Add first-chunk timeout

src/main/kotlin/ModelService.kt

@@ -14,6 +14,7 @@ import kotlinx.coroutines.currentCoroutineContext
 import kotlinx.coroutines.flow.Flow
 import kotlinx.coroutines.flow.flow
 import kotlinx.coroutines.isActive
+import kotlinx.coroutines.withTimeout
 import kotlinx.serialization.json.*
 import kotlin.time.Duration
 
@@ -21,15 +22,16 @@ class ModelService(
     val baseUrl: String,
     val token: String,
     val timeout: Duration,
+    val firstChunkTimeout: Duration,
     val extraBody: JsonObject? = null
 ) {
     val httpClient: HttpClient by lazy {
         HttpClient(OkHttp) {
             install(HttpTimeout) {
-                val millis = timeout.inWholeMilliseconds
-                requestTimeoutMillis = millis
-                connectTimeoutMillis = millis
-                socketTimeoutMillis = millis
+                // 总请求/socket 超时保持长值，允许慢速流式输出；连接握手则用短超时。
+                requestTimeoutMillis = timeout.inWholeMilliseconds
+                socketTimeoutMillis = timeout.inWholeMilliseconds
+                connectTimeoutMillis = firstChunkTimeout.inWholeMilliseconds
             }
             defaultRequest {
                 url(baseUrl)
@@ -66,17 +68,35 @@ class ModelService(
             }.let { response ->
                 val channel: ByteReadChannel = response.body()
                 try {
-                    while (currentCoroutineContext().isActive && !channel.isClosedForRead) {
-                        val line = channel.readUTF8Line() ?: continue
-                        when {
-                            line.startsWith("data: [DONE]") -> break
-                            line.startsWith("data: ") -> {
-                                val chunk = json.decodeFromString<ChatCompletionChunk>(
-                                    line.removePrefix("data: ")
-                                )
-                                emit(chunk)
+                    // 首块 data: 必须在 firstChunkTimeout 内到达，否则抛 TimeoutCancellationException
+                    // 走 JChatGPT 的重试流程；之后的流式读取不再有应用层超时，由 socketTimeoutMillis 兜底。
+                    val firstDataLine: String? = withTimeout(firstChunkTimeout) {
+                        var found: String? = null
+                        while (currentCoroutineContext().isActive && !channel.isClosedForRead) {
+                            val line = channel.readUTF8Line() ?: continue
+                            if (line.startsWith("data: ")) {
+                                found = line
+                                break
+                            }
+                            // 心跳/空行/注释行，不计为首块，继续等
+                        }
+                        found
+                    }
+
+                    if (firstDataLine != null) {
+                        if (!firstDataLine.startsWith("data: [DONE]")) {
+                            emit(json.decodeFromString(firstDataLine.removePrefix("data: ")))
+
+                            while (currentCoroutineContext().isActive && !channel.isClosedForRead) {
+                                val line = channel.readUTF8Line() ?: continue
+                                when {
+                                    line.startsWith("data: [DONE]") -> break
+                                    line.startsWith("data: ") -> {
+                                        emit(json.decodeFromString(line.removePrefix("data: ")))
+                                    }
+                                    else -> continue
+                                }
                             }
-                            else -> continue
                         }
                     }
                 } finally {