diff --git a/orchestrator/orchestrator/builder_upgrade.py b/orchestrator/orchestrator/builder_upgrade.py
index b7a5d06..4405acb 100644
--- a/orchestrator/orchestrator/builder_upgrade.py
+++ b/orchestrator/orchestrator/builder_upgrade.py
@@ -159,7 +159,7 @@ def free_gb(h: BuilderHost) -> int:
 
 
 def reclaimable_gb(h: BuilderHost) -> int:
-    out = _sh(h, "sudo -n du -sg /opt/worker/cache /opt/worker/tasks /Users/cltbld/caches 2>/dev/null | awk '{s+=$1} END{print s+0}'", check=False)
+    out = _sh(h, "sudo -n du -sg /opt/worker/cache /opt/worker/tasks /Users/cltbld/caches 2>/dev/null | awk '{s+=$1} END{print s+0}'", check=False, timeout=600)
     return int(out or 0)
 
 
@@ -295,8 +295,10 @@ def step_clear_caches(h: BuilderHost) -> None:
     staging dir, and the Xcode the previous run moved aside."""
     ui.step("CLEAR CACHES", "worker cache + tasks, stale caches, leftover installer data")
     before = free_gb(h)
-    _sh(h, "sudo sh -c 'rm -rf /opt/worker/cache/* /opt/worker/tasks/* && echo {} > /opt/worker/directory-caches.json'")
-    _sh(h, "sudo rm -rf /Users/cltbld/caches '/System/Volumes/Data/macOS Install Data' /Applications/Xcode.app.pre-*", check=False)
+    # Deleting ~20G of hg stores outlasts ssh.run's 120s default (r8-138, 2026-10-22): the rm
+    # finished on the host but the tool had already given up.
+    _sh(h, "sudo sh -c 'rm -rf /opt/worker/cache/* /opt/worker/tasks/* && echo {} > /opt/worker/directory-caches.json'", timeout=1800)
+    _sh(h, "sudo rm -rf /Users/cltbld/caches '/System/Volumes/Data/macOS Install Data' /Applications/Xcode.app.pre-*", check=False, timeout=1800)
     ui.ok(f"{before}G -> {free_gb(h)}G free")
 
 
diff --git a/orchestrator/tests/test_builder_upgrade.py b/orchestrator/tests/test_builder_upgrade.py
index 88ad206..56e1d2a 100644
--- a/orchestrator/tests/test_builder_upgrade.py
+++ b/orchestrator/tests/test_builder_upgrade.py
@@ -188,6 +188,19 @@ def test_stop_worker_fails_if_task_user_comes_straight_back():
             bu.step_stop_worker(_host(), logout_wait=0)
 
 
+def test_clear_caches_allows_a_long_rm():
+    timeouts = {}
+
+    def run(_h, command, *, stdin=None, check=True, timeout=None, user=None):
+        timeouts[command] = timeout
+        return _cp("90")
+
+    with patch("orchestrator.builder_upgrade.ssh.run", run):
+        bu.step_clear_caches(_host())
+    rm = [t for c, t in timeouts.items() if "rm -rf" in c]
+    assert rm and all(t and t >= 1800 for t in rm)
+
+
 # --- os update ---