feat: тайминг-лог запросов ([llm-proxy] chat model=... status=... в Nms)

src/main/kotlin/pw/binom/llmproxy/Main.kt

@@ -103,6 +103,10 @@ private suspend fun handleChat(
         return
     }
 
+    val start = System.currentTimeMillis()
+    val model = try {
+        json.parseToJsonElement(patched).jsonObject["model"]?.jsonPrimitive?.content ?: "?"
+    } catch (e: Exception) { "?" }
     val req = HttpRequest.newBuilder()
         .uri(URI.create(upstream.trimEnd('/') + "/chat/completions"))
         .header("Authorization", "Bearer $apiKey")
@@ -111,17 +115,27 @@ private suspend fun handleChat(
         .build()
 
     val stream = json.parseToJsonElement(raw).jsonObject["stream"]?.jsonPrimitive?.content == "true"
-    if (stream) {
-        // SSE-стрим: транслируем как есть, чанк за чанком.
-        val resp = http.send(req, HttpResponse.BodyHandlers.ofInputStream())
-        val ct = resp.headers().firstValue("content-type").orElse("text/event-stream")
-        call.respondOutputStream(ContentType.parse(ct), HttpStatusCode.fromValue(resp.statusCode())) {
-            resp.body().use { input -> input.copyTo(this, 8192) }
+    try {
+        if (stream) {
+            // SSE-стрим: транслируем как есть, чанк за чанком.
+            val resp = http.send(req, HttpResponse.BodyHandlers.ofInputStream())
+            val ct = resp.headers().firstValue("content-type").orElse("text/event-stream")
+            call.respondOutputStream(ContentType.parse(ct), HttpStatusCode.fromValue(resp.statusCode())) {
+                resp.body().use { input -> input.copyTo(this, 8192) }
+            }
+            println("[llm-proxy] chat model=$model status=${resp.statusCode()} в ${System.currentTimeMillis() - start}ms stream=true")
+        } else {
+            val resp = http.send(req, HttpResponse.BodyHandlers.ofByteArray())
+            val ct = resp.headers().firstValue("content-type").orElse("application/json")
+            call.respondBytes(resp.body(), ContentType.parse(ct), HttpStatusCode.fromValue(resp.statusCode()))
+            println("[llm-proxy] chat model=$model status=${resp.statusCode()} в ${System.currentTimeMillis() - start}ms stream=false")
         }
-    } else {
-        val resp = http.send(req, HttpResponse.BodyHandlers.ofByteArray())
-        val ct = resp.headers().firstValue("content-type").orElse("application/json")
-        call.respondBytes(resp.body(), ContentType.parse(ct), HttpStatusCode.fromValue(resp.statusCode()))
+    } catch (e: Exception) {
+        println("[llm-proxy] chat model=$model ОШИБКА: ${e.message} в ${System.currentTimeMillis() - start}ms")
+        call.respondBytes(
+            """{"error":{"message":"upstream: ${e.message}"}}""".toByteArray(),
+            ContentType.Application.Json, HttpStatusCode.BadGateway,
+        )
     }
 }