From dfcea2c1866630313ec3083794a922ff6971583a Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 28 Aug 2026 17:16:00 -0700
Subject: [PATCH 01/25] fix(policy_engine): execute post_call guardrail
 pipelines on responses

---
 litellm/proxy/utils.py                        |  47 +++++-
 .../proxy_logging/test_guardrail_pipeline.py  | 152 +++++++++++++++++-
 2 files changed, 196 insertions(+), 3 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index d880b529727..bd8ba6ae9f0 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -455,6 +455,30 @@ def _pipeline_managed_guardrail_names(data: Mapping[str, object]) -> frozenset[s
     )
 
 
+def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object]) -> None:
+    if data.get("stream") is not True:
+        return
+    post_call_policies: Final = tuple(
+        policy_name for policy_name, pipeline in _policy_pipelines(data) if pipeline.mode == "post_call"
+    )
+    if not post_call_policies:
+        return
+    raise HTTPException(
+        status_code=400,
+        detail={
+            "error": {
+                "message": (
+                    "Policies with post_call guardrail pipelines cannot govern streaming responses yet: "
+                    f"{', '.join(post_call_policies)}. Retry with stream=false, or move these policies' output "
+                    "guardrails from pipeline steps to guardrails.add, which scans streamed output."
+                ),
+                "type": "guardrail_pipeline_error",
+                "policies": list(post_call_policies),
+            }
+        },
+    )
+
+
 def _prompt_block_text(block: object) -> str:
     if isinstance(block, str):
         return block
@@ -1578,6 +1602,7 @@ class ProxyLogging:
         call_type: str,
         event_hook: str,
         raw_request_snapshot: dict | None = None,  # mutable-ok: same request-payload shape as data
+        response: LLMResponseTypes | None = None,
     ) -> dict:
         """
         Execute guardrail pipelines if any are configured for this request.
@@ -1596,6 +1621,8 @@ class ProxyLogging:
         if not pipelines:
             return data
 
+        step_input: Final = {**data, "response": response} if response is not None else data
+
         for policy_name, pipeline in pipelines:
             if pipeline.mode != event_hook:
                 continue
@@ -1603,7 +1630,7 @@ class ProxyLogging:
             result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
                 steps=pipeline.steps,
                 mode=pipeline.mode,
-                data=data,
+                data=step_input,
                 user_api_key_dict=user_api_key_dict,
                 call_type=call_type,
                 policy_name=policy_name,
@@ -1614,6 +1641,7 @@ class ProxyLogging:
                 result=result,
                 data=data,
                 policy_name=policy_name,
+                original_response=response,
             )
 
         return data
@@ -1623,14 +1651,18 @@ class ProxyLogging:
         result: PipelineExecutionResult,
         data: dict,
         policy_name: str,
+        original_response: LLMResponseTypes | None = None,
     ) -> dict:
         """
         Handle a PipelineExecutionResult — allow, block, or modify_response.
 
         Returns data dict if allowed, raises on block/modify_response.
+        ``original_response`` is set on the post_call path, where allowed
+        modifications land on the response object in place, so the request
+        payload (already sent upstream) is left untouched.
         """
         if result.terminal_action == "allow":
-            if result.modified_data is not None:
+            if result.modified_data is not None and original_response is None:
                 data.update(result.modified_data)
             return data
 
@@ -1671,6 +1703,7 @@ class ProxyLogging:
                 request_data=data,
                 guardrail_name=f"pipeline:{policy_name}",
                 detection_info=None,
+                original_response=original_response,
             )
 
         return data
@@ -1786,6 +1819,8 @@ class ProxyLogging:
         )
 
         try:
+            _raise_for_streaming_post_call_pipelines(data)
+
             # Execute guardrail pipelines before the normal callback loop
             data = await self._maybe_execute_pipelines(
                 data=data,
@@ -2774,6 +2809,14 @@ class ProxyLogging:
         from litellm.proxy.proxy_server import llm_router
         from litellm.types.guardrails import GuardrailEventHooks
 
+        await self._maybe_execute_pipelines(
+            data=data,
+            user_api_key_dict=user_api_key_dict,
+            call_type=getattr(data.get("litellm_logging_obj"), "call_type", None) or "acompletion",
+            event_hook="post_call",
+            response=response,
+        )
+
         guardrail_callbacks: Final[list[CustomGuardrail]] = []
         other_callbacks: Final[list[CustomLogger]] = []
         try:
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index e99e34d65d4..b7e86b68fe4 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -23,7 +23,7 @@ from litellm.integrations.custom_guardrail import (
     ModifyResponseException,
 )
 from litellm.integrations.prometheus import PrometheusLogger
-from litellm.proxy.utils import ProxyLogging
+from litellm.proxy.utils import ProxyLogging, _raise_for_streaming_post_call_pipelines
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.proxy.policy_engine.pipeline_types import (
     GuardrailPipeline,
@@ -865,3 +865,153 @@ async def test_process_prompt_template_aresponses_swaps_model_and_merges_input(p
     hook_kwargs = logging_obj.async_get_chat_completion_prompt.await_args.kwargs
     assert hook_kwargs["messages"] == [{"role": "user", "content": "Who are you?"}]
     assert hook_kwargs["prompt_spec"] is prompt_spec
+
+
+# ---------------------------------------------------------------------------
+# post_call pipeline execution (LIT-6410)
+# ---------------------------------------------------------------------------
+
+
+def _post_call_pipeline_data(guardrail: str = "gr-post", **extra: Any) -> Dict[str, Any]:
+    pipeline = GuardrailPipeline(
+        mode="post_call",
+        steps=[PipelineStep(guardrail=guardrail, on_pass="allow", on_fail="block")],
+    )
+    return {
+        "model": "m",
+        "messages": [{"role": "user", "content": "hi"}],
+        "metadata": {
+            "_guardrail_pipelines": [("response-governance", pipeline)],
+            "_pipeline_managed_guardrails": {guardrail},
+        },
+        **extra,
+    }
+
+
+@pytest.mark.asyncio
+async def test_post_call_success_hook_runs_post_call_pipeline_and_reraises_block(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+
+    class OutputBlockingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["response"] = response
+            raise HTTPException(status_code=400, detail={"error": "output blocked"})
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [OutputBlockingGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+    response = litellm.ModelResponse()
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.post_call_success_hook(
+            data=data, response=response, user_api_key_dict=make_user_api_key_auth()
+        )
+
+    assert info.value.detail["error"] == "output blocked"
+    assert seen["response"] is response
+
+
+@pytest.mark.asyncio
+async def test_post_call_pipeline_pass_runs_once_and_leaves_request_data_untouched(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class RecordingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["count"] += 1
+            seen["response"] = response
+            return None
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [RecordingGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+    response = litellm.ModelResponse()
+
+    out = await proxy_logging.post_call_success_hook(
+        data=data, response=response, user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert out is response
+    assert seen["response"] is response
+    assert seen["count"] == 1
+    assert "response" not in data
+    assert "guardrails" not in data["metadata"]
+
+
+def test_handle_pipeline_result_modify_response_carries_original_response():
+    result = MagicMock()
+    result.terminal_action = "modify_response"
+    result.modify_response_message = "filtered"
+    response = litellm.ModelResponse()
+
+    with pytest.raises(ModifyResponseException) as info:
+        ProxyLogging._handle_pipeline_result(
+            result=result, data={"model": "m"}, policy_name="p", original_response=response
+        )
+
+    assert info.value.original_response is response
+
+
+def test_handle_pipeline_result_allow_discards_modifications_on_post_call():
+    data = {"a": 1, "metadata": {"guardrails": ["other"]}}
+    result = MagicMock()
+    result.terminal_action = "allow"
+    result.modified_data = {"metadata": {"guardrails": ["gr-post"]}, "response": object()}
+
+    out = ProxyLogging._handle_pipeline_result(
+        result=result, data=data, policy_name="p", original_response=litellm.ModelResponse()
+    )
+
+    assert out is data
+    assert data == {"a": 1, "metadata": {"guardrails": ["other"]}}
+
+
+@pytest.mark.asyncio
+async def test_pre_call_hook_rejects_streaming_request_with_post_call_pipeline(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    monkeypatch.setattr(litellm, "callbacks", [])
+    data = _post_call_pipeline_data(stream=True)
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=make_user_api_key_auth(),
+            data=data,
+            call_type="completion",
+            guardrails_only=True,
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["policies"] == ["response-governance"]
+    assert "stream=false" in info.value.detail["error"]["message"]
+
+
+def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_call():
+    post_call = GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
+    pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
+
+    assert (
+        _raise_for_streaming_post_call_pipelines(
+            {"stream": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}
+        )
+        is None
+    )
+    assert _raise_for_streaming_post_call_pipelines({"metadata": {"_guardrail_pipelines": [("p", post_call)]}}) is None
+    assert (
+        _raise_for_streaming_post_call_pipelines(
+            {"stream": True, "metadata": {"_guardrail_pipelines": [("p", pre_call)]}}
+        )
+        is None
+    )
+    assert _raise_for_streaming_post_call_pipelines({"stream": True}) is None

From e6edd62f5d0f010d34c203d9df8192462a1622c0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 28 Aug 2026 17:40:48 -0700
Subject: [PATCH 02/25] fix(policy_engine): propagate post_call pipeline
 replacement responses to the client

---
 .../proxy/policy_engine/pipeline_executor.py  | 19 +++--
 litellm/proxy/utils.py                        | 32 +++++---
 .../proxy_logging/test_guardrail_pipeline.py  | 81 ++++++++++++++++++-
 .../utils/proxy_logging/test_pre_call_hook.py |  2 +-
 4 files changed, 112 insertions(+), 22 deletions(-)

diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index a5619821197..190784a2a60 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -108,8 +108,10 @@ class PipelineExecutor:
                 action,
             )
 
-            # Forward modified data to next step if pass_data is True
-            if step.pass_data and modified_data is not None:
+            # Forward modified data to the next step if pass_data is True;
+            # post_call response replacements always chain, matching the flat
+            # callback loop where each hook sees the previous hook's response
+            if modified_data is not None and (step.pass_data or mode == "post_call"):
                 working_data = {**working_data, **modified_data}
 
             # Handle terminal actions
@@ -227,11 +229,14 @@ class PipelineExecutor:
             # same contract as run_in_parallel/scan_raw_request elsewhere: any
             # data it returned is discarded, since applying it on top of the
             # raw snapshot would silently undo whatever an earlier step in
-            # this pipeline already did.
-            modified_data = None
-            if response is not None and isinstance(response, dict) and not scans_raw_request:
-                modified_data = response
-            return ("pass", modified_data, None, None)
+            # this pipeline already did. A post_call hook's non-None return is
+            # a replacement response (the flat callback-loop contract), carried
+            # under the same "response" key the step input uses.
+            if response is None or scans_raw_request:
+                return ("pass", None, None, None)
+            if mode == "post_call":
+                return ("pass", {"response": response}, None, None)
+            return ("pass", response if isinstance(response, dict) else None, None, None)
 
         except Exception as e:
             if CustomGuardrail._is_guardrail_intervention(e):
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index bd8ba6ae9f0..25c1068cc6d 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -1603,7 +1603,7 @@ class ProxyLogging:
         event_hook: str,
         raw_request_snapshot: dict | None = None,  # mutable-ok: same request-payload shape as data
         response: LLMResponseTypes | None = None,
-    ) -> dict:
+    ) -> tuple[dict, LLMResponseTypes | None]:
         """
         Execute guardrail pipelines if any are configured for this request.
 
@@ -1615,18 +1615,21 @@ class ProxyLogging:
         ``scan_raw_request`` evaluates the pristine request, not whatever an
         earlier ``pass_data`` step in the same pipeline already rewrote.
 
-        Returns the (possibly modified) data dict.
+        Returns the (possibly modified) data dict, plus the replacement
+        response when a post_call pipeline step returned one (None when the
+        response is unchanged), matching the flat callback-loop contract.
         """
         pipelines: Final = _policy_pipelines(data)
         if not pipelines:
-            return data
-
-        step_input: Final = {**data, "response": response} if response is not None else data
+            return data, None
 
+        current_response = response  # rebind-ok: chains each pipeline's replacement response into the next
         for policy_name, pipeline in pipelines:
             if pipeline.mode != event_hook:
                 continue
 
+            step_input: dict = {**data, "response": current_response} if current_response is not None else data
+
             result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
                 steps=pipeline.steps,
                 mode=pipeline.mode,
@@ -1641,10 +1644,13 @@ class ProxyLogging:
                 result=result,
                 data=data,
                 policy_name=policy_name,
-                original_response=response,
+                original_response=current_response,
             )
 
-        return data
+            if current_response is not None and result.modified_data is not None:
+                current_response = result.modified_data.get("response", current_response)
+
+        return data, current_response if current_response is not response else None
 
     @staticmethod
     def _handle_pipeline_result(
@@ -1657,9 +1663,9 @@ class ProxyLogging:
         Handle a PipelineExecutionResult — allow, block, or modify_response.
 
         Returns data dict if allowed, raises on block/modify_response.
-        ``original_response`` is set on the post_call path, where allowed
-        modifications land on the response object in place, so the request
-        payload (already sent upstream) is left untouched.
+        ``original_response`` is set on the post_call path, where the request
+        payload (already sent upstream) must stay untouched; a replacement
+        response carried in ``modified_data`` is adopted by the caller.
         """
         if result.terminal_action == "allow":
             if result.modified_data is not None and original_response is None:
@@ -1822,7 +1828,7 @@ class ProxyLogging:
             _raise_for_streaming_post_call_pipelines(data)
 
             # Execute guardrail pipelines before the normal callback loop
-            data = await self._maybe_execute_pipelines(
+            data, _ = await self._maybe_execute_pipelines(
                 data=data,
                 user_api_key_dict=user_api_key_dict,
                 call_type=call_type,
@@ -2809,13 +2815,15 @@ class ProxyLogging:
         from litellm.proxy.proxy_server import llm_router
         from litellm.types.guardrails import GuardrailEventHooks
 
-        await self._maybe_execute_pipelines(
+        _, pipeline_response = await self._maybe_execute_pipelines(
             data=data,
             user_api_key_dict=user_api_key_dict,
             call_type=getattr(data.get("litellm_logging_obj"), "call_type", None) or "acompletion",
             event_hook="post_call",
             response=response,
         )
+        if pipeline_response is not None:
+            response = pipeline_response  # rebind-ok: adopt the pipeline's replacement response, same contract as the callback loops below
 
         guardrail_callbacks: Final[list[CustomGuardrail]] = []
         other_callbacks: Final[list[CustomLogger]] = []
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index b7e86b68fe4..8bc71f6e178 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -326,13 +326,14 @@ def test_process_guardrail_metadata_invalid_data_raises(proxy_logging):
 @pytest.mark.asyncio
 async def test_maybe_execute_pipelines_no_pipelines_returns_data(proxy_logging, make_user_api_key_auth):
     data = {"messages": [{"role": "user"}], "model": "m", "temperature": 0.1}
-    out = await proxy_logging._maybe_execute_pipelines(
+    out, replacement = await proxy_logging._maybe_execute_pipelines(
         data=data,
         user_api_key_dict=make_user_api_key_auth(),
         call_type="completion",
         event_hook="pre_call",
     )
     assert out == {"messages": [{"role": "user"}], "model": "m", "temperature": 0.1}
+    assert replacement is None
 
 
 @pytest.mark.asyncio
@@ -344,7 +345,7 @@ async def test_maybe_execute_pipelines_skips_pipelines_with_other_mode(proxy_log
     monkeypatch.setattr(
         "litellm.proxy.policy_engine.pipeline_executor.PipelineExecutor.execute_steps", executed
     )
-    out = await proxy_logging._maybe_execute_pipelines(
+    out, replacement = await proxy_logging._maybe_execute_pipelines(
         data=data,
         user_api_key_dict=make_user_api_key_auth(),
         call_type="completion",
@@ -352,6 +353,7 @@ async def test_maybe_execute_pipelines_skips_pipelines_with_other_mode(proxy_log
     )
     executed.assert_not_called()
     assert out is data
+    assert replacement is None
 
 
 @pytest.mark.parametrize(
@@ -949,6 +951,81 @@ async def test_post_call_pipeline_pass_runs_once_and_leaves_request_data_untouch
     assert "guardrails" not in data["metadata"]
 
 
+@pytest.mark.asyncio
+async def test_post_call_pipeline_replacement_response_reaches_caller(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    masked = litellm.ModelResponse()
+
+    class MaskingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            return masked
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [MaskingGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+
+    out = await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert out is masked
+    assert "response" not in data
+
+
+@pytest.mark.asyncio
+async def test_post_call_pipeline_replacement_chains_to_next_step_without_pass_data(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    masked = litellm.ModelResponse()
+    seen: Dict[str, Any] = {}
+
+    class MaskingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            return masked
+
+    class RecordingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["response"] = response
+            return None
+
+    pipeline = GuardrailPipeline(
+        mode="post_call",
+        steps=[
+            PipelineStep(guardrail="gr-mask", on_pass="next", on_fail="block"),
+            PipelineStep(guardrail="gr-audit", on_pass="allow", on_fail="block"),
+        ],
+    )
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [
+            MaskingGuardrail(guardrail_name="gr-mask", event_hook=GuardrailEventHooks.post_call, default_on=False),
+            RecordingGuardrail(guardrail_name="gr-audit", event_hook=GuardrailEventHooks.post_call, default_on=False),
+        ],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = {
+        "model": "m",
+        "messages": [{"role": "user", "content": "hi"}],
+        "metadata": {
+            "_guardrail_pipelines": [("response-governance", pipeline)],
+            "_pipeline_managed_guardrails": {"gr-mask", "gr-audit"},
+        },
+    }
+
+    out = await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert out is masked
+    assert seen["response"] is masked
+
+
 def test_handle_pipeline_result_modify_response_carries_original_response():
     result = MagicMock()
     result.terminal_action = "modify_response"
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_pre_call_hook.py b/tests/test_litellm/proxy/utils/proxy_logging/test_pre_call_hook.py
index 0971ce09d79..06cf328a20c 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_pre_call_hook.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_pre_call_hook.py
@@ -660,7 +660,7 @@ async def test_scan_raw_request_snapshot_taken_before_pipelines(
         for msg in data.get("messages", []):
             if "SECRET" in msg.get("content", ""):
                 msg["content"] = msg["content"].replace("SECRET", "[REDACTED]")
-        return data
+        return data, None
 
     monkeypatch.setattr(ProxyLogging, "_maybe_execute_pipelines", fake_pipelines)
     monkeypatch.setattr(litellm, "callbacks", [_BlockOnSecretGuardrail(scan_raw_request=True)])

From aeac6a412c98c07cce64c5ddbd74d005f2e60e7e Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 28 Aug 2026 17:59:15 -0700
Subject: [PATCH 03/25] fix(policy_engine): skip pipeline-managed guardrails in
 the response-path guardrail loop

---
 litellm/proxy/utils.py                        |  9 ++++++-
 .../proxy_logging/test_guardrail_pipeline.py  | 26 +++++++++++++++++++
 2 files changed, 34 insertions(+), 1 deletion(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 25c1068cc6d..75cc5c259a7 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -2825,6 +2825,7 @@ class ProxyLogging:
         if pipeline_response is not None:
             response = pipeline_response  # rebind-ok: adopt the pipeline's replacement response, same contract as the callback loops below
 
+        pipeline_managed: Final = _pipeline_managed_guardrail_names(data)
         guardrail_callbacks: Final[list[CustomGuardrail]] = []
         other_callbacks: Final[list[CustomLogger]] = []
         try:
@@ -2849,12 +2850,18 @@ class ProxyLogging:
             guardrail_data: Final = _check_and_merge_model_level_guardrails(data=data, llm_router=llm_router)
 
             parallel_guardrails: Final[tuple[CustomGuardrail, ...]] = tuple(
-                callback for callback in guardrail_callbacks if getattr(callback, "run_in_parallel", False)
+                callback
+                for callback in guardrail_callbacks
+                if getattr(callback, "run_in_parallel", False)
+                and not (callback.guardrail_name and callback.guardrail_name in pipeline_managed)
             )
 
             for callback in guardrail_callbacks:
                 # Main - V2 Guardrails implementation
 
+                if callback.guardrail_name and callback.guardrail_name in pipeline_managed:
+                    continue
+
                 if getattr(callback, "run_in_parallel", False):
                     continue
 
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 8bc71f6e178..c92f5fa6c55 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -951,6 +951,32 @@ async def test_post_call_pipeline_pass_runs_once_and_leaves_request_data_untouch
     assert "guardrails" not in data["metadata"]
 
 
+@pytest.mark.asyncio
+async def test_post_call_pipeline_managed_default_on_guardrail_runs_exactly_once(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class CountingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["count"] += 1
+            return None
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [CountingGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=True)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+
+    await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert seen["count"] == 1
+
+
 @pytest.mark.asyncio
 async def test_post_call_pipeline_replacement_response_reaches_caller(
     proxy_logging, make_user_api_key_auth, monkeypatch

From 55569729b05d601c139e43b8faba447983e89f74 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 28 Aug 2026 18:16:35 -0700
Subject: [PATCH 04/25] fix(policy_engine): scope pipeline-managed guardrail
 skips to the pipeline's mode

---
 litellm/proxy/utils.py                        | 20 ++---
 .../proxy_logging/test_guardrail_pipeline.py  | 73 +++++++++++++++++++
 2 files changed, 84 insertions(+), 9 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 75cc5c259a7..d3f2e1d7301 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -11,7 +11,7 @@ import sys
 import threading
 import time
 import traceback
-from collections.abc import AsyncGenerator, Awaitable, Callable, Collection, Coroutine, Mapping, Sequence
+from collections.abc import AsyncGenerator, Awaitable, Callable, Coroutine, Mapping, Sequence
 from dataclasses import dataclass, field
 from datetime import date, datetime, timedelta, timezone
 from email.mime.multipart import MIMEMultipart
@@ -446,12 +446,14 @@ def _policy_pipelines(data: Mapping[str, object]) -> tuple[tuple[str, "Guardrail
     )
 
 
-def _pipeline_managed_guardrail_names(data: Mapping[str, object]) -> frozenset[str]:
-    managed: Final = _policy_state_metadata(data).get("_pipeline_managed_guardrails")
-    return (
-        frozenset(cast("Collection[str]", managed))  # cast-ok: the policy engine wrote these guardrail names
-        if managed
-        else frozenset()
+def _pipeline_managed_guardrail_names(
+    data: Mapping[str, object], mode: Literal["pre_call", "post_call"]
+) -> frozenset[str]:
+    return frozenset(
+        step.guardrail
+        for _policy_name, pipeline in _policy_pipelines(data)
+        if pipeline.mode == mode
+        for step in pipeline.steps
     )
 
 
@@ -1837,7 +1839,7 @@ class ProxyLogging:
             )
 
             # Get pipeline-managed guardrails to skip in normal loop
-            pipeline_managed: Final = _pipeline_managed_guardrail_names(data)
+            pipeline_managed: Final = _pipeline_managed_guardrail_names(data, "pre_call")
 
             caps: Final = ProxyLogging._callback_capabilities()
             # Skip the per-request callback walk entirely when nothing in
@@ -2825,7 +2827,7 @@ class ProxyLogging:
         if pipeline_response is not None:
             response = pipeline_response  # rebind-ok: adopt the pipeline's replacement response, same contract as the callback loops below
 
-        pipeline_managed: Final = _pipeline_managed_guardrail_names(data)
+        pipeline_managed: Final = _pipeline_managed_guardrail_names(data, "post_call")
         guardrail_callbacks: Final[list[CustomGuardrail]] = []
         other_callbacks: Final[list[CustomLogger]] = []
         try:
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index c92f5fa6c55..4e1ccf71c5c 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -977,6 +977,79 @@ async def test_post_call_pipeline_managed_default_on_guardrail_runs_exactly_once
     assert seen["count"] == 1
 
 
+@pytest.mark.asyncio
+async def test_post_call_hook_still_runs_guardrail_managed_only_by_pre_call_pipeline(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class DualStageGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["count"] += 1
+            return None
+
+    pre_call_pipeline = GuardrailPipeline(
+        mode="pre_call",
+        steps=[PipelineStep(guardrail="gr-dual", on_pass="allow", on_fail="block")],
+    )
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [DualStageGuardrail(guardrail_name="gr-dual", event_hook=["pre_call", "post_call"], default_on=True)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = {
+        "model": "m",
+        "messages": [{"role": "user", "content": "hi"}],
+        "metadata": {
+            "_guardrail_pipelines": [("request-governance", pre_call_pipeline)],
+            "_pipeline_managed_guardrails": {"gr-dual"},
+        },
+    }
+
+    await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert seen["count"] == 1
+
+
+@pytest.mark.asyncio
+async def test_pre_call_hook_still_runs_guardrail_managed_only_by_post_call_pipeline(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class DualStageGuardrail(CustomGuardrail):
+        async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type):
+            seen["count"] += 1
+            return data
+
+    post_call_pipeline = GuardrailPipeline(
+        mode="post_call",
+        steps=[PipelineStep(guardrail="gr-dual", on_pass="allow", on_fail="block")],
+    )
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [DualStageGuardrail(guardrail_name="gr-dual", event_hook=["pre_call", "post_call"], default_on=True)],
+    )
+    data = {
+        "model": "m",
+        "messages": [{"role": "user", "content": "hi"}],
+        "metadata": {
+            "_guardrail_pipelines": [("response-governance", post_call_pipeline)],
+            "_pipeline_managed_guardrails": {"gr-dual"},
+        },
+    }
+
+    await proxy_logging.pre_call_hook(
+        user_api_key_dict=make_user_api_key_auth(), data=data, call_type="completion"
+    )
+
+    assert seen["count"] == 1
+
+
 @pytest.mark.asyncio
 async def test_post_call_pipeline_replacement_response_reaches_caller(
     proxy_logging, make_user_api_key_auth, monkeypatch

From 996019cd23423c7b2a35dbd50f4b3b3571362cd3 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 00:14:23 -0700
Subject: [PATCH 05/25] fix(policy_engine): keep post_call pipeline guardrail
 logging and reject background bypass

Post_call pipelines run step hooks against a copied request dict, so guardrail
writes into the metadata bucket (applied_guardrails for the response header,
standard_logging_guardrail_information for spend logs) were dropped when the
guardrail was the first writer. Merge those writes back onto the request on the
post_call allow path, keeping the request payload and the executor's per-step
guardrails activation flag out of it.

Background /v1/responses requests dodge the streaming 400: pre_call sees stream
unset, then the polling task forces stream=true with pre-call logic skipped and
the streaming branch returns before post_call_success_hook, silently bypassing
post_call pipelines. Reject background=true at pre_call the same way as
stream=true.

Also pin the run_in_parallel pipeline-managed exclusion in both hook loops with
regression tests.
---
 litellm/proxy/utils.py                        |  49 +++++-
 .../proxy_logging/test_guardrail_pipeline.py  | 148 +++++++++++++++++-
 2 files changed, 187 insertions(+), 10 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index d3f2e1d7301..0f97473312c 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -457,8 +457,37 @@ def _pipeline_managed_guardrail_names(
     )
 
 
+def _merge_pipeline_metadata_bucket(data: dict, bucket_key: str, modified_bucket_value: object) -> None:
+    if not isinstance(modified_bucket_value, dict):
+        return
+    modified_bucket: Final = cast("dict[str, object]", modified_bucket_value)  # cast-ok: metadata buckets are str-keyed
+    surviving_writes: Final = {key: value for key, value in modified_bucket.items() if key != "guardrails"}
+    existing_bucket: Final = data.get(bucket_key)
+    if isinstance(existing_bucket, dict):
+        cast("dict[str, object]", existing_bucket).update(surviving_writes)  # cast-ok: metadata buckets are str-keyed
+    else:
+        data[bucket_key] = surviving_writes
+
+
+def _merge_pipeline_metadata_writes(data: dict, modified_data: Mapping[str, object]) -> None:
+    """
+    Copy metadata-bucket writes from a pipeline's working copy back onto the request.
+
+    Post_call pipelines run step hooks against a copied request dict so the payload
+    already sent upstream stays untouched, but hooks record proxy-internal logging
+    state in the metadata buckets (``applied_guardrails`` for response headers,
+    ``standard_logging_guardrail_information`` for spend logs), and those writes
+    must reach the request dict the proxy keeps reading after the pipeline returns.
+
+    The ``guardrails`` key is the executor's per-step activation flag for
+    ``should_run_guardrail``, not a hook write, so it stays in the working copy.
+    """
+    for bucket_key in ("metadata", "litellm_metadata"):
+        _merge_pipeline_metadata_bucket(data, bucket_key, modified_data.get(bucket_key))
+
+
 def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object]) -> None:
-    if data.get("stream") is not True:
+    if data.get("stream") is not True and data.get("background") is not True:
         return
     post_call_policies: Final = tuple(
         policy_name for policy_name, pipeline in _policy_pipelines(data) if pipeline.mode == "post_call"
@@ -470,9 +499,10 @@ def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object]) -> None
         detail={
             "error": {
                 "message": (
-                    "Policies with post_call guardrail pipelines cannot govern streaming responses yet: "
-                    f"{', '.join(post_call_policies)}. Retry with stream=false, or move these policies' output "
-                    "guardrails from pipeline steps to guardrails.add, which scans streamed output."
+                    "Policies with post_call guardrail pipelines cannot govern streaming or background "
+                    f"responses yet: {', '.join(post_call_policies)}. Retry with stream=false and "
+                    "background=false, or move these policies' output guardrails from pipeline steps to "
+                    "guardrails.add, which scans streamed output."
                 ),
                 "type": "guardrail_pipeline_error",
                 "policies": list(post_call_policies),
@@ -1667,11 +1697,16 @@ class ProxyLogging:
         Returns data dict if allowed, raises on block/modify_response.
         ``original_response`` is set on the post_call path, where the request
         payload (already sent upstream) must stay untouched; a replacement
-        response carried in ``modified_data`` is adopted by the caller.
+        response carried in ``modified_data`` is adopted by the caller, and
+        metadata-bucket writes (applied guardrails, guardrail logging info)
+        are merged back so headers and spend logs still see them.
         """
         if result.terminal_action == "allow":
-            if result.modified_data is not None and original_response is None:
-                data.update(result.modified_data)
+            if result.modified_data is not None:
+                if original_response is None:
+                    data.update(result.modified_data)
+                else:
+                    _merge_pipeline_metadata_writes(data, result.modified_data)
             return data
 
         if result.terminal_action == "block":
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 4e1ccf71c5c..34a25d75bc0 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -23,6 +23,7 @@ from litellm.integrations.custom_guardrail import (
     ModifyResponseException,
 )
 from litellm.integrations.prometheus import PrometheusLogger
+from litellm.proxy.common_utils.callback_utils import add_guardrail_to_applied_guardrails_header
 from litellm.proxy.utils import ProxyLogging, _raise_for_streaming_post_call_pipelines
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.proxy.policy_engine.pipeline_types import (
@@ -1139,18 +1140,132 @@ def test_handle_pipeline_result_modify_response_carries_original_response():
     assert info.value.original_response is response
 
