@astropods/messaging 0.0.2 → 0.0.3

package/dist/messaging-client.d.ts

@@ -45,6 +45,9 @@ export interface AgentResponse {
     threadMetadata?: ThreadMetadata;
     error?: ErrorResponse;
     contextRequest?: ThreadHistoryRequest;
+    transcript?: Transcript;
+    audioConfig?: AudioStreamConfig;
+    audioChunk?: AudioChunk;
 }
 export interface StatusUpdate {
     status: 'THINKING' | 'SEARCHING' | 'GENERATING' | 'PROCESSING' | 'ANALYZING' | 'CUSTOM';
@@ -77,6 +80,11 @@ export interface ErrorResponse {
     details?: string;
     retryable?: boolean;
 }
+export interface Transcript {
+    text: string;
+    messageId?: string;
+    language?: string;
+}
 export interface ThreadHistoryRequest {
     conversationId: string;
     maxMessages?: number;
@@ -129,11 +137,54 @@ export interface AgentConfig {
     systemPrompt: string;
     tools: AgentToolConfig[];
 }
+/**
+ * Supported audio encoding formats. Matches the AudioEncoding protobuf enum.
+ *
+ * Common sources:
+ * - LINEAR16: Universal PCM baseline (any platform)
+ * - MULAW: Twilio / telephony (G.711 mu-law, 8kHz)
+ * - WEBM_OPUS: Browser MediaRecorder default
+ * - AAC: iOS native recording
+ */
+export type AudioEncoding = 'LINEAR16' | 'MULAW' | 'OPUS' | 'MP3' | 'WEBM_OPUS' | 'OGG_OPUS' | 'FLAC' | 'AAC';
+/**
+ * Configuration sent at the start of an audio segment to describe the format.
+ * Maps to the AudioStreamConfig protobuf message.
+ */
+export interface AudioStreamConfig {
+    encoding: AudioEncoding;
+    sampleRate: number;
+    channels: number;
+    language?: string;
+    conversationId: string;
+    source?: string;
+}
+/**
+ * A chunk of raw audio bytes. Maps to the AudioChunk protobuf message.
+ *
+ * Chunks arrive sequentially during a segment. When done=true, the segment
+ * is complete and the agent should run STT on the accumulated audio.
+ */
+export interface AudioChunk {
+    data: Buffer | Uint8Array;
+    sequence?: number;
+    done?: boolean;
+}
+/**
+ * Maps an AudioEncoding to the filetype string expected by Mastra's voice.listen().
+ *
+ * Usage:
+ *   const filetype = audioEncodingToFiletype(config.encoding);
+ *   const transcript = await agent.voice.listen(audioStream, { filetype });
+ */
+export declare function audioEncodingToFiletype(encoding: AudioEncoding): string;
 export interface ConversationRequest {
     message?: Message;
     feedback?: any;
     agentConfig?: AgentConfig;
     agentResponse?: AgentResponse;
+    audioConfig?: AudioStreamConfig;
+    audio?: AudioChunk;
 }
 export interface ReconnectOptions {
     /** Maximum number of reconnect attempts. Default: Infinity */
@@ -244,6 +295,61 @@ export declare class ConversationStream extends EventEmitter {
      * Send a status update for a conversation
      */
     sendStatusUpdate(conversationId: string, status: StatusUpdate): void;
+    /**
+     * Send a transcript of the user's audio input back to the platform.
+     *
+     * After the agent runs STT on the audio, it calls this to send the transcribed
+     * text back to the platform (web adapter). The platform uses it to replace the
+     * "[audio]" placeholder message with the actual spoken text in the chat UI.
+     *
+     * @param conversationId - The conversation this transcript belongs to
+     * @param text - The transcribed text from STT
+     * @param messageId - Optional: the original "[audio]" message ID to update
+     * @param language - Optional: BCP-47 language detected by STT (e.g. "en-US")
+     */
+    sendTranscript(conversationId: string, text: string, messageId?: string, language?: string): void;
+    /**
+     * Send an audio stream config through the bidi stream.
+     * Must be called before sendAudioChunk() so the receiver knows the encoding.
+     */
+    sendAudioConfig(config: AudioStreamConfig): void;
+    /**
+     * Send a raw audio chunk through the bidi stream.
+     * The chunk's sequence number should increase monotonically.
+     */
+    sendAudioChunk(chunk: AudioChunk): void;
+    /**
+     * Signal end of the current audio segment by sending an empty chunk with done=true.
+     * The receiver should process all accumulated audio (e.g. run STT).
+     * After this, more audio can follow — either new config or more chunks.
+     */
+    endAudio(): void;
+    /**
+     * Converts incoming audioChunk events into a Web Streams API ReadableStream.
+     *
+     * This is the primary integration point with Mastra's voice system. The agent
+     * listens for the 'audioConfig' event to know the format, then calls this
+     * method to get a stream it can pass directly to voice.listen():
+     *
+     * ```typescript
+     * conversation.on('audioConfig', async (config) => {
+     *   const audioStream = conversation.audioAsReadable();
+     *   const filetype = audioEncodingToFiletype(config.encoding);
+     *   const transcript = await agent.voice.listen(audioStream, { filetype });
+     *   // ... process transcript
+     * });
+     * ```
+     *
+     * The ReadableStream:
+     * - Yields Uint8Array chunks as audioChunk events arrive
+     * - Closes when an AudioChunk with done=true arrives (end of segment)
+     * - Closes when the ConversationStream emits 'end' (intentional close)
+     * - Errors when the ConversationStream emits 'error'
+     * - Properly cleans up all event listeners on close, error, or cancel
+     *
+     * @returns A ReadableStream<Uint8Array> suitable for Mastra voice.listen()
+     */
+    audioAsReadable(): ReadableStream<Uint8Array>;
     /**
      * End the stream intentionally. Emits 'end' and prevents any further reconnects.
      */

