The best way to control Claude Code, Codex and more from any computer or phone
| 1 | import { |
| 2 | AuthOrchestrationOperateScope, |
| 3 | SPEECH_STREAM_PATH, |
| 4 | SPEECH_STREAM_MAX_CHUNK_BYTES, |
| 5 | SpeechStreamCommand, |
| 6 | SPEECH_MAX_OPTIONS_BYTES, |
| 7 | type SpeechStreamEvent, |
| 8 | } from "@t3tools/contracts"; |
| 9 | import * as Effect from "effect/Effect"; |
| 10 | import * as Layer from "effect/Layer"; |
| 11 | import * as Queue from "effect/Queue"; |
| 12 | import * as Schema from "effect/Schema"; |
| 13 | import * as Scope from "effect/Scope"; |
| 14 | import { HttpRouter, HttpServerRequest, HttpServerResponse } from "effect/http"; |
| 15 | import * as Socket from "effect/socket/Socket"; |
| 16 | import * as EnvironmentAuth from "../auth/EnvironmentAuth.ts"; |
| 17 | import * as SpeechService from "./SpeechService.ts"; |
| 18 | |
| 19 | const decodeCommand = Schema.decodeUnknownEffect(Schema.fromJsonString(SpeechStreamCommand)); |
| 20 | |
| 21 | /** One socket owns one stream. Audio is processed and acknowledged in arrival order. */ |
| 22 | export const runSpeechSocket = Effect.fn("speech.runSocket")(function* ( |
| 23 | socket: Socket.Socket, |
| 24 | speech: Pick<SpeechService.SpeechService["Service"], "startStream">, |
| 25 | ) { |
| 26 | const scope = yield* Scope.Scope; |
| 27 | const writer = yield* socket.writer; |
| 28 | const send = (event: SpeechStreamEvent) => writer.write(JSON.stringify(event)); |
| 29 | let stream: SpeechService.SpeechStream | undefined; |
| 30 | let ended = false; |
| 31 | const fail = Effect.fn("speech.failSocket")(function* (message: string) { |
| 32 | if (ended) return; |
| 33 | ended = true; |
| 34 | yield* send({ type: "error", message }); |
| 35 | yield* writer.write(new Socket.CloseEvent(1008, "Speech stream ended")); |
| 36 | }); |
| 37 | const receive = (data: string | Uint8Array) => { |
| 38 | if (ended) return Effect.void; |
| 39 | return Effect.gen(function* () { |
| 40 | if (typeof data === "string") { |
| 41 | if (new TextEncoder().encode(data).byteLength > SPEECH_MAX_OPTIONS_BYTES) |
| 42 | return yield* fail("Invalid speech stream command."); |
| 43 | const command = yield* decodeCommand(data); |
| 44 | if (command.type === "start") { |
| 45 | if (stream) return yield* fail("Speech stream already started."); |
| 46 | stream = yield* speech.startStream(command.options).pipe(Scope.provide(scope)); |
| 47 | yield* send({ type: "ready" }); |
| 48 | return; |
| 49 | } |
| 50 | if (!stream) return yield* fail("Start speech transcription before finishing."); |
| 51 | const text = yield* stream.finish; |
| 52 | ended = true; |
| 53 | yield* send({ type: "finished", text }); |
| 54 | yield* writer.write(new Socket.CloseEvent(1000)); |
| 55 | } else { |
| 56 | if (data.byteLength > SPEECH_STREAM_MAX_CHUNK_BYTES) |
| 57 | return yield* fail("Speech audio chunk is too large."); |
| 58 | if (!stream) return yield* fail("Start speech transcription before sending audio."); |
| 59 | const update = yield* stream.feed(data); |
| 60 | yield* send({ type: "update", ...update }); |
| 61 | } |
| 62 | }).pipe( |
| 63 | Effect.catch((cause) => |
| 64 | Effect.logError("live transcription operation failed", { cause }).pipe( |
| 65 | Effect.andThen(fail("Live transcription failed. Please try recording again.")), |
| 66 | ), |
| 67 | ), |
| 68 | ); |
| 69 | }; |
| 70 | const reader = yield* socket.reader; |
| 71 | yield* Effect.gen(function* () { |
| 72 | const pending = yield* Queue.bounded<string | Uint8Array>(2); |
| 73 | yield* Effect.forever(Queue.take(pending).pipe(Effect.flatMap(receive))).pipe( |
| 74 | Effect.forkScoped, |
| 75 | ); |
| 76 | while (true) { |
| 77 | const frames = yield* reader.pull; |
| 78 | yield* Effect.forEach(frames, (frame) => Queue.offer(pending, frame), { |
| 79 | discard: true, |
| 80 | }); |
| 81 | } |
| 82 | }).pipe( |
| 83 | Effect.timeoutOrElse({ |
| 84 | duration: "10 minutes", |
| 85 | // The socket read loop is already interrupted here; its writer can no longer send. |
| 86 | orElse: () => Effect.void, |
| 87 | }), |
| 88 | Effect.ignore, |
| 89 | ); |
| 90 | }); |
| 91 | |
| 92 | export const speechStreamRouteLayer = Layer.unwrap( |
| 93 | Effect.gen(function* () { |
| 94 | const auth = yield* EnvironmentAuth.EnvironmentAuth; |
| 95 | const speech = yield* SpeechService.SpeechService; |
| 96 | return HttpRouter.add( |
| 97 | "GET", |
| 98 | SPEECH_STREAM_PATH, |
| 99 | Effect.gen(function* () { |
| 100 | const request = yield* HttpServerRequest.HttpServerRequest; |
| 101 | const session = yield* auth |
| 102 | .authenticateWebSocketUpgrade(request, "media") |
| 103 | .pipe(Effect.result); |
| 104 | if (session._tag === "Failure") return HttpServerResponse.empty({ status: 401 }); |
| 105 | if (!session.success.scopes.includes(AuthOrchestrationOperateScope)) |
| 106 | return HttpServerResponse.empty({ status: 403 }); |
| 107 | const socket = yield* request.upgrade; |
| 108 | yield* runSpeechSocket(socket, speech).pipe(Effect.scoped); |
| 109 | return HttpServerResponse.empty(); |
| 110 | }), |
| 111 | ); |
| 112 | }), |
| 113 | ); |