 
-def test_handle_pipeline_result_allow_discards_modifications_on_post_call():
+def test_handle_pipeline_result_allow_on_post_call_keeps_metadata_writes_only():
     data = {"a": 1, "metadata": {"guardrails": ["other"]}}
     result = MagicMock()
     result.terminal_action = "allow"
-    result.modified_data = {"metadata": {"guardrails": ["gr-post"]}, "response": object()}
+    result.modified_data = {
+        "a": 2,
+        "metadata": {"guardrails": ["other"], "applied_guardrails": ["gr-post"]},
+        "response": object(),
+    }
 
     out = ProxyLogging._handle_pipeline_result(
         result=result, data=data, policy_name="p", original_response=litellm.ModelResponse()
     )
 
     assert out is data
-    assert data == {"a": 1, "metadata": {"guardrails": ["other"]}}
+    assert data["a"] == 1
+    assert "response" not in data
+    assert data["metadata"] == {"guardrails": ["other"], "applied_guardrails": ["gr-post"]}
+
+
+@pytest.mark.asyncio
+async def test_post_call_pipeline_guardrail_metadata_writes_reach_request_data(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    class HeaderWritingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            add_guardrail_to_applied_guardrails_header(request_data=data, guardrail_name="gr-post")
+            self.add_standard_logging_guardrail_information_to_request_data(
+                guardrail_json_response={"verdict": "pass"},
+                request_data=data,
+                guardrail_status="success",
+            )
+            return None
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [HeaderWritingGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+
+    await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert data["metadata"]["applied_guardrails"] == ["gr-post"]
+    slg_entries = data["metadata"]["standard_logging_guardrail_information"]
+    assert len(slg_entries) == 1
+    assert slg_entries[0]["guardrail_name"] == "gr-post"
+
+
+@pytest.mark.asyncio
+async def test_post_call_pipeline_managed_parallel_guardrail_runs_exactly_once(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class CountingGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            seen["count"] += 1
+            return None
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [
+            CountingGuardrail(
+                guardrail_name="gr-post",
+                event_hook=GuardrailEventHooks.post_call,
+                default_on=True,
+                run_in_parallel=True,
+            )
+        ],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+
+    await proxy_logging.post_call_success_hook(
+        data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+    )
+
+    assert seen["count"] == 1
+
+
+@pytest.mark.asyncio
+async def test_pre_call_pipeline_managed_parallel_guardrail_runs_exactly_once(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {"count": 0}
+
+    class CountingGuardrail(CustomGuardrail):
+        async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type):
+            seen["count"] += 1
+            return data
+
+    pre_call_pipeline = GuardrailPipeline(
+        mode="pre_call",
+        steps=[PipelineStep(guardrail="gr-pre", on_pass="allow", on_fail="block")],
+    )
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [
+            CountingGuardrail(
+                guardrail_name="gr-pre",
+                event_hook=GuardrailEventHooks.pre_call,
+                default_on=True,
+                run_in_parallel=True,
+            )
+        ],
+    )
+    data = {
+        "model": "m",
+        "messages": [{"role": "user", "content": "hi"}],
+        "metadata": {
+            "_guardrail_pipelines": [("request-governance", pre_call_pipeline)],
+            "_pipeline_managed_guardrails": {"gr-pre"},
+        },
+    }
+
+    await proxy_logging.pre_call_hook(
+        user_api_key_dict=make_user_api_key_auth(), data=data, call_type="completion"
+    )
+
+    assert seen["count"] == 1
 
 
 @pytest.mark.asyncio
@@ -1173,6 +1288,26 @@ async def test_pre_call_hook_rejects_streaming_request_with_post_call_pipeline(
     assert "stream=false" in info.value.detail["error"]["message"]
 
 
+@pytest.mark.asyncio
+async def test_pre_call_hook_rejects_background_request_with_post_call_pipeline(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    monkeypatch.setattr(litellm, "callbacks", [])
+    data = _post_call_pipeline_data(background=True)
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=make_user_api_key_auth(),
+            data=data,
+            call_type="aresponses",
+            guardrails_only=True,
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["policies"] == ["response-governance"]
+    assert "background=false" in info.value.detail["error"]["message"]
+
+
 def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_call():
     post_call = GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
     pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
@@ -1183,6 +1318,12 @@ def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_c
         )
         is None
     )
+    assert (
+        _raise_for_streaming_post_call_pipelines(
+            {"background": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}
+        )
+        is None
+    )
     assert _raise_for_streaming_post_call_pipelines({"metadata": {"_guardrail_pipelines": [("p", post_call)]}}) is None
     assert (
         _raise_for_streaming_post_call_pipelines(
@@ -1191,3 +1332,4 @@ def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_c
         is None
     )
     assert _raise_for_streaming_post_call_pipelines({"stream": True}) is None
+    assert _raise_for_streaming_post_call_pipelines({"background": True}) is None

From c5bcf3a73594ce5fad662a47781d89dbe7955718 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 12:06:43 -0700
Subject: [PATCH 06/25] feat(policy_engine): execute post_call guardrail
 pipelines on streaming responses

---
 .../unified_guardrail/unified_guardrail.py    |  46 ++++-
 .../proxy/policy_engine/pipeline_executor.py  |  47 ++++-
 litellm/proxy/utils.py                        | 187 ++++++++++++++++--
 .../test_unified_guardrail.py                 |   2 +-
 .../proxy_logging/test_guardrail_pipeline.py  | 168 +++++++++++++++-
 5 files changed, 417 insertions(+), 33 deletions(-)

diff --git a/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py b/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
index e95e97bfe74..60b4444e1d4 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
@@ -19,7 +19,7 @@ from litellm.cost_calculator import _infer_call_type
 from litellm.integrations.custom_guardrail import CustomGuardrail
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.api_route_to_call_types import get_call_types_for_route
-from litellm.llms import load_guardrail_translation_mappings
+from litellm.llms import get_guardrail_translation_mapping, load_guardrail_translation_mappings
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.utils import (
@@ -62,6 +62,36 @@ def _as_endpoint_translation(translation: _EndpointTranslation) -> _EndpointTran
     return translation
 
 
+def resolve_endpoint_translation(
+    user_api_key_dict: UserAPIKeyAuth, first_response_item: object | None
+) -> "tuple[str, BaseTranslation] | None":
+    """
+    Resolve the endpoint guardrail translation for a streamed response: the
+    request route wins, falling back to inferring the call type from the first
+    response chunk (the same resolution order the streaming iterator hook uses).
+    Returns None when the call type is unresolvable or has no translation.
+    """
+    route_call_types: Final = (
+        get_call_types_for_route(user_api_key_dict.request_route) if user_api_key_dict.request_route else None
+    )
+    call_type: Final = (
+        route_call_types[0].value
+        if route_call_types
+        else (
+            _infer_call_type(call_type=None, completion_response=first_response_item)
+            if first_response_item is not None
+            else None
+        )
+    )
+    if call_type is None:
+        return None
+    try:
+        handler_cls: Final = get_guardrail_translation_mapping(CallTypes(call_type))
+    except ValueError:
+        return None
+    return call_type, handler_cls()
+
+
 def _chunk_choices(item: object) -> Sequence[object]:
     choices: Final[Sequence[object]] = getattr(item, "choices", None) or []
     return choices
@@ -346,7 +376,7 @@ class UnifiedLLMGuardrails(CustomLogger):
 
         return response
 
-    async def _handle_streaming_block(
+    async def handle_streaming_block(
         self,
         exc: "ModifyResponseException",
         endpoint_translation: _EndpointTranslation,
@@ -402,7 +432,7 @@ class UnifiedLLMGuardrails(CustomLogger):
             return None
         return call_type
 
-    async def _emit_streaming_http_error(
+    async def emit_streaming_http_error(
         self,
         exc: HTTPException,
         call_type: str | None,
@@ -577,7 +607,7 @@ class UnifiedLLMGuardrails(CustomLogger):
         except ModifyResponseException as e:
             if e.original_response is None:
                 e.original_response = responses_so_far
-            async for block_chunk in self._handle_streaming_block(
+            async for block_chunk in self.handle_streaming_block(
                 e,
                 endpoint_translation,
                 stream_started=bool(responses_yielded),
@@ -586,7 +616,7 @@ class UnifiedLLMGuardrails(CustomLogger):
                 yield block_chunk
             raise _StreamTerminated()
         except HTTPException as e:
-            async for error_item in self._emit_streaming_http_error(e, call_type, responses_so_far, request_data):
+            async for error_item in self.emit_streaming_http_error(e, call_type, responses_so_far, request_data):
                 yield error_item
             raise _StreamTerminated()
 
@@ -758,7 +788,7 @@ class UnifiedLLMGuardrails(CustomLogger):
         except ModifyResponseException as e:
             if e.original_response is None:
                 e.original_response = responses_so_far
-            async for block_chunk in self._handle_streaming_block(
+            async for block_chunk in self.handle_streaming_block(
                 e,
                 endpoint_translation,
                 stream_started=bool(responses_yielded),
@@ -1060,7 +1090,7 @@ class UnifiedLLMGuardrails(CustomLogger):
                     # The current chunk was appended to responses_so_far but not
                     # yet yielded, so exclude it: the continuation must reflect
                     # only what the client has actually received.
-                    async for block_chunk in self._handle_streaming_block(
+                    async for block_chunk in self.handle_streaming_block(
                         e,
                         endpoint_translation,
                         stream_started=chunks_yielded,
@@ -1124,7 +1154,7 @@ class UnifiedLLMGuardrails(CustomLogger):
                 # terminating SSE sequence with the block message rather than
                 # propagating into a bare error blob that truncates the stream.
                 # The withheld original chunks are never released.
-                async for block_chunk in self._handle_streaming_block(
+                async for block_chunk in self.handle_streaming_block(
                     e,
                     endpoint_translation,
                     stream_started=bool(responses_yielded),
diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index 190784a2a60..c422a7c0964 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -6,7 +6,7 @@ pass/fail actions (allow, block, next, modify_response) and data forwarding.
 """
 
 import time
-from typing import Any, Final, Literal
+from typing import TYPE_CHECKING, Any, Final, Literal
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -25,6 +25,11 @@ from litellm.types.proxy.policy_engine.pipeline_types import (
     PipelineStepResult,
 )
 
+if TYPE_CHECKING:
+    from litellm.llms.base_llm.guardrail_translation.base_translation import (
+        BaseTranslation,
+    )
+
 try:
     from fastapi.exceptions import HTTPException
 except ImportError:
@@ -43,6 +48,8 @@ class PipelineExecutor:
         call_type: str,
         policy_name: str,
         raw_request_snapshot: dict | None = None,  # mutable-ok: same request-payload shape as data
+        streaming_chunks: list[Any] | None = None,  # mutable-ok: shared buffered-stream chunks, read per step
+        endpoint_translation: "BaseTranslation | None" = None,
     ) -> PipelineExecutionResult:
         """
         Execute pipeline steps sequentially with conditional actions.
@@ -59,6 +66,12 @@ class PipelineExecutor:
                 step whose guardrail opted into ``scan_raw_request`` evaluates
                 the original request instead of whatever an earlier
                 ``pass_data`` step in this same pipeline already rewrote.
+            streaming_chunks: buffered chunks of a completed stream. When set
+                (with ``endpoint_translation``), post_call steps scan the
+                assembled streamed output through the endpoint translation
+                instead of calling ``async_post_call_success_hook``.
+            endpoint_translation: the guardrail translation for the streamed
+                endpoint, resolved by the caller.
 
         Returns:
             PipelineExecutionResult with terminal action and step results
@@ -83,6 +96,8 @@ class PipelineExecutor:
                 user_api_key_dict=user_api_key_dict,
                 call_type=call_type,
                 raw_request_snapshot=raw_request_snapshot,
+                streaming_chunks=streaming_chunks,
+                endpoint_translation=endpoint_translation,
             )
 
             duration = time.perf_counter() - start_time
@@ -154,6 +169,8 @@ class PipelineExecutor:
         user_api_key_dict: Any,
         call_type: str,
         raw_request_snapshot: dict | None = None,  # mutable-ok: same request-payload shape as data
+        streaming_chunks: list[Any] | None = None,  # mutable-ok: shared buffered-stream chunks, read per step
+        endpoint_translation: "BaseTranslation | None" = None,
     ) -> tuple[
         Literal["pass", "fail", "error"],
         dict | None,
@@ -198,10 +215,8 @@ class PipelineExecutor:
 
             # Use unified_guardrail path if callback implements apply_guardrail
             target: CustomLogger = callback
-            use_unified: Final = (
-                "apply_guardrail" in type(callback).__dict__ and not callback.use_native_lifecycle_hooks
-            )
-            if use_unified:
+            use_unified: Final = PipelineExecutor.supports_unified_execution(callback)
+            if use_unified and streaming_chunks is None:
                 hook_input["guardrail_to_apply"] = callback
                 target = UnifiedLLMGuardrails()
 
@@ -216,6 +231,22 @@ class PipelineExecutor:
                     callback.mark_pre_call_hook_ran(data)
                     if isinstance(response, dict):
                         callback.mark_pre_call_hook_ran(response)
+            elif mode == "post_call" and streaming_chunks is not None:
+                if not use_unified or endpoint_translation is None:
+                    return (
+                        "error",
+                        None,
+                        f"Guardrail '{step.guardrail}' does not support streaming pipeline execution",
+                        None,
+                    )
+                await endpoint_translation.process_output_streaming_response(
+                    responses_so_far=streaming_chunks,
+                    guardrail_to_apply=callback,
+                    litellm_logging_obj=data.get("litellm_logging_obj"),
+                    user_api_key_dict=user_api_key_dict,
+                    request_data=hook_input,
+                )
+                response = None
             elif mode == "post_call":
                 response = await target.async_post_call_success_hook(
                     user_api_key_dict=user_api_key_dict,
@@ -246,6 +277,12 @@ class PipelineExecutor:
                 verbose_proxy_logger.error("Pipeline: unexpected error from guardrail '%s': %s", step.guardrail, e)
                 return ("error", None, str(e), e)
 
+    @staticmethod
+    def supports_unified_execution(callback: CustomGuardrail) -> bool:
+        """Whether this guardrail runs through the unified apply_guardrail path,
+        the interface streaming pipeline execution requires."""
+        return "apply_guardrail" in type(callback).__dict__ and not callback.use_native_lifecycle_hooks
+
     @staticmethod
     def find_guardrail_callback(guardrail_name: str) -> CustomGuardrail | None:
         """Look up an initialized guardrail callback by name from litellm.callbacks."""
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 80e991640a9..5642eb5383e 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -19,7 +19,7 @@ from email.mime.text import MIMEText
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, ClassVar, Final, Literal, Optional, Protocol, TypeVar, Union, cast, overload
 
-from typing_extensions import ReadOnly, TypedDict
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 from litellm import _custom_logger_compatible_callbacks_literal
 from litellm.constants import (
@@ -486,29 +486,80 @@ def _merge_pipeline_metadata_writes(data: dict, modified_data: Mapping[str, obje
         _merge_pipeline_metadata_bucket(data, bucket_key, modified_data.get(bucket_key))
 
 
+def _pipeline_step_supports_streaming(guardrail_name: str) -> bool:
+    callback: Final = PipelineExecutor.find_guardrail_callback(guardrail_name)
+    return callback is not None and PipelineExecutor.supports_unified_execution(callback)
+
+
+class _PipelineErrorBody(TypedDict):
+    message: ReadOnly[str]
+    type: ReadOnly[str]
+    policies: ReadOnly[tuple[str, ...]]
+    guardrails: NotRequired[ReadOnly[tuple[str, ...]]]
+
+
+class _PipelineErrorDetail(TypedDict):
+    error: ReadOnly[_PipelineErrorBody]
+
+
 def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object]) -> None:
-    if data.get("stream") is not True and data.get("background") is not True:
+    """
+    Reject up front the requests whose post_call pipelines could never run.
+
+    Background responses skip the post_call hooks entirely, so a pipeline
+    governing one would silently never execute. Streaming responses execute
+    pipelines against the buffered stream through the endpoint guardrail
+    translations, which requires every step's guardrail to support the unified
+    apply_guardrail interface; steps that cannot (native-lifecycle guardrails,
+    or guardrails not registered at all) keep the 400 rather than letting
+    ungoverned output stream through.
+    """
+    is_stream: Final = data.get("stream") is True
+    is_background: Final = data.get("background") is True
+    if not is_stream and not is_background:
         return
-    post_call_policies: Final = tuple(
-        policy_name for policy_name, pipeline in _policy_pipelines(data) if pipeline.mode == "post_call"
+    post_call_pipelines: Final = tuple(
+        (policy_name, pipeline) for policy_name, pipeline in _policy_pipelines(data) if pipeline.mode == "post_call"
     )
-    if not post_call_policies:
+    if not post_call_pipelines:
         return
-    raise HTTPException(
-        status_code=400,
-        detail={
+    post_call_policies: Final = tuple(policy_name for policy_name, _pipeline in post_call_pipelines)
+    if is_background:
+        background_detail: Final[_PipelineErrorDetail] = {
             "error": {
                 "message": (
-                    "Policies with post_call guardrail pipelines cannot govern streaming or background "
-                    f"responses yet: {', '.join(post_call_policies)}. Retry with stream=false and "
-                    "background=false, or move these policies' output guardrails from pipeline steps to "
-                    "guardrails.add, which scans streamed output."
+                    "Policies with post_call guardrail pipelines cannot govern background "
+                    f"responses: {', '.join(post_call_policies)}. Retry with background=false."
                 ),
                 "type": "guardrail_pipeline_error",
-                "policies": list(post_call_policies),
+                "policies": post_call_policies,
             }
-        },
+        }
+        raise HTTPException(status_code=400, detail=background_detail)
+    unsupported_guardrails: Final = tuple(
+        dict.fromkeys(
+            step.guardrail
+            for _policy_name, pipeline in post_call_pipelines
+            for step in pipeline.steps
+            if not _pipeline_step_supports_streaming(step.guardrail)
+        )
     )
+    if not unsupported_guardrails:
+        return
+    unsupported_detail: Final[_PipelineErrorDetail] = {
+        "error": {
+            "message": (
+                "Policies with post_call guardrail pipelines cannot govern streaming responses "
+                "because these pipeline guardrails do not support the unified apply_guardrail "
+                f"interface: {', '.join(unsupported_guardrails)}. Retry with stream=false, or move "
+                "them from pipeline steps to guardrails.add, which scans streamed output."
+            ),
+            "type": "guardrail_pipeline_error",
+            "policies": post_call_policies,
+            "guardrails": unsupported_guardrails,
+        }
+    }
+    raise HTTPException(status_code=400, detail=unsupported_detail)
 
 
 def _prompt_block_text(block: object) -> str:
@@ -1689,7 +1740,7 @@ class ProxyLogging:
         result: PipelineExecutionResult,
         data: dict,
         policy_name: str,
-        original_response: LLMResponseTypes | None = None,
+        original_response: "LLMResponseTypes | Sequence[object] | None" = None,
     ) -> dict:
         """
         Handle a PipelineExecutionResult — allow, block, or modify_response.
@@ -1699,7 +1750,9 @@ class ProxyLogging:
         payload (already sent upstream) must stay untouched; a replacement
         response carried in ``modified_data`` is adopted by the caller, and
         metadata-bucket writes (applied guardrails, guardrail logging info)
-        are merged back so headers and spend logs still see them.
+        are merged back so headers and spend logs still see them. On the
+        streaming path it is the buffered chunk list, carried into
+        ``ModifyResponseException.original_response`` for usage reporting.
         """
         if result.terminal_action == "allow":
             if result.modified_data is not None:
@@ -3195,11 +3248,16 @@ class ProxyLogging:
             # dict lookups + llm_router.get_deployment() per callback per chunk.
             _cached_guardrail_data: dict | None = None
             _guardrail_data_computed = False
+            pipeline_managed: Final = (
+                _pipeline_managed_guardrail_names(data, "post_call") if caps.has_guardrail else frozenset()
+            )
 
             for callback in litellm.callbacks:
                 try:
                     _callback: CustomLogger | None = None
                     if isinstance(callback, CustomGuardrail):
+                        if callback.guardrail_name in pipeline_managed:
+                            continue
                         # Main - V2 Guardrails implementation
                         from litellm.types.guardrails import GuardrailEventHooks
 
@@ -3256,12 +3314,17 @@ class ProxyLogging:
         1. /chat/completions
         """
         caps: Final = ProxyLogging._callback_capabilities()
+        post_call_pipelines: Final = tuple(
+            (policy_name, pipeline)
+            for policy_name, pipeline in _policy_pipelines(request_data)
+            if pipeline.mode == "post_call"
+        )
         # Fast path: no real overrides. Internal proxy CustomLogger callbacks
         # (e.g. _PROXY_MaxBudgetLimiter, ManagedFiles) inherit the default
         # ``async for chunk: yield chunk`` body, so wrapping the iterator
         # through each of them adds N pass-through trampolines per chunk for
         # zero behavior change. Skip the chain entirely and stream through.
-        if not caps.iterator_overrides:
+        if not caps.iterator_overrides and not post_call_pipelines:
             try:
                 async for chunk in response:
                     yield chunk
@@ -3281,8 +3344,11 @@ class ProxyLogging:
         current_response = response
         stream_needs_translation: Final = ProxyLogging._stream_requires_guardrail_translation(user_api_key_dict)
 
+        pipeline_managed_names: Final = _pipeline_managed_guardrail_names(request_data, "post_call")
         for resolved_callback, kind in caps.iterator_overrides:
             if isinstance(resolved_callback, CustomGuardrail):
+                if resolved_callback.guardrail_name in pipeline_managed_names:
+                    continue
                 if (
                     resolved_callback.should_run_guardrail(data=request_data, event_type=GuardrailEventHooks.post_call)
                     is not True
@@ -3322,6 +3388,17 @@ class ProxyLogging:
                     ),
                 )
 
+        # Policy pipelines run last, over the fully buffered stream, so a
+        # pipeline verdict covers whatever the flat guardrail chain above
+        # already let through.
+        if post_call_pipelines:
+            current_response = self._pipeline_gated_stream(
+                response=current_response,
+                user_api_key_dict=user_api_key_dict,
+                request_data=request_data,
+                pipelines=post_call_pipelines,
+            )
+
         try:
             async for chunk in current_response:
                 yield chunk
@@ -3337,6 +3414,82 @@ class ProxyLogging:
         # we reach this point the metadata is fully populated.
         ProxyLogging._fire_deferred_stream_logging(request_data)
 
+    async def _pipeline_gated_stream(
+        self,
+        response: "AsyncGenerator[object, None]",
+        user_api_key_dict: UserAPIKeyAuth,
+        request_data: dict,
+        pipelines: "tuple[tuple[str, GuardrailPipeline], ...]",
+    ) -> "AsyncGenerator[Any, None]":
+        """
+        Execute post_call policy pipelines against a streamed response.
+
+        Buffers the whole stream (nothing reaches the client until every
+        pipeline allows it), then runs each pipeline's steps against the
+        assembled output through the endpoint guardrail translation, the same
+        machinery flat post_call guardrails use at end of stream. An allow
+        releases the buffered chunks verbatim; a block or modify_response
+        terminates with the translation's block chunks or the raised error.
+        """
+        from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import (
+            resolve_endpoint_translation,
+        )
+
+        buffered: Final[list[object]] = []  # mutable-ok: accumulates the stream before the pipeline verdict
+        async for item in response:
+            buffered.append(item)
+        if not buffered:
+            return
+
+        resolved: Final = resolve_endpoint_translation(user_api_key_dict, buffered[0])
+        if resolved is None:
+            policy_names: Final = tuple(policy_name for policy_name, _pipeline in pipelines)
+            unresolvable_detail: Final[_PipelineErrorDetail] = {
+                "error": {
+                    "message": (
+                        "Policy pipelines could not govern this streaming response shape; "
+                        f"the response was withheld: {', '.join(policy_names)}."
+                    ),
+                    "type": "guardrail_pipeline_error",
+                    "policies": policy_names,
+                }
+            }
+            raise HTTPException(status_code=500, detail=unresolvable_detail)
+        call_type, endpoint_translation = resolved
+
+        for policy_name, pipeline in pipelines:
+            result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
+                steps=pipeline.steps,
+                mode="post_call",
+                data=request_data,
+                user_api_key_dict=user_api_key_dict,
+                call_type=call_type,
+                policy_name=policy_name,
+                streaming_chunks=buffered,
+                endpoint_translation=endpoint_translation,
+            )
+            try:
+                ProxyLogging._handle_pipeline_result(
+                    result, data=request_data, policy_name=policy_name, original_response=buffered
+                )
+            except ModifyResponseException as e:
+                if e.original_response is None:
+                    e.original_response = buffered
+                async for block_chunk in unified_guardrail.handle_streaming_block(
+                    e, endpoint_translation, stream_started=False, responses_so_far=()
+                ):
+                    yield block_chunk
+                return
+            except HTTPException as e:
+                async for error_chunk in unified_guardrail.emit_streaming_http_error(
+                    e, call_type, buffered, request_data
+                ):
+                    yield error_chunk
+                return
+
+        for buffered_item in buffered:
+            yield buffered_item
+
     @staticmethod
     def _fire_deferred_stream_logging(request_data: dict) -> None:
         """
diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/unified_guardrails/test_unified_guardrail.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/unified_guardrails/test_unified_guardrail.py
index 8b9ecfbbeee..fe4acbf8277 100644
--- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/unified_guardrails/test_unified_guardrail.py
+++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/unified_guardrails/test_unified_guardrail.py
@@ -1026,7 +1026,7 @@ class TestStreamingTransform:
         )
 
         emitted = []
-        async for item in handler._emit_streaming_http_error(
+        async for item in handler.emit_streaming_http_error(
             exc,
             call_type=CallTypes.asend_message.value,
             responses_so_far=[{"id": "req-1"}],
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 34a25d75bc0..2b36ae111dc 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -1284,7 +1284,8 @@ async def test_pre_call_hook_rejects_streaming_request_with_post_call_pipeline(
         )
 
     assert info.value.status_code == 400
-    assert info.value.detail["error"]["policies"] == ["response-governance"]
+    assert info.value.detail["error"]["policies"] == ("response-governance",)
+    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
     assert "stream=false" in info.value.detail["error"]["message"]
 
 
@@ -1304,7 +1305,7 @@ async def test_pre_call_hook_rejects_background_request_with_post_call_pipeline(
         )
 
     assert info.value.status_code == 400
-    assert info.value.detail["error"]["policies"] == ["response-governance"]
+    assert info.value.detail["error"]["policies"] == ("response-governance",)
     assert "background=false" in info.value.detail["error"]["message"]
 
 
@@ -1333,3 +1334,166 @@ def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_c
     )
     assert _raise_for_streaming_post_call_pipelines({"stream": True}) is None
     assert _raise_for_streaming_post_call_pipelines({"background": True}) is None
+
+
+# ---------------------------------------------------------------------------
+# post_call pipelines on streaming responses
+# ---------------------------------------------------------------------------
+
+
+def _unified_stream_guardrail(seen: Dict[str, Any], block: bool = False) -> CustomGuardrail:
+    class UnifiedStreamGuardrail(CustomGuardrail):
+        async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+            seen["count"] = seen.get("count", 0) + 1
+            seen["input_type"] = input_type
+            if block:
+                raise HTTPException(status_code=400, detail={"error": "output blocked"})
+            return inputs
+
+    return UnifiedStreamGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)
+
+
+def _stream_chunks() -> List[Any]:
+    return [
+        litellm.ModelResponseStream(choices=[{"index": 0, "delta": {"content": "hello "}, "finish_reason": None}]),
+        litellm.ModelResponseStream(choices=[{"index": 0, "delta": {"content": "world"}, "finish_reason": "stop"}]),
+    ]
+
+
+async def _async_chunk_iter(chunks: List[Any]):
+    for chunk in chunks:
+        yield chunk
+
+
+@pytest.mark.asyncio
+async def test_pre_call_hook_allows_streaming_when_pipeline_guardrail_supports_unified(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
+    data = _post_call_pipeline_data(stream=True)
+
+    out = await proxy_logging.pre_call_hook(
+        user_api_key_dict=make_user_api_key_auth(),
+        data=data,
+        call_type="completion",
+        guardrails_only=True,
+    )
+
+    assert out is not None
+    assert out.get("stream") is True
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("native_lifecycle", [False, True])
+async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_lacks_unified_support(
+    proxy_logging, make_user_api_key_auth, monkeypatch, native_lifecycle
+):
+    if native_lifecycle:
+
+        class NativeOnlyGuardrail(CustomGuardrail):
+            use_native_lifecycle_hooks = True
+
+            async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+                return inputs
+
+    else:
+
+        class NativeOnlyGuardrail(CustomGuardrail):
+            pass
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [NativeOnlyGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)],
+    )
+    data = _post_call_pipeline_data(stream=True)
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=make_user_api_key_auth(),
+            data=data,
+            call_type="completion",
+            guardrails_only=True,
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
+    assert "apply_guardrail" in info.value.detail["error"]["message"]
+
+
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_allow_releases_buffered_chunks(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    chunks = _stream_chunks()
+
+    delivered = [
+        item
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        )
+    ]
+
+    assert [id(item) for item in delivered] == [id(chunk) for chunk in chunks]
+    assert seen["count"] == 1
+    assert seen["input_type"] == "response"
+
+
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_block_withholds_all_chunks(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen, block=True)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    delivered: List[Any] = []
+
+    async def _drain() -> None:
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(_stream_chunks()),
+            request_data=data,
+        ):
+            delivered.append(item)
+
+    with pytest.raises(HTTPException) as info:
+        await _drain()
+
+    assert delivered == []
+    assert info.value.status_code == 400
+    assert "output blocked" in str(info.value.detail)
+
+
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_withholds_unresolvable_response_shape(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    delivered: List[Any] = []
+
+    async def _drain() -> None:
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(),
+            response=_async_chunk_iter([object(), object()]),
+            request_data=data,
+        ):
+            delivered.append(item)
+
+    with pytest.raises(HTTPException) as info:
+        await _drain()
+
+    assert delivered == []
+    assert info.value.status_code == 500
+    assert "withheld" in info.value.detail["error"]["message"]
+    assert seen.get("count") is None

From fa5a10941e713968fbe7251a99d867ef0050dd69 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 13:28:11 -0700
Subject: [PATCH 07/25] fix(policy_engine): fail closed on streaming for
 content-rewriting pipeline steps and untranslatable routes

---
 .../unified_guardrail/unified_guardrail.py    |  19 ++-
 litellm/proxy/utils.py                        | 111 ++++++++++++------
 .../proxy_logging/test_guardrail_pipeline.py  |  84 +++++++++++--
 3 files changed, 155 insertions(+), 59 deletions(-)

diff --git a/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py b/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
index 60b4444e1d4..89527c05eb6 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py
@@ -876,6 +876,14 @@ class UnifiedLLMGuardrails(CustomLogger):
         choices: Final = _chunk_choices(item)
         return any(getattr(choice, "finish_reason", None) is not None for choice in choices)
 
+    def resolve_streaming_flag(self, guardrail_to_apply: CustomGuardrail | None, name: str, default: object) -> object:
+        """Streaming flag resolution order (later wins): default < guardrail
+        attribute < guardrail_config dict < this callback's optional_params."""
+        attribute_value: Final = default if guardrail_to_apply is None else getattr(guardrail_to_apply, name, default)
+        config: Final = None if guardrail_to_apply is None else getattr(guardrail_to_apply, "guardrail_config", None)
+        config_value: Final = config.get(name, attribute_value) if isinstance(config, dict) else attribute_value
+        return self.optional_params.get(name, config_value)
+
     async def async_post_call_streaming_iterator_hook(
         self,
         user_api_key_dict: UserAPIKeyAuth,
@@ -906,17 +914,8 @@ class UnifiedLLMGuardrails(CustomLogger):
         if guardrail_to_apply is None:
             guardrail_to_apply = request_data.pop("guardrail_to_apply", None)
 
-        # Get streaming configuration. Resolution order (later wins): default
-        # < guardrail attribute < guardrail_config dict < this callback's
-        # optional_params.
         def _streaming_flag(name: str, default: object) -> Any:
-            value = default
-            if guardrail_to_apply is not None:
-                value = getattr(guardrail_to_apply, name, value)
-                config: Final[Mapping[str, object]] = getattr(guardrail_to_apply, "guardrail_config", {})
-                if isinstance(config, dict):
-                    value = config.get(name, value)
-            return self.optional_params.get(name, value)
+            return self.resolve_streaming_flag(guardrail_to_apply, name, default)
 
         sampling_rate: Final[int] = _streaming_flag("streaming_sampling_rate", 5)
         # Only apply the guardrail at end of stream (not per chunk).
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 5642eb5383e..8b2c38d457f 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -139,6 +139,7 @@ from litellm.proxy.db.token_auth import (
 )
 from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import (
     UnifiedLLMGuardrails,
+    resolve_endpoint_translation,
 )
 from litellm.proxy.hooks import PROXY_HOOKS, get_proxy_hook
 from litellm.proxy.hooks.cache_control_check import _PROXY_CacheControlCheck
@@ -486,11 +487,19 @@ def _merge_pipeline_metadata_writes(data: dict, modified_data: Mapping[str, obje
         _merge_pipeline_metadata_bucket(data, bucket_key, modified_data.get(bucket_key))
 
 
-def _pipeline_step_supports_streaming(guardrail_name: str) -> bool:
+def _pipeline_step_supports_unified_streaming(guardrail_name: str) -> bool:
     callback: Final = PipelineExecutor.find_guardrail_callback(guardrail_name)
     return callback is not None and PipelineExecutor.supports_unified_execution(callback)
 
 
+def _pipeline_step_rewrites_streamed_content(guardrail_name: str) -> bool:
+    callback: Final = PipelineExecutor.find_guardrail_callback(guardrail_name)
+    if callback is None:
+        return False
+    transform_mode: Final = unified_guardrail.resolve_streaming_flag(callback, "streaming_transform_mode", "block_only")
+    return callback.mask_response_content or transform_mode == "incremental_diff"
+
+
 class _PipelineErrorBody(TypedDict):
     message: ReadOnly[str]
     type: ReadOnly[str]
@@ -502,17 +511,20 @@ class _PipelineErrorDetail(TypedDict):
     error: ReadOnly[_PipelineErrorBody]
 
 
-def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object]) -> None:
+def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_api_key_dict: UserAPIKeyAuth) -> None:
     """
     Reject up front the requests whose post_call pipelines could never run.
 
