Sibyl-Research-Team
diff --git a/‎sibyl/gpu_scheduler.py‎
Lines changed: 14 additions & 0 deletions b/‎sibyl/gpu_scheduler.py‎
Lines changed: 14 additions & 0 deletions
diff --git a/‎tests/test_gpu_scheduler.py‎
Lines changed: 61 additions & 0 deletions b/‎tests/test_gpu_scheduler.py‎
Lines changed: 61 additions & 0 deletions
@@ -147,15 +147,29 @@ def _lease_entry_matches_running(gpu_id: int, entry: dict) -> bool:
     return False
 
 
+_LEASE_TTL_SEC = 3600  # 1 hour — if lease is older AND workspace can't confirm, remove
+
+
 def _clean_global_gpu_leases_unlocked(leases: dict[str, dict]) -> dict[str, dict]:
     cleaned: dict[str, dict] = {}
+    now = time.time()
     for gpu_key, entry in leases.items():
         try:
             gpu_id = int(gpu_key)
         except (TypeError, ValueError):
             continue
+        # Fast path: if lease is recent, keep it without expensive workspace check
+        claimed_at = entry.get("claimed_at", 0)
+        if now - claimed_at < 60:  # Less than 1 minute old — always keep
+            cleaned[str(gpu_id)] = entry
+            continue
+        # Check if workspace still confirms this lease
         if _lease_entry_matches_running(gpu_id, entry):
             cleaned[str(gpu_id)] = entry
+        elif now - claimed_at < _LEASE_TTL_SEC:
+            # Workspace doesn't confirm but lease is recent — keep (may be in transition)
+            cleaned[str(gpu_id)] = entry
+        # else: workspace doesn't confirm AND lease is old → drop
     return cleaned
 
 
 
@@ -1183,6 +1183,67 @@ def test_script_includes_dispatch_logic(self):
         assert "DISPATCH" in script
 
 
+# ══════════════════════════════════════════════
+# TTL-based stale lease cleanup
+# ══════════════════════════════════════════════
+
+def test_stale_leases_cleaned_by_ttl(tmp_path, monkeypatch):
+    """Leases older than TTL should be cleaned even if workspace is gone."""
+    import time as time_mod
+    from sibyl import gpu_scheduler
+
+    monkeypatch.setattr(gpu_scheduler, "_global_gpu_leases_path",
+                        lambda: tmp_path / "gpu_leases.json")
+
+    old_lease = {
+        "0": {
+            "workspace_root": "/nonexistent/path",
+            "task_ids": ["old_task"],
+            "claimed_at": time_mod.time() - 7200,  # 2 hours ago
+        }
+    }
+    cleaned = gpu_scheduler._clean_global_gpu_leases_unlocked(old_lease)
+    assert "0" not in cleaned  # Should be removed — workspace gone + old
+
+
+def test_recent_leases_kept_without_workspace_check(tmp_path, monkeypatch):
+    """Leases less than 60s old should always be kept."""
+    import time as time_mod
+    from sibyl import gpu_scheduler
+
+    monkeypatch.setattr(gpu_scheduler, "_global_gpu_leases_path",
+                        lambda: tmp_path / "gpu_leases.json")
+
+    recent_lease = {
+        "0": {
+            "workspace_root": "/nonexistent/path",
+            "task_ids": ["new_task"],
+            "claimed_at": time_mod.time() - 10,  # 10 seconds ago
+        }
+    }
+    cleaned = gpu_scheduler._clean_global_gpu_leases_unlocked(recent_lease)
+    assert "0" in cleaned  # Should be kept — very recent
+
+
+def test_mid_age_lease_kept_when_workspace_gone(tmp_path, monkeypatch):
+    """Leases newer than TTL but older than 60s, with gone workspace, are kept."""
+    import time as time_mod
+    from sibyl import gpu_scheduler
+
+    monkeypatch.setattr(gpu_scheduler, "_global_gpu_leases_path",
+                        lambda: tmp_path / "gpu_leases.json")
+
+    mid_lease = {
+        "0": {
+            "workspace_root": "/nonexistent/path",
+            "task_ids": ["mid_task"],
+            "claimed_at": time_mod.time() - 600,  # 10 minutes ago (< 1hr TTL)
+        }
+    }
+    cleaned = gpu_scheduler._clean_global_gpu_leases_unlocked(mid_lease)
+    assert "0" in cleaned  # Should be kept — within TTL even though workspace is gone
+
+
 # ══════════════════════════════════════════════
 # Failed tasks don't block pipeline
 # ══════════════════════════════════════════════