package/dist/messaging-client.js

@@ -34,10 +34,31 @@ var __importStar = (this && this.__importStar) || (function () {
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.Helpers = exports.MessageStream = exports.ConversationStream = exports.MessagingClient = void 0;
+exports.audioEncodingToFiletype = audioEncodingToFiletype;
 const grpc = __importStar(require("@grpc/grpc-js"));
 const protoLoader = __importStar(require("@grpc/proto-loader"));
 const path_1 = require("path");
 const events_1 = require("events");
+/**
+ * Maps an AudioEncoding to the filetype string expected by Mastra's voice.listen().
+ *
+ * Usage:
+ *   const filetype = audioEncodingToFiletype(config.encoding);
+ *   const transcript = await agent.voice.listen(audioStream, { filetype });
+ */
+function audioEncodingToFiletype(encoding) {
+    const map = {
+        LINEAR16: 'wav',
+        MULAW: 'wav',
+        OPUS: 'opus',
+        MP3: 'mp3',
+        WEBM_OPUS: 'webm',
+        OGG_OPUS: 'ogg',
+        FLAC: 'flac',
+        AAC: 'm4a',
+    };
+    return map[encoding] ?? 'wav';
+}
 // gRPC status codes: DEADLINE_EXCEEDED=4, INTERNAL=13, UNAVAILABLE=14, RESOURCE_EXHAUSTED=8
 const DEFAULT_RETRYABLE_STATUS_CODES = [4, 8, 13, 14];
 function resolveReconnectOptions(options) {
@@ -234,6 +255,17 @@ class ConversationStream extends events_1.EventEmitter {
     attachHandlers(stream) {
         stream.on('data', (response) => {
             this.retryCount = 0;
+            // Emit audio-specific events if present.
+            // The server sends audio data through the bidi stream as AgentResponse
+            // messages with audioConfig or audioChunk payloads. We emit dedicated
+            // events for these so the agent can handle audio separately from text,
+            // while still emitting the generic 'response' event for observability.
+            if (response.audioConfig) {
+                this.emit('audioConfig', response.audioConfig);
+            }
+            else if (response.audioChunk) {
+                this.emit('audioChunk', response.audioChunk);
+            }
             this.emit('response', response);
         });
         stream.on('error', (error) => {
@@ -345,6 +377,131 @@ class ConversationStream extends events_1.EventEmitter {
             status,
         });
     }
+    /**
+     * Send a transcript of the user's audio input back to the platform.
+     *
+     * After the agent runs STT on the audio, it calls this to send the transcribed
+     * text back to the platform (web adapter). The platform uses it to replace the
+     * "[audio]" placeholder message with the actual spoken text in the chat UI.
+     *
+     * @param conversationId - The conversation this transcript belongs to
+     * @param text - The transcribed text from STT
+     * @param messageId - Optional: the original "[audio]" message ID to update
+     * @param language - Optional: BCP-47 language detected by STT (e.g. "en-US")
+     */
+    sendTranscript(conversationId, text, messageId, language) {
+        this.sendAgentResponse({
+            conversationId,
+            transcript: { text, messageId, language },
+        });
+    }
+    // --- Audio support ---
+    //
+    // These methods handle sending audio data through the gRPC bidi stream.
+    // Two directions:
+    //   - Agent → Server (sendAudioConfig/sendAudioChunk/endAudio): used when the
+    //     agent needs to forward audio upstream (less common)
+    //   - Server → Agent (audioConfig/audioChunk events + audioAsReadable): the main
+    //     path where the server forwards client mic audio to the agent for STT
+    /**
+     * Send an audio stream config through the bidi stream.