     Background responses skip the post_call hooks entirely, so a pipeline
     governing one would silently never execute. Streaming responses execute
     pipelines against the buffered stream through the endpoint guardrail
-    translations, which requires every step's guardrail to support the unified
-    apply_guardrail interface; steps that cannot (native-lifecycle guardrails,
-    or guardrails not registered at all) keep the 400 rather than letting
-    ungoverned output stream through.
+    translation of the request route, releasing the buffered chunks verbatim
+    on allow. That needs every step's guardrail to support the unified
+    apply_guardrail interface and to only allow or block (a step that rewrites
+    streamed content, via mask_response_content or
+    streaming_transform_mode=incremental_diff, would have its rewrite silently
+    dropped), and needs the route to have a translation at all; anything else
+    keeps the 400 rather than letting ungoverned output stream through.
     """
     is_stream: Final = data.get("stream") is True
     is_background: Final = data.get("background") is True
@@ -536,30 +548,61 @@ def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object]) -> None
             }
         }
         raise HTTPException(status_code=400, detail=background_detail)
-    unsupported_guardrails: Final = tuple(
-        dict.fromkeys(
-            step.guardrail
-            for _policy_name, pipeline in post_call_pipelines
-            for step in pipeline.steps
-            if not _pipeline_step_supports_streaming(step.guardrail)
-        )
+    step_guardrails: Final = tuple(
+        dict.fromkeys(step.guardrail for _policy_name, pipeline in post_call_pipelines for step in pipeline.steps)
     )
-    if not unsupported_guardrails:
+    unsupported_guardrails: Final = tuple(
+        guardrail for guardrail in step_guardrails if not _pipeline_step_supports_unified_streaming(guardrail)
+    )
+    if unsupported_guardrails:
+        unsupported_detail: Final[_PipelineErrorDetail] = {
+            "error": {
+                "message": (
+                    "Policies with post_call guardrail pipelines cannot govern streaming responses "
+                    "because these pipeline guardrails do not support the unified apply_guardrail "
+                    f"interface: {', '.join(unsupported_guardrails)}. Retry with stream=false, or drop "
+                    "them from the pipeline steps so guardrails.add scans them on streamed output."
+                ),
+                "type": "guardrail_pipeline_error",
+                "policies": post_call_policies,
+                "guardrails": unsupported_guardrails,
+            }
+        }
+        raise HTTPException(status_code=400, detail=unsupported_detail)
+    rewriting_guardrails: Final = tuple(
+        guardrail for guardrail in step_guardrails if _pipeline_step_rewrites_streamed_content(guardrail)
+    )
+    if rewriting_guardrails:
+        rewriting_detail: Final[_PipelineErrorDetail] = {
+            "error": {
+                "message": (
+                    "Policies with post_call guardrail pipelines cannot govern streaming responses "
+                    "because these pipeline guardrails rewrite streamed content (mask_response_content "
+                    "or streaming_transform_mode=incremental_diff), which pipeline steps would release "
+                    f"unmodified: {', '.join(rewriting_guardrails)}. Retry with stream=false, or drop "
+                    "them from the pipeline steps so guardrails.add applies them to streamed output."
+                ),
+                "type": "guardrail_pipeline_error",
+                "policies": post_call_policies,
+                "guardrails": rewriting_guardrails,
+            }
+        }
+        raise HTTPException(status_code=400, detail=rewriting_detail)
+    route: Final = user_api_key_dict.request_route
+    if not route or resolve_endpoint_translation(user_api_key_dict, None) is not None:
         return
-    unsupported_detail: Final[_PipelineErrorDetail] = {
+    route_detail: Final[_PipelineErrorDetail] = {
         "error": {
             "message": (
-                "Policies with post_call guardrail pipelines cannot govern streaming responses "
-                "because these pipeline guardrails do not support the unified apply_guardrail "
-                f"interface: {', '.join(unsupported_guardrails)}. Retry with stream=false, or move "
-                "them from pipeline steps to guardrails.add, which scans streamed output."
+                "Policies with post_call guardrail pipelines cannot govern streaming responses on "
+                f"route {route} because it has no endpoint guardrail translation to scan the stream "
+                f"through: {', '.join(post_call_policies)}. Retry with stream=false."
             ),
             "type": "guardrail_pipeline_error",
             "policies": post_call_policies,
-            "guardrails": unsupported_guardrails,
         }
     }
-    raise HTTPException(status_code=400, detail=unsupported_detail)
+    raise HTTPException(status_code=400, detail=route_detail)
 
 
 def _prompt_block_text(block: object) -> str:
@@ -1915,7 +1958,7 @@ class ProxyLogging:
         )
 
         try:
-            _raise_for_streaming_post_call_pipelines(data)
+            _raise_for_streaming_post_call_pipelines(data, user_api_key_dict)
 
             # Execute guardrail pipelines before the normal callback loop
             data, _ = await self._maybe_execute_pipelines(
@@ -3431,10 +3474,6 @@ class ProxyLogging:
         releases the buffered chunks verbatim; a block or modify_response
         terminates with the translation's block chunks or the raised error.
         """
-        from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import (
-            resolve_endpoint_translation,
-        )
-
         buffered: Final[list[object]] = []  # mutable-ok: accumulates the stream before the pipeline verdict
         async for item in response:
             buffered.append(item)
@@ -3444,17 +3483,15 @@ class ProxyLogging:
         resolved: Final = resolve_endpoint_translation(user_api_key_dict, buffered[0])
         if resolved is None:
             policy_names: Final = tuple(policy_name for policy_name, _pipeline in pipelines)
-            unresolvable_detail: Final[_PipelineErrorDetail] = {
-                "error": {
-                    "message": (
-                        "Policy pipelines could not govern this streaming response shape; "
-                        f"the response was withheld: {', '.join(policy_names)}."
-                    ),
-                    "type": "guardrail_pipeline_error",
-                    "policies": policy_names,
-                }
-            }
-            raise HTTPException(status_code=500, detail=unresolvable_detail)
+            raise ProxyException(
+                message=(
+                    "Policy pipelines could not govern this streaming response shape; "
+                    f"the response was withheld: {', '.join(policy_names)}."
+                ),
+                type="guardrail_pipeline_error",
+                param=None,
+                code=500,
+            )
         call_type, endpoint_translation = resolved
 
         for policy_name, pipeline in pipelines:
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 2b36ae111dc..6c90637158e 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -23,6 +23,7 @@ from litellm.integrations.custom_guardrail import (
     ModifyResponseException,
 )
 from litellm.integrations.prometheus import PrometheusLogger
+from litellm.proxy._types import ProxyException
 from litellm.proxy.common_utils.callback_utils import add_guardrail_to_applied_guardrails_header
 from litellm.proxy.utils import ProxyLogging, _raise_for_streaming_post_call_pipelines
 from litellm.types.guardrails import GuardrailEventHooks
@@ -1309,31 +1310,35 @@ async def test_pre_call_hook_rejects_background_request_with_post_call_pipeline(
     assert "background=false" in info.value.detail["error"]["message"]
 
 
-def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_call():
+def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_call(make_user_api_key_auth):
     post_call = GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
     pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
+    auth = make_user_api_key_auth(request_route="/custom/stream")
 
     assert (
         _raise_for_streaming_post_call_pipelines(
-            {"stream": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}
+            {"stream": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}, auth
         )
         is None
     )
     assert (
         _raise_for_streaming_post_call_pipelines(
-            {"background": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}
+            {"background": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}, auth
         )
         is None
     )
-    assert _raise_for_streaming_post_call_pipelines({"metadata": {"_guardrail_pipelines": [("p", post_call)]}}) is None
+    assert (
+        _raise_for_streaming_post_call_pipelines({"metadata": {"_guardrail_pipelines": [("p", post_call)]}}, auth)
+        is None
+    )
     assert (
         _raise_for_streaming_post_call_pipelines(
-            {"stream": True, "metadata": {"_guardrail_pipelines": [("p", pre_call)]}}
+            {"stream": True, "metadata": {"_guardrail_pipelines": [("p", pre_call)]}}, auth
         )
         is None
     )
-    assert _raise_for_streaming_post_call_pipelines({"stream": True}) is None
-    assert _raise_for_streaming_post_call_pipelines({"background": True}) is None
+    assert _raise_for_streaming_post_call_pipelines({"stream": True}, auth) is None
+    assert _raise_for_streaming_post_call_pipelines({"background": True}, auth) is None
 
 
 # ---------------------------------------------------------------------------
@@ -1366,15 +1371,16 @@ async def _async_chunk_iter(chunks: List[Any]):
 
 
 @pytest.mark.asyncio
+@pytest.mark.parametrize("request_route", [None, "/v1/chat/completions"])
 async def test_pre_call_hook_allows_streaming_when_pipeline_guardrail_supports_unified(
-    proxy_logging, make_user_api_key_auth, monkeypatch
+    proxy_logging, make_user_api_key_auth, monkeypatch, request_route
 ):
     seen: Dict[str, Any] = {}
     monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
     data = _post_call_pipeline_data(stream=True)
 
     out = await proxy_logging.pre_call_hook(
-        user_api_key_dict=make_user_api_key_auth(),
+        user_api_key_dict=make_user_api_key_auth(request_route=request_route),
         data=data,
         call_type="completion",
         guardrails_only=True,
@@ -1422,6 +1428,60 @@ async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_lacks_uni
     assert "apply_guardrail" in info.value.detail["error"]["message"]
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "rewrite_attribute, value",
+    [
+        ("mask_response_content", True),
+        ("streaming_transform_mode", "incremental_diff"),
+        ("guardrail_config", {"streaming_transform_mode": "incremental_diff"}),
+    ],
+)
+async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_rewrites_streamed_content(
+    proxy_logging, make_user_api_key_auth, monkeypatch, rewrite_attribute, value
+):
+    seen: Dict[str, Any] = {}
+    guardrail = _unified_stream_guardrail(seen)
+    setattr(guardrail, rewrite_attribute, value)
+    monkeypatch.setattr(litellm, "callbacks", [guardrail])
+    data = _post_call_pipeline_data(stream=True)
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            data=data,
+            call_type="completion",
+            guardrails_only=True,
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
+    assert "rewrite streamed content" in info.value.detail["error"]["message"]
+    assert seen.get("count") is None
+
+
+@pytest.mark.asyncio
+async def test_pre_call_hook_rejects_streaming_when_route_has_no_guardrail_translation(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
+    data = _post_call_pipeline_data(stream=True)
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/custom/stream"),
+            data=data,
+            call_type="completion",
+            guardrails_only=True,
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["policies"] == ("response-governance",)
+    assert "/custom/stream" in info.value.detail["error"]["message"]
+    assert seen.get("count") is None
+
+
 @pytest.mark.asyncio
 async def test_streaming_iterator_hook_pipeline_allow_releases_buffered_chunks(
     proxy_logging, make_user_api_key_auth, monkeypatch
@@ -1490,10 +1550,10 @@ async def test_streaming_iterator_hook_pipeline_withholds_unresolvable_response_
         ):
             delivered.append(item)
 
-    with pytest.raises(HTTPException) as info:
+    with pytest.raises(ProxyException) as info:
         await _drain()
 
     assert delivered == []
-    assert info.value.status_code == 500
-    assert "withheld" in info.value.detail["error"]["message"]
+    assert info.value.code == "500"
+    assert "withheld" in info.value.message
     assert seen.get("count") is None

From 1bed9bae43a7c28e46d6b38d4b58d10d35d87c58 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 13:41:33 -0700
Subject: [PATCH 08/25] chore(policy_engine): drop control-flow comment flagged
 in review

---
 litellm/proxy/utils.py | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 8b2c38d457f..231d4f18aa0 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -3431,9 +3431,6 @@ class ProxyLogging:
                     ),
                 )
 
-        # Policy pipelines run last, over the fully buffered stream, so a
-        # pipeline verdict covers whatever the flat guardrail chain above
-        # already let through.
         if post_call_pipelines:
             current_response = self._pipeline_gated_stream(
                 response=current_response,

From d51198fdeb3ccc7fcf70fbb116f94d8f360ef4d6 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 14:07:34 -0700
Subject: [PATCH 09/25] test(policy_engine): cover streaming pipeline gate
 branches

Adds regression tests for the modify_response block on the Anthropic route,
the gate with no iterator overrides, and the per-chunk hook skipping
pipeline-managed guardrails. Corrects the gate docstring: an allow releases
the chunks as the endpoint translation left them, not verbatim
---
 litellm/proxy/utils.py                        |   7 +-
 .../proxy_logging/test_guardrail_pipeline.py  | 110 ++++++++++++++++++
 2 files changed, 115 insertions(+), 2 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 231d4f18aa0..a64b57c1388 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -3468,8 +3468,11 @@ class ProxyLogging:
         pipeline allows it), then runs each pipeline's steps against the
         assembled output through the endpoint guardrail translation, the same
         machinery flat post_call guardrails use at end of stream. An allow
-        releases the buffered chunks verbatim; a block or modify_response
-        terminates with the translation's block chunks or the raised error.
+        releases the buffered chunks as that machinery left them (the
+        Responses and A2A translations write guardrail output back into the
+        final chunk, exactly as they do for flat guardrails); a block or
+        modify_response terminates with the translation's block chunks or the
+        raised error.
         """
         buffered: Final[list[object]] = []  # mutable-ok: accumulates the stream before the pipeline verdict
         async for item in response:
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 6c90637158e..a258442c79a 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -10,6 +10,7 @@ Covers ``_should_use_guardrail_load_balancing``, ``_execute_guardrail_hook``,
 from __future__ import annotations
 
 import asyncio
+import json
 from typing import Any, Dict, List
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -1557,3 +1558,112 @@ async def test_streaming_iterator_hook_pipeline_withholds_unresolvable_response_
     assert info.value.code == "500"
     assert "withheld" in info.value.message
     assert seen.get("count") is None
+
+
+def _anthropic_sse_chunks() -> List[bytes]:
+    events = [
+        ("message_start", {"type": "message_start", "message": {"id": "msg_1", "type": "message", "role": "assistant", "model": "m", "content": [], "stop_reason": None, "usage": {"input_tokens": 1, "output_tokens": 0}}}),
+        ("content_block_start", {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}),
+        ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "hello world"}}),
+        ("content_block_stop", {"type": "content_block_stop", "index": 0}),
+        ("message_delta", {"type": "message_delta", "delta": {"stop_reason": "end_turn", "stop_sequence": None}, "usage": {"output_tokens": 2}}),
+        ("message_stop", {"type": "message_stop"}),
+    ]
+    return [f"event: {name}\ndata: {json.dumps(payload)}\n\n".encode() for name, payload in events]
+
+
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_modify_response_emits_translated_block(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen, block=True)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    pipeline = GuardrailPipeline(
+        mode="post_call",
+        steps=[
+            PipelineStep(
+                guardrail="gr-post",
+                on_pass="allow",
+                on_fail="modify_response",
+                modify_response_message="content policy block",
+            )
+        ],
+    )
+    data = _post_call_pipeline_data(stream=True)
+    data["metadata"]["_guardrail_pipelines"] = [("response-governance", pipeline)]
+    chunks = _anthropic_sse_chunks()
+
+    delivered = [
+        item
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/messages"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        )
+    ]
+
+    raw = b"".join(delivered).decode()
+    assert seen["count"] == 1
+    assert "content policy block" in raw
+    assert "hello world" not in raw
+    assert not any(item is chunk for item in delivered for chunk in chunks)
+
+
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_gates_without_iterator_overrides(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    monkeypatch.setattr(litellm, "callbacks", [])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    delivered: List[Any] = []
+
+    async def _drain() -> None:
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(_stream_chunks()),
+            request_data=data,
+        ):
+            delivered.append(item)
+
+    with pytest.raises(HTTPException) as info:
+        await _drain()
+
+    assert delivered == []
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["pipeline_context"]["step_results"] == [
+        {"guardrail": "gr-post", "outcome": "error", "action": "block"}
+    ]
+
+
+@pytest.mark.asyncio
+async def test_per_chunk_streaming_hook_skips_pipeline_managed_guardrail(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+
+    class RecordingGuardrail(CustomGuardrail):
+        async def async_post_call_streaming_hook(self, user_api_key_dict, response):
+            seen[self.guardrail_name] = seen.get(self.guardrail_name, 0) + 1
+            return None
+
+    managed = RecordingGuardrail(
+        guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=True
+    )
+    free = RecordingGuardrail(
+        guardrail_name="gr-free", event_hook=GuardrailEventHooks.post_call, default_on=True
+    )
+    monkeypatch.setattr(litellm, "callbacks", [managed, free])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+
+    result = await proxy_logging.async_post_call_streaming_hook(
+        data=data,
+        response=_stream_chunks()[0],
+        user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+    )
+
+    assert result is not None
+    assert seen.get("gr-post") is None
+    assert seen["gr-free"] == 1

From 0c1f33dff7084559cf1612a38b5dde4ea0afe9b8 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 14:19:21 -0700
Subject: [PATCH 10/25] fix(policy_engine): fail closed on content filter MASK
 steps for streaming pipelines

A litellm_content_filter step with a MASK action masks chat streams through
its own iterator hook under guardrails.add, which pipeline-managed guardrails
skip, so the pipeline path released the stream unmasked. CustomGuardrail now
declares rewrites_streamed_output (mask_response_content by default, any MASK
action for the content filter) and the upfront streaming check names such
steps in the same 400 it gives mask_response_content and incremental_diff
---
 litellm/integrations/custom_guardrail.py      |  3 ++
 .../litellm_content_filter/content_filter.py  |  7 ++++
 litellm/proxy/utils.py                        | 14 ++++----
 .../content_filter/test_content_filter.py     | 36 +++++++++++++++++++
 .../proxy_logging/test_guardrail_pipeline.py  | 34 +++++++++++++++++-
 5 files changed, 86 insertions(+), 8 deletions(-)

diff --git a/litellm/integrations/custom_guardrail.py b/litellm/integrations/custom_guardrail.py
index 8dc6881d23e..a6e3d000120 100644
--- a/litellm/integrations/custom_guardrail.py
+++ b/litellm/integrations/custom_guardrail.py
@@ -762,6 +762,9 @@ class CustomGuardrail(CustomLogger):
     def uses_apply_guardrail_interface(self) -> bool:
         return type(self).apply_guardrail is not CustomGuardrail.apply_guardrail
 
+    def rewrites_streamed_output(self) -> bool:
+        return self.mask_response_content
+
     def _deployment_pre_call_target(self) -> "CustomLogger":
         if not self.uses_apply_guardrail_interface() or self.use_native_lifecycle_hooks:
             return self
diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
index 722f96ef814..bd31882841e 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
@@ -1947,6 +1947,13 @@ class ContentFilterGuardrail(CustomGuardrail):
                 exception_str=exception_str,
             )
 
+    def rewrites_streamed_output(self) -> bool:
+        return (
+            super().rewrites_streamed_output()
+            or any(entry["action"] == ContentFilterAction.MASK for entry in self.compiled_patterns)
+            or any(action == ContentFilterAction.MASK for action, _ in self.blocked_words.values())
+        )
+
     async def async_post_call_streaming_iterator_hook(
         self,
         user_api_key_dict: UserAPIKeyAuth,
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index a64b57c1388..138f272af42 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -497,7 +497,7 @@ def _pipeline_step_rewrites_streamed_content(guardrail_name: str) -> bool:
     if callback is None:
         return False
     transform_mode: Final = unified_guardrail.resolve_streaming_flag(callback, "streaming_transform_mode", "block_only")
-    return callback.mask_response_content or transform_mode == "incremental_diff"
+    return callback.rewrites_streamed_output() or transform_mode == "incremental_diff"
 
 
 class _PipelineErrorBody(TypedDict):
@@ -518,10 +518,10 @@ def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_ap
     Background responses skip the post_call hooks entirely, so a pipeline
     governing one would silently never execute. Streaming responses execute
     pipelines against the buffered stream through the endpoint guardrail
-    translation of the request route, releasing the buffered chunks verbatim
-    on allow. That needs every step's guardrail to support the unified
-    apply_guardrail interface and to only allow or block (a step that rewrites
-    streamed content, via mask_response_content or
+    translation of the request route, releasing the buffered chunks on allow.
+    That needs every step's guardrail to support the unified apply_guardrail
+    interface and to only allow or block (a step that rewrites streamed
+    content, via mask_response_content, a MASK action, or
     streaming_transform_mode=incremental_diff, would have its rewrite silently
     dropped), and needs the route to have a translation at all; anything else
     keeps the 400 rather than letting ungoverned output stream through.
@@ -577,8 +577,8 @@ def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_ap
             "error": {
                 "message": (
                     "Policies with post_call guardrail pipelines cannot govern streaming responses "
-                    "because these pipeline guardrails rewrite streamed content (mask_response_content "
-                    "or streaming_transform_mode=incremental_diff), which pipeline steps would release "
+                    "because these pipeline guardrails rewrite streamed content (mask_response_content, "
+                    "a MASK action, or streaming_transform_mode=incremental_diff), which pipeline steps would release "
                     f"unmodified: {', '.join(rewriting_guardrails)}. Retry with stream=false, or drop "
                     "them from the pipeline steps so guardrails.add applies them to streamed output."
                 ),
diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
index be55ac47bde..ffbedfa43ff 100644
--- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
+++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
@@ -3068,3 +3068,39 @@ class TestContentFilterToolCallArguments:
                 request_data={},
                 input_type="response",
             )
+
+
+class TestRewritesStreamedOutput:
+    def test_block_only_rules_do_not_rewrite(self):
+        guardrail = ContentFilterGuardrail(
+            guardrail_name="cf",
+            patterns=[ContentFilterPattern(pattern_type="prebuilt", pattern_name="us_ssn", action=ContentFilterAction.BLOCK)],
+            blocked_words=[BlockedWord(keyword="kumquat", action=ContentFilterAction.BLOCK)],
+        )
+
+        assert guardrail.rewrites_streamed_output() is False
+
+    def test_mask_blocked_word_rewrites(self):
+        guardrail = ContentFilterGuardrail(
+            guardrail_name="cf",
+            blocked_words=[BlockedWord(keyword="persimmon", action=ContentFilterAction.MASK)],
+        )
+
+        assert guardrail.rewrites_streamed_output() is True
+
+    def test_mask_pattern_rewrites(self):
+        guardrail = ContentFilterGuardrail(
+            guardrail_name="cf",
+            patterns=[ContentFilterPattern(pattern_type="prebuilt", pattern_name="us_ssn", action=ContentFilterAction.MASK)],
+        )
+
+        assert guardrail.rewrites_streamed_output() is True
+
+    def test_mask_response_content_rewrites(self):
+        guardrail = ContentFilterGuardrail(
+            guardrail_name="cf",
+            blocked_words=[BlockedWord(keyword="kumquat", action=ContentFilterAction.BLOCK)],
+            mask_response_content=True,
+        )
+
+        assert guardrail.rewrites_streamed_output() is True
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index a258442c79a..ee0a9a10172 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -27,7 +27,8 @@ from litellm.integrations.prometheus import PrometheusLogger
 from litellm.proxy._types import ProxyException
 from litellm.proxy.common_utils.callback_utils import add_guardrail_to_applied_guardrails_header
 from litellm.proxy.utils import ProxyLogging, _raise_for_streaming_post_call_pipelines
-from litellm.types.guardrails import GuardrailEventHooks
+from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import ContentFilterGuardrail
+from litellm.types.guardrails import BlockedWord, ContentFilterAction, GuardrailEventHooks
 from litellm.types.proxy.policy_engine.pipeline_types import (
     GuardrailPipeline,
     PipelineStep,
@@ -1461,6 +1462,37 @@ async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_rewrites_
     assert seen.get("count") is None
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("action, rejected", [(ContentFilterAction.MASK, True), (ContentFilterAction.BLOCK, False)])
+async def test_pre_call_hook_rejects_streaming_only_when_content_filter_step_masks(
+    proxy_logging, make_user_api_key_auth, monkeypatch, action, rejected
+):
+    guardrail = ContentFilterGuardrail(
+        guardrail_name="gr-post",
+        event_hook=GuardrailEventHooks.post_call,
+        blocked_words=[BlockedWord(keyword="persimmon", action=action)],
+    )
+    monkeypatch.setattr(litellm, "callbacks", [guardrail])
+    data = _post_call_pipeline_data(stream=True)
+    user_api_key_dict = make_user_api_key_auth(request_route="/v1/chat/completions")
+
+    if not rejected:
+        out = await proxy_logging.pre_call_hook(
+            user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
+        )
+        assert out is not None and out.get("stream") is True
+        return
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
+    assert "a MASK action" in info.value.detail["error"]["message"]
+
+
 @pytest.mark.asyncio
 async def test_pre_call_hook_rejects_streaming_when_route_has_no_guardrail_translation(
     proxy_logging, make_user_api_key_auth, monkeypatch

From 90c8031dd76c5565c25b2adfe301a4ce715a6964 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 14:32:35 -0700
Subject: [PATCH 11/25] fix(policy_engine): fail closed on content filter
 category MASK steps for streaming pipelines

---
 .../litellm_content_filter/content_filter.py  |  2 ++
 .../content_filter/test_content_filter.py     | 20 +++++++++++++++++
 .../proxy_logging/test_guardrail_pipeline.py  | 22 +++++++++++++++++++
 3 files changed, 44 insertions(+)

diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
index bd31882841e..85eb50c78e7 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
@@ -1952,6 +1952,8 @@ class ContentFilterGuardrail(CustomGuardrail):
             super().rewrites_streamed_output()
             or any(entry["action"] == ContentFilterAction.MASK for entry in self.compiled_patterns)
             or any(action == ContentFilterAction.MASK for action, _ in self.blocked_words.values())
+            or any(action == ContentFilterAction.MASK for _, _, action in self.category_keywords.values())
+            or any(action == ContentFilterAction.MASK for _, _, action in self.always_block_category_keywords.values())
         )
 
     async def async_post_call_streaming_iterator_hook(
diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
index ffbedfa43ff..73020fe3e6f 100644
--- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
+++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
@@ -3104,3 +3104,23 @@ class TestRewritesStreamedOutput:
         )
 
         assert guardrail.rewrites_streamed_output() is True
+
+    @pytest.mark.parametrize("action, expected", [("MASK", True), ("BLOCK", False)])
+    def test_category_keywords_follow_the_category_action(self, action, expected):
+        guardrail = ContentFilterGuardrail(
+            guardrail_name="cf",
+            categories=[{"category": "bias_gender", "enabled": True, "action": action}],
+        )
+
+        assert guardrail.category_keywords and not guardrail.always_block_category_keywords
+        assert guardrail.rewrites_streamed_output() is expected
+
+    @pytest.mark.parametrize("action, expected", [("MASK", True), ("BLOCK", False)])
+    def test_always_block_category_keywords_follow_the_category_action(self, action, expected):
+        guardrail = ContentFilterGuardrail(
+            guardrail_name="cf",
+            categories=[{"category": "age_discrimination", "enabled": True, "action": action}],
+        )
+
+        assert guardrail.always_block_category_keywords and not guardrail.category_keywords
+        assert guardrail.rewrites_streamed_output() is expected
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index ee0a9a10172..895a986f348 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -1493,6 +1493,28 @@ async def test_pre_call_hook_rejects_streaming_only_when_content_filter_step_mas
     assert "a MASK action" in info.value.detail["error"]["message"]
 
 
+@pytest.mark.asyncio
+async def test_pre_call_hook_rejects_streaming_when_content_filter_category_masks(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    guardrail = ContentFilterGuardrail(
+        guardrail_name="gr-post",
+        event_hook=GuardrailEventHooks.post_call,
+        categories=[{"category": "bias_gender", "enabled": True, "action": "MASK"}],
+    )
+    monkeypatch.setattr(litellm, "callbacks", [guardrail])
+    data = _post_call_pipeline_data(stream=True)
+    user_api_key_dict = make_user_api_key_auth(request_route="/v1/chat/completions")
+
+    with pytest.raises(HTTPException) as info:
+        await proxy_logging.pre_call_hook(
+            user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
+        )
+
+    assert info.value.status_code == 400
+    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
+
+
 @pytest.mark.asyncio
 async def test_pre_call_hook_rejects_streaming_when_route_has_no_guardrail_translation(
     proxy_logging, make_user_api_key_auth, monkeypatch

From 2247fbc66df9769c3e3661b457b0b132513c1bd5 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 15:44:08 -0700
Subject: [PATCH 12/25] fix(policy_engine): withhold streams when a pipeline
 guardrail rewrites output at runtime

---
 .../proxy/policy_engine/pipeline_executor.py  | 114 +++++++++++++++---
 litellm/proxy/utils.py                        |  60 ++++++---
 .../policy_engine/test_pipeline_executor.py   |  63 +++++++++-
 .../proxy_logging/test_guardrail_pipeline.py  | 103 +++++++++++++++-
 4 files changed, 302 insertions(+), 38 deletions(-)

diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index c422a7c0964..acd2c2c973a 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -6,8 +6,11 @@ pass/fail actions (allow, block, next, modify_response) and data forwarding.
 """
 
 import time
+from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Any, Final, Literal
 
+from pydantic import BaseModel
+
 import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.integrations.custom_guardrail import (
@@ -24,8 +27,10 @@ from litellm.types.proxy.policy_engine.pipeline_types import (
     PipelineStep,
     PipelineStepResult,
 )
+from litellm.types.utils import GenericGuardrailAPIInputs
 
 if TYPE_CHECKING:
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
     from litellm.llms.base_llm.guardrail_translation.base_translation import (
         BaseTranslation,
     )
@@ -36,6 +41,90 @@ except ImportError:
     HTTPException = None
 
 
+class UndeliverableStreamRewrite(Exception):
+    def __init__(self, guardrail_name: str) -> None:
+        super().__init__(
+            f"Guardrail '{guardrail_name}' rewrote the streamed response, which streaming pipelines cannot deliver"
+        )
+        self.guardrail_name: Final = guardrail_name
+
+
+def _tool_call_shape(tool_call: object) -> tuple[object, object]:
+    plain: Final = tool_call.model_dump() if isinstance(tool_call, BaseModel) else tool_call
+    function: Final = plain.get("function") if isinstance(plain, Mapping) else None
+    if not isinstance(function, Mapping):
+        return (None, None)
+    return (function.get("name"), function.get("arguments"))
+
+
+def _rewrote_texts(sent: Sequence[str] | None, returned: Sequence[str] | None) -> bool:
+    return sent is not None and returned is not None and list(returned) != list(sent)
+
+
+def _rewrote_tool_calls(sent: Sequence[object] | None, returned: Sequence[object] | None) -> bool:
+    if sent is None or returned is None:
+        return False
+    return [_tool_call_shape(tool_call) for tool_call in returned] != [
+        _tool_call_shape(tool_call) for tool_call in sent
+    ]
+
+
+class _StreamRewriteObserver(CustomGuardrail):
+    """Stand-in handed to the endpoint translation in place of a streaming pipeline step's
+    guardrail. Translations cannot rewrite every buffered chunk consistently, so the gate
+    withholds the stream whenever the guardrail returned different output than it was given,
+    which for guardrails like Bedrock's ANONYMIZED action is only known at runtime."""
+
+    def __init__(self, inner: CustomGuardrail) -> None:
+        super().__init__(guardrail_name=inner.guardrail_name)
+        self.inner: Final = inner
+        self.rewrote = False
+
+    def structured_messages_cover_full_request(self) -> bool:
+        return self.inner.structured_messages_cover_full_request()
+
+    async def apply_guardrail(
+        self,
+        inputs: GenericGuardrailAPIInputs,
+        request_data: dict,  # mutable-ok: matches CustomGuardrail.apply_guardrail
+        input_type: Literal["request", "response"],
+        logging_obj: "LiteLLMLoggingObj | None" = None,
+    ) -> GenericGuardrailAPIInputs:
+        outputs: Final = await self.inner.apply_guardrail(
+            inputs=inputs, request_data=request_data, input_type=input_type, logging_obj=logging_obj
+        )
+        self.rewrote = (
+            self.rewrote
+            or _rewrote_texts(inputs.get("texts"), outputs.get("texts"))
+            or _rewrote_tool_calls(inputs.get("tool_calls"), outputs.get("tool_calls"))
+        )
+        return outputs
+
+
+def _prepare_hook_input(
+    step: PipelineStep,
+    callback: CustomLogger,
+    data: dict,  # mutable-ok: same request-payload shape the hooks mutate
+    raw_request_snapshot: dict | None,  # mutable-ok: same request-payload shape as data
+) -> tuple[dict, bool]:  # mutable-ok: returns that same request-payload dict
+    """Inject the step's guardrail name into metadata so should_run_guardrail() allows it,
+    and pick the payload the step scans: a scan_raw_request step evaluates the pristine
+    pre-pipeline snapshot instead of `data` (which earlier pass_data steps in this same
+    pipeline may have already rewritten), same reason the normal sequential/parallel
+    guardrail loops do this."""
+    if "metadata" not in data:
+        data["metadata"] = {}
+    data["metadata"]["guardrails"] = [step.guardrail]
+
+    scans_raw_request: Final = getattr(callback, "scan_raw_request", False)
+    hook_input: Final[dict] = (  # mutable-ok: same request-payload shape as data
+        independent_snapshot(raw_request_snapshot) if scans_raw_request and raw_request_snapshot is not None else data
+    )
+    if hook_input is not data:
+        hook_input.setdefault("metadata", {})["guardrails"] = [step.guardrail]
+    return hook_input, scans_raw_request
+
+
 class PipelineExecutor:
     """Executes guardrail pipelines with ordered, conditional step logic."""
 
