComplete(cancelled);\n        } catch (error) {\n          // Barge-in cancels the run; that's not a failure — the turn still completed\n          // (interrupted), so the hook still fires for memory reconciliation.\n          if (cancelled) {\n            emitTurnComplete(true);\n            return;\n          }\n          controller.error(error);\n        }\n      },\n      cancel: () => {\n        cancelled = true;\n        // Barge-in fires this synchronously; swallow any rejection from the run cancellation so a\n        // failed cancel can't surface as an unhandled promise rejection.\n        void Promise.resolve(run.cancel()).catch(error => {\n          console.warn('@mastra/livekit: failed to cancel the workflow run on barge-in', error);\n        });\n      },\n    });\n  };\n}\n"],"mappings":";;;AACA,MAAa,6BAA6B;;;ACc1C,SAAgB,qBAAqB,KAAwD;CAC3F,IAAI,CAAC,KAAK,OAAO,CAAC;CAClB,IAAI;EACF,MAAM,SAAS,KAAK,MAAM,GAAG;EAC7B,IAAI,UAAU,OAAO,WAAW,YAAY,CAAC,MAAM,QAAQ,MAAM,GAC/D,OAAO;CAEX,QAAQ,CAER;CACA,OAAO,CAAC;AACV;AAEA,SAAgB,yBAAyB,UAA0C;CACjF,OAAO,KAAK,UAAU,QAAQ;AAChC;;;;;;;;;ACsBA,SAAgB,eAAe,QAAqC;CAClE,IAAI,OAAO,WAAW,UAAU,OAAO;CACvC,IAAI,UAAU,OAAO,WAAW,YAAa,OAA8B,SAAS,cAAc;EAChG,MAAM,OAAQ,OAA4C,SAAS;EACnE,OAAO,OAAO,SAAS,WAAW,OAAO,KAAA;CAC3C;AAEF;;;;;;AAOA,SAAgB,mBAAmB,QAA4C;CAC7E,IAAI,CAAC,UAAU,OAAO,WAAW,YAAa,OAA8B,SAAS,aAAa,OAAO,KAAA;CACzG,MAAM,UAAW,OAAsF;CACvG,IAAI,WAAW,OAAO,QAAQ,eAAe,YAAY,OAAO,QAAQ,aAAa,UACnF,OAAO;EAAE,YAAY,QAAQ;EAAY,UAAU,QAAQ;EAAU,MAAM,QAAQ;CAAK;AAG5F;;;;;;;;;;;;;;;;;;;;;;;;;AA+BA,SAAgB,uBACd,aACA,QACiB;CACjB,IAAI,OAAO;CAoBX,OAAO,IAjBe,eAAwB,EAC5C,OAAO,OAAM,eAAc;EACzB,WAAW,MAAM,SAAS,YAAY,YAAY;GAChD,MAAM,OAAQ,OAA8B;GAC5C,IAAI,SAAS,cAAc;IACzB,MAAM,QAAS,MAA2C,SAAS;IACnE,IAAI,OAAO,UAAU,YAAY,OAAO;KACtC,QAAQ;KACR,WAAW,QAAQ,KAAK;IAC1B;GACF,OAAO,IAAI,SAAS,aAClB,WAAW,QAAQ,KAAK;EAE5B;EACA,WAAW,MAAM;CACnB,EACF,CACe,CAAC,CAAC,OAAO,MAAM,CAAC,CAAC,WAAW,IAAI;AACjD;;;;;;;;;;;;;;AAeA,SAAgB,6BAA6B,SAA6D;CACxG,MAAM,EAAE,UAAU,eAAe,WAAW,YAAY,cAAc,mBAAmB;CACzF,OAAO,OAAM,QAAO;EAClB,MAAM,YAAY,MAAM,cAAc,GAAG;EACzC,MAAM,MAAM,MAAM,SAAS,UAAU;EAErC,MAAM,aAAuG,EAC3G,UACF;EACA,IAAI,IAAI,gBAAgB,WAAW,iBAAiB,IAAI;EAExD,IAAI,IAAI,gBAAgB,WAAW,iBAAiB,IAAI;EACxD,MAAM,SAAS,IAAI,OAAO,UAAU;EAEpC,IAAI,YAAY;EAEhB,IAAI,YAAY;EAChB,MAAM,YAA6B,CAAC;EAIpC,MAAM,oBAAoB,gBAAyB;GACjD,IAAI,CAAC,gBAAgB;GACrB,MAAM,cAAwC;IAAE,GAAG;IAAK,QAAQ;KAAE,MAAM;KAAW;KAAW;IAAY;GAAE;GAC5G,QAAQ,QAAQ,CAAC,CACd,WAAW,eAAe,WAAW,CAAC,CAAC,CACvC,OAAM,UAAS;IACd,QAAQ,KAAK,8CAA8C,KAAK;GAClE,CAAC;EACL;EAEA,OAAO,IAAI,eAAuB;GAChC,OAAO,OAAM,eAAc;IACzB,IAAI,cAAc;IAClB,IAAI;KACF,WAAW,MAAM,SAAS,OAAO,YAAY;MAC3C,IAAI,WAAW;MACf,IAAI,MAAM,SAAS,wBAAwB;MAC3C,MAAM,UAAU,MAAM;MACtB,IAAI,aAAa,QAAQ,aAAa,WAAW;MACjD,MAAM,OAAO,eAAe,QAAQ,MAAM;MAC1C,IAAI,MAAM;OACR,cAAc;OACd,aAAa;OACb,WAAW,QAAQ,IAAI;OACvB;MACF;MAGA,MAAM,WAAW,mBAAmB,QAAQ,MAAM;MAClD,IAAI,UAAU;OACZ,UAAU,KAAK,QAAQ;OACvB,IAAI,cAAc;QAChB,MAAM,SAAS,aAAa,QAAQ;QACpC,IAAI,QAAQ,WAAW,QAAQ,OAAO,SAAS,GAAG,IAAI,SAAS,GAAG,OAAO,EAAE;OAC7E;MACF;KACF;KACA,IAAI,CAAC,aAAa,CAAC,eAAe,YAAY;MAC5C,MAAM,YAAY,WAAW,MAAM,OAAO,MAAM;MAChD,IAAI,WAAW;OACb,aAAa;OACb,WAAW,QAAQ,SAAS;MAC9B;KACF;KACA,IAAI,CAAC,WAAW,WAAW,MAAM;KAEjC,iBAAiB,SAAS;IAC5B,SAAS,OAAO;KAGd,IAAI,WAAW;MACb,iBAAiB,IAAI;MACrB;KACF;KACA,WAAW,MAAM,KAAK;IACxB;GACF;GACA,cAAc;IACZ,YAAY;IAGZ,QAAa,QAAQ,IAAI,OAAO,CAAC,CAAC,CAAC,OAAM,UAAS;KAChD,QAAQ,KAAK,kEAAkE,KAAK;IACtF,CAAC;GACH;EACF,CAAC;CACH;AACF"}

package/package.json

@@ -1,13 +1,12 @@
 {
   "name": "@mastra/livekit",
-  "version": "0.3.0",
+  "version": "0.3.1-alpha.1",
   "description": "LiveKit voice integration for Mastra agents — realtime voice with semantic turn detection and barge-in",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",
   "files": [
-    "dist",
-    "CHANGELOG.md"
+    "dist"
   ],
   "exports": {
     ".": {
@@ -84,23 +83,24 @@
     }
   },
   "devDependencies": {
-    "@livekit/agents": "^1.4.5",
-    "@livekit/agents-plugin-livekit": "^1.4.5",
-    "@livekit/agents-plugin-silero": "^1.4.5",
+    "@livekit/agents": "^1.7.1",
+    "@livekit/agents-plugin-livekit": "^1.7.1",
+    "@livekit/agents-plugin-silero": "^1.7.1",
     "@types/node": "22.20.1",
-    "eslint": "^10.4.1",
-    "tsup": "^8.5.1",
-    "typescript": "^6.0.3",
+    "eslint": "^10.7.0",
+    "tsdown": "0.22.9",
+    "typescript": "^7.0.2",
     "vitest": "4.1.10",
     "zod": "^4.4.3",
-    "@internal/types-builder": "0.0.89",
-    "@internal/lint": "0.0.114",
-    "@mastra/core": "1.51.0"
+    "@internal/types-builder": "0.0.104",
+    "@internal/lint": "0.0.129",
+    "@mastra/core": "1.64.0-alpha.5"
   },
   "scripts": {
-    "build:lib": "tsup --silent --config tsup.config.ts",
+    "build:lib": "tsdown --silent --config tsdown.config.ts",
     "build:watch": "pnpm build:lib --watch",
-    "lint": "eslint .",
+    "lint": "oxlint . && eslint .",
+    "lint:fix": "oxlint --fix . && eslint --fix .",
     "test": "vitest run"
   }
 }