+     * Must be called before sendAudioChunk() so the receiver knows the encoding.
+     */
+    sendAudioConfig(config) {
+        this.write({ audioConfig: config });
+    }
+    /**
+     * Send a raw audio chunk through the bidi stream.
+     * The chunk's sequence number should increase monotonically.
+     */
+    sendAudioChunk(chunk) {
+        this.write({ audio: chunk });
+    }
+    /**
+     * Signal end of the current audio segment by sending an empty chunk with done=true.
+     * The receiver should process all accumulated audio (e.g. run STT).
+     * After this, more audio can follow — either new config or more chunks.
+     */
+    endAudio() {
+        this.write({ audio: { data: Buffer.alloc(0), done: true } });
+    }
+    /**
+     * Converts incoming audioChunk events into a Web Streams API ReadableStream.
+     *
+     * This is the primary integration point with Mastra's voice system. The agent
+     * listens for the 'audioConfig' event to know the format, then calls this
+     * method to get a stream it can pass directly to voice.listen():
+     *
+     * ```typescript
+     * conversation.on('audioConfig', async (config) => {
+     *   const audioStream = conversation.audioAsReadable();
+     *   const filetype = audioEncodingToFiletype(config.encoding);
+     *   const transcript = await agent.voice.listen(audioStream, { filetype });
+     *   // ... process transcript
+     * });
+     * ```
+     *
+     * The ReadableStream:
+     * - Yields Uint8Array chunks as audioChunk events arrive
+     * - Closes when an AudioChunk with done=true arrives (end of segment)
+     * - Closes when the ConversationStream emits 'end' (intentional close)
+     * - Errors when the ConversationStream emits 'error'
+     * - Properly cleans up all event listeners on close, error, or cancel
+     *
+     * @returns A ReadableStream<Uint8Array> suitable for Mastra voice.listen()
+     */
+    audioAsReadable() {
+        // Centralized cleanup to prevent listener leaks. Called on:
+        // - done=true chunk (normal completion)
+        // - stream 'end' event (intentional close)
+        // - stream 'error' event
+        // - ReadableStream cancel() (consumer gave up, e.g. reader.cancel())
+        const cleanup = () => {
+            this.removeListener('audioChunk', onChunk);
+            this.removeListener('end', onEnd);
+            this.removeListener('error', onError);
+        };
+        const onChunk = (chunk) => {
+            if (chunk.done) {
+                cleanup();
+                try {
+                    controller.close();
+                }
+                catch { }
+            }
+            else {
+                controller.enqueue(new Uint8Array(chunk.data));
+            }
+        };
+        const onEnd = () => {
+            cleanup();
+            try {
+                controller.close();
+            }
+            catch { }
+        };
+        const onError = (err) => {
+            cleanup();
+            try {
+                controller.error(err);
+            }
+            catch { }
+        };
+        let controller;
+        return new ReadableStream({
+            start: (ctrl) => {
+                controller = ctrl;
+                this.on('audioChunk', onChunk);
+                this.once('end', onEnd);
+                this.once('error', onError);
+            },
+            cancel: () => {
+                // Consumer cancelled (e.g. reader.cancel()) — remove all listeners
+                // to prevent memory leaks
+                cleanup();
+            },
+        });
+    }
     /**
      * End the stream intentionally. Emits 'end' and prevents any further reconnects.
      */