@@ -195,23 +284,7 @@ class PipelineExecutor:
             return ("error", None, f"Guardrail '{step.guardrail}' not found", None)
 
         try:
-            # Inject guardrail name into metadata so should_run_guardrail() allows it
-            if "metadata" not in data:
-                data["metadata"] = {}
-            data["metadata"]["guardrails"] = [step.guardrail]
-
-            # A scan_raw_request step evaluates the pristine pre-pipeline
-            # snapshot instead of `data` (which earlier pass_data steps in
-            # this same pipeline may have already rewritten), same reason
-            # the normal sequential/parallel guardrail loops do this.
-            scans_raw_request: Final = getattr(callback, "scan_raw_request", False)
-            hook_input: Final[dict] = (  # mutable-ok: same request-payload shape as data
-                independent_snapshot(raw_request_snapshot)
-                if scans_raw_request and raw_request_snapshot is not None
-                else data
-            )
-            if hook_input is not data:
-                hook_input.setdefault("metadata", {})["guardrails"] = [step.guardrail]
+            hook_input, scans_raw_request = _prepare_hook_input(step, callback, data, raw_request_snapshot)
 
             # Use unified_guardrail path if callback implements apply_guardrail
             target: CustomLogger = callback
@@ -239,13 +312,16 @@ class PipelineExecutor:
                         f"Guardrail '{step.guardrail}' does not support streaming pipeline execution",
                         None,
                     )
+                observer: Final = _StreamRewriteObserver(callback)
                 await endpoint_translation.process_output_streaming_response(
                     responses_so_far=streaming_chunks,
-                    guardrail_to_apply=callback,
+                    guardrail_to_apply=observer,
                     litellm_logging_obj=data.get("litellm_logging_obj"),
                     user_api_key_dict=user_api_key_dict,
                     request_data=hook_input,
                 )
+                if observer.rewrote:
+                    raise UndeliverableStreamRewrite(step.guardrail)
                 response = None
             elif mode == "post_call":
                 response = await target.async_post_call_success_hook(
@@ -269,6 +345,8 @@ class PipelineExecutor:
                 return ("pass", {"response": response}, None, None)
             return ("pass", response if isinstance(response, dict) else None, None, None)
 
+        except UndeliverableStreamRewrite:
+            raise
         except Exception as e:
             if CustomGuardrail._is_guardrail_intervention(e):
                 error_msg: Final = _extract_error_message(e)
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 138f272af42..6c990222e51 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -155,7 +155,7 @@ from litellm.proxy.hooks.sensitive_data_routing import (
 )
 from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
 from litellm.proxy.management_helpers.key_settings_audit import with_settings_updated_at
-from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor
+from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor, UndeliverableStreamRewrite
 from litellm.repositories.budget_repository import BudgetRepository
 from litellm.repositories.config_repository import ConfigRepository
 from litellm.repositories.table_repositories import (
@@ -511,6 +511,23 @@ class _PipelineErrorDetail(TypedDict):
     error: ReadOnly[_PipelineErrorBody]
 
 
+def _undeliverable_stream_rewrite_error(policy_name: str, guardrail_name: str) -> HTTPException:
+    detail: Final[_PipelineErrorDetail] = {
+        "error": {
+            "message": (
+                f"Streaming response withheld by policy pipeline '{policy_name}' because guardrail "
+                f"'{guardrail_name}' rewrote the streamed output, and streaming pipelines cannot deliver "
+                "rewrites. Retry with stream=false, or drop it from the pipeline steps so guardrails.add "
+                "applies it to streamed output."
+            ),
+            "type": "guardrail_pipeline_error",
+            "policies": (policy_name,),
+            "guardrails": (guardrail_name,),
+        }
+    }
+    return HTTPException(status_code=400, detail=detail)
+
+
 def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_api_key_dict: UserAPIKeyAuth) -> None:
     """
     Reject up front the requests whose post_call pipelines could never run.
@@ -3468,11 +3485,12 @@ class ProxyLogging:
         pipeline allows it), then runs each pipeline's steps against the
         assembled output through the endpoint guardrail translation, the same
         machinery flat post_call guardrails use at end of stream. An allow
-        releases the buffered chunks as that machinery left them (the
-        Responses and A2A translations write guardrail output back into the
-        final chunk, exactly as they do for flat guardrails); a block or
-        modify_response terminates with the translation's block chunks or the
-        raised error.
+        releases the buffered chunks verbatim; a step whose guardrail rewrote
+        the output withholds the stream with a 400 instead, since no
+        translation rewrites every buffered chunk consistently and some
+        rewrites (Bedrock's ANONYMIZED action, for one) are only decided at
+        runtime; a block or modify_response terminates with the translation's
+        block chunks or the raised error.
         """
         buffered: Final[list[object]] = []  # mutable-ok: accumulates the stream before the pipeline verdict
         async for item in response:
@@ -3495,16 +3513,26 @@ class ProxyLogging:
         call_type, endpoint_translation = resolved
 
         for policy_name, pipeline in pipelines:
-            result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
-                steps=pipeline.steps,
-                mode="post_call",
-                data=request_data,
-                user_api_key_dict=user_api_key_dict,
-                call_type=call_type,
-                policy_name=policy_name,
-                streaming_chunks=buffered,
-                endpoint_translation=endpoint_translation,
-            )
+            try:
+                result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
+                    steps=pipeline.steps,
+                    mode="post_call",
+                    data=request_data,
+                    user_api_key_dict=user_api_key_dict,
+                    call_type=call_type,
+                    policy_name=policy_name,
+                    streaming_chunks=buffered,
+                    endpoint_translation=endpoint_translation,
+                )
+            except UndeliverableStreamRewrite as rewrite:
+                async for error_chunk in unified_guardrail.emit_streaming_http_error(
+                    _undeliverable_stream_rewrite_error(policy_name, rewrite.guardrail_name),
+                    call_type,
+                    buffered,
+                    request_data,
+                ):
+                    yield error_chunk
+                return
             try:
                 ProxyLogging._handle_pipeline_result(
                     result, data=request_data, policy_name=policy_name, original_response=buffered
diff --git a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
index 054a5af4148..52fd8777a19 100644
--- a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
+++ b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
@@ -13,7 +13,7 @@ from litellm.integrations.custom_guardrail import CustomGuardrail
 from litellm.proxy.guardrails.guardrail_hooks.custom_code.custom_code_guardrail import (
     CustomCodeGuardrail,
 )
-from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor
+from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor, UndeliverableStreamRewrite
 from litellm.types.proxy.policy_engine.pipeline_types import (
     GuardrailPipeline,
     PipelineStep,
@@ -811,3 +811,64 @@ async def test_pipeline_step_keeps_native_hook_when_opted_out(monkeypatch):
     assert outcome == "pass"
     assert guardrail.native_pre_call_ran is True
     assert "guardrail_to_apply" not in data
+
+
+class _TextReturningGuardrail(CustomGuardrail):
+    def __init__(self, returned_texts):
+        super().__init__(guardrail_name="masker", event_hook="post_call", default_on=True)
+        self.returned_texts = returned_texts
+
+    async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+        return {**inputs, "texts": self.returned_texts}
+
+
+class _TextTranslation:
+    def __init__(self):
+        self.seen_guardrail_names = []
+
+    async def process_output_streaming_response(
+        self, responses_so_far, guardrail_to_apply, litellm_logging_obj=None, user_api_key_dict=None, request_data=None
+    ):
+        self.seen_guardrail_names.append(guardrail_to_apply.guardrail_name)
+        await guardrail_to_apply.apply_guardrail(
+            inputs={"texts": ["hello world"]},
+            request_data=request_data or {},
+            input_type="response",
+            logging_obj=litellm_logging_obj,
+        )
+        return responses_so_far
+
+
+async def _run_streaming_step(returned_texts, translation):
+    return await PipelineExecutor.execute_steps(
+        steps=[PipelineStep(guardrail="masker", on_pass="allow", on_fail="next", on_error="next")],
+        mode="post_call",
+        data={"model": "m"},
+        user_api_key_dict=MagicMock(),
+        call_type="completion",
+        policy_name="p",
+        streaming_chunks=[object()],
+        endpoint_translation=translation,
+    )
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_rewrite_escapes_execute_steps_regardless_of_step_actions(monkeypatch):
+    monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(["hello [MASKED]"])])
+    translation = _TextTranslation()
+
+    with pytest.raises(UndeliverableStreamRewrite) as info:
+        await _run_streaming_step(["hello [MASKED]"], translation)
+
+    assert info.value.guardrail_name == "masker"
+    assert translation.seen_guardrail_names == ["masker"]
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_unchanged_texts_in_another_container_allow(monkeypatch):
+    monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(("hello world",))])
+
+    result = await _run_streaming_step(("hello world",), _TextTranslation())
+
+    assert result.terminal_action == "allow"
+    assert [step.outcome for step in result.step_results] == ["pass"]
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 895a986f348..74bd2e483cc 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -11,7 +11,7 @@ from __future__ import annotations
 
 import asyncio
 import json
-from typing import Any, Dict, List
+from typing import Any, Callable, Dict, List
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
@@ -878,10 +878,12 @@ async def test_process_prompt_template_aresponses_swaps_model_and_merges_input(p
 # ---------------------------------------------------------------------------
 
 
-def _post_call_pipeline_data(guardrail: str = "gr-post", **extra: Any) -> Dict[str, Any]:
+def _post_call_pipeline_data(
+    guardrail: str = "gr-post", step: PipelineStep | None = None, **extra: Any
+) -> Dict[str, Any]:
     pipeline = GuardrailPipeline(
         mode="post_call",
-        steps=[PipelineStep(guardrail=guardrail, on_pass="allow", on_fail="block")],
+        steps=[step or PipelineStep(guardrail=guardrail, on_pass="allow", on_fail="block")],
     )
     return {
         "model": "m",
@@ -1587,6 +1589,101 @@ async def test_streaming_iterator_hook_pipeline_block_withholds_all_chunks(
     assert "output blocked" in str(info.value.detail)
 
 
+def _rewriting_stream_guardrail(transform: Callable[[Dict[str, Any]], Dict[str, Any]]) -> CustomGuardrail:
+    class RewritingStreamGuardrail(CustomGuardrail):
+        async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+            return {**inputs, **transform(inputs)}
+
+    return RewritingStreamGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)
+
+
+def _tool_call_stream_chunks() -> List[Any]:
+    tool_call = {
+        "index": 0,
+        "id": "call_1",
+        "type": "function",
+        "function": {"name": "lookup", "arguments": '{"ssn": "123"}'},
+    }
+    return [
+        litellm.ModelResponseStream(
+            choices=[{"index": 0, "delta": {"tool_calls": [tool_call]}, "finish_reason": None}]
+        ),
+        litellm.ModelResponseStream(choices=[{"index": 0, "delta": {}, "finish_reason": "tool_calls"}]),
+    ]
+
+
+def _echoed_tool_call_dicts(arguments: str) -> List[Dict[str, Any]]:
+    return [{"id": "call_1", "type": "function", "function": {"name": "lookup", "arguments": arguments}}]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("on_fail, on_error", [("block", None), ("next", "next")])
+@pytest.mark.parametrize(
+    "make_chunks, transform",
+    [
+        (_stream_chunks, lambda inputs: {"texts": ["hello [MASKED]"]}),
+        (_tool_call_stream_chunks, lambda inputs: {"tool_calls": _echoed_tool_call_dicts('{"ssn": "[MASKED]"}')}),
+    ],
+    ids=["texts", "tool_calls"],
+)
+async def test_streaming_iterator_hook_pipeline_withholds_runtime_rewrite(
+    proxy_logging, make_user_api_key_auth, monkeypatch, make_chunks, transform, on_fail, on_error
+):
+    monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    step = PipelineStep(guardrail="gr-post", on_pass="allow", on_fail=on_fail, on_error=on_error)
+    data = _post_call_pipeline_data(step=step, stream=True)
+    delivered: List[Any] = []
+
+    async def _drain() -> None:
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(make_chunks()),
+            request_data=data,
+        ):
+            delivered.append(item)
+
+    with pytest.raises(HTTPException) as info:
+        await _drain()
+
+    error = info.value.detail["error"]
+    assert delivered == []
+    assert info.value.status_code == 400
+    assert error["type"] == "guardrail_pipeline_error"
+    assert error["policies"] == ("response-governance",)
+    assert error["guardrails"] == ("gr-post",)
+    assert "stream=false" in error["message"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "make_chunks, transform",
+    [
+        (_stream_chunks, lambda inputs: {"texts": tuple(inputs["texts"])}),
+        (_tool_call_stream_chunks, lambda inputs: {"tool_calls": _echoed_tool_call_dicts('{"ssn": "123"}')}),
+    ],
+    ids=["texts_as_tuple", "tool_calls_as_dicts"],
+)
+async def test_streaming_iterator_hook_pipeline_releases_stream_echoed_in_another_shape(
+    proxy_logging, make_user_api_key_auth, monkeypatch, make_chunks, transform
+):
+    monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    chunks = make_chunks()
+
+    delivered = [
+        item
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        )
+    ]
+
+    assert [id(item) for item in delivered] == [id(chunk) for chunk in chunks]
+
+
 @pytest.mark.asyncio
 async def test_streaming_iterator_hook_pipeline_withholds_unresolvable_response_shape(
     proxy_logging, make_user_api_key_auth, monkeypatch

From badefa395cbfea283e2bac3d7ba545638a0792d0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 21:30:31 -0700
Subject: [PATCH 13/25] fix(policy_engine): snapshot guardrail inputs before
 apply_guardrail so in-place stream rewrites are withheld

---
 .../proxy/policy_engine/pipeline_executor.py  | 22 ++++++++++---------
 .../policy_engine/test_pipeline_executor.py   | 22 +++++++++++++++++++
 2 files changed, 34 insertions(+), 10 deletions(-)

diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index 21f3aca3f58..4c192a50096 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -57,16 +57,16 @@ def _tool_call_shape(tool_call: object) -> tuple[object, object]:
     return (function.get("name"), function.get("arguments"))
 
 
-def _rewrote_texts(sent: Sequence[str] | None, returned: Sequence[str] | None) -> bool:
-    return sent is not None and returned is not None and tuple(returned) != tuple(sent)
+def _text_snapshot(texts: Sequence[str] | None) -> tuple[str, ...] | None:
+    return None if texts is None else tuple(texts)
 
 
-def _rewrote_tool_calls(sent: Sequence[object] | None, returned: Sequence[object] | None) -> bool:
-    if sent is None or returned is None:
-        return False
-    return tuple(_tool_call_shape(tool_call) for tool_call in returned) != tuple(
-        _tool_call_shape(tool_call) for tool_call in sent
-    )
+def _tool_call_shapes(tool_calls: Sequence[object] | None) -> tuple[tuple[object, object], ...] | None:
+    return None if tool_calls is None else tuple(_tool_call_shape(tool_call) for tool_call in tool_calls)
+
+
+def _rewrote(sent: tuple[object, ...] | None, returned: tuple[object, ...] | None) -> bool:
+    return sent is not None and returned is not None and returned != sent
 
 
 class _StreamRewriteObserver(CustomGuardrail):
@@ -90,13 +90,15 @@ class _StreamRewriteObserver(CustomGuardrail):
         input_type: Literal["request", "response"],
         logging_obj: "LiteLLMLoggingObj | None" = None,
     ) -> GenericGuardrailAPIInputs:
+        sent_texts: Final = _text_snapshot(inputs.get("texts"))
+        sent_tool_shapes: Final = _tool_call_shapes(inputs.get("tool_calls"))
         outputs: Final = await self.inner.apply_guardrail(
             inputs=inputs, request_data=request_data, input_type=input_type, logging_obj=logging_obj
         )
         self.rewrote = (
             self.rewrote
-            or _rewrote_texts(inputs.get("texts"), outputs.get("texts"))
-            or _rewrote_tool_calls(inputs.get("tool_calls"), outputs.get("tool_calls"))
+            or _rewrote(sent_texts, _text_snapshot(outputs.get("texts")))
+            or _rewrote(sent_tool_shapes, _tool_call_shapes(outputs.get("tool_calls")))
         )
         return outputs
 
diff --git a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
index 52fd8777a19..ef5d206f1d8 100644
--- a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
+++ b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
@@ -872,3 +872,25 @@ async def test_streaming_step_unchanged_texts_in_another_container_allow(monkeyp
 
     assert result.terminal_action == "allow"
     assert [step.outcome for step in result.step_results] == ["pass"]
+
+
+class _InPlaceMutatingGuardrail(CustomGuardrail):
+    """Rewrites like bedrock/presidio do: rebinds inputs["texts"] on the dict it was handed
+    and returns that same dict, so a post-call comparison against inputs sees no change."""
+
+    def __init__(self):
+        super().__init__(guardrail_name="masker", event_hook="post_call", default_on=True)
+
+    async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+        inputs["texts"] = ["hello [MASKED]"]
+        return inputs
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_in_place_rewrite_still_withholds_stream(monkeypatch):
+    monkeypatch.setattr(litellm, "callbacks", [_InPlaceMutatingGuardrail()])
+
+    with pytest.raises(UndeliverableStreamRewrite) as info:
+        await _run_streaming_step(["hello [MASKED]"], _TextTranslation())
+
+    assert info.value.guardrail_name == "masker"

From bcee01a7a7a3a29c5f6e54a0045ff3688d2dbdef Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 21:52:52 -0700
Subject: [PATCH 14/25] fix(policy_engine): merge guardrail metadata writes
 back on block and modify_response so failure spend records keep guardrail
 cost and status

---
 .../proxy/policy_engine/pipeline_executor.py  |  2 +
 litellm/proxy/utils.py                        |  7 +++-
 .../proxy_logging/test_guardrail_pipeline.py  | 38 +++++++++++++++++++
 3 files changed, 46 insertions(+), 1 deletion(-)

diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index 4c192a50096..0c3ceb53707 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -236,6 +236,7 @@ class PipelineExecutor:
                     step_results=step_results,
                     error_message=error_detail,
                     original_exception=original_exception,
+                    modified_data=working_data if working_data != data else None,
                 )
 
             if action == "modify_response":
@@ -243,6 +244,7 @@ class PipelineExecutor:
                     terminal_action="modify_response",
                     step_results=step_results,
                     modify_response_message=step.modify_response_message or error_detail,
+                    modified_data=working_data if working_data != data else None,
                 )
 
             # action == "next" → continue to next step
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index f9d018bc452..dbcea834d50 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -1838,7 +1838,9 @@ class ProxyLogging:
         payload (already sent upstream) must stay untouched; a replacement
         response carried in ``modified_data`` is adopted by the caller, and
         metadata-bucket writes (applied guardrails, guardrail logging info)
-        are merged back so headers and spend logs still see them. On the
+        are merged back so headers and spend logs still see them, on block
+        and modify_response too, so failure spend records keep guardrail
+        cost and status. On the
         streaming path it is the buffered chunk list, carried into
         ``ModifyResponseException.original_response`` for usage reporting.
         """
@@ -1850,6 +1852,9 @@ class ProxyLogging:
                     _merge_pipeline_metadata_writes(data, result.modified_data)
             return data
 
+        if result.modified_data is not None:
+            _merge_pipeline_metadata_writes(data, result.modified_data)
+
         if result.terminal_action == "block":
             original_exception: Final = result.original_exception
             if original_exception is not None and not _exception_changes_request_flow(original_exception):
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 74bd2e483cc..d9b3578c966 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -1197,6 +1197,44 @@ async def test_post_call_pipeline_guardrail_metadata_writes_reach_request_data(
     assert slg_entries[0]["guardrail_name"] == "gr-post"
 
 
+@pytest.mark.asyncio
+async def test_post_call_pipeline_block_keeps_guardrail_metadata_writes(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    class BlockingWriterGuardrail(CustomGuardrail):
+        async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+            add_guardrail_to_applied_guardrails_header(request_data=data, guardrail_name="gr-post")
+            self.add_standard_logging_guardrail_information_to_request_data(
+                guardrail_json_response={"verdict": "fail"},
+                request_data=data,
+                guardrail_status="guardrail_intervened",
+            )
+            raise HTTPException(status_code=400, detail={"error": "output blocked"})
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [
+            BlockingWriterGuardrail(
+                guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False
+            )
+        ],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data()
+
+    with pytest.raises(HTTPException):
+        await proxy_logging.post_call_success_hook(
+            data=data, response=litellm.ModelResponse(), user_api_key_dict=make_user_api_key_auth()
+        )
+
+    assert data["metadata"]["applied_guardrails"] == ["gr-post"]
+    slg_entries = data["metadata"]["standard_logging_guardrail_information"]
+    assert len(slg_entries) == 1
+    assert slg_entries[0]["guardrail_name"] == "gr-post"
+    assert slg_entries[0]["guardrail_status"] == "guardrail_intervened"
+
+
 @pytest.mark.asyncio
 async def test_post_call_pipeline_managed_parallel_guardrail_runs_exactly_once(
     proxy_logging, make_user_api_key_auth, monkeypatch

From 673d1743a66363022777f0b3b261142ef77964ab Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 22:12:59 -0700
Subject: [PATCH 15/25] fix(policy_engine): apply post_call pipeline text
 rewrites on streams

Buffered streams governed by post_call policy pipelines now deliver text
rewrites back into the stream per surface (chat SSE, responses SSE,
anthropic messages SSE) instead of rejecting the request with a 400
upfront. Rewrites chain across pipeline steps; tool-call rewrites and
translations without stream write-back still withhold the stream.
---
 .../chat/guardrail_translation/handler.py     |  74 +++++-
 .../guardrail_translation/base_translation.py |  15 +-
 .../chat/guardrail_translation/handler.py     |  91 ++++++-
 .../guardrail_translation/handler.py          |  83 ++++++-
 .../proxy/policy_engine/pipeline_executor.py  |  90 +++++--
 litellm/proxy/utils.py                        |  61 ++---
 .../test_anthropic_guardrail_handler.py       |  65 +++++
 .../test_openai_guardrail_handler.py          |  55 +++++
 ...test_openai_responses_guardrail_handler.py |  83 +++++++
 .../policy_engine/test_pipeline_executor.py   |   2 +
 .../proxy_logging/test_guardrail_pipeline.py  | 222 ++++++++++++++----
 11 files changed, 711 insertions(+), 130 deletions(-)

diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py
index b9ca18c7843..89c8431dfe4 100644
--- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py
+++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py
@@ -13,9 +13,10 @@ Pattern Overview:
 """
 
 import json
-from collections.abc import Mapping, Sequence
+from collections.abc import Iterator, Mapping, Sequence
 from copy import deepcopy
 from dataclasses import dataclass
+from itertools import chain, repeat
 from typing import TYPE_CHECKING, Any, Final, cast
 
 from typing_extensions import assert_never
@@ -120,6 +121,8 @@ class AnthropicMessagesHandler(BaseTranslation):
     them through guardrail rewrites; downstream provider handling is out of scope.
     """
 
+    delivers_ended_stream_text_rewrites = True
+
     def __init__(self):
         super().__init__()
         self.adapter = LiteLLMAnthropicMessagesAdapter()
@@ -931,11 +934,14 @@ class AnthropicMessagesHandler(BaseTranslation):
         litellm_logging_obj: "LiteLLMLoggingObj | None" = None,
         user_api_key_dict: "UserAPIKeyAuth | None" = None,
         request_data: dict | None = None,
+        deliver_ended_stream_rewrites: bool = False,
     ) -> list[Any]:
         """
         Process output streaming response by applying guardrails to text content.
 
         Get the string so far, check the apply guardrail to the string so far, and return the list of responses so far.
+        With ``deliver_ended_stream_rewrites``, an ended stream whose guardrail rewrote the text gets the rewrite
+        written back across the buffered chunks (full rewritten text in the first ``text_delta``, the rest blanked).
         """
         from litellm.integrations.custom_guardrail import ModifyResponseException
 
@@ -982,6 +988,15 @@ class AnthropicMessagesHandler(BaseTranslation):
                             responses_so_far, request_data
                         )
                     raise
+                guardrailed_texts: Final = _guardrailed_inputs.get("texts")
+                if (
+                    deliver_ended_stream_rewrites
+                    and isinstance(string_so_far, str)
+                    and string_so_far
+                    and guardrailed_texts
+                    and guardrailed_texts[0] != string_so_far
+                ):
+                    self._write_ended_stream_text_rewrite(responses_so_far, guardrailed_texts[0])
             else:
                 verbose_proxy_logger.debug("Skipping output guardrail - model response has no choices")
             return responses_so_far
@@ -1093,6 +1108,63 @@ class AnthropicMessagesHandler(BaseTranslation):
             inputs["model"] = response_model
         return inputs
 
+    @staticmethod
+    def _write_ended_stream_text_rewrite(
+        responses_so_far: list[Any],  # mutable-ok: rewrites the caller's buffered chunks in place
+        rewritten_text: str,
+    ) -> None:
+        """Deliver an ended-stream guardrail text rewrite by rewriting the
+        buffered chunks in place: the first ``text_delta`` carries the full
+        rewritten text and every later one is blanked, leaving the surrounding
+        message and content-block framing untouched. Handles both chunk formats
+        this stream carries (parsed event dicts and raw SSE bytes)."""
+        replacements: Final = chain((rewritten_text,), repeat(""))
+        for idx, item in enumerate(responses_so_far):
+            if isinstance(item, dict):
+                delta = item.get("delta")
+                if item.get("type") == "content_block_delta" and isinstance(delta, dict):
+                    if delta.get("type") == "text_delta":
+                        delta["text"] = next(replacements)
+            elif isinstance(item, (bytes, bytearray)):
+                responses_so_far[idx] = (  # rebind-ok: delivers the rewrite into the caller's buffer
+                    AnthropicMessagesHandler._rewrite_sse_text_deltas(bytes(item), replacements)
+                )
+
+    @staticmethod
+    def _rewrite_sse_text_deltas(sse_bytes: bytes, replacements: "Iterator[str]") -> bytes:
+        """Rewrite every ``text_delta`` data line in one SSE chunk with the next
+        replacement text, leaving all other events and framing byte-identical."""
+        try:
+            decoded: Final = sse_bytes.decode("utf-8")
+        except UnicodeDecodeError:
+            return sse_bytes
+        return "\n\n".join(
+            AnthropicMessagesHandler._rewrite_sse_block(block, replacements) for block in decoded.split("\n\n")
+        ).encode("utf-8")
+
+    @staticmethod
+    def _rewrite_sse_block(block: str, replacements: "Iterator[str]") -> str:
+        return "\n".join(AnthropicMessagesHandler._rewrite_sse_line(line, replacements) for line in block.split("\n"))
+
+    @staticmethod
+    def _rewrite_sse_line(line: str, replacements: "Iterator[str]") -> str:
+        if not line.startswith("data:"):
+            return line
+        try:
+            data: Final[str | int | float | bool | None | Sequence[object] | Mapping[str, object]] = json.loads(
+                line[len("data:") :].strip()
+            )
+        except json.JSONDecodeError:
+            return line
+        if not isinstance(data, dict) or data.get("type") != "content_block_delta":
+            return line
+        delta: Final = data.get("delta")
+        if not isinstance(delta, dict) or delta.get("type") != "text_delta":
+            return line
+        return "data: " + json.dumps(
+            {**data, "delta": {**delta, "text": next(replacements)}}  # mutable-ok: json.dumps needs plain dicts
+        )
+
     def get_streaming_string_so_far(self, responses_so_far: list[Any]) -> str:
         """
         Parse streaming responses and extract accumulated text content.
diff --git a/litellm/llms/base_llm/guardrail_translation/base_translation.py b/litellm/llms/base_llm/guardrail_translation/base_translation.py
index ba96ab3dc99..4b0cc0fd97c 100644
--- a/litellm/llms/base_llm/guardrail_translation/base_translation.py
+++ b/litellm/llms/base_llm/guardrail_translation/base_translation.py
@@ -1,6 +1,6 @@
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Any, Final, Optional
+from typing import TYPE_CHECKING, Any, ClassVar, Final, Optional
 
 if TYPE_CHECKING:
     from litellm.integrations.custom_guardrail import (
@@ -33,6 +33,13 @@ class StreamTransformSink:
 
 
 class BaseTranslation(ABC):
+    delivers_ended_stream_text_rewrites: ClassVar[bool] = False
+    """Whether ``process_output_streaming_response`` accepts
+    ``deliver_ended_stream_rewrites=True`` and, on an ended (fully buffered)
+    stream, writes guardrail text rewrites back across ``responses_so_far`` so
+    a buffered pipeline can release rewritten chunks instead of withholding the
+    stream. Tool-call rewrites stay undeliverable everywhere."""
+
     @staticmethod
     def transform_user_api_key_dict_to_metadata(
         user_api_key_dict: Any | None,
@@ -113,6 +120,7 @@ class BaseTranslation(ABC):
         user_api_key_dict: Optional["UserAPIKeyAuth"] = None,
         request_data: dict | None = None,
         stream_transform_sink: StreamTransformSink | None = None,
+        deliver_ended_stream_rewrites: bool = False,
     ) -> Any:
         """
         Process output streaming response with guardrails.
@@ -120,6 +128,11 @@ class BaseTranslation(ABC):
         Optional to override in subclasses. ``stream_transform_sink`` is the
         out-parameter used by handlers that support streaming text
         transformations (see ``StreamTransformSink``); base handlers ignore it.
+        ``deliver_ended_stream_rewrites`` is passed True only when the caller
+        holds the whole buffered stream and the subclass declares
+        ``delivers_ended_stream_text_rewrites``: the handler then writes
+        guardrail text rewrites back across ``responses_so_far`` instead of
+        discarding them.
         """
         return responses_so_far
 
diff --git a/litellm/llms/openai/chat/guardrail_translation/handler.py b/litellm/llms/openai/chat/guardrail_translation/handler.py
index 54673c77f80..1358cf7c37a 100644
--- a/litellm/llms/openai/chat/guardrail_translation/handler.py
+++ b/litellm/llms/openai/chat/guardrail_translation/handler.py
@@ -61,6 +61,8 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
     Methods can be overridden to customize behavior for different message formats.
     """
 
+    delivers_ended_stream_text_rewrites = True
+
     def get_structured_messages(self, data: dict) -> list[AllMessageValues] | None:
         """
         Convert chat completions request data to OpenAI-spec structured messages.
@@ -440,6 +442,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         user_api_key_dict: Any | None = None,
         request_data: dict | None = None,
         stream_transform_sink: StreamTransformSink | None = None,
+        deliver_ended_stream_rewrites: bool = False,
     ) -> list["ModelResponseStream"]:
         """
         Process output streaming responses by applying guardrails to text content.
@@ -454,6 +457,10 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
                 accumulated text (``responses_so_far`` is left untouched so it stays
                 a correct raw accumulator across rounds) and the guardrailed text
                 plus requested holdback are reported per choice on the sink.
+            deliver_ended_stream_rewrites: When True and the buffered stream has
+                ended, guardrail text rewrites are written back across
+                ``responses_so_far`` (full rewritten text in each choice's first
+                content-carrying chunk, the rest blanked) instead of discarded.
 
