From 3a26f43b9d078630a5224d7259ce83ef080233e5 Mon Sep 17 00:00:00 2001
From: "macroscopeapp[bot]"
 <170038800+macroscopeapp[bot]@users.noreply.github.com>
Date: Fri, 21 Aug 2026 03:10:07 +0000
Subject: [PATCH] Halve memory-derived warm pool size to prevent OOM during
 successor provisioning

---
 crates/preloop-orchestrator/src/lib.rs | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/crates/preloop-orchestrator/src/lib.rs b/crates/preloop-orchestrator/src/lib.rs
index 8d70fb9a..59dd9a7d 100644
--- a/crates/preloop-orchestrator/src/lib.rs
+++ b/crates/preloop-orchestrator/src/lib.rs
@@ -2371,10 +2371,11 @@ impl<P: VmProvider + 'static> RunnerPool<P> {
         // at startup is safer than OOMing mid-run; `PRELOOP_RUNNER_POOL_SIZE`
         // remains an explicit override that wins.
         let warm_size = match host_memory_mib() {
-            Some(total) => self.config.size.min(on_demand_memory_cap(
-                total,
-                u64::from(self.config.memory_mib),
-            )),
+            Some(total) => self.config.size.min(
+                on_demand_memory_cap(total, u64::from(self.config.memory_mib))
+                    .saturating_div(2)
+                    .max(1),
+            ),
             None => self.config.size,
         };
         if warm_size < self.config.size {