package/dist/proto/astro/messaging/v1/audio.proto

@@ -0,0 +1,71 @@
+// Audio streaming types for the messaging system.
+//
+// These types enable raw audio input from any frontend (browser, phone, mobile app)
+// to be streamed through the messaging server to an agent. The messaging system is
+// a pass-through — it does NOT perform speech-to-text, transcoding, or voice activity
+// detection. The agent handles STT via Mastra's voice provider abstraction.
+//
+// Data flow:
+//   Client (mic) → WebSocket → Server → gRPC (these types) → Agent → Mastra voice.listen()
+//
+// Two ways audio enters the system:
+//   1. ProcessAudioStream RPC: dedicated audio-only streaming (AudioStreamRequest)
+//   2. ProcessConversation RPC: audio mixed into the bidi stream (ConversationRequest.audio_config/audio)
+//
+// Both converge on the same types: AudioStreamConfig describes the format,
+// AudioChunk carries the bytes, and done=true signals end of an utterance.
+
+syntax = "proto3";
+
+package astro.messaging.v1;
+
+option go_package = "github.com/postman/astro/messaging/v1;messagingv1";
+
+// Audio encoding format — covers browser, telephony, and mobile sources.
+// The agent uses this to configure the STT provider (e.g. Whisper, Deepgram).
+enum AudioEncoding {
+  AUDIO_ENCODING_UNSPECIFIED = 0;
+  LINEAR16 = 1;          // PCM signed 16-bit little-endian — universal baseline, any platform
+  MULAW = 2;             // G.711 mu-law — Twilio and traditional telephony (8kHz)
+  OPUS = 3;              // Raw Opus frames — low-latency codec
+  MP3 = 4;               // MP3 — batch uploads, pre-recorded audio
+  WEBM_OPUS = 5;         // WebM container with Opus — browser MediaRecorder default
+  OGG_OPUS = 6;          // OGG container with Opus — Firefox MediaRecorder
+  FLAC = 7;              // FLAC lossless — high-quality uploads
+  AAC = 8;               // AAC — iOS native recording
+}
+
+// Sent once at the start of an audio segment to tell the agent what format
+// the subsequent AudioChunk bytes are in. Without this, the agent can't
+// decode the raw bytes.
+message AudioStreamConfig {
+  AudioEncoding encoding = 1;     // What codec the audio bytes use
+  int32 sample_rate = 2;          // Hz: 8000 (telephony), 16000 (speech), 48000 (browser)
+  int32 channels = 3;             // 1 = mono (speech default), 2 = stereo
+  string language = 4;            // BCP-47 hint for STT, e.g. "en-US" (optional)
+  string conversation_id = 5;     // Links this audio to an existing conversation
+
+  // Source metadata — helps the agent pick the right STT config.
+  // Examples: "browser", "twilio", "vonage", "mobile", "upload"
+  string source = 6;
+}
+
+// A chunk of raw audio bytes in the encoding specified by AudioStreamConfig.
+//
+// Chunks arrive in order during a segment. When done=true, the segment is
+// complete and the agent should run STT on all accumulated chunks.
+// The data field may be empty on the final done=true chunk.
+message AudioChunk {
+  bytes data = 1;                 // Raw audio bytes (pass-through, no transcoding)
+  int64 sequence = 2;             // Monotonic sequence number for ordering
+  bool done = 3;                  // true = end of segment, process accumulated audio
+}
+
+// Wrapper for the ProcessAudioStream RPC (dedicated audio streaming).
+// The first message MUST be config, all subsequent messages are audio chunks.
+message AudioStreamRequest {
+  oneof request {
+    AudioStreamConfig config = 1;    // First message: tells agent the audio format
+    AudioChunk audio = 2;            // Subsequent: raw audio bytes
+  }
+}