         Returns:
             The (unmodified) list of responses.
@@ -479,6 +486,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
             litellm_logging_obj=litellm_logging_obj,
             user_api_key_dict=user_api_key_dict,
             request_data=request_data,
+            deliver_ended_stream_rewrites=deliver_ended_stream_rewrites,
         )
 
     async def _process_streaming_block_only(
@@ -489,10 +497,12 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         litellm_logging_obj: "LiteLLMLoggingObj | None",
         user_api_key_dict: Any | None,
         request_data: dict | None,
+        deliver_ended_stream_rewrites: bool = False,
     ) -> list["ModelResponseStream"]:
         """Block-only streaming path: run the guardrail so an in-flight BLOCK can
         terminate the stream. Text rewrites are not propagated to the client here
-        (see ``_process_streaming_transform`` for the incremental_diff path)."""
+        (see ``_process_streaming_transform`` for the incremental_diff path) unless
+        ``deliver_ended_stream_rewrites`` opts the ended-stream branch in."""
         # check if the stream has ended
         has_stream_ended = False
         for chunk in responses_so_far:
@@ -501,20 +511,14 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
                 break
 
         if has_stream_ended:
-            # convert to model response
-            model_response: Final = cast(
-                ModelResponse,
-                stream_chunk_builder(chunks=responses_so_far, logging_obj=litellm_logging_obj),
-            )
-            # run process_output_response
-            await self.process_output_response(
-                response=model_response,
+            await self._process_ended_stream(
+                responses_so_far=responses_so_far,
                 guardrail_to_apply=guardrail_to_apply,
                 litellm_logging_obj=litellm_logging_obj,
                 user_api_key_dict=user_api_key_dict,
                 request_data=request_data,
+                deliver_ended_stream_rewrites=deliver_ended_stream_rewrites,
             )
-
             return responses_so_far
 
         # Step 0: Check if any response has text content to process
@@ -591,6 +595,38 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
 
         return responses_so_far
 
+    async def _process_ended_stream(
+        self,
+        *,
+        responses_so_far: list["ModelResponseStream"],  # mutable-ok: rewrites the caller's buffered chunks in place
+        guardrail_to_apply: "CustomGuardrail",
+        litellm_logging_obj: "LiteLLMLoggingObj | None",
+        user_api_key_dict: object,
+        request_data: dict[str, object] | None,  # mutable-ok: same request-payload shape the hooks take
+        deliver_ended_stream_rewrites: bool,
+    ) -> None:
+        """Ended-stream path: rebuild the full response, run the non-streaming
+        output guardrail against it, and (when opted in) write any text rewrite
+        back across the buffered chunks."""
+        model_response: Final = cast(
+            ModelResponse,
+            stream_chunk_builder(chunks=responses_so_far, logging_obj=litellm_logging_obj),
+        )
+        pre_guardrail_texts: Final = self._string_choice_contents(model_response)
+        await self.process_output_response(
+            response=model_response,
+            guardrail_to_apply=guardrail_to_apply,
+            litellm_logging_obj=litellm_logging_obj,
+            user_api_key_dict=user_api_key_dict,
+            request_data=request_data,
+        )
+        if deliver_ended_stream_rewrites:
+            await self._write_ended_stream_text_rewrites(
+                responses_so_far=responses_so_far,
+                guardrailed_response=model_response,
+                pre_guardrail_texts=pre_guardrail_texts,
+            )
+
     @staticmethod
     def _accumulate_string_content_by_choice_index(
         responses_so_far: list["ModelResponseStream"],
@@ -922,6 +958,41 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
                             if "name" in func_dict:
                                 existing_tool_call.function.name = func_dict["name"]
 
+    @staticmethod
+    def _string_choice_contents(response: "ModelResponse") -> tuple[str | None, ...]:
+        return tuple(
+            choice.message.content if isinstance(choice.message.content, str) else None for choice in response.choices
+        )
+
+    async def _write_ended_stream_text_rewrites(
+        self,
+        responses_so_far: list["ModelResponseStream"],  # mutable-ok: rewrites the caller's buffered chunks in place
+        guardrailed_response: "ModelResponse",
+        pre_guardrail_texts: tuple[str | None, ...],
+    ) -> None:
+        """Write ended-stream guardrail text rewrites back across the buffered
+        chunks: each rewritten choice's full text lands in its first
+        content-carrying chunk and the rest are blanked, the same shape the
+        in-flight write-back uses. Chunks carrying only finish_reason or usage
+        stay untouched."""
+        post_guardrail_texts: Final = self._string_choice_contents(guardrailed_response)
+        changed: Final = tuple(
+            (choice_idx, after)
+            for choice_idx, (before, after) in enumerate(zip(pre_guardrail_texts, post_guardrail_texts))
+            if before is not None and after is not None and after != before
+        )
+        if not changed:
+            return
+        await self._apply_guardrail_responses_to_output_streaming(
+            responses=responses_so_far,
+            guardrailed_texts=[
+                after for _choice_idx, after in changed
+            ],  # mutable-ok: the callee's signature predates this change and takes lists
+            task_mappings=[
+                (choice_idx, None) for choice_idx, _after in changed
+            ],  # mutable-ok: the callee's signature predates this change and takes lists
+        )
+
     async def _apply_guardrail_responses_to_output_streaming(
         self,
         responses: list["ModelResponseStream"],
diff --git a/litellm/llms/openai/responses/guardrail_translation/handler.py b/litellm/llms/openai/responses/guardrail_translation/handler.py
index 7c5d8ac99ad..0f475aa04c8 100644
--- a/litellm/llms/openai/responses/guardrail_translation/handler.py
+++ b/litellm/llms/openai/responses/guardrail_translation/handler.py
@@ -28,7 +28,9 @@ Output: response.output is List[GenericResponseOutputItem] where each has:
     - text: str
 """
 
-from collections.abc import Sequence
+from collections.abc import Mapping, Sequence
+from itertools import chain, repeat
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Union, cast
 
 from openai.types.responses.response_function_tool_call import ResponseFunctionToolCall
@@ -91,6 +93,8 @@ class OpenAIResponsesHandler(BaseTranslation):
     Methods can be overridden to customize behavior for different message formats.
     """
 
+    delivers_ended_stream_text_rewrites = True
+
     def get_structured_messages(self, data: dict) -> list[AllMessageValues] | None:
         """
         Convert Responses API request data to OpenAI-spec structured messages.
@@ -482,6 +486,7 @@ class OpenAIResponsesHandler(BaseTranslation):
         litellm_logging_obj: "LiteLLMLoggingObj | None" = None,
         user_api_key_dict: "UserAPIKeyAuth | None" = None,
         request_data: dict | None = None,
+        deliver_ended_stream_rewrites: bool = False,
     ) -> list[Any]:
         """
         Process output streaming response by applying guardrails to text content.
@@ -493,7 +498,11 @@ class OpenAIResponsesHandler(BaseTranslation):
         For ``response.completed`` events (the normal end-of-stream signal) we
         use the same per-item extraction + task-mapping approach as
         ``process_output_response`` so that unmasking / blocking works correctly
-        for every output item.
+        for every output item. With ``deliver_ended_stream_rewrites`` the earlier
+        text-carrying events (``response.output_text.delta`` / ``.done``,
+        ``response.content_part.done``, ``response.output_item.done``) are synced
+        to the rewritten completed response too, so a client reading deltas sees
+        the rewrite instead of the raw model output.
         """
         if not responses_so_far:
             return responses_so_far
@@ -562,6 +571,19 @@ class OpenAIResponsesHandler(BaseTranslation):
                     responses=guardrailed_texts,
                     task_mappings=task_mappings,
                 )
+                if deliver_ended_stream_rewrites:
+                    rewrites_by_position: Final = MappingProxyType(
+                        {
+                            task_mappings[task_idx]: rewritten
+                            for task_idx, rewritten in enumerate(guardrailed_texts)
+                            if task_idx < len(texts_to_check) and rewritten != texts_to_check[task_idx]
+                        }
+                    )
+                    if rewrites_by_position:
+                        self._sync_stream_events_with_rewrites(
+                            stream_events=responses_so_far[:-1],
+                            rewrites_by_position=rewrites_by_position,
+                        )
 
             return responses_so_far
 
@@ -607,6 +629,63 @@ class OpenAIResponsesHandler(BaseTranslation):
             )
         return responses_so_far
 
+    @staticmethod
+    def _write_event_field(event: object, field: str, value: str) -> None:
+        if isinstance(event, dict):
+            event[field] = value  # rebind-ok: delivering the rewrite means editing the buffered event in place
+        else:
+            setattr(event, field, value)
+
+    def _sync_stream_events_with_rewrites(
+        self,
+        stream_events: Sequence[Any],
+        rewrites_by_position: Mapping[tuple[int, int], str],
+    ) -> None:
+        """Sync pre-completion stream events with the rewritten completed
+        response, keyed by ``(output_index, content_index)``: the first
+        ``output_text.delta`` for a rewritten item carries the full rewritten
+        text and the rest are blanked, while ``output_text.done``,
+        ``content_part.done``, and ``output_item.done`` events carry the full
+        rewritten text, so every event a client may read agrees with the
+        rewritten ``response.completed`` payload."""
+        delta_replacements: Final = MappingProxyType(
+            {position: chain((rewritten,), repeat("")) for position, rewritten in rewrites_by_position.items()}
+        )
+        for event in stream_events:
+            if not (isinstance(event, dict) or hasattr(event, "get")):
+                continue
+            event_type = event.get("type")
+            output_index = event.get("output_index")
+            content_index = event.get("content_index")
+            if event_type == "response.output_item.done" and isinstance(output_index, int):
+                self._sync_output_item_done_event(event.get("item"), output_index, rewrites_by_position)
+                continue
+            if not isinstance(output_index, int) or not isinstance(content_index, int):
+                continue
+            position = (output_index, content_index)
+            if event_type == "response.output_text.delta" and position in delta_replacements:
+                self._write_event_field(event, "delta", next(delta_replacements[position]))
+            elif event_type == "response.output_text.done" and position in rewrites_by_position:
+                self._write_event_field(event, "text", rewrites_by_position[position])
+            elif event_type == "response.content_part.done" and position in rewrites_by_position:
+                part = event.get("part")
+                if isinstance(part, dict) or hasattr(part, "text"):
+                    self._write_event_field(part, "text", rewrites_by_position[position])
+
+    @staticmethod
+    def _sync_output_item_done_event(
+        item: object,
+        output_index: int,
+        rewrites_by_position: Mapping[tuple[int, int], str],
+    ) -> None:
+        content: Final = item.get("content") if isinstance(item, dict) else getattr(item, "content", None)
+        if not isinstance(content, list):
+            return
+        for (item_idx, content_idx), rewritten in rewrites_by_position.items():
+            if item_idx != output_index or content_idx >= len(content):
+                continue
+            OpenAIResponsesHandler._write_event_field(content[content_idx], "text", rewritten)
+
     def _check_streaming_has_ended(self, responses_so_far: Sequence[ResponsesStreamChunk]) -> bool:
         """
         Check if the streaming has ended.
diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index acd2c2c973a..1264cdfc14a 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -34,6 +34,7 @@ if TYPE_CHECKING:
     from litellm.llms.base_llm.guardrail_translation.base_translation import (
         BaseTranslation,
     )
+    from litellm.proxy._types import UserAPIKeyAuth
 
 try:
     from fastapi.exceptions import HTTPException
@@ -44,7 +45,8 @@ except ImportError:
 class UndeliverableStreamRewrite(Exception):
     def __init__(self, guardrail_name: str) -> None:
         super().__init__(
-            f"Guardrail '{guardrail_name}' rewrote the streamed response, which streaming pipelines cannot deliver"
+            f"Guardrail '{guardrail_name}' rewrote the streamed response in a way this endpoint's "
+            "streaming pipeline cannot deliver"
         )
         self.guardrail_name: Final = guardrail_name
 
@@ -57,28 +59,31 @@ def _tool_call_shape(tool_call: object) -> tuple[object, object]:
     return (function.get("name"), function.get("arguments"))
 
 
-def _rewrote_texts(sent: Sequence[str] | None, returned: Sequence[str] | None) -> bool:
-    return sent is not None and returned is not None and list(returned) != list(sent)
+def _text_snapshot(texts: Sequence[str] | None) -> tuple[str, ...] | None:
+    return None if texts is None else tuple(texts)
 
 
-def _rewrote_tool_calls(sent: Sequence[object] | None, returned: Sequence[object] | None) -> bool:
-    if sent is None or returned is None:
-        return False
-    return [_tool_call_shape(tool_call) for tool_call in returned] != [
-        _tool_call_shape(tool_call) for tool_call in sent
-    ]
+def _tool_call_shapes(tool_calls: Sequence[object] | None) -> tuple[tuple[object, object], ...] | None:
+    return None if tool_calls is None else tuple(_tool_call_shape(tool_call) for tool_call in tool_calls)
+
+
+def _rewrote(sent: tuple[object, ...] | None, returned: tuple[object, ...] | None) -> bool:
+    return sent is not None and returned is not None and returned != sent
 
 
 class _StreamRewriteObserver(CustomGuardrail):
     """Stand-in handed to the endpoint translation in place of a streaming pipeline step's
-    guardrail. Translations cannot rewrite every buffered chunk consistently, so the gate
-    withholds the stream whenever the guardrail returned different output than it was given,
-    which for guardrails like Bedrock's ANONYMIZED action is only known at runtime."""
+    guardrail. It records whether the guardrail returned different output than it was given,
+    which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text
+    rewrites are deliverable on translations that write them back across the buffered chunks
+    (``delivers_ended_stream_text_rewrites``); tool-call rewrites and text rewrites on any
+    other translation make the gate withhold the stream."""
 
     def __init__(self, inner: CustomGuardrail) -> None:
         super().__init__(guardrail_name=inner.guardrail_name)
         self.inner: Final = inner
-        self.rewrote = False
+        self.rewrote_texts = False
+        self.rewrote_tool_calls = False
 
     def structured_messages_cover_full_request(self) -> bool:
         return self.inner.structured_messages_cover_full_request()
@@ -90,13 +95,14 @@ class _StreamRewriteObserver(CustomGuardrail):
         input_type: Literal["request", "response"],
         logging_obj: "LiteLLMLoggingObj | None" = None,
     ) -> GenericGuardrailAPIInputs:
+        sent_texts: Final = _text_snapshot(inputs.get("texts"))
+        sent_tool_shapes: Final = _tool_call_shapes(inputs.get("tool_calls"))
         outputs: Final = await self.inner.apply_guardrail(
             inputs=inputs, request_data=request_data, input_type=input_type, logging_obj=logging_obj
         )
-        self.rewrote = (
-            self.rewrote
-            or _rewrote_texts(inputs.get("texts"), outputs.get("texts"))
-            or _rewrote_tool_calls(inputs.get("tool_calls"), outputs.get("tool_calls"))
+        self.rewrote_texts = self.rewrote_texts or _rewrote(sent_texts, _text_snapshot(outputs.get("texts")))
+        self.rewrote_tool_calls = self.rewrote_tool_calls or _rewrote(
+            sent_tool_shapes, _tool_call_shapes(outputs.get("tool_calls"))
         )
         return outputs
 
@@ -250,6 +256,41 @@ class PipelineExecutor:
             modified_data=working_data if working_data != data else None,
         )
 
+    @staticmethod
+    async def _run_streaming_step(
+        step: PipelineStep,
+        callback: CustomGuardrail,
+        endpoint_translation: "BaseTranslation",
+        streaming_chunks: list[object],  # mutable-ok: shared buffered-stream chunks the translation rewrites in place
+        hook_input: dict[str, object],  # mutable-ok: same request-payload shape as data
+        user_api_key_dict: "UserAPIKeyAuth | None",
+        litellm_logging_obj: "LiteLLMLoggingObj | None",
+    ) -> None:
+        """Run one streaming post_call step through the endpoint translation, delivering
+        text rewrites on translations that support ended-stream write-back and raising
+        ``UndeliverableStreamRewrite`` for any rewrite that cannot reach the client."""
+        observer: Final = _StreamRewriteObserver(callback)
+        deliver_rewrites: Final = type(endpoint_translation).delivers_ended_stream_text_rewrites
+        if deliver_rewrites:
+            await endpoint_translation.process_output_streaming_response(
+                responses_so_far=streaming_chunks,
+                guardrail_to_apply=observer,
+                litellm_logging_obj=litellm_logging_obj,
+                user_api_key_dict=user_api_key_dict,
+                request_data=hook_input,
+                deliver_ended_stream_rewrites=True,
+            )
+        else:
+            await endpoint_translation.process_output_streaming_response(
+                responses_so_far=streaming_chunks,
+                guardrail_to_apply=observer,
+                litellm_logging_obj=litellm_logging_obj,
+                user_api_key_dict=user_api_key_dict,
+                request_data=hook_input,
+            )
+        if observer.rewrote_tool_calls or (observer.rewrote_texts and not deliver_rewrites):
+            raise UndeliverableStreamRewrite(step.guardrail)
+
     @staticmethod
     async def _run_step(
         step: PipelineStep,
@@ -312,16 +353,15 @@ class PipelineExecutor:
                         f"Guardrail '{step.guardrail}' does not support streaming pipeline execution",
                         None,
                     )
-                observer: Final = _StreamRewriteObserver(callback)
-                await endpoint_translation.process_output_streaming_response(
-                    responses_so_far=streaming_chunks,
-                    guardrail_to_apply=observer,
-                    litellm_logging_obj=data.get("litellm_logging_obj"),
+                await PipelineExecutor._run_streaming_step(
+                    step=step,
+                    callback=callback,
+                    endpoint_translation=endpoint_translation,
+                    streaming_chunks=streaming_chunks,
+                    hook_input=hook_input,
                     user_api_key_dict=user_api_key_dict,
-                    request_data=hook_input,
+                    litellm_logging_obj=data.get("litellm_logging_obj"),
                 )
-                if observer.rewrote:
-                    raise UndeliverableStreamRewrite(step.guardrail)
                 response = None
             elif mode == "post_call":
                 response = await target.async_post_call_success_hook(
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 6c990222e51..7f3c3aecd87 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -492,14 +492,6 @@ def _pipeline_step_supports_unified_streaming(guardrail_name: str) -> bool:
     return callback is not None and PipelineExecutor.supports_unified_execution(callback)
 
 
-def _pipeline_step_rewrites_streamed_content(guardrail_name: str) -> bool:
-    callback: Final = PipelineExecutor.find_guardrail_callback(guardrail_name)
-    if callback is None:
-        return False
-    transform_mode: Final = unified_guardrail.resolve_streaming_flag(callback, "streaming_transform_mode", "block_only")
-    return callback.rewrites_streamed_output() or transform_mode == "incremental_diff"
-
-
 class _PipelineErrorBody(TypedDict):
     message: ReadOnly[str]
     type: ReadOnly[str]
@@ -516,9 +508,10 @@ def _undeliverable_stream_rewrite_error(policy_name: str, guardrail_name: str) -
         "error": {
             "message": (
                 f"Streaming response withheld by policy pipeline '{policy_name}' because guardrail "
-                f"'{guardrail_name}' rewrote the streamed output, and streaming pipelines cannot deliver "
-                "rewrites. Retry with stream=false, or drop it from the pipeline steps so guardrails.add "
-                "applies it to streamed output."
+                f"'{guardrail_name}' rewrote the streamed output in a way this endpoint's streaming "
+                "pipeline cannot deliver (a tool-call rewrite, or a text rewrite on a route without "
+                "stream write-back). Retry with stream=false, or drop it from the pipeline steps so "
+                "guardrails.add applies it to streamed output."
             ),
             "type": "guardrail_pipeline_error",
             "policies": (policy_name,),
@@ -535,13 +528,13 @@ def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_ap
     Background responses skip the post_call hooks entirely, so a pipeline
     governing one would silently never execute. Streaming responses execute
     pipelines against the buffered stream through the endpoint guardrail
-    translation of the request route, releasing the buffered chunks on allow.
-    That needs every step's guardrail to support the unified apply_guardrail
-    interface and to only allow or block (a step that rewrites streamed
-    content, via mask_response_content, a MASK action, or
-    streaming_transform_mode=incremental_diff, would have its rewrite silently
-    dropped), and needs the route to have a translation at all; anything else
-    keeps the 400 rather than letting ungoverned output stream through.
+    translation of the request route, releasing the buffered chunks on allow
+    (rewritten in place when a guardrail rewrote text and the translation
+    delivers ended-stream rewrites; a rewrite the translation cannot deliver
+    fails closed at runtime instead). That needs every step's guardrail to
+    support the unified apply_guardrail interface, and needs the route to have
+    a translation at all; anything else keeps the 400 rather than letting
+    ungoverned output stream through.
     """
     is_stream: Final = data.get("stream") is True
     is_background: Final = data.get("background") is True
@@ -586,25 +579,6 @@ def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_ap
             }
         }
         raise HTTPException(status_code=400, detail=unsupported_detail)
-    rewriting_guardrails: Final = tuple(
-        guardrail for guardrail in step_guardrails if _pipeline_step_rewrites_streamed_content(guardrail)
-    )
-    if rewriting_guardrails:
-        rewriting_detail: Final[_PipelineErrorDetail] = {
-            "error": {
-                "message": (
-                    "Policies with post_call guardrail pipelines cannot govern streaming responses "
-                    "because these pipeline guardrails rewrite streamed content (mask_response_content, "
-                    "a MASK action, or streaming_transform_mode=incremental_diff), which pipeline steps would release "
-                    f"unmodified: {', '.join(rewriting_guardrails)}. Retry with stream=false, or drop "
-                    "them from the pipeline steps so guardrails.add applies them to streamed output."
-                ),
-                "type": "guardrail_pipeline_error",
-                "policies": post_call_policies,
-                "guardrails": rewriting_guardrails,
-            }
-        }
-        raise HTTPException(status_code=400, detail=rewriting_detail)
     route: Final = user_api_key_dict.request_route
     if not route or resolve_endpoint_translation(user_api_key_dict, None) is not None:
         return
@@ -3485,12 +3459,13 @@ class ProxyLogging:
         pipeline allows it), then runs each pipeline's steps against the
         assembled output through the endpoint guardrail translation, the same
         machinery flat post_call guardrails use at end of stream. An allow
-        releases the buffered chunks verbatim; a step whose guardrail rewrote
-        the output withholds the stream with a 400 instead, since no
-        translation rewrites every buffered chunk consistently and some
-        rewrites (Bedrock's ANONYMIZED action, for one) are only decided at
-        runtime; a block or modify_response terminates with the translation's
-        block chunks or the raised error.
+        releases the buffered chunks: verbatim when no guardrail rewrote the
+        output, rewritten in place when one rewrote text and the translation
+        delivers ended-stream rewrites (later steps then re-scan the rewritten
+        chunks, so rewrites chain). A rewrite the translation cannot deliver
+        (a tool-call rewrite, or a text rewrite on a route without write-back)
+        withholds the stream with a 400; a block or modify_response terminates
+        with the translation's block chunks or the raised error.
         """
         buffered: Final[list[object]] = []  # mutable-ok: accumulates the stream before the pipeline verdict
         async for item in response:
diff --git a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py
index af3ccd65b11..ba26da50bc8 100644
--- a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py
+++ b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py
@@ -263,6 +263,71 @@ class TestAnthropicMessagesHandlerStreamingOutputProcessing:
             # Should return the responses unchanged
             assert result == responses_so_far
 
+    @staticmethod
+    def _ended_sse_chunks() -> list:
+        events = [
+            ("message_start", {"type": "message_start", "message": {"id": "msg_1", "type": "message", "role": "assistant", "model": "claude-sonnet-4-5", "content": [], "stop_reason": None, "usage": {"input_tokens": 1, "output_tokens": 0}}}),
+            ("content_block_start", {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}}),
+            ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "hello "}}),
+            ("content_block_delta", {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "world"}}),
+            ("content_block_stop", {"type": "content_block_stop", "index": 0}),
+            ("message_delta", {"type": "message_delta", "delta": {"stop_reason": "end_turn", "stop_sequence": None}, "usage": {"output_tokens": 2}}),
+            ("message_stop", {"type": "message_stop"}),
+        ]
+        return [f"event: {name}\ndata: {json.dumps(payload)}\n\n".encode() for name, payload in events]
+
+    @staticmethod
+    def _masking_guardrail() -> CustomGuardrail:
+        class MaskWorld(CustomGuardrail):
+            async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+                return {**inputs, "texts": [text.replace("world", "[MASKED]") for text in inputs.get("texts", [])]}
+
+        return MaskWorld(guardrail_name="test")
+
+    @staticmethod
+    def _delta_texts(chunks: list) -> list:
+        texts = []
+        for chunk in chunks:
+            for line in chunk.decode().split("\n"):
+                if not line.startswith("data:"):
+                    continue
+                data = json.loads(line[len("data:") :].strip())
+                if data.get("type") == "content_block_delta":
+                    texts.append(data["delta"]["text"])
+        return texts
+
+    @pytest.mark.asyncio
+    async def test_deliver_ended_stream_rewrites_writes_text_back_into_sse_chunks(self):
+        handler = AnthropicMessagesHandler()
+        chunks = self._ended_sse_chunks()
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=MagicMock(),
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is chunks
+        assert self._delta_texts(chunks) == ["hello [MASKED]", ""]
+        raw = b"".join(chunks).decode()
+        assert "event: message_start" in raw and "event: message_stop" in raw
+        assert '"stop_reason": "end_turn"' in raw
+
+    @pytest.mark.asyncio
+    async def test_ended_stream_rewrite_leaves_chunks_untouched_by_default(self):
+        handler = AnthropicMessagesHandler()
+        chunks = self._ended_sse_chunks()
+        original = [bytes(chunk) for chunk in chunks]
+
+        await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=MagicMock(),
+        )
+
+        assert chunks == original
+
 
 class TestAnthropicMessagesHandlerInputProcessing:
     """Test input processing preserves litellm_metadata for dynamic guardrails."""
diff --git a/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py b/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py
index a29e0be4655..26442a4a6ed 100644
--- a/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py
+++ b/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py
@@ -1073,6 +1073,61 @@ class TestOpenAIChatCompletionsHandlerStreamingOutput:
         # Should return the responses
         assert result == responses_so_far
 
+    @staticmethod
+    def _ended_stream_chunks() -> list:
+        from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
+
+        return [
+            ModelResponseStream(
+                id="chatcmpl-123",
+                created=1234567890,
+                model="gpt-4",
+                object="chat.completion.chunk",
+                choices=[StreamingChoices(index=0, delta=Delta(content="Hello"), finish_reason=None)],
+            ),
+            ModelResponseStream(
+                id="chatcmpl-123",
+                created=1234567890,
+                model="gpt-4",
+                object="chat.completion.chunk",
+                choices=[StreamingChoices(index=0, delta=Delta(content=" world"), finish_reason="stop")],
+            ),
+        ]
+
+    @pytest.mark.asyncio
+    async def test_deliver_ended_stream_rewrites_writes_text_back_into_chunks(self):
+        handler = OpenAIChatCompletionsHandler()
+        guardrail = MockGuardrail(guardrail_name="test")
+        chunks = self._ended_stream_chunks()
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=guardrail,
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is chunks
+        assert chunks[0].choices[0].delta.content == "HELLO WORLD"
+        assert chunks[1].choices[0].delta.content in (None, "")
+        assert chunks[1].choices[0].finish_reason == "stop"
+
+    @pytest.mark.asyncio
+    async def test_ended_stream_rewrite_leaves_chunks_untouched_by_default(self):
+        handler = OpenAIChatCompletionsHandler()
+        guardrail = MockGuardrail(guardrail_name="test")
+        chunks = self._ended_stream_chunks()
+
+        await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=guardrail,
+            litellm_logging_obj=None,
+        )
+
+        assert chunks[0].choices[0].delta.content == "Hello"
+        assert chunks[1].choices[0].delta.content == " world"
+        assert chunks[1].choices[0].finish_reason == "stop"
+
 
 class TestGetStructuredMessages:
     """Test the get_structured_messages method."""
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
index 447175b09a6..4f95e08cb71 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
@@ -1104,6 +1104,89 @@ class TestOpenAIResponsesHandlerStreamingOutputProcessing:
         output_text = result[-1]["response"]["output"][0]["content"][0]["text"]
         assert output_text == original_text
 
