From 3ff9c8281a42fac9d033c4860e50635399b5006b Mon Sep 17 00:00:00 2001
From: kami <kamixtrash@gmail.com>
Date: Sun, 12 Jul 2026 12:49:46 +0400
Subject: [PATCH] fix(inference): single ContentNegotiation config +
 firstOrNull on empty choices

---
 .../llama/cpp/LlamaCppInferenceProvider.kt     | 18 ++++++------------
 1 file changed, 6 insertions(+), 12 deletions(-)

diff --git a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt
index f0712812..54ba0906 100644
--- a/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt
+++ b/infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppInferenceProvider.kt
@@ -42,22 +42,14 @@ private fun defaultHttpClient(): HttpClient = HttpClient(CIO) {
             Json {
                 ignoreUnknownKeys = true
                 isLenient = true
-                encodeDefaults = false
+                explicitNulls = false
+                encodeDefaults = true
             },
         )
     }
     install(HttpTimeout) {
         requestTimeoutMillis = DEFAULT_REQUEST_TIMEOUT_MS
     }
-    install(ContentNegotiation) {
-        json(
-            Json {
-                ignoreUnknownKeys = true
-                explicitNulls = false
-                encodeDefaults = true
-            },
-        )
-    }
 }
 private val json = Json {
     ignoreUnknownKeys = true
@@ -199,7 +191,9 @@ class LlamaCppInferenceProvider(
 
         log.debug("got response from llm: {}", response)
 
-        val message = response.choices.first().message
+        val choice = response.choices.firstOrNull()
+            ?: error("llama-server returned no choices in the completion response")
+        val message = choice.message
         // Some local models emit the tool call as a JSON blob in `content` instead of the native
         // tool_calls array; salvage it so the orchestrator gets a real call rather than treating
         // the blob as a (failing) artifact and retry-looping to exhaustion.
@@ -207,7 +201,7 @@ class LlamaCppInferenceProvider(
         val toolCalls = message.toolCalls.ifEmpty { salvaged }
         val finishReason = when {
             toolCalls.isNotEmpty() -> FinishReason.ToolCall
-            response.choices.first().finishReason.lowercase() == "length" -> FinishReason.Length
+            choice.finishReason.lowercase() == "length" -> FinishReason.Length
             else -> FinishReason.Stop
         }