package/dist/proto/astro/messaging/v1/response.proto

@@ -4,6 +4,7 @@ package astro.messaging.v1;
 
 import "google/protobuf/timestamp.proto";
 import "astro/messaging/v1/message.proto";
+import "astro/messaging/v1/audio.proto";
 
 option go_package = "github.com/postman/astro/messaging/v1;messagingv1";
 
@@ -22,6 +23,9 @@ message AgentResponse {
     ThreadMetadata thread_metadata = 7;   // Thread title, creation
     ErrorResponse error = 8;              // Error during processing
     ThreadHistoryRequest context_request = 9;  // Request cached context (optional)
+    Transcript transcript = 10;           // Audio transcript (agent → platform)
+    AudioStreamConfig audio_config = 11;  // Audio session config (server → agent)
+    AudioChunk audio_chunk = 12;          // Audio data (server → agent)
   }
 }
 
@@ -144,6 +148,14 @@ message ThreadMetadata {
   bool create_new = 3;                    // Create new thread
 }
 
+// Transcript of user audio input (agent → platform after STT)
+// Used to update a placeholder message with the actual transcribed text
+message Transcript {
+  string text = 1;                        // Transcribed text
+  string message_id = 2;                  // User message ID to update (optional)
+  string language = 3;                    // Detected language BCP-47 (optional)
+}
+
 // Error response from agent
 message ErrorResponse {
   enum ErrorCode {

package/dist/proto/astro/messaging/v1/service.proto

@@ -5,6 +5,7 @@ package astro.messaging.v1;
 import "astro/messaging/v1/message.proto";
 import "astro/messaging/v1/response.proto";
 import "astro/messaging/v1/feedback.proto";
+import "astro/messaging/v1/audio.proto";
 import "astro/messaging/v1/config.proto";
 import "google/protobuf/timestamp.proto";
 
@@ -29,6 +30,11 @@ service AgentMessaging {
   rpc GetConversationMetadata(ConversationMetadataRequest)
       returns (ConversationMetadataResponse);
 
+  // Audio: client streams raw audio, server responds with text
+  // First message MUST be AudioStreamConfig, rest are AudioChunks
+  rpc ProcessAudioStream(stream AudioStreamRequest)
+      returns (stream AgentResponse);
+
   // Health check
   rpc HealthCheck(HealthCheckRequest)
       returns (HealthCheckResponse);
@@ -41,6 +47,8 @@ message ConversationRequest {
     PlatformFeedback feedback = 2;
     AgentConfig agent_config = 3;
     AgentResponse agent_response = 4;
+    AudioStreamConfig audio_config = 5;  // Start audio within conversation
+    AudioChunk audio = 6;                // Audio data within conversation
   }
 }
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@astropods/messaging",
   "license": "Apache-2.0",
-  "version": "0.0.2",
+  "version": "0.0.3",
   "description": "TypeScript SDK for Astro Messaging",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -10,7 +10,7 @@
   ],
   "scripts": {
     "postinstall": "rm -rf proto && ln -sf ../../proto proto",
-    "build": "tsc && cp -r ../../proto dist/proto",
+    "build": "tsc && rm -rf dist/proto && cp -r ../../proto dist/proto",
     "watch": "tsc --watch",
     "test": "bun test",
     "test:watch": "bun test --watch"