+    @staticmethod
+    def _ended_stream_events() -> List[dict]:
+        content = [{"type": "output_text", "text": "hello world"}]
+        item = {
+            "type": "message",
+            "id": "msg_123",
+            "status": "completed",
+            "role": "assistant",
+            "content": content,
+        }
+        return [
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "hello "},
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "world"},
+            {"type": "response.output_text.done", "output_index": 0, "content_index": 0, "text": "hello world"},
+            {
+                "type": "response.content_part.done",
+                "output_index": 0,
+                "content_index": 0,
+                "part": {"type": "output_text", "text": "hello world"},
+            },
+            {"type": "response.output_item.done", "output_index": 0, "item": {**item, "content": [dict(c) for c in content]}},
+            {
+                "type": "response.completed",
+                "response": {
+                    "id": "resp_123",
+                    "model": "gpt-4o",
+                    "output": [{**item, "content": [dict(c) for c in content]}],
+                    "status": "completed",
+                },
+            },
+        ]
+
+    @staticmethod
+    def _masking_guardrail() -> CustomGuardrail:
+        class MaskWorld(CustomGuardrail):
+            async def apply_guardrail(
+                self,
+                inputs: GenericGuardrailAPIInputs,
+                request_data: dict,
+                input_type: Literal["request", "response"],
+                logging_obj: Optional[Any] = None,
+            ) -> GenericGuardrailAPIInputs:
+                texts = inputs.get("texts", [])
+                return {**inputs, "texts": [t.replace("world", "[MASKED]") for t in texts]}
+
+        return MaskWorld(guardrail_name="test-mask")
+
+    @pytest.mark.asyncio
+    async def test_deliver_ended_stream_rewrites_syncs_all_stream_events(self):
+        handler = OpenAIResponsesHandler()
+        events = self._ended_stream_events()
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is events
+        assert events[0]["delta"] == "hello [MASKED]"
+        assert events[1]["delta"] == ""
+        assert events[2]["text"] == "hello [MASKED]"
+        assert events[3]["part"]["text"] == "hello [MASKED]"
+        assert events[4]["item"]["content"][0]["text"] == "hello [MASKED]"
+        assert events[5]["response"]["output"][0]["content"][0]["text"] == "hello [MASKED]"
+
+    @pytest.mark.asyncio
+    async def test_ended_stream_rewrite_leaves_delta_events_untouched_by_default(self):
+        handler = OpenAIResponsesHandler()
+        events = self._ended_stream_events()
+
+        await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=None,
+        )
+
+        assert events[0]["delta"] == "hello "
+        assert events[1]["delta"] == "world"
+        assert events[2]["text"] == "hello world"
+        assert events[5]["response"]["output"][0]["content"][0]["text"] == "hello [MASKED]"
+
 
 class TestGetStructuredMessages:
     """Test the get_structured_messages method for Responses API handler."""
diff --git a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
index 52fd8777a19..908c9f12c9e 100644
--- a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
+++ b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
@@ -823,6 +823,8 @@ class _TextReturningGuardrail(CustomGuardrail):
 
 
 class _TextTranslation:
+    delivers_ended_stream_text_rewrites = False
+
     def __init__(self):
         self.seen_guardrail_names = []
 
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 74bd2e483cc..73270ec5671 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -24,7 +24,7 @@ from litellm.integrations.custom_guardrail import (
     ModifyResponseException,
 )
 from litellm.integrations.prometheus import PrometheusLogger
-from litellm.proxy._types import ProxyException
+from litellm.proxy._types import ProxyException, UserAPIKeyAuth
 from litellm.proxy.common_utils.callback_utils import add_guardrail_to_applied_guardrails_header
 from litellm.proxy.utils import ProxyLogging, _raise_for_streaming_post_call_pipelines
 from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import ContentFilterGuardrail
@@ -1441,7 +1441,7 @@ async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_lacks_uni
         ("guardrail_config", {"streaming_transform_mode": "incremental_diff"}),
     ],
 )
-async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_rewrites_streamed_content(
+async def test_pre_call_hook_allows_streaming_when_pipeline_guardrail_rewrites_streamed_content(
     proxy_logging, make_user_api_key_auth, monkeypatch, rewrite_attribute, value
 ):
     seen: Dict[str, Any] = {}
@@ -1450,24 +1450,21 @@ async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_rewrites_
     monkeypatch.setattr(litellm, "callbacks", [guardrail])
     data = _post_call_pipeline_data(stream=True)
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
-            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
-            data=data,
-            call_type="completion",
-            guardrails_only=True,
-        )
+    out = await proxy_logging.pre_call_hook(
+        user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+        data=data,
+        call_type="completion",
+        guardrails_only=True,
+    )
 
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
-    assert "rewrite streamed content" in info.value.detail["error"]["message"]
-    assert seen.get("count") is None
+    assert out is not None
+    assert out.get("stream") is True
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("action, rejected", [(ContentFilterAction.MASK, True), (ContentFilterAction.BLOCK, False)])
-async def test_pre_call_hook_rejects_streaming_only_when_content_filter_step_masks(
-    proxy_logging, make_user_api_key_auth, monkeypatch, action, rejected
+@pytest.mark.parametrize("action", [ContentFilterAction.MASK, ContentFilterAction.BLOCK])
+async def test_pre_call_hook_allows_streaming_when_content_filter_step_masks_or_blocks(
+    proxy_logging, make_user_api_key_auth, monkeypatch, action
 ):
     guardrail = ContentFilterGuardrail(
         guardrail_name="gr-post",
@@ -1478,25 +1475,15 @@ async def test_pre_call_hook_rejects_streaming_only_when_content_filter_step_mas
     data = _post_call_pipeline_data(stream=True)
     user_api_key_dict = make_user_api_key_auth(request_route="/v1/chat/completions")
 
-    if not rejected:
-        out = await proxy_logging.pre_call_hook(
-            user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
-        )
-        assert out is not None and out.get("stream") is True
-        return
+    out = await proxy_logging.pre_call_hook(
+        user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
+    )
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
-            user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
-        )
-
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
-    assert "a MASK action" in info.value.detail["error"]["message"]
+    assert out is not None and out.get("stream") is True
 
 
 @pytest.mark.asyncio
-async def test_pre_call_hook_rejects_streaming_when_content_filter_category_masks(
+async def test_pre_call_hook_allows_streaming_when_content_filter_category_masks(
     proxy_logging, make_user_api_key_auth, monkeypatch
 ):
     guardrail = ContentFilterGuardrail(
@@ -1508,13 +1495,11 @@ async def test_pre_call_hook_rejects_streaming_when_content_filter_category_mask
     data = _post_call_pipeline_data(stream=True)
     user_api_key_dict = make_user_api_key_auth(request_route="/v1/chat/completions")
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
-            user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
-        )
+    out = await proxy_logging.pre_call_hook(
+        user_api_key_dict=user_api_key_dict, data=data, call_type="completion", guardrails_only=True
+    )
 
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
+    assert out is not None and out.get("stream") is True
 
 
 @pytest.mark.asyncio
@@ -1618,17 +1603,10 @@ def _echoed_tool_call_dicts(arguments: str) -> List[Dict[str, Any]]:
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("on_fail, on_error", [("block", None), ("next", "next")])
-@pytest.mark.parametrize(
-    "make_chunks, transform",
-    [
-        (_stream_chunks, lambda inputs: {"texts": ["hello [MASKED]"]}),
-        (_tool_call_stream_chunks, lambda inputs: {"tool_calls": _echoed_tool_call_dicts('{"ssn": "[MASKED]"}')}),
-    ],
-    ids=["texts", "tool_calls"],
-)
-async def test_streaming_iterator_hook_pipeline_withholds_runtime_rewrite(
-    proxy_logging, make_user_api_key_auth, monkeypatch, make_chunks, transform, on_fail, on_error
+async def test_streaming_iterator_hook_pipeline_withholds_runtime_tool_call_rewrite(
+    proxy_logging, make_user_api_key_auth, monkeypatch, on_fail, on_error
 ):
+    transform = lambda inputs: {"tool_calls": _echoed_tool_call_dicts('{"ssn": "[MASKED]"}')}  # noqa: E731
     monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
     monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
     step = PipelineStep(guardrail="gr-post", on_pass="allow", on_fail=on_fail, on_error=on_error)
@@ -1638,7 +1616,7 @@ async def test_streaming_iterator_hook_pipeline_withholds_runtime_rewrite(
     async def _drain() -> None:
         async for item in proxy_logging.async_post_call_streaming_iterator_hook(
             user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
-            response=_async_chunk_iter(make_chunks()),
+            response=_async_chunk_iter(_tool_call_stream_chunks()),
             request_data=data,
         ):
             delivered.append(item)
@@ -1655,6 +1633,81 @@ async def test_streaming_iterator_hook_pipeline_withholds_runtime_rewrite(
     assert "stream=false" in error["message"]
 
 
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_delivers_runtime_text_rewrite(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    transform = lambda inputs: {"texts": ["hello [MASKED]"]}  # noqa: E731
+    monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    chunks = _stream_chunks()
+
+    delivered = [
+        item
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        )
+    ]
+
+    assert [id(item) for item in delivered] == [id(chunk) for chunk in chunks]
+    assert delivered[0].choices[0].delta.content == "hello [MASKED]"
+    assert delivered[1].choices[0].delta.content in (None, "")
+    assert delivered[1].choices[0].finish_reason == "stop"
+
+
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_chains_text_rewrites_across_steps(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    second_step_saw: Dict[str, Any] = {}
+
+    class FirstMask(CustomGuardrail):
+        async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+            return {**inputs, "texts": [text.replace("world", "[MASKED]") for text in inputs["texts"]]}
+
+    class SecondMask(CustomGuardrail):
+        async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+            second_step_saw["texts"] = list(inputs["texts"])
+            return {**inputs, "texts": [text.replace("hello", "[GREETING]") for text in inputs["texts"]]}
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [
+            FirstMask(guardrail_name="gr-first", event_hook=GuardrailEventHooks.post_call, default_on=False),
+            SecondMask(guardrail_name="gr-second", event_hook=GuardrailEventHooks.post_call, default_on=False),
+        ],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    pipeline = GuardrailPipeline(
+        mode="post_call",
+        steps=[
+            PipelineStep(guardrail="gr-first", on_pass="next", on_fail="block"),
+            PipelineStep(guardrail="gr-second", on_pass="allow", on_fail="block"),
+        ],
+    )
+    data = _post_call_pipeline_data(stream=True)
+    data["metadata"]["_guardrail_pipelines"] = [("response-governance", pipeline)]
+    chunks = _stream_chunks()
+
+    delivered = [
+        item
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        )
+    ]
+
+    assert second_step_saw["texts"] == ["hello [MASKED]"]
+    assert delivered[0].choices[0].delta.content == "[GREETING] [MASKED]"
+    assert delivered[1].choices[0].delta.content in (None, "")
+    assert delivered[1].choices[0].finish_reason == "stop"
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
     "make_chunks, transform",
@@ -1761,6 +1814,79 @@ async def test_streaming_iterator_hook_pipeline_modify_response_emits_translated
     assert not any(item is chunk for item in delivered for chunk in chunks)
 
 
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_pipeline_delivers_text_rewrite_on_anthropic_sse(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    transform = lambda inputs: {"texts": ["hello [MASKED]"]}  # noqa: E731
+    monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+    chunks = _anthropic_sse_chunks()
+
+    delivered = [
+        item
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/messages"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        )
+    ]
+
+    raw = b"".join(delivered).decode()
+    assert "hello [MASKED]" in raw
+    assert "hello world" not in raw
+    assert raw.count("event: content_block_delta") == 1
+    for expected_event in ("message_start", "content_block_start", "content_block_stop", "message_delta", "message_stop"):
+        assert f"event: {expected_event}" in raw
+
+
+@pytest.mark.asyncio
+async def test_pipeline_executor_withholds_text_rewrite_when_translation_lacks_write_back(monkeypatch):
+    from litellm.llms.base_llm.guardrail_translation.base_translation import BaseTranslation
+    from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor, UndeliverableStreamRewrite
+
+    class NoWriteBackTranslation(BaseTranslation):
+        async def process_input_messages(self, data, guardrail_to_apply, litellm_logging_obj):
+            return data
+
+        async def process_output_response(self, response, guardrail_to_apply, litellm_logging_obj, **kwargs):
+            return response
+
+        async def process_output_streaming_response(
+            self,
+            responses_so_far,
+            guardrail_to_apply,
+            litellm_logging_obj,
+            user_api_key_dict=None,
+            request_data=None,
+            stream_transform_sink=None,
+            deliver_ended_stream_rewrites=False,
+        ):
+            assert deliver_ended_stream_rewrites is False
+            await guardrail_to_apply.apply_guardrail(
+                inputs={"texts": ["hello world"]},
+                request_data=request_data or {},
+                input_type="response",
+            )
+            return responses_so_far
+
+    transform = lambda inputs: {"texts": ["hello [MASKED]"]}  # noqa: E731
+    monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
+
+    with pytest.raises(UndeliverableStreamRewrite):
+        await PipelineExecutor.execute_steps(
+            steps=[PipelineStep(guardrail="gr-post", on_pass="allow", on_fail="block")],
+            mode="post_call",
+            data={"metadata": {}},
+            user_api_key_dict=UserAPIKeyAuth(api_key="sk-test"),
+            call_type="acompletion",
+            policy_name="response-governance",
+            streaming_chunks=_stream_chunks(),
+            endpoint_translation=NoWriteBackTranslation(),
+        )
+
+
 @pytest.mark.asyncio
 async def test_streaming_iterator_hook_pipeline_gates_without_iterator_overrides(
     proxy_logging, make_user_api_key_auth, monkeypatch

From 85fea1a6752ac5f0c4c0415c5d3da4f6bc542911 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 1 Sep 2026 17:15:11 -0700
Subject: [PATCH 16/25] fix(policy_engine): move mutable-ok suppressions onto
 the flagged lines

---
 litellm/llms/openai/chat/guardrail_translation/handler.py | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/litellm/llms/openai/chat/guardrail_translation/handler.py b/litellm/llms/openai/chat/guardrail_translation/handler.py
index fcb447d5277..02206a36b0e 100644
--- a/litellm/llms/openai/chat/guardrail_translation/handler.py
+++ b/litellm/llms/openai/chat/guardrail_translation/handler.py
@@ -989,12 +989,8 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
             return
         await self._apply_guardrail_responses_to_output_streaming(
             responses=responses_so_far,
-            guardrailed_texts=[
-                after for _choice_idx, after in changed
-            ],  # mutable-ok: the callee's signature predates this change and takes lists
-            task_mappings=[
-                (choice_idx, None) for choice_idx, _after in changed
-            ],  # mutable-ok: the callee's signature predates this change and takes lists
+            guardrailed_texts=[after for _choice_idx, after in changed],  # mutable-ok: callee takes lists
+            task_mappings=[(choice_idx, None) for choice_idx, _after in changed],  # mutable-ok: callee takes lists
         )
 
     async def _apply_guardrail_responses_to_output_streaming(

From 4fbe4ce2e225940b009d390543bd76694eece343 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 1 Sep 2026 17:29:52 -0700
Subject: [PATCH 17/25] fix(guardrail_translation): deliver stream rewrites on
 incomplete and failed responses terminals

---
 .../guardrail_translation/handler.py          | 60 ++++++++++++-------
 ...test_openai_responses_guardrail_handler.py | 56 +++++++++++++++++
 2 files changed, 93 insertions(+), 23 deletions(-)

diff --git a/litellm/llms/openai/responses/guardrail_translation/handler.py b/litellm/llms/openai/responses/guardrail_translation/handler.py
index ac894401628..3c902f1b829 100644
--- a/litellm/llms/openai/responses/guardrail_translation/handler.py
+++ b/litellm/llms/openai/responses/guardrail_translation/handler.py
@@ -86,6 +86,15 @@ class ResponsesStreamChunk(TypedDict, total=False):
     text: ReadOnly[str]
 
 
+_TERMINAL_ENVELOPE_EVENT_TYPES: Final = frozenset(
+    {
+        ResponsesAPIStreamEvents.RESPONSE_COMPLETED.value,
+        ResponsesAPIStreamEvents.RESPONSE_FAILED.value,
+        ResponsesAPIStreamEvents.RESPONSE_INCOMPLETE.value,
+    }
+)
+
+
 def _next_stream_sequence_number(responses_so_far: Sequence[Any] | None) -> int:
     sequence_numbers: Final = (
         item.get("sequence_number") if isinstance(item, dict) else getattr(item, "sequence_number", None)
@@ -507,14 +516,17 @@ class OpenAIResponsesHandler(BaseTranslation):
         chunk, apply the guardrail, then write the result back in-place so the
         caller sees the modified content (e.g. PII tokens replaced).
 
-        For ``response.completed`` events (the normal end-of-stream signal) we
-        use the same per-item extraction + task-mapping approach as
-        ``process_output_response`` so that unmasking / blocking works correctly
-        for every output item. With ``deliver_ended_stream_rewrites`` the earlier
-        text-carrying events (``response.output_text.delta`` / ``.done``,
+        For terminal envelope events (``response.completed``, and equally
+        ``response.incomplete`` / ``response.failed``, whose envelopes carry the
+        partial output) we use the same per-item extraction + task-mapping
+        approach as ``process_output_response`` so that unmasking / blocking
+        works correctly for every output item. With
+        ``deliver_ended_stream_rewrites`` the earlier text-carrying events
+        (``response.output_text.delta`` / ``.done``,
         ``response.content_part.done``, ``response.output_item.done``) are synced
-        to the rewritten completed response too, so a client reading deltas sees
-        the rewrite instead of the raw model output.
+        to the rewritten envelope too, so a client reading deltas sees the
+        rewrite instead of the raw model output; a rewrite observed where no
+        write-back is possible fails closed instead of releasing raw output.
         """
         if not responses_so_far:
             return responses_so_far
@@ -526,14 +538,16 @@ class OpenAIResponsesHandler(BaseTranslation):
             return responses_so_far
 
         # ------------------------------------------------------------------ #
-        # Case 1: response.completed — full response is available in the      #
-        # final chunk; iterate output items, apply guardrail, write back.     #
+        # Case 1: terminal envelope events (completed/incomplete/failed).     #
+        # the accumulated response is available in the final chunk; iterate   #
+        # output items, apply guardrail, write back. Falls through to the     #
+        # string fallback when the envelope yields nothing to check.          #
         # ------------------------------------------------------------------ #
-        if final_chunk.get("type") == "response.completed":
+        if final_chunk.get("type") in _TERMINAL_ENVELOPE_EVENT_TYPES:
             response_obj: Final[ResponseOutputEnvelope] = final_chunk.get("response") or {}
-            if not hasattr(response_obj, "get"):
-                return responses_so_far
-            outputs: Final[Sequence[object]] = response_obj.get("output") or []
+            outputs: Final[Sequence[object]] = (
+                (response_obj.get("output") or []) if hasattr(response_obj, "get") else []
+            )
 
             texts_to_check: Final[list[str]] = []
             tool_calls_to_check: Final[list[ChatCompletionToolCallChunk]] = []
@@ -596,8 +610,7 @@ class OpenAIResponsesHandler(BaseTranslation):
                             stream_events=responses_so_far[:-1],
                             rewrites_by_position=rewrites_by_position,
                         )
-
-            return responses_so_far
+                return responses_so_far
 
         # ------------------------------------------------------------------ #
         # Case 2: response.output_item.done — extract tool calls only.        #
@@ -623,7 +636,8 @@ class OpenAIResponsesHandler(BaseTranslation):
         # ------------------------------------------------------------------ #
         # Fallback: apply guardrail to the accumulated text string.           #
         # No structured write-back is possible here; guardrails that only     #
-        # need to block/flag (not rewrite) still work correctly.             #
+        # need to block/flag (not rewrite) still work correctly, and a        #
+        # rewrite a caller expects delivered fails closed instead.            #
         # ------------------------------------------------------------------ #
         string_so_far: Final = self.get_streaming_string_so_far(responses_so_far)
         if string_so_far:
@@ -633,12 +647,17 @@ class OpenAIResponsesHandler(BaseTranslation):
             )
             if response_model:
                 fallback_inputs["model"] = response_model
-            await guardrail_to_apply.apply_guardrail(
+            fallback_outputs: Final = await guardrail_to_apply.apply_guardrail(
                 inputs=fallback_inputs,
                 request_data=request_data if request_data is not None else {},
                 input_type="response",
                 logging_obj=litellm_logging_obj,
             )
+            fallback_texts: Final = fallback_outputs.get("texts")
+            if deliver_ended_stream_rewrites and fallback_texts and tuple(fallback_texts) != (string_so_far,):
+                from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+                raise UndeliverableStreamRewrite(guardrail_to_apply.guardrail_name or "unknown")
         return responses_so_far
 
     @staticmethod
@@ -704,12 +723,7 @@ class OpenAIResponsesHandler(BaseTranslation):
         """
         if not responses_so_far:
             return False
-        terminal_types: Final = {
-            ResponsesAPIStreamEvents.RESPONSE_COMPLETED.value,
-            ResponsesAPIStreamEvents.RESPONSE_FAILED.value,
-            ResponsesAPIStreamEvents.RESPONSE_INCOMPLETE.value,
-        }
-        return responses_so_far[-1].get("type") in terminal_types
+        return responses_so_far[-1].get("type") in _TERMINAL_ENVELOPE_EVENT_TYPES
 
     def build_stream_error_items(
         self,
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
index 4f95e08cb71..dfd6352f9ec 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
@@ -1171,6 +1171,62 @@ class TestOpenAIResponsesHandlerStreamingOutputProcessing:
         assert events[4]["item"]["content"][0]["text"] == "hello [MASKED]"
         assert events[5]["response"]["output"][0]["content"][0]["text"] == "hello [MASKED]"
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("terminal_type", ["response.incomplete", "response.failed"])
+    async def test_deliver_ended_stream_rewrites_syncs_non_completed_terminals(self, terminal_type):
+        handler = OpenAIResponsesHandler()
+        events = self._ended_stream_events()
+        events[-1]["type"] = terminal_type
+        events[-1]["response"]["status"] = terminal_type.split(".")[-1]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is events
+        assert events[0]["delta"] == "hello [MASKED]"
+        assert events[1]["delta"] == ""
+        assert events[2]["text"] == "hello [MASKED]"
+        assert events[3]["part"]["text"] == "hello [MASKED]"
+        assert events[4]["item"]["content"][0]["text"] == "hello [MASKED]"
+        assert events[5]["response"]["output"][0]["content"][0]["text"] == "hello [MASKED]"
+
+    @pytest.mark.asyncio
+    async def test_fallback_rewrite_with_delivery_expected_fails_closed(self):
+        from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+        handler = OpenAIResponsesHandler()
+        events = [
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "hello "},
+            {"type": "response.output_text.done", "output_index": 0, "content_index": 0, "text": "hello world"},
+        ]
+
+        with pytest.raises(UndeliverableStreamRewrite):
+            await handler.process_output_streaming_response(
+                responses_so_far=events,
+                guardrail_to_apply=self._masking_guardrail(),
+                litellm_logging_obj=None,
+                deliver_ended_stream_rewrites=True,
+            )
+
+    @pytest.mark.asyncio
+    async def test_fallback_rewrite_without_delivery_expected_does_not_raise(self):
+        handler = OpenAIResponsesHandler()
+        events = [
+            {"type": "response.output_text.done", "output_index": 0, "content_index": 0, "text": "hello world"},
+        ]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=None,
+        )
+
+        assert result is events
+
     @pytest.mark.asyncio
     async def test_ended_stream_rewrite_leaves_delta_events_untouched_by_default(self):
         handler = OpenAIResponsesHandler()

From c9435b5ff3f72c6d9c5ebfa30175498bc5928daf Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 1 Sep 2026 18:00:42 -0700
Subject: [PATCH 18/25] fix(guardrails): key stream rewrites by choice index
 and scan delta-only responses buffers

Chat streaming write-backs now match chunks by the choice's index field
instead of its list position, delivering rewrites to the right choice on
n>1 streams; an ended-stream rewrite on a multi-choice buffer fails
closed since stream_chunk_builder collapses the choices. The Responses
fallback joins output_text.delta events when delivery is expected, so a
delta-only buffer is guardrail-checked instead of released raw.
---
 .../chat/guardrail_translation/handler.py     | 50 ++++++----
 .../guardrail_translation/handler.py          | 31 +++++--
 .../test_openai_guardrail_handler.py          | 93 +++++++++++++++++++
 ...test_openai_responses_guardrail_handler.py | 71 ++++++++++++++
 4 files changed, 222 insertions(+), 23 deletions(-)

diff --git a/litellm/llms/openai/chat/guardrail_translation/handler.py b/litellm/llms/openai/chat/guardrail_translation/handler.py
index 02206a36b0e..182dec81937 100644
--- a/litellm/llms/openai/chat/guardrail_translation/handler.py
+++ b/litellm/llms/openai/chat/guardrail_translation/handler.py
@@ -620,6 +620,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
                 responses_so_far=responses_so_far,
                 guardrailed_response=model_response,
                 pre_guardrail_texts=pre_guardrail_texts,
+                guardrail_name=guardrail_to_apply.guardrail_name or "unknown",
             )
 
     def build_stream_error_items(
@@ -747,8 +748,8 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         """
         combined_texts: Final[dict[tuple[int, int | None], str]] = {}
 
-        for response_idx, response in enumerate(responses_so_far):
-            for choice_idx, choice in enumerate(response.choices):
+        for response in responses_so_far:
+            for choice in response.choices:
                 if isinstance(choice, litellm.StreamingChoices):
                     content = choice.delta.content
                 elif isinstance(choice, litellm.Choices):
@@ -761,7 +762,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
 
                 if isinstance(content, str):
                     # String content - accumulate for this choice
-                    str_key: tuple[int, int | None] = (choice_idx, None)
+                    str_key: tuple[int, int | None] = (choice.index, None)
                     if str_key not in combined_texts:
                         combined_texts[str_key] = ""
                     combined_texts[str_key] += content
@@ -772,7 +773,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
                         text_str = content_item.get("text")
                         if text_str:
                             list_key: tuple[int, int | None] = (
-                                choice_idx,
+                                choice.index,
                                 content_idx,
                             )
                             if list_key not in combined_texts:
@@ -973,24 +974,38 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         responses_so_far: list["ModelResponseStream"],  # mutable-ok: rewrites the caller's buffered chunks in place
         guardrailed_response: "ModelResponse",
         pre_guardrail_texts: tuple[str | None, ...],
+        guardrail_name: str,
     ) -> None:
         """Write ended-stream guardrail text rewrites back across the buffered
-        chunks: each rewritten choice's full text lands in its first
+        chunks: the full rewritten text lands in the choice's first
         content-carrying chunk and the rest are blanked, the same shape the
         in-flight write-back uses. Chunks carrying only finish_reason or usage
-        stay untouched."""
+        stay untouched. A rewrite on a stream carrying more than one distinct
+        choice index fails closed."""
         post_guardrail_texts: Final = self._string_choice_contents(guardrailed_response)
         changed: Final = tuple(
-            (choice_idx, after)
-            for choice_idx, (before, after) in enumerate(zip(pre_guardrail_texts, post_guardrail_texts))
+            after
+            for before, after in zip(pre_guardrail_texts, post_guardrail_texts)
             if before is not None and after is not None and after != before
         )
         if not changed:
             return
+        stream_choice_indices: Final = frozenset(
+            choice.index for response in responses_so_far for choice in response.choices
+        )
+        if len(stream_choice_indices) != 1:
+            # stream_chunk_builder collapses every choice into one index-0
+            # choice, so a rewrite of the rebuilt response cannot be attributed
+            # back to a single choice on an n>1 stream: withhold the stream
+            # rather than deliver the rewrite on the wrong choice
+            from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+            raise UndeliverableStreamRewrite(guardrail_name)
+        target_choice_index: Final = next(iter(stream_choice_indices))
         await self._apply_guardrail_responses_to_output_streaming(
             responses=responses_so_far,
-            guardrailed_texts=[after for _choice_idx, after in changed],  # mutable-ok: callee takes lists
-            task_mappings=[(choice_idx, None) for choice_idx, _after in changed],  # mutable-ok: callee takes lists
+            guardrailed_texts=list(changed),  # mutable-ok: callee takes lists
+            task_mappings=[(target_choice_index, None) for _ in changed],  # mutable-ok: callee takes lists
         )
 
     async def _apply_guardrail_responses_to_output_streaming(
@@ -1008,7 +1023,8 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         Args:
             responses: List of ModelResponseStream objects to modify
             guardrailed_texts: List of guardrailed text responses (combined from all chunks)
-            task_mappings: List of tuples (choice_idx, content_idx)
+            task_mappings: List of tuples (choice_idx, content_idx), where choice_idx
+                is the choice's ``index`` field, not its position in a chunk's list
 
         Override this method to customize how responses are applied to streaming responses.
         """
@@ -1024,9 +1040,11 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         # Key: (choice_idx, content_idx), Value: boolean (True if already set)
         already_set: Final[dict[tuple[int, int | None], bool]] = {}
 
-        # Iterate through all responses and update content
-        for response_idx, response in enumerate(responses):
-            for choice_idx_in_response, choice in enumerate(response.choices):
+        # Iterate through all responses and update content, matching each chunk's
+        # choice by its index field: on n>1 streams a chunk usually carries one
+        # choice at list position 0 whose index names the logical choice.
+        for response in responses:
+            for choice in response.choices:
                 if isinstance(choice, litellm.StreamingChoices):
                     content = choice.delta.content
                 elif isinstance(choice, litellm.Choices):
@@ -1039,7 +1057,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
 
                 if isinstance(content, str):
                     # String content
-                    str_key: tuple[int, int | None] = (choice_idx_in_response, None)
+                    str_key: tuple[int, int | None] = (choice.index, None)
                     if str_key in guardrail_map:
                         if str_key not in already_set:
                             # First chunk - set the complete guardrailed text
@@ -1060,7 +1078,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
                     for content_idx, content_item in enumerate(content):
                         if "text" in content_item:
                             list_key: tuple[int, int | None] = (
-                                choice_idx_in_response,
+                                choice.index,
                                 content_idx,
                             )
                             if list_key in guardrail_map:
diff --git a/litellm/llms/openai/responses/guardrail_translation/handler.py b/litellm/llms/openai/responses/guardrail_translation/handler.py
index 3c902f1b829..ebd1070a8e8 100644
--- a/litellm/llms/openai/responses/guardrail_translation/handler.py
+++ b/litellm/llms/openai/responses/guardrail_translation/handler.py
@@ -634,14 +634,20 @@ class OpenAIResponsesHandler(BaseTranslation):
             return responses_so_far
 
         # ------------------------------------------------------------------ #
-        # Fallback: apply guardrail to the accumulated text string.           #
-        # No structured write-back is possible here; guardrails that only     #
-        # need to block/flag (not rewrite) still work correctly, and a        #
-        # rewrite a caller expects delivered fails closed instead.            #
+        # Fallback: apply guardrail to the accumulated text string. When a    #
+        # caller expects rewrites delivered and only output_text.delta events #
+        # carried the text (a stream cut off before any .done or terminal     #
+        # envelope), the delta text is scanned instead so nothing escapes     #
+        # unchecked. No structured write-back is possible here; guardrails    #
+        # that only need to block/flag (not rewrite) still work correctly,    #
+        # and a rewrite a caller expects delivered fails closed instead.      #
         # ------------------------------------------------------------------ #
         string_so_far: Final = self.get_streaming_string_so_far(responses_so_far)
-        if string_so_far:
-            fallback_inputs: Final = GenericGuardrailAPIInputs(texts=[string_so_far])
+        text_to_check: Final = string_so_far or (
+            self._delta_text_so_far(responses_so_far) if deliver_ended_stream_rewrites else ""
+        )
+        if text_to_check:
+            fallback_inputs: Final = GenericGuardrailAPIInputs(texts=[text_to_check])
             response_model = (
                 final_chunk.get("response", {}).get("model") if isinstance(final_chunk.get("response"), dict) else None
             )
@@ -654,7 +660,7 @@ class OpenAIResponsesHandler(BaseTranslation):
                 logging_obj=litellm_logging_obj,
             )
             fallback_texts: Final = fallback_outputs.get("texts")
-            if deliver_ended_stream_rewrites and fallback_texts and tuple(fallback_texts) != (string_so_far,):
+            if deliver_ended_stream_rewrites and fallback_texts and tuple(fallback_texts) != (text_to_check,):
                 from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
 
                 raise UndeliverableStreamRewrite(guardrail_to_apply.guardrail_name or "unknown")
@@ -754,6 +760,17 @@ class OpenAIResponsesHandler(BaseTranslation):
         """
         return "".join([response.get("text", "") for response in responses_so_far])
 
