From bc1f2b47991c460ce5238770c0a5c80872695fd2 Mon Sep 17 00:00:00 2001
From: Roy Lin <roylin@a3s.box>
Date: Fri, 12 Jun 2026 14:18:44 +0800
Subject: [PATCH] fix(vm): event-style log drain for run_deferred_main

Replace the fixed 150ms post-exit sleep with a bounded poll: wait until
container.json stops growing for one 40ms interval (1s cap). Faster when the
shim's log processor has already drained, safer when it lags.
---
 src/runtime/src/vm/mod.rs | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/src/runtime/src/vm/mod.rs b/src/runtime/src/vm/mod.rs
index 41de141..770449d 100644
--- a/src/runtime/src/vm/mod.rs
+++ b/src/runtime/src/vm/mod.rs
@@ -530,8 +530,25 @@ impl VmManager {
         };
 
         // Let the shim's log processor finish draining console.log into the json
-        // file (it flushes as the VM halts) before reading the captured output.
-        tokio::time::sleep(std::time::Duration::from_millis(150)).await;
+        // file (it flushes as the VM halts): poll until container.json stops
+        // growing for one interval (bounded at 1s) instead of a fixed sleep —
+        // fast when the drain is already done, safe when it lags.
+        let json_path = self
+            .home_dir
+            .join("boxes")
+            .join(&self.box_id)
+            .join("logs")
+            .join("container.json");
+        let drain_start = std::time::Instant::now();
+        let mut last_len = u64::MAX;
+        loop {
+            let len = std::fs::metadata(&json_path).map(|m| m.len()).unwrap_or(0);
+            if len == last_len || drain_start.elapsed() >= std::time::Duration::from_secs(1) {
+                break;
+            }
+            last_len = len;
+            tokio::time::sleep(std::time::Duration::from_millis(40)).await;
+        }
         let (stdout, stderr) = self.read_container_logs();
         Ok(a3s_box_core::exec::ExecOutput {
             stdout,