loop-benchmarking

commit 10d037e6ee747ffe4d2c8d725b8a8e46ebf5e596
parent 296bd543e30783c961ea89001e7d1df52e986345
Author: Brian Graham <brian@buildingbetterteams.de>
Date:   Sat,  4 Apr 2026 10:25:10 +0200

Auto-refresh OAuth token during sweeps

keep-auth-alive.sh runs a trivial non-bare claude command every 5
minutes, triggering OAuth token refresh in the credentials file.
The --bare experiment runs then read the fresh token.

run.py starts the keepalive automatically at sweep start and kills
it when the sweep finishes. No more token expiry mid-sweep.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Diffstat:
A
harness/lib/keep-auth-alive.sh
 | 
23
+++++++++++++++++++++++
M
harness/run.py
 | 
12
++++++++++++

2 files changed, 35 insertions(+), 0 deletions(-)
diff --git a/harness/lib/keep-auth-alive.sh b/harness/lib/keep-auth-alive.sh
@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Keep OAuth token fresh by running a trivial claude command every 5 minutes.
+# Run this in the background before starting a sweep:
+#   bash harness/lib/keep-auth-alive.sh &
+#   KEEP_AUTH_PID=$!
+#   python3 harness/run.py grid.yaml main_effects -j 2
+#   kill $KEEP_AUTH_PID
+
+INTERVAL=${1:-300}  # seconds, default 5 minutes
+
+echo "[keep-auth] Starting token refresh loop (every ${INTERVAL}s)"
+echo "[keep-auth] PID: $$"
+
+while true; do
+  sleep "$INTERVAL"
+  # Run a trivial non-bare command so claude refreshes the OAuth token
+  claude -p "reply with ok" --max-turns 1 --output-format json > /dev/null 2>&1
+  if [ $? -eq 0 ]; then
+    echo "[keep-auth] Token refreshed at $(date -u +%H:%M:%S)"
+  else
+    echo "[keep-auth] WARNING: refresh failed at $(date -u +%H:%M:%S)"
+  fi
+done
diff --git a/harness/run.py b/harness/run.py
@@ -672,6 +672,14 @@ def main():
     print(f"Total jobs:  {len(jobs)}")
     print()
 
+    # Start auth keepalive in background (refreshes OAuth token every 5 min)
+    auth_keepalive = subprocess.Popen(
+        ["bash", str(SCRIPT_DIR / "lib" / "keep-auth-alive.sh"), "300"],
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+    )
+    print(f"Auth keepalive started (PID {auth_keepalive.pid})")
+
     completed = 0
     skipped = 0
     failed = 0
@@ -714,6 +722,10 @@ def main():
                     total_done = completed + skipped + failed
                     log(f"  Progress: {total_done}/{len(jobs)} ({completed} completed, {skipped} skipped, {failed} failed)")
 
+    # Stop auth keepalive
+    auth_keepalive.terminate()
+    auth_keepalive.wait()
+
     print()
     print("=" * 40)
     print("All runs complete.")