+    @staticmethod
+    def _delta_text_so_far(responses_so_far: Sequence[ResponsesStreamChunk]) -> str:
+        """Accumulate the text carried by ``response.output_text.delta`` events,
+        for buffers where no ``.done`` event or terminal envelope repeats it."""
+        deltas: Final = (
+            response.get("delta")
+            for response in responses_so_far
+            if response.get("type") == ResponsesAPIStreamEvents.OUTPUT_TEXT_DELTA.value
+        )
+        return "".join(delta for delta in deltas if isinstance(delta, str))
+
     def _has_text_content(self, response: "ResponsesAPIResponse") -> bool:
         """
         Check if response has any text content to process.
diff --git a/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py b/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py
index 26442a4a6ed..b177d4a73d4 100644
--- a/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py
+++ b/tests/test_litellm/llms/openai/chat/guardrail_translation/test_openai_guardrail_handler.py
@@ -1128,6 +1128,99 @@ class TestOpenAIChatCompletionsHandlerStreamingOutput:
         assert chunks[1].choices[0].delta.content == " world"
         assert chunks[1].choices[0].finish_reason == "stop"
 
+    @staticmethod
+    def _two_choice_stream_chunks() -> list:
+        from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
+
+        def chunk(index: int, content: str, finish_reason: Optional[str] = None) -> ModelResponseStream:
+            return ModelResponseStream(
+                id="chatcmpl-123",
+                created=1234567890,
+                model="gpt-4",
+                object="chat.completion.chunk",
+                choices=[StreamingChoices(index=index, delta=Delta(content=content), finish_reason=finish_reason)],
+            )
+
+        return [
+            chunk(0, "safe "),
+            chunk(1, "hello "),
+            chunk(0, "text", "stop"),
+            chunk(1, "world", "stop"),
+        ]
+
+    @staticmethod
+    def _world_masking_guardrail() -> CustomGuardrail:
+        class MaskWorld(CustomGuardrail):
+            async def apply_guardrail(
+                self,
+                inputs: GenericGuardrailAPIInputs,
+                request_data: dict,
+                input_type: Literal["request", "response"],
+                logging_obj: Optional[Any] = None,
+            ) -> GenericGuardrailAPIInputs:
+                texts = inputs.get("texts", [])
+                return {**inputs, "texts": [t.replace("world", "[MASKED]") for t in texts]}
+
+        return MaskWorld(guardrail_name="test-mask")
+
+    @pytest.mark.asyncio
+    async def test_deliver_ended_stream_rewrite_on_multi_choice_stream_fails_closed(self):
+        from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+        handler = OpenAIChatCompletionsHandler()
+        chunks = self._two_choice_stream_chunks()
+
+        with pytest.raises(UndeliverableStreamRewrite):
+            await handler.process_output_streaming_response(
+                responses_so_far=chunks,
+                guardrail_to_apply=self._world_masking_guardrail(),
+                litellm_logging_obj=None,
+                deliver_ended_stream_rewrites=True,
+            )
+
+    @pytest.mark.asyncio
+    async def test_deliver_ended_stream_clean_multi_choice_stream_released_untouched(self):
+        handler = OpenAIChatCompletionsHandler()
+        chunks = self._two_choice_stream_chunks()
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=MockPassThroughGuardrail(guardrail_name="test"),
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is chunks
+        assert [c.choices[0].delta.content for c in chunks] == ["safe ", "hello ", "text", "world"]
+
+    @pytest.mark.asyncio
+    async def test_deliver_ended_stream_rewrite_lands_on_nonzero_choice_index(self):
+        from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
+
+        handler = OpenAIChatCompletionsHandler()
+
+        def chunk(content: str, finish_reason: Optional[str]) -> ModelResponseStream:
+            return ModelResponseStream(
+                id="chatcmpl-123",
+                created=1234567890,
+                model="gpt-4",
+                object="chat.completion.chunk",
+                choices=[StreamingChoices(index=1, delta=Delta(content=content), finish_reason=finish_reason)],
+            )
+
+        chunks = [chunk("hello ", None), chunk("world", "stop")]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=self._world_masking_guardrail(),
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is chunks
+        assert chunks[0].choices[0].delta.content == "hello [MASKED]"
+        assert chunks[1].choices[0].delta.content in (None, "")
+
 
 class TestGetStructuredMessages:
     """Test the get_structured_messages method."""
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
index dfd6352f9ec..ae3700b1123 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
@@ -1212,6 +1212,77 @@ class TestOpenAIResponsesHandlerStreamingOutputProcessing:
                 deliver_ended_stream_rewrites=True,
             )
 
+    @staticmethod
+    def _recording_guardrail() -> "tuple[CustomGuardrail, List[List[str]]]":
+        seen: List[List[str]] = []
+
+        class Recorder(CustomGuardrail):
+            async def apply_guardrail(
+                self,
+                inputs: GenericGuardrailAPIInputs,
+                request_data: dict,
+                input_type: Literal["request", "response"],
+                logging_obj: Optional[Any] = None,
+            ) -> GenericGuardrailAPIInputs:
+                seen.append(list(inputs.get("texts", [])))
+                return inputs
+
+        return Recorder(guardrail_name="recorder"), seen
+
+    @pytest.mark.asyncio
+    async def test_fallback_delta_only_rewrite_with_delivery_expected_fails_closed(self):
+        from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+        handler = OpenAIResponsesHandler()
+        events = [
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "hello "},
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "world"},
+        ]
+
+        with pytest.raises(UndeliverableStreamRewrite):
+            await handler.process_output_streaming_response(
+                responses_so_far=events,
+                guardrail_to_apply=self._masking_guardrail(),
+                litellm_logging_obj=None,
+                deliver_ended_stream_rewrites=True,
+            )
+
+    @pytest.mark.asyncio
+    async def test_fallback_scans_delta_text_when_delivery_expected(self):
+        handler = OpenAIResponsesHandler()
+        guardrail, seen = self._recording_guardrail()
+        events = [
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "hello "},
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "there"},
+        ]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=guardrail,
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is events
+        assert seen == [["hello there"]]
+
+    @pytest.mark.asyncio
+    async def test_fallback_ignores_delta_text_without_delivery_expected(self):
+        handler = OpenAIResponsesHandler()
+        guardrail, seen = self._recording_guardrail()
+        events = [
+            {"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "hello world"},
+        ]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=guardrail,
+            litellm_logging_obj=None,
+        )
+
+        assert result is events
+        assert seen == []
+
     @pytest.mark.asyncio
     async def test_fallback_rewrite_without_delivery_expected_does_not_raise(self):
         handler = OpenAIResponsesHandler()

From c09db7c7a3d77d3ccd6a1a2163778b6f22f29a8f Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 1 Sep 2026 22:01:04 -0700
Subject: [PATCH 19/25] Fail closed on rewrites for buffers that never reached
 their terminal event

An Anthropic buffer without a stop_reason only ran the flat text scan, so a
rewrite there was dropped while the executor trusted the translation to have
delivered it. A Responses buffer ending at response.output_item.done returned
after the tool-call scan without ever checking the text. Both now reach the
flat scan and raise UndeliverableStreamRewrite when a caller expects the
rewrite delivered, matching the existing Responses no-envelope fallback.
---
 .../chat/guardrail_translation/handler.py     |  8 +++-
 .../guardrail_translation/handler.py          |  7 ++-
 .../test_anthropic_guardrail_handler.py       | 46 +++++++++++++++++++
 ...test_openai_responses_guardrail_handler.py | 46 +++++++++++++++++++
 4 files changed, 104 insertions(+), 3 deletions(-)

diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py
index 091ad8ecba9..fc276dd7fe6 100644
--- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py
+++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py
@@ -1020,7 +1020,8 @@ class AnthropicMessagesHandler(BaseTranslation):
 
         Get the string so far, check the apply guardrail to the string so far, and return the list of responses so far.
         With ``deliver_ended_stream_rewrites``, an ended stream whose guardrail rewrote the text gets the rewrite
-        written back across the buffered chunks (full rewritten text in the first ``text_delta``, the rest blanked).
+        written back across the buffered chunks (full rewritten text in the first ``text_delta``, the rest blanked);
+        a rewrite on a stream that never reported a ``stop_reason`` has no write-back and fails closed instead.
         """
         from litellm.integrations.custom_guardrail import ModifyResponseException
 
@@ -1098,6 +1099,11 @@ class AnthropicMessagesHandler(BaseTranslation):
             if e.original_response is None:
                 e.original_response = self._build_streaming_usage_response(responses_so_far, request_data)
             raise
+        unended_texts: Final = _guardrailed_inputs.get("texts")
+        if deliver_ended_stream_rewrites and unended_texts and tuple(unended_texts) != (string_so_far,):
+            from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+            raise UndeliverableStreamRewrite(guardrail_to_apply.guardrail_name or "unknown")
         return responses_so_far
 
     def _prepare_request_data(
diff --git a/litellm/llms/openai/responses/guardrail_translation/handler.py b/litellm/llms/openai/responses/guardrail_translation/handler.py
index 8b808348c79..724a0a1d2f0 100644
--- a/litellm/llms/openai/responses/guardrail_translation/handler.py
+++ b/litellm/llms/openai/responses/guardrail_translation/handler.py
@@ -638,7 +638,9 @@ class OpenAIResponsesHandler(BaseTranslation):
                 return responses_so_far
 
         # ------------------------------------------------------------------ #
-        # Case 2: response.output_item.done — extract tool calls only.        #
+        # Case 2: response.output_item.done — extract tool calls only, then  #
+        # fall through to the text fallback when a caller expects rewrites   #
+        # delivered, so a buffer truncated here still fails closed on text.  #
         # ------------------------------------------------------------------ #
         if final_chunk.get("type") == "response.output_item.done":
             model_response_stream: Final = (
@@ -656,7 +658,8 @@ class OpenAIResponsesHandler(BaseTranslation):
                     input_type="response",
                     logging_obj=litellm_logging_obj,
                 )
-            return responses_so_far
+            if not deliver_ended_stream_rewrites:
+                return responses_so_far
 
         # ------------------------------------------------------------------ #
         # Fallback: apply guardrail to the accumulated text string.           #
diff --git a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py
index f60655cd7ec..274c351ebc7 100644
--- a/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py
+++ b/tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py
@@ -328,6 +328,52 @@ class TestAnthropicMessagesHandlerStreamingOutputProcessing:
 
         assert chunks == original
 
+    @pytest.mark.asyncio
+    async def test_unended_stream_rewrite_with_delivery_expected_fails_closed(self):
+        from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+        handler = AnthropicMessagesHandler()
+        chunks = self._ended_sse_chunks()[:-2]
+
+        with pytest.raises(UndeliverableStreamRewrite):
+            await handler.process_output_streaming_response(
+                responses_so_far=chunks,
+                guardrail_to_apply=self._masking_guardrail(),
+                litellm_logging_obj=MagicMock(),
+                deliver_ended_stream_rewrites=True,
+            )
+
+    @pytest.mark.asyncio
+    async def test_unended_stream_without_rewrite_is_released_with_delivery_expected(self):
+        handler = AnthropicMessagesHandler()
+        chunks = self._ended_sse_chunks()[:-2]
+        original = [bytes(chunk) for chunk in chunks]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=MockPassThroughGuardrail(guardrail_name="test"),
+            litellm_logging_obj=MagicMock(),
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is chunks
+        assert chunks == original
+
+    @pytest.mark.asyncio
+    async def test_unended_stream_rewrite_without_delivery_expected_does_not_raise(self):
+        handler = AnthropicMessagesHandler()
+        chunks = self._ended_sse_chunks()[:-2]
+        original = [bytes(chunk) for chunk in chunks]
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=chunks,
+            guardrail_to_apply=self._masking_guardrail(),
+            litellm_logging_obj=MagicMock(),
+        )
+
+        assert result is chunks
+        assert chunks == original
+
 
 class TestAnthropicMessagesHandlerInputProcessing:
     """Test input processing preserves litellm_metadata for dynamic guardrails."""
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
index 155b9ef9810..bd07e924d12 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_guardrail_handler.py
@@ -1248,6 +1248,52 @@ class TestOpenAIResponsesHandlerStreamingOutputProcessing:
                 deliver_ended_stream_rewrites=True,
             )
 
+    @pytest.mark.asyncio
+    async def test_output_item_done_last_rewrite_with_delivery_expected_fails_closed(self):
+        from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+        handler = OpenAIResponsesHandler()
+        events = self._ended_stream_events()[:-1]
+
+        with pytest.raises(UndeliverableStreamRewrite):
+            await handler.process_output_streaming_response(
+                responses_so_far=events,
+                guardrail_to_apply=self._masking_guardrail(),
+                litellm_logging_obj=None,
+                deliver_ended_stream_rewrites=True,
+            )
+
+    @pytest.mark.asyncio
+    async def test_output_item_done_last_scans_text_with_delivery_expected(self):
+        handler = OpenAIResponsesHandler()
+        events = self._ended_stream_events()[:-1]
+        guardrail = MockRecordingGuardrail(guardrail_name="test")
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=guardrail,
+            litellm_logging_obj=None,
+            deliver_ended_stream_rewrites=True,
+        )
+
+        assert result is events
+        assert [inputs.get("texts") for inputs in guardrail.seen_inputs] == [["hello world"]]
+
+    @pytest.mark.asyncio
+    async def test_output_item_done_last_without_delivery_expected_skips_text(self):
+        handler = OpenAIResponsesHandler()
+        events = self._ended_stream_events()[:-1]
+        guardrail = MockRecordingGuardrail(guardrail_name="test")
+
+        result = await handler.process_output_streaming_response(
+            responses_so_far=events,
+            guardrail_to_apply=guardrail,
+            litellm_logging_obj=None,
+        )
+
+        assert result is events
+        assert guardrail.seen_inputs == []
+
     @pytest.mark.asyncio
     async def test_fallback_rewrite_without_delivery_expected_does_not_raise(self):
         handler = OpenAIResponsesHandler()

From 192ea9ec80ed97d771f9258320ac948080e87d2e Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 7 Sep 2026 17:55:37 -0700
Subject: [PATCH 20/25] fix(policy_engine): fail open on streaming shapes
 post_call pipelines cannot govern yet

A post_call pipeline now releases the original stream instead of refusing the
request on every shape it has no handler for: a background request, a pipeline
guardrail without the unified apply_guardrail interface, a route with no
endpoint translation, a buffered stream no translation resolves, and a rewrite
the translation cannot write back (tool-call edits, text edits on translations
without write-back, n>1 chat, an unended Anthropic stream, a Responses dump
with no event envelope). Each case logs a warning naming the policy and
guardrail. Real blocks and writable text masks are unchanged.
---
 .../chat/guardrail_translation/handler.py     |   3 +-
 .../guardrail_translation/base_translation.py |   5 +-
 .../chat/guardrail_translation/handler.py     |   5 +-
 .../guardrail_translation/handler.py          |   7 +-
 .../proxy/policy_engine/pipeline_executor.py  |  66 ++--
 litellm/proxy/utils.py                        | 226 ++++++--------
 .../policy_engine/test_pipeline_executor.py   | 137 ++++++++-
 .../proxy_logging/test_guardrail_pipeline.py  | 282 +++++++++++-------
 8 files changed, 437 insertions(+), 294 deletions(-)

diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py
index da30d85e26b..e486be12fe2 100644
--- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py
+++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py
@@ -1027,7 +1027,8 @@ class AnthropicMessagesHandler(BaseTranslation):
         Get the string so far, check the apply guardrail to the string so far, and return the list of responses so far.
         With ``deliver_ended_stream_rewrites``, an ended stream whose guardrail rewrote the text gets the rewrite
         written back across the buffered chunks (full rewritten text in the first ``text_delta``, the rest blanked);
-        a rewrite on a stream that never reported a ``stop_reason`` has no write-back and fails closed instead.
+        a rewrite on a stream that never reported a ``stop_reason`` has no write-back and is reported as
+        undeliverable, so the pipeline executor discards it and releases the original chunks.
         """
         from litellm.integrations.custom_guardrail import ModifyResponseException
 
diff --git a/litellm/llms/base_llm/guardrail_translation/base_translation.py b/litellm/llms/base_llm/guardrail_translation/base_translation.py
index 770fac6e443..afd8e0f67f7 100644
--- a/litellm/llms/base_llm/guardrail_translation/base_translation.py
+++ b/litellm/llms/base_llm/guardrail_translation/base_translation.py
@@ -56,8 +56,9 @@ class BaseTranslation(ABC):
     """Whether ``process_output_streaming_response`` accepts
     ``deliver_ended_stream_rewrites=True`` and, on an ended (fully buffered)
     stream, writes guardrail text rewrites back across ``responses_so_far`` so
-    a buffered pipeline can release rewritten chunks instead of withholding the
-    stream. Tool-call rewrites stay undeliverable everywhere."""
+    a buffered pipeline can release rewritten chunks. Tool-call rewrites, and
+    text rewrites on every other translation, are undeliverable: the pipeline
+    executor discards them and releases the original chunks."""
 
     @staticmethod
     def transform_user_api_key_dict_to_metadata(
diff --git a/litellm/llms/openai/chat/guardrail_translation/handler.py b/litellm/llms/openai/chat/guardrail_translation/handler.py
index 4b7b1cc2700..80292aef2cf 100644
--- a/litellm/llms/openai/chat/guardrail_translation/handler.py
+++ b/litellm/llms/openai/chat/guardrail_translation/handler.py
@@ -1015,7 +1015,8 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         content-carrying chunk and the rest are blanked, the same shape the
         in-flight write-back uses. Chunks carrying only finish_reason or usage
         stay untouched. A rewrite on a stream carrying more than one distinct
-        choice index fails closed."""
+        choice index is reported as undeliverable, so the pipeline executor
+        discards it and releases the original chunks."""
         post_guardrail_texts: Final = self._string_choice_contents(guardrailed_response)
         changed: Final = tuple(
             after
@@ -1030,7 +1031,7 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         if len(stream_choice_indices) != 1:
             # stream_chunk_builder collapses every choice into one index-0
             # choice, so a rewrite of the rebuilt response cannot be attributed
-            # back to a single choice on an n>1 stream: withhold the stream
+            # back to a single choice on an n>1 stream: report it undeliverable
             # rather than deliver the rewrite on the wrong choice
             from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
 
diff --git a/litellm/llms/openai/responses/guardrail_translation/handler.py b/litellm/llms/openai/responses/guardrail_translation/handler.py
index b543c7f1e17..b0f79552bc5 100644
--- a/litellm/llms/openai/responses/guardrail_translation/handler.py
+++ b/litellm/llms/openai/responses/guardrail_translation/handler.py
@@ -699,7 +699,8 @@ class OpenAIResponsesHandler(BaseTranslation):
         ``response.content_part.done``, ``response.output_item.done``) are synced
         to the rewritten envelope too, so a client reading deltas sees the
         rewrite instead of the raw model output; a rewrite observed where no
-        write-back is possible fails closed instead of releasing raw output.
+        write-back is possible is reported as undeliverable, so the pipeline
+        executor discards it and releases the original events.
         """
         if not responses_so_far:
             return responses_so_far
@@ -788,7 +789,7 @@ class OpenAIResponsesHandler(BaseTranslation):
         # ------------------------------------------------------------------ #
         # Case 2: response.output_item.done — extract tool calls only, then  #
         # fall through to the text fallback when a caller expects rewrites   #
-        # delivered, so a buffer truncated here still fails closed on text.  #
+        # delivered, so a truncated buffer still reports text undeliverable. #
         # ------------------------------------------------------------------ #
         if final_chunk.get("type") == "response.output_item.done":
             model_response_stream: Final = (
@@ -813,7 +814,7 @@ class OpenAIResponsesHandler(BaseTranslation):
         # Fallback: apply guardrail to the accumulated text string.           #
         # No structured write-back is possible here; guardrails that only     #
         # need to block/flag (not rewrite) still work correctly, and a        #
-        # rewrite a caller expects delivered fails closed instead.            #
+        # rewrite a caller expects delivered is reported undeliverable.       #
         # ------------------------------------------------------------------ #
         string_so_far: Final = self.get_streaming_string_so_far(responses_so_far)
         if string_so_far:
diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index 160938b723c..970ac20487c 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -5,6 +5,7 @@ Runs guardrails sequentially per pipeline step definitions, handling
 pass/fail actions (allow, block, next, modify_response) and data forwarding.
 """
 
+import copy
 import time
 from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Any, Final, Literal
@@ -77,7 +78,7 @@ class _StreamRewriteObserver(CustomGuardrail):
     which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text
     rewrites are deliverable on translations that write them back across the buffered chunks
     (``delivers_ended_stream_text_rewrites``); tool-call rewrites and text rewrites on any
-    other translation make the gate withhold the stream."""
+    other translation are discarded by the executor, which releases the original chunks."""
 
     def __init__(self, inner: CustomGuardrail) -> None:
         super().__init__(guardrail_name=inner.guardrail_name)
@@ -133,6 +134,19 @@ def _prepare_hook_input(
     return hook_input, scans_raw_request
 
 
+def _release_original_chunks(
+    guardrail_name: str,
+    streaming_chunks: list[object],  # mutable-ok: shared buffered-stream chunks, restored in place
+    originals: Sequence[object],
+) -> None:
+    streaming_chunks[:] = originals  # rebind-ok: the caller's buffer is the stream the client receives
+    verbose_proxy_logger.warning(
+        "Pipeline: guardrail '%s' rewrote the streamed response in a way this endpoint's streaming "
+        "pipeline cannot deliver yet; the rewrite was discarded and the original stream released",
+        guardrail_name,
+    )
+
+
 class PipelineExecutor:
     """Executes guardrail pipelines with ordered, conditional step logic."""
 
@@ -263,29 +277,37 @@ class PipelineExecutor:
         litellm_logging_obj: "LiteLLMLoggingObj | None",
     ) -> None:
         """Run one streaming post_call step through the endpoint translation, delivering
-        text rewrites on translations that support ended-stream write-back and raising
-        ``UndeliverableStreamRewrite`` for any rewrite that cannot reach the client."""
+        text rewrites on translations that support ended-stream write-back. A rewrite that
+        cannot reach the client yet (a tool-call rewrite, a text rewrite on a translation
+        without write-back, or one the translation refused with
+        ``UndeliverableStreamRewrite``) is discarded: the buffered chunks go back to the
+        originals and the step passes, so the client gets the stream the merge base sent."""
         observer: Final = _StreamRewriteObserver(callback)
         deliver_rewrites: Final = type(endpoint_translation).delivers_ended_stream_text_rewrites
-        if deliver_rewrites:
-            await endpoint_translation.process_output_streaming_response(
-                responses_so_far=streaming_chunks,
-                guardrail_to_apply=observer,
-                litellm_logging_obj=litellm_logging_obj,
-                user_api_key_dict=user_api_key_dict,
-                request_data=hook_input,
-                deliver_ended_stream_rewrites=True,
-            )
-        else:
-            await endpoint_translation.process_output_streaming_response(
-                responses_so_far=streaming_chunks,
-                guardrail_to_apply=observer,
-                litellm_logging_obj=litellm_logging_obj,
-                user_api_key_dict=user_api_key_dict,
-                request_data=hook_input,
-            )
+        originals: Final = copy.deepcopy(streaming_chunks)
+        try:
+            if deliver_rewrites:
+                await endpoint_translation.process_output_streaming_response(
+                    responses_so_far=streaming_chunks,
+                    guardrail_to_apply=observer,
+                    litellm_logging_obj=litellm_logging_obj,
+                    user_api_key_dict=user_api_key_dict,
+                    request_data=hook_input,
+                    deliver_ended_stream_rewrites=True,
+                )
+            else:
+                await endpoint_translation.process_output_streaming_response(
+                    responses_so_far=streaming_chunks,
+                    guardrail_to_apply=observer,
+                    litellm_logging_obj=litellm_logging_obj,
+                    user_api_key_dict=user_api_key_dict,
+                    request_data=hook_input,
+                )
+        except UndeliverableStreamRewrite:
+            _release_original_chunks(step.guardrail, streaming_chunks, originals)
+            return
         if observer.rewrote_tool_calls or (observer.rewrote_texts and not deliver_rewrites):
-            raise UndeliverableStreamRewrite(step.guardrail)
+            _release_original_chunks(step.guardrail, streaming_chunks, originals)
 
     @staticmethod
     async def _run_step(
@@ -386,8 +408,6 @@ class PipelineExecutor:
                 )  # mutable-ok: modified-data contract is a plain dict
             return ("pass", response if isinstance(response, dict) else None, None, None)
 
-        except UndeliverableStreamRewrite:
-            raise
         except Exception as e:
             if CustomGuardrail._is_guardrail_intervention(e):
                 error_msg: Final = _extract_error_message(e)
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index ca8e48b5dfa..d7bf832bca4 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -19,7 +19,7 @@ from email.mime.text import MIMEText
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, ClassVar, Final, Literal, Optional, Protocol, TypeVar, Union, cast, overload
 
-from typing_extensions import NotRequired, ReadOnly, TypedDict
+from typing_extensions import ReadOnly, TypedDict
 
 from litellm import _custom_logger_compatible_callbacks_literal
 from litellm.constants import (
@@ -155,7 +155,7 @@ from litellm.proxy.hooks.sensitive_data_routing import (
 )
 from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup, add_guardrails_from_auth_metadata
 from litellm.proxy.management_helpers.key_settings_audit import with_settings_updated_at
-from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor, UndeliverableStreamRewrite
+from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor
 from litellm.repositories.budget_repository import BudgetRepository
 from litellm.repositories.config_repository import ConfigRepository
 from litellm.repositories.table_repositories import (
@@ -518,108 +518,72 @@ def _pipeline_step_supports_unified_streaming(guardrail_name: str) -> bool:
     return callback is not None and PipelineExecutor.supports_unified_execution(callback)
 
 
-class _PipelineErrorBody(TypedDict):
-    message: ReadOnly[str]
-    type: ReadOnly[str]
-    policies: ReadOnly[tuple[str, ...]]
-    guardrails: NotRequired[ReadOnly[tuple[str, ...]]]
-
-
-class _PipelineErrorDetail(TypedDict):
-    error: ReadOnly[_PipelineErrorBody]
-
-
-def _undeliverable_stream_rewrite_error(policy_name: str, guardrail_name: str) -> HTTPException:
-    detail: Final[_PipelineErrorDetail] = {
-        "error": {
-            "message": (
-                f"Streaming response withheld by policy pipeline '{policy_name}' because guardrail "
-                f"'{guardrail_name}' rewrote the streamed output in a way this endpoint's streaming "
-                "pipeline cannot deliver (a tool-call rewrite, or a text rewrite on a route without "
-                "stream write-back). Retry with stream=false, or drop it from the pipeline steps so "
-                "guardrails.add applies it to streamed output."
-            ),
-            "type": "guardrail_pipeline_error",
-            "policies": (policy_name,),
-            "guardrails": (guardrail_name,),
-        }
-    }
-    return HTTPException(status_code=400, detail=detail)
-
-
-def _raise_for_streaming_post_call_pipelines(data: Mapping[str, object], user_api_key_dict: UserAPIKeyAuth) -> None:
-    """
-    Reject up front the requests whose post_call pipelines could never run.
-
-    Background responses skip the post_call hooks entirely, so a pipeline
-    governing one would silently never execute. Streaming responses execute
-    pipelines against the buffered stream through the endpoint guardrail
-    translation of the request route, releasing the buffered chunks on allow
-    (rewritten in place when a guardrail rewrote text and the translation
-    delivers ended-stream rewrites; a rewrite the translation cannot deliver
-    fails closed at runtime instead). That needs every step's guardrail to
-    support the unified apply_guardrail interface, and needs the route to have
-    a translation at all; anything else keeps the 400 rather than letting
-    ungoverned output stream through.
-    """
-    is_stream: Final = data.get("stream") is True
-    is_background: Final = data.get("background") is True
-    if not is_stream and not is_background:
-        return
-    post_call_pipelines: Final = tuple(
+def _post_call_pipelines(data: Mapping[str, object]) -> tuple[tuple[str, "GuardrailPipeline"], ...]:
+    return tuple(
         (policy_name, pipeline) for policy_name, pipeline in _policy_pipelines(data) if pipeline.mode == "post_call"
     )
+
+
+def _warn_background_skips_post_call_pipelines(data: Mapping[str, object]) -> None:
+    if data.get("background") is not True:
+        return
+    policy_names: Final = tuple(policy_name for policy_name, _pipeline in _post_call_pipelines(data))
+    if not policy_names:
+        return
+    verbose_proxy_logger.warning(
+        "Policies with post_call guardrail pipelines do not run on background responses yet; "
+        "the response is released ungoverned by them: %s",
+        ", ".join(policy_names),
+    )
+
+
+def _pipeline_is_streamable(policy_name: str, pipeline: "GuardrailPipeline") -> bool:
+    unsupported: Final = tuple(
+        dict.fromkeys(
+            step.guardrail for step in pipeline.steps if not _pipeline_step_supports_unified_streaming(step.guardrail)
+        )
+    )
+    if not unsupported:
+        return True
+    verbose_proxy_logger.warning(
+        "Policy '%s' has post_call pipeline guardrails without the unified apply_guardrail interface, "
+        "which streaming pipelines need; the stream is released ungoverned by it: %s",
+        policy_name,
+        ", ".join(unsupported),
+    )
+    return False
+
+
+def _streamable_post_call_pipelines(
+    request_data: Mapping[str, object], user_api_key_dict: UserAPIKeyAuth
+) -> tuple[tuple[str, "GuardrailPipeline"], ...]:
+    """
+    The post_call pipelines a streaming response can be gated through.
+
+    Streaming pipelines scan the buffered stream through the endpoint guardrail
+    translation of the request route, so every step's guardrail needs the
+    unified apply_guardrail interface and the route needs a translation. A
+    pipeline that cannot be run that way yet is left out and the stream is
+    released the way it was before pipelines ran on streams at all, with a
+    warning naming what went ungoverned.
+    """
+    post_call_pipelines: Final = _post_call_pipelines(request_data)
     if not post_call_pipelines:
-        return
-    post_call_policies: Final = tuple(policy_name for policy_name, _pipeline in post_call_pipelines)
-    if is_background:
-        background_detail: Final[_PipelineErrorDetail] = {
-            "error": {
-                "message": (
-                    "Policies with post_call guardrail pipelines cannot govern background "
-                    f"responses: {', '.join(post_call_policies)}. Retry with background=false."
-                ),
-                "type": "guardrail_pipeline_error",
-                "policies": post_call_policies,
-            }
-        }
-        raise HTTPException(status_code=400, detail=background_detail)
-    step_guardrails: Final = tuple(
-        dict.fromkeys(step.guardrail for _policy_name, pipeline in post_call_pipelines for step in pipeline.steps)
-    )
-    unsupported_guardrails: Final = tuple(
-        guardrail for guardrail in step_guardrails if not _pipeline_step_supports_unified_streaming(guardrail)
-    )
-    if unsupported_guardrails:
-        unsupported_detail: Final[_PipelineErrorDetail] = {
-            "error": {
-                "message": (
-                    "Policies with post_call guardrail pipelines cannot govern streaming responses "
-                    "because these pipeline guardrails do not support the unified apply_guardrail "
-                    f"interface: {', '.join(unsupported_guardrails)}. Retry with stream=false, or drop "
-                    "them from the pipeline steps so guardrails.add scans them on streamed output."
-                ),
-                "type": "guardrail_pipeline_error",
-                "policies": post_call_policies,
-                "guardrails": unsupported_guardrails,
-            }
-        }
-        raise HTTPException(status_code=400, detail=unsupported_detail)
+        return ()
     route: Final = user_api_key_dict.request_route
-    if not route or resolve_endpoint_translation(user_api_key_dict, None) is not None:
-        return
-    route_detail: Final[_PipelineErrorDetail] = {
-        "error": {
-            "message": (
-                "Policies with post_call guardrail pipelines cannot govern streaming responses on "
-                f"route {route} because it has no endpoint guardrail translation to scan the stream "
-                f"through: {', '.join(post_call_policies)}. Retry with stream=false."
-            ),
-            "type": "guardrail_pipeline_error",
-            "policies": post_call_policies,
-        }
-    }
-    raise HTTPException(status_code=400, detail=route_detail)
+    if route and resolve_endpoint_translation(user_api_key_dict, None) is None:
+        verbose_proxy_logger.warning(
+            "Policies with post_call guardrail pipelines cannot scan streaming responses on route %s yet "
+            "(no endpoint guardrail translation); the stream is released ungoverned by them: %s",
+            route,
+            ", ".join(policy_name for policy_name, _pipeline in post_call_pipelines),
+        )
+        return ()
+    return tuple(
+        (policy_name, pipeline)
+        for policy_name, pipeline in post_call_pipelines
+        if _pipeline_is_streamable(policy_name, pipeline)
+    )
 
 
 def _prompt_block_text(block: object) -> str:
@@ -1990,7 +1954,7 @@ class ProxyLogging:
         )
 
         try:
-            _raise_for_streaming_post_call_pipelines(data, user_api_key_dict)
+            _warn_background_skips_post_call_pipelines(data)
 
