diff --git a/tests/test_tokens.py b/tests/test_tokens.py
index 133cbd2..bf2cdc0 100644
--- a/tests/test_tokens.py
+++ b/tests/test_tokens.py
@@ -115,6 +115,34 @@ def test_missing_usage_defaults_to_zero():
     assert c.prompt == 0 and c.completion == 0 and c.cache_read == 0 and c.cache_creation == 0
 
 
+def test_malformed_usage_values_default_to_zero():
+    """Malformed usage values should not break trace emission."""
+    msg = {
+        "type": "assistant",
+        "requestId": "r_bad_usage",
+        "timestamp": "2026-06-22T00:00:00Z",
+        "message": {
+            "id": "r_bad_usage",
+            "model": "claude-haiku",
+            "content": [{"type": "text", "text": "ok"}],
+            "usage": {
+                "input_tokens": "unknown",
+                "output_tokens": "NaN",
+                "cache_read_input_tokens": None,
+                "cache_creation_input_tokens": {"bad": "shape"},
+            },
+        },
+    }
+    calls = tokens.llm_calls([msg])
+    assert len(calls) == 1
+    c = calls[0]
+    assert c.prompt == 0
+    assert c.completion == 0
+    assert c.cache_read == 0
+    assert c.cache_creation == 0
+    assert c.text == "ok"
+
+
 def test_start_and_end_timestamps():
     """start_ts = first frame's timestamp, end_ts = last frame's timestamp."""
     msgs = [
diff --git a/traceroot_observability/tokens.py b/traceroot_observability/tokens.py
index a45136e..6233355 100644
--- a/traceroot_observability/tokens.py
+++ b/traceroot_observability/tokens.py
@@ -68,6 +68,14 @@ def _tool_calls(content) -> list:
     return out
 
 
+def _usage_int(value) -> int:
+    """Return a token count as int, defaulting malformed values to zero."""
+    try:
+        return int(value or 0)
+    except (TypeError, ValueError):
+        return 0
+
+
 def llm_calls(assistant_msgs: list[dict]) -> list[LlmCall]:
     """
     Collapse a flat list of assistant transcript messages into per-requestId LlmCall
@@ -118,16 +126,16 @@ def llm_calls(assistant_msgs: list[dict]) -> list[LlmCall]:
         u = msg.get("usage") or {}
         if rid not in seen_input:
             # First sighting: count input and cache tokens exactly once.
-            inp = int(u.get("input_tokens") or 0)
-            cr = int(u.get("cache_read_input_tokens") or 0)
-            cc = int(u.get("cache_creation_input_tokens") or 0)
+            inp = _usage_int(u.get("input_tokens"))
+            cr = _usage_int(u.get("cache_read_input_tokens"))
+            cc = _usage_int(u.get("cache_creation_input_tokens"))
             call.prompt += inp + cr + cc
             call.cache_read += cr
             call.cache_creation += cc
             seen_input.add(rid)
 
         # Output: add only the delta over the running max for this requestId.
-        out = int(u.get("output_tokens") or 0)
+        out = _usage_int(u.get("output_tokens"))
         prior = out_max.get(rid, 0)
         if out > prior:
             call.completion += out - prior