Merge 486a7e33b6 into eb0e3f8c18

litellm/proxy/guardrails/guardrail_hooks/headroom/headroom.py

@@ -390,6 +390,24 @@ class HeadroomGuardrail(CustomGuardrail):
         value: Final = headers.get(BYPASS_HEADER)
         return str(value).lower() == "true"
 
+    def _record_nothing_to_compress(self, request_data: dict[str, object]) -> None:
+        """Log a zero-savings run when every message is protected.
+
+        Without an entry, a request headroom looked at and found nothing to
+        compress is indistinguishable in the spend logs and on the cost
+        optimization dashboard from one where headroom never ran at all.
+        """
+        now = time.time()
+        self.add_standard_logging_guardrail_information_to_request_data(
+            guardrail_json_response={"tokens_before": 0, "tokens_after": 0, "tokens_saved": 0},
+            request_data=request_data,
+            guardrail_status="success",
+            guardrail_provider=HEADROOM_GUARDRAIL_PROVIDER,
+            start_time=now,
+            end_time=now,
+            duration=0.0,
+        )
+
     def _request_headers(self) -> dict[str, str]:
         headers: Final[dict[str, str]] = {"Content-Type": "application/json"}
         if self.headroom_api_key:
@@ -643,6 +661,7 @@ class HeadroomGuardrail(CustomGuardrail):
         protected_indices: Final = _protected_indices(messages)
         compressible: Final = [m for i, m in enumerate(messages) if i not in protected_indices]
         if not compressible:
+            self._record_nothing_to_compress(request_data)
             return inputs
 
         model: Final = self.headroom_model or request_data.get("model")

tests/test_litellm/proxy/guardrails/guardrail_hooks/test_headroom.py

@@ -2021,6 +2021,31 @@ async def test_nothing_compressible_returns_inputs_untouched(guardrail: Headroom
     assert result is inputs
 
 
+@pytest.mark.asyncio
+async def test_nothing_compressible_is_logged_as_a_zero_savings_run(guardrail: HeadroomGuardrail):
+    """A single-turn request never reaches /v1/compress, so without an entry the
+    cost optimization dashboard cannot tell it apart from a request headroom was
+    never applied to; both read as no compression at all."""
+    request_data: dict = {"model": "gpt-4o"}
+    inputs = GenericGuardrailAPIInputs(
+        texts=["A" * 5000],
+        structured_messages=[
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "A" * 5000},
+        ],
+    )
+
+    with patch.object(guardrail.async_handler, "post", new_callable=AsyncMock):
+        await guardrail.apply_guardrail(inputs=inputs, request_data=request_data, input_type="request")
+
+    entries = _recorded_guardrail_entries(request_data)
+    assert len(entries) == 1
+    assert entries[0]["guardrail_provider"] == "headroom"
+    assert entries[0]["guardrail_status"] == "success"
+    assert entries[0]["guardrail_response"] == {"tokens_before": 0, "tokens_after": 0, "tokens_saved": 0}
+    assert extract_compression_saved_tokens({"guardrail_information": entries}) == 0
+
+
 @pytest.mark.asyncio
 async def test_fail_open_returns_the_caller_inputs_object():
     """Translation handlers detect a rewrite by object identity, so a request