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 68d40cac..6ea2441a 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
@@ -31,7 +31,10 @@ import kotlinx.serialization.encodeToString
 import kotlinx.serialization.json.Json
 import org.slf4j.LoggerFactory
 
-private const val DEFAULT_REQUEST_TIMEOUT_MS = 600_000L
+// Aligned with the orchestrator's per-stage timeout (1_800_000ms). A slow local reasoning model
+// (large prompt + a multi-thousand-token reasoning budget) can take >10min for a single call;
+// a shorter HTTP timeout here fails the call before the stage timeout ever applies.
+private const val DEFAULT_REQUEST_TIMEOUT_MS = 1_800_000L
 
 private fun defaultHttpClient(): HttpClient = HttpClient(CIO) {
     install(ContentNegotiation) {