             # Execute guardrail pipelines before the normal callback loop
             data, _ = await self._maybe_execute_pipelines(  # rebind-ok: pipeline edits feed the callback loop below
@@ -3371,11 +3335,7 @@ class ProxyLogging:
         1. /chat/completions
         """
         caps: Final = ProxyLogging._callback_capabilities()
-        post_call_pipelines: Final = tuple(
-            (policy_name, pipeline)
-            for policy_name, pipeline in _policy_pipelines(request_data)
-            if pipeline.mode == "post_call"
-        )
+        post_call_pipelines: Final = _streamable_post_call_pipelines(request_data, user_api_key_dict)
         # Fast path: no real overrides. Internal proxy CustomLogger callbacks
         # (e.g. _PROXY_MaxBudgetLimiter, ManagedFiles) inherit the default
         # ``async for chunk: yield chunk`` body, so wrapping the iterator
@@ -3486,9 +3446,11 @@ class ProxyLogging:
         output, rewritten in place when one rewrote text and the translation
         delivers ended-stream rewrites (later steps then re-scan the rewritten
         chunks, so rewrites chain). A rewrite the translation cannot deliver
-        (a tool-call rewrite, or a text rewrite on a route without write-back)
-        withholds the stream with a 400; a block or modify_response terminates
-        with the translation's block chunks or the raised error.
+        yet (a tool-call rewrite, or a text rewrite on a route without
+        write-back) is discarded by the executor and the original chunks are
+        released, as is a buffered shape no translation resolves; a block or
+        modify_response terminates with the translation's block chunks or the
+        raised error.
         """
         buffered: Final[list[object]] = []  # mutable-ok: accumulates the stream before the pipeline verdict
         async for item in response:
@@ -3498,39 +3460,27 @@ class ProxyLogging:
 
         resolved: Final = resolve_endpoint_translation(user_api_key_dict, buffered[0])
         if resolved is None:
-            policy_names: Final = tuple(policy_name for policy_name, _pipeline in pipelines)
-            raise ProxyException(
-                message=(
-                    "Policy pipelines could not govern this streaming response shape; "
-                    f"the response was withheld: {', '.join(policy_names)}."
-                ),
-                type="guardrail_pipeline_error",
-                param=None,
-                code=500,
+            verbose_proxy_logger.warning(
+                "Policies with post_call guardrail pipelines cannot scan this streaming response shape yet; "
+                "the stream is released ungoverned by them: %s",
+                ", ".join(policy_name for policy_name, _pipeline in pipelines),
             )
+            for buffered_item in buffered:
+                yield buffered_item
+            return
         call_type, endpoint_translation = resolved
 
         for policy_name, pipeline in pipelines:
-            try:
-                result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
-                    steps=pipeline.steps,
-                    mode="post_call",
-                    data=request_data,
-                    user_api_key_dict=user_api_key_dict,
-                    call_type=call_type,
-                    policy_name=policy_name,
-                    streaming_chunks=buffered,
-                    endpoint_translation=endpoint_translation,
-                )
-            except UndeliverableStreamRewrite as rewrite:
-                async for error_chunk in unified_guardrail.emit_streaming_http_error(
-                    _undeliverable_stream_rewrite_error(policy_name, rewrite.guardrail_name),
-                    call_type,
-                    buffered,
-                    request_data,
-                ):
-                    yield error_chunk
-                return
+            result: PipelineExecutionResult = await PipelineExecutor.execute_steps(
+                steps=pipeline.steps,
+                mode="post_call",
+                data=request_data,
+                user_api_key_dict=user_api_key_dict,
+                call_type=call_type,
+                policy_name=policy_name,
+                streaming_chunks=buffered,
+                endpoint_translation=endpoint_translation,
+            )
             try:
                 ProxyLogging._handle_pipeline_result(
                     result, data=request_data, policy_name=policy_name, original_response=buffered
diff --git a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
index d399b4f01cb..0685bc6aa1e 100644
--- a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
+++ b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
@@ -4,6 +4,7 @@ Tests for the pipeline executor.
 Uses mock guardrails to validate pipeline execution without external services.
 """
 
+import logging
 from unittest.mock import MagicMock
 
 import pytest
@@ -941,7 +942,55 @@ class _TextTranslation:
         return responses_so_far
 
 
-async def _run_streaming_step(returned_texts, translation):
+class _WritingTranslation:
+    """Writes the guardrail's text (and tool-call) outputs back into the buffered chunks the way the
+    chat/Responses/Messages handlers do on an ended stream."""
+
+    delivers_ended_stream_text_rewrites = True
+
+    async def process_output_streaming_response(
+        self,
+        responses_so_far,
+        guardrail_to_apply,
+        litellm_logging_obj=None,
+        user_api_key_dict=None,
+        request_data=None,
+        deliver_ended_stream_rewrites=False,
+    ):
+        assert deliver_ended_stream_rewrites is True
+        outputs = await guardrail_to_apply.apply_guardrail(
+            inputs={"texts": [responses_so_far[0]["text"]], "tool_calls": [dict(responses_so_far[0]["tool_call"])]},
+            request_data=request_data or {},
+            input_type="response",
+            logging_obj=litellm_logging_obj,
+        )
+        responses_so_far[0]["text"] = outputs["texts"][0]
+        responses_so_far[0]["tool_call"] = outputs["tool_calls"][0]
+        return responses_so_far
+
+
+class _RefusingTranslation:
+    delivers_ended_stream_text_rewrites = True
+
+    async def process_output_streaming_response(
+        self,
+        responses_so_far,
+        guardrail_to_apply,
+        litellm_logging_obj=None,
+        user_api_key_dict=None,
+        request_data=None,
+        deliver_ended_stream_rewrites=False,
+    ):
+        responses_so_far[0]["text"] = "half-written"
+        raise UndeliverableStreamRewrite(guardrail_to_apply.guardrail_name)
+
+
+def _chunk():
+    return {"text": "hello world", "tool_call": {"function": {"name": "lookup", "arguments": '{"ssn": "123"}'}}}
+
+
+async def _run_streaming_step(translation, streaming_chunks=None):
+    chunks = [object()] if streaming_chunks is None else streaming_chunks
     return await PipelineExecutor.execute_steps(
         steps=[PipelineStep(guardrail="masker", on_pass="allow", on_fail="next", on_error="next")],
         mode="post_call",
@@ -949,31 +998,41 @@ async def _run_streaming_step(returned_texts, translation):
         user_api_key_dict=MagicMock(),
         call_type="completion",
         policy_name="p",
-        streaming_chunks=[object()],
+        streaming_chunks=chunks,
         endpoint_translation=translation,
     )
 
 
+def _assert_passed_with_discard_warning(result, caplog):
+    assert result.terminal_action == "allow"
+    assert [step.outcome for step in result.step_results] == ["pass"]
+    assert any("'masker'" in record.getMessage() and "discarded" in record.getMessage() for record in caplog.records)
+
+
 @pytest.mark.asyncio
-async def test_streaming_step_rewrite_escapes_execute_steps_regardless_of_step_actions(monkeypatch):
+async def test_streaming_step_discards_text_rewrite_when_translation_lacks_write_back(monkeypatch, caplog):
     monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(["hello [MASKED]"])])
     translation = _TextTranslation()
+    chunks = [_chunk()]
 
-    with pytest.raises(UndeliverableStreamRewrite) as info:
-        await _run_streaming_step(["hello [MASKED]"], translation)
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await _run_streaming_step(translation, chunks)
 
-    assert info.value.guardrail_name == "masker"
+    _assert_passed_with_discard_warning(result, caplog)
+    assert chunks == [_chunk()]
     assert translation.seen_guardrail_names == ["masker"]
 
 
 @pytest.mark.asyncio
-async def test_streaming_step_unchanged_texts_in_another_container_allow(monkeypatch):
+async def test_streaming_step_unchanged_texts_in_another_container_allow(monkeypatch, caplog):
     monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(("hello world",))])
 
-    result = await _run_streaming_step(("hello world",), _TextTranslation())
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await _run_streaming_step(_TextTranslation())
 
     assert result.terminal_action == "allow"
     assert [step.outcome for step in result.step_results] == ["pass"]
+    assert not any("discarded" in record.getMessage() for record in caplog.records)
 
 
 class _InPlaceMutatingGuardrail(CustomGuardrail):
@@ -989,10 +1048,64 @@ class _InPlaceMutatingGuardrail(CustomGuardrail):
 
 
 @pytest.mark.asyncio
-async def test_streaming_step_in_place_rewrite_still_withholds_stream(monkeypatch):
+async def test_streaming_step_in_place_rewrite_is_discarded_without_write_back(monkeypatch, caplog):
     monkeypatch.setattr(litellm, "callbacks", [_InPlaceMutatingGuardrail()])
+    chunks = [_chunk()]
 
-    with pytest.raises(UndeliverableStreamRewrite) as info:
-        await _run_streaming_step(["hello [MASKED]"], _TextTranslation())
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await _run_streaming_step(_TextTranslation(), chunks)
 
-    assert info.value.guardrail_name == "masker"
+    _assert_passed_with_discard_warning(result, caplog)
+    assert chunks == [_chunk()]
+
+
+class _TextAndToolCallRewritingGuardrail(CustomGuardrail):
+    def __init__(self, rewrite_tool_call):
+        super().__init__(guardrail_name="masker", event_hook="post_call", default_on=True)
+        self.rewrite_tool_call = rewrite_tool_call
+
+    async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+        tool_calls = (
+            [{"function": {"name": "lookup", "arguments": '{"ssn": "[MASKED]"}'}}]
+            if self.rewrite_tool_call
+            else inputs["tool_calls"]
+        )
+        return {**inputs, "texts": ["hello [MASKED]"], "tool_calls": tool_calls}
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_delivers_text_rewrite_through_writing_translation(monkeypatch, caplog):
+    monkeypatch.setattr(litellm, "callbacks", [_TextAndToolCallRewritingGuardrail(rewrite_tool_call=False)])
+    chunks = [_chunk()]
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await _run_streaming_step(_WritingTranslation(), chunks)
+
+    assert result.terminal_action == "allow"
+    assert chunks[0]["text"] == "hello [MASKED]"
+    assert chunks[0]["tool_call"]["function"]["arguments"] == '{"ssn": "123"}'
+    assert not any("discarded" in record.getMessage() for record in caplog.records)
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_discards_tool_call_rewrite_and_restores_written_text(monkeypatch, caplog):
+    monkeypatch.setattr(litellm, "callbacks", [_TextAndToolCallRewritingGuardrail(rewrite_tool_call=True)])
+    chunks = [_chunk()]
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await _run_streaming_step(_WritingTranslation(), chunks)
+
+    _assert_passed_with_discard_warning(result, caplog)
+    assert chunks == [_chunk()]
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_restores_chunks_when_translation_refuses_the_rewrite(monkeypatch, caplog):
+    monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(["hello [MASKED]"])])
+    chunks = [_chunk()]
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await _run_streaming_step(_RefusingTranslation(), chunks)
+
+    _assert_passed_with_discard_warning(result, caplog)
+    assert chunks == [_chunk()]
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 2e73bebb07c..ad2b4a0efbb 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -11,6 +11,7 @@ from __future__ import annotations
 
 import asyncio
 import json
+import logging
 from typing import Any, Callable, Dict, List
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -24,9 +25,9 @@ from litellm.integrations.custom_guardrail import (
     ModifyResponseException,
 )
 from litellm.integrations.prometheus import PrometheusLogger
-from litellm.proxy._types import ProxyException, UserAPIKeyAuth
+from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.common_utils.callback_utils import add_guardrail_to_applied_guardrails_header
-from litellm.proxy.utils import ProxyLogging, _raise_for_streaming_post_call_pipelines
+from litellm.proxy.utils import ProxyLogging, _streamable_post_call_pipelines
 from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import ContentFilterGuardrail
 from litellm.types.guardrails import BlockedWord, ContentFilterAction, GuardrailEventHooks
 from litellm.types.proxy.policy_engine.pipeline_types import (
@@ -1384,76 +1385,87 @@ async def test_pre_call_pipeline_managed_parallel_guardrail_runs_exactly_once(
     assert seen["count"] == 1
 
 
+def _warnings(caplog: pytest.LogCaptureFixture) -> List[str]:
+    return [record.getMessage() for record in caplog.records if record.levelno >= logging.WARNING]
+
+
 @pytest.mark.asyncio
-async def test_pre_call_hook_rejects_streaming_request_with_post_call_pipeline(
-    proxy_logging, make_user_api_key_auth, monkeypatch
+async def test_streaming_request_whose_pipeline_guardrail_is_missing_streams_verbatim(
+    proxy_logging, make_user_api_key_auth, monkeypatch, caplog
 ):
     monkeypatch.setattr(litellm, "callbacks", [])
     data = _post_call_pipeline_data(stream=True)
+    chunks = _stream_chunks()
+    delivered: List[Any] = []
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        out = await proxy_logging.pre_call_hook(
             user_api_key_dict=make_user_api_key_auth(),
             data=data,
             call_type="completion",
             guardrails_only=True,
         )
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        ):
+            delivered.append(item)
 
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["policies"] == ("response-governance",)
-    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
-    assert "stream=false" in info.value.detail["error"]["message"]
+    assert out is not None
+    assert out.get("stream") is True
+    assert [item is chunk for item, chunk in zip(delivered, chunks)] == [True, True]
+    assert len(delivered) == 2
+    assert any("response-governance" in message and "gr-post" in message for message in _warnings(caplog))
 
 
 @pytest.mark.asyncio
-async def test_pre_call_hook_rejects_background_request_with_post_call_pipeline(
-    proxy_logging, make_user_api_key_auth, monkeypatch
+async def test_pre_call_hook_accepts_background_request_with_post_call_pipeline(
+    proxy_logging, make_user_api_key_auth, monkeypatch, caplog
 ):
     monkeypatch.setattr(litellm, "callbacks", [])
     data = _post_call_pipeline_data(background=True)
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        out = await proxy_logging.pre_call_hook(
             user_api_key_dict=make_user_api_key_auth(),
             data=data,
             call_type="aresponses",
             guardrails_only=True,
         )
 
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["policies"] == ("response-governance",)
-    assert "background=false" in info.value.detail["error"]["message"]
+    assert out is not None
+    assert out.get("background") is True
+    assert any("response-governance" in message and "background" in message for message in _warnings(caplog))
 
 
-def test_raise_for_streaming_post_call_pipelines_ignores_non_streaming_and_pre_call(make_user_api_key_auth):
-    post_call = GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
-    pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
-    auth = make_user_api_key_auth(request_route="/custom/stream")
+@pytest.mark.asyncio
+async def test_pre_call_hook_stays_quiet_on_background_request_without_post_call_pipeline(
+    proxy_logging, make_user_api_key_auth, monkeypatch, caplog
+):
+    seen: Dict[str, Any] = {}
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
+    pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="gr-post", on_fail="block")])
+    data = {
+        "model": "m",
+        "messages": [{"role": "user", "content": "hi"}],
+        "background": True,
+        "metadata": {
+            "_guardrail_pipelines": [("request-governance", pre_call)],
+            "_pipeline_managed_guardrails": {"gr-post"},
+        },
+    }
 
-    assert (
-        _raise_for_streaming_post_call_pipelines(
-            {"stream": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}, auth
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        out = await proxy_logging.pre_call_hook(
+            user_api_key_dict=make_user_api_key_auth(),
+            data=data,
+            call_type="aresponses",
+            guardrails_only=True,
         )
-        is None
-    )
-    assert (
-        _raise_for_streaming_post_call_pipelines(
-            {"background": False, "metadata": {"_guardrail_pipelines": [("p", post_call)]}}, auth
-        )
-        is None
-    )
-    assert (
-        _raise_for_streaming_post_call_pipelines({"metadata": {"_guardrail_pipelines": [("p", post_call)]}}, auth)
-        is None
-    )
-    assert (
-        _raise_for_streaming_post_call_pipelines(
-            {"stream": True, "metadata": {"_guardrail_pipelines": [("p", pre_call)]}}, auth
-        )
-        is None
-    )
-    assert _raise_for_streaming_post_call_pipelines({"stream": True}, auth) is None
-    assert _raise_for_streaming_post_call_pipelines({"background": True}, auth) is None
+
+    assert out is not None
+    assert not any("background" in message for message in _warnings(caplog))
 
 
 # ---------------------------------------------------------------------------
@@ -1485,6 +1497,56 @@ async def _async_chunk_iter(chunks: List[Any]):
         yield chunk
 
 
+def test_streamable_post_call_pipelines_keeps_supported_and_drops_unsupported(
+    make_user_api_key_auth, monkeypatch, caplog
+):
+    class NativeOnlyGuardrail(CustomGuardrail):
+        pass
+
+    supported = _unified_stream_guardrail({})
+    native_only = NativeOnlyGuardrail(guardrail_name="gr-native", event_hook=GuardrailEventHooks.post_call)
+    monkeypatch.setattr(litellm, "callbacks", [supported, native_only])
+    governed = GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="gr-post", on_fail="block")])
+    ungoverned = GuardrailPipeline(
+        mode="post_call",
+        steps=[PipelineStep(guardrail="gr-post", on_fail="next"), PipelineStep(guardrail="gr-native", on_fail="block")],
+    )
+    pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="gr-native", on_fail="block")])
+    data = {"metadata": {"_guardrail_pipelines": [("governed", governed), ("ungoverned", ungoverned), ("req", pre_call)]}}
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        streamable = _streamable_post_call_pipelines(data, make_user_api_key_auth(request_route="/v1/chat/completions"))
+
+    assert streamable == (("governed", governed),)
+    assert any("'ungoverned'" in message and "gr-native" in message for message in _warnings(caplog))
+    assert not any("'governed'" in message for message in _warnings(caplog))
+
+
+def test_streamable_post_call_pipelines_is_empty_on_route_without_translation(
+    make_user_api_key_auth, monkeypatch, caplog
+):
+    monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail({})])
+    governed = GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="gr-post", on_fail="block")])
+    data = {"metadata": {"_guardrail_pipelines": [("governed", governed)]}}
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        streamable = _streamable_post_call_pipelines(data, make_user_api_key_auth(request_route="/custom/stream"))
+
+    assert streamable == ()
+    assert any("/custom/stream" in message and "governed" in message for message in _warnings(caplog))
+
+
+def test_streamable_post_call_pipelines_is_empty_without_post_call_pipelines(make_user_api_key_auth, caplog):
+    pre_call = GuardrailPipeline(mode="pre_call", steps=[PipelineStep(guardrail="g", on_fail="block")])
+    auth = make_user_api_key_auth(request_route="/custom/stream")
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        assert _streamable_post_call_pipelines({"metadata": {"_guardrail_pipelines": [("p", pre_call)]}}, auth) == ()
+        assert _streamable_post_call_pipelines({"stream": True}, auth) == ()
+
+    assert _warnings(caplog) == []
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("request_route", [None, "/v1/chat/completions"])
 async def test_pre_call_hook_allows_streaming_when_pipeline_guardrail_supports_unified(
@@ -1507,21 +1569,25 @@ async def test_pre_call_hook_allows_streaming_when_pipeline_guardrail_supports_u
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("native_lifecycle", [False, True])
-async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_lacks_unified_support(
-    proxy_logging, make_user_api_key_auth, monkeypatch, native_lifecycle
+async def test_streaming_iterator_hook_releases_stream_when_pipeline_guardrail_lacks_unified_support(
+    proxy_logging, make_user_api_key_auth, monkeypatch, native_lifecycle, caplog
 ):
+    seen: Dict[str, Any] = {}
     if native_lifecycle:
 
         class NativeOnlyGuardrail(CustomGuardrail):
             use_native_lifecycle_hooks = True
 
             async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+                seen["count"] = seen.get("count", 0) + 1
                 return inputs
 
     else:
 
         class NativeOnlyGuardrail(CustomGuardrail):
-            pass
+            async def async_post_call_success_hook(self, data, user_api_key_dict, response):
+                seen["count"] = seen.get("count", 0) + 1
+                return response
 
     monkeypatch.setattr(
         litellm,
@@ -1529,18 +1595,29 @@ async def test_pre_call_hook_rejects_streaming_when_pipeline_guardrail_lacks_uni
         [NativeOnlyGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=False)],
     )
     data = _post_call_pipeline_data(stream=True)
+    chunks = _stream_chunks()
+    delivered: List[Any] = []
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        out = await proxy_logging.pre_call_hook(
             user_api_key_dict=make_user_api_key_auth(),
             data=data,
             call_type="completion",
             guardrails_only=True,
         )
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        ):
+            delivered.append(item)
 
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["guardrails"] == ("gr-post",)
-    assert "apply_guardrail" in info.value.detail["error"]["message"]
+    assert out is not None
+    assert out.get("stream") is True
+    assert [item is chunk for item, chunk in zip(delivered, chunks)] == [True, True]
+    assert len(delivered) == 2
+    assert seen.get("count") is None
+    assert any("'response-governance'" in message and "gr-post" in message for message in _warnings(caplog))
 
 
 @pytest.mark.asyncio
@@ -1614,25 +1691,34 @@ async def test_pre_call_hook_allows_streaming_when_content_filter_category_masks
 
 
 @pytest.mark.asyncio
-async def test_pre_call_hook_rejects_streaming_when_route_has_no_guardrail_translation(
-    proxy_logging, make_user_api_key_auth, monkeypatch
+async def test_streaming_iterator_hook_releases_stream_when_route_has_no_guardrail_translation(
+    proxy_logging, make_user_api_key_auth, monkeypatch, caplog
 ):
     seen: Dict[str, Any] = {}
     monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
     data = _post_call_pipeline_data(stream=True)
+    chunks = _stream_chunks()
+    delivered: List[Any] = []
 
-    with pytest.raises(HTTPException) as info:
-        await proxy_logging.pre_call_hook(
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        out = await proxy_logging.pre_call_hook(
             user_api_key_dict=make_user_api_key_auth(request_route="/custom/stream"),
             data=data,
             call_type="completion",
             guardrails_only=True,
         )
+        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+            user_api_key_dict=make_user_api_key_auth(request_route="/custom/stream"),
+            response=_async_chunk_iter(chunks),
+            request_data=data,
+        ):
+            delivered.append(item)
 
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["policies"] == ("response-governance",)
-    assert "/custom/stream" in info.value.detail["error"]["message"]
+    assert out is not None
+    assert [item is chunk for item, chunk in zip(delivered, chunks)] == [True, True]
+    assert len(delivered) == 2
     assert seen.get("count") is None
+    assert any("/custom/stream" in message and "response-governance" in message for message in _warnings(caplog))
 
 
 @pytest.mark.asyncio
@@ -1714,8 +1800,8 @@ def _echoed_tool_call_dicts(arguments: str) -> List[Dict[str, Any]]:
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("on_fail, on_error", [("block", None), ("next", "next")])
-async def test_streaming_iterator_hook_pipeline_withholds_runtime_tool_call_rewrite(
-    proxy_logging, make_user_api_key_auth, monkeypatch, on_fail, on_error
+async def test_streaming_iterator_hook_pipeline_releases_originals_on_runtime_tool_call_rewrite(
+    proxy_logging, make_user_api_key_auth, monkeypatch, on_fail, on_error, caplog
 ):
     transform = lambda inputs: {"tool_calls": _echoed_tool_call_dicts('{"ssn": "[MASKED]"}')}  # noqa: E731
     monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
@@ -1724,7 +1810,7 @@ async def test_streaming_iterator_hook_pipeline_withholds_runtime_tool_call_rewr
     data = _post_call_pipeline_data(step=step, stream=True)
     delivered: List[Any] = []
 
-    async def _drain() -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
         async for item in proxy_logging.async_post_call_streaming_iterator_hook(
             user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
             response=_async_chunk_iter(_tool_call_stream_chunks()),
@@ -1732,16 +1818,10 @@ async def test_streaming_iterator_hook_pipeline_withholds_runtime_tool_call_rewr
         ):
             delivered.append(item)
 
-    with pytest.raises(HTTPException) as info:
-        await _drain()
-
-    error = info.value.detail["error"]
-    assert delivered == []
-    assert info.value.status_code == 400
-    assert error["type"] == "guardrail_pipeline_error"
-    assert error["policies"] == ("response-governance",)
-    assert error["guardrails"] == ("gr-post",)
-    assert "stream=false" in error["message"]
+    assert len(delivered) == 2
+    assert delivered[0].choices[0].delta.tool_calls[0].function.arguments == '{"ssn": "123"}'
+    assert delivered[1].choices[0].finish_reason == "tool_calls"
+    assert any("'gr-post'" in message and "discarded" in message for message in _warnings(caplog))
 
 
 @pytest.mark.asyncio
@@ -1849,30 +1929,28 @@ async def test_streaming_iterator_hook_pipeline_releases_stream_echoed_in_anothe
 
 
 @pytest.mark.asyncio
-async def test_streaming_iterator_hook_pipeline_withholds_unresolvable_response_shape(
-    proxy_logging, make_user_api_key_auth, monkeypatch
+async def test_streaming_iterator_hook_pipeline_releases_originals_on_unresolvable_response_shape(
+    proxy_logging, make_user_api_key_auth, monkeypatch, caplog
 ):
     seen: Dict[str, Any] = {}
     monkeypatch.setattr(litellm, "callbacks", [_unified_stream_guardrail(seen)])
     monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
     data = _post_call_pipeline_data(stream=True)
+    chunks = [object(), object()]
     delivered: List[Any] = []
 
-    async def _drain() -> None:
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
         async for item in proxy_logging.async_post_call_streaming_iterator_hook(
             user_api_key_dict=make_user_api_key_auth(),
-            response=_async_chunk_iter([object(), object()]),
+            response=_async_chunk_iter(chunks),
             request_data=data,
         ):
             delivered.append(item)
 
-    with pytest.raises(ProxyException) as info:
-        await _drain()
-
-    assert delivered == []
-    assert info.value.code == "500"
-    assert "withheld" in info.value.message
+    assert [item is chunk for item, chunk in zip(delivered, chunks)] == [True, True]
+    assert len(delivered) == 2
     assert seen.get("count") is None
+    assert any("response-governance" in message and "shape" in message for message in _warnings(caplog))
 
 
 def _anthropic_sse_chunks() -> List[bytes]:
@@ -1953,9 +2031,9 @@ async def test_streaming_iterator_hook_pipeline_delivers_text_rewrite_on_anthrop
 
 
 @pytest.mark.asyncio
-async def test_pipeline_executor_withholds_text_rewrite_when_translation_lacks_write_back(monkeypatch):
+async def test_pipeline_executor_discards_text_rewrite_when_translation_lacks_write_back(monkeypatch, caplog):
     from litellm.llms.base_llm.guardrail_translation.base_translation import BaseTranslation
-    from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor, UndeliverableStreamRewrite
+    from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor
 
     class NoWriteBackTranslation(BaseTranslation):
         async def process_input_messages(self, data, guardrail_to_apply, litellm_logging_obj):
@@ -1984,45 +2062,23 @@ async def test_pipeline_executor_withholds_text_rewrite_when_translation_lacks_w
 
     transform = lambda inputs: {"texts": ["hello [MASKED]"]}  # noqa: E731
     monkeypatch.setattr(litellm, "callbacks", [_rewriting_stream_guardrail(transform)])
+    chunks = _stream_chunks()
 
-    with pytest.raises(UndeliverableStreamRewrite):
-        await PipelineExecutor.execute_steps(
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        result = await PipelineExecutor.execute_steps(
             steps=[PipelineStep(guardrail="gr-post", on_pass="allow", on_fail="block")],
             mode="post_call",
             data={"metadata": {}},
             user_api_key_dict=UserAPIKeyAuth(api_key="sk-test"),
             call_type="acompletion",
             policy_name="response-governance",
-            streaming_chunks=_stream_chunks(),
+            streaming_chunks=chunks,
             endpoint_translation=NoWriteBackTranslation(),
         )
 
-
-@pytest.mark.asyncio
-async def test_streaming_iterator_hook_pipeline_gates_without_iterator_overrides(
-    proxy_logging, make_user_api_key_auth, monkeypatch
-):
-    monkeypatch.setattr(litellm, "callbacks", [])
-    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
-    data = _post_call_pipeline_data(stream=True)
-    delivered: List[Any] = []
-
-    async def _drain() -> None:
-        async for item in proxy_logging.async_post_call_streaming_iterator_hook(
-            user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
-            response=_async_chunk_iter(_stream_chunks()),
-            request_data=data,
-        ):
-            delivered.append(item)
-
-    with pytest.raises(HTTPException) as info:
-        await _drain()
-
-    assert delivered == []
-    assert info.value.status_code == 400
-    assert info.value.detail["error"]["pipeline_context"]["step_results"] == [
-        {"guardrail": "gr-post", "outcome": "error", "action": "block"}
-    ]
+    assert result.terminal_action == "allow"
+    assert [chunk.choices[0].delta.content for chunk in chunks] == ["hello ", "world"]
+    assert any("'gr-post'" in message and "discarded" in message for message in _warnings(caplog))
 
 
 @pytest.mark.asyncio

From 69d2ac1edb83336723d4c9ce5024b93612230e5d Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 7 Sep 2026 21:46:00 -0700
Subject: [PATCH 21/25] fix(policy_engine): run iterator-hook guardrails whose
 post_call pipeline cannot stream

The streaming loop skipped every guardrail stepped by a post_call pipeline, even when the pipeline was dropped from the stream for lacking the unified apply_guardrail interface, so a default_on guardrail that only implements async_post_call_streaming_iterator_hook stopped governing streams it governed on the merge base. The skip set now comes from the pipelines that will gate the stream
---
 litellm/proxy/utils.py                        | 26 +++++++-------
 .../proxy_logging/test_guardrail_pipeline.py  | 36 +++++++++++++++++++
 2 files changed, 50 insertions(+), 12 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index d7bf832bca4..8b36061c6be 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -447,14 +447,15 @@ def _policy_pipelines(data: Mapping[str, object]) -> tuple[tuple[str, "Guardrail
     )
 
 
+def _pipeline_step_guardrail_names(pipelines: Sequence[tuple[str, "GuardrailPipeline"]]) -> frozenset[str]:
+    return frozenset(step.guardrail for _policy_name, pipeline in pipelines for step in pipeline.steps)
+
+
 def _pipeline_managed_guardrail_names(
     data: Mapping[str, object], mode: Literal["pre_call", "post_call"]
 ) -> frozenset[str]:
-    return frozenset(
-        step.guardrail
-        for _policy_name, pipeline in _policy_pipelines(data)
-        if pipeline.mode == mode
-        for step in pipeline.steps
+    return _pipeline_step_guardrail_names(
+        tuple((policy_name, pipeline) for policy_name, pipeline in _policy_pipelines(data) if pipeline.mode == mode)
     )
 
 
@@ -547,7 +548,7 @@ def _pipeline_is_streamable(policy_name: str, pipeline: "GuardrailPipeline") ->
         return True
     verbose_proxy_logger.warning(
         "Policy '%s' has post_call pipeline guardrails without the unified apply_guardrail interface, "
-        "which streaming pipelines need; the stream is released ungoverned by it: %s",
+        "which streaming pipelines need; the stream skips the pipeline and its guardrails run on their own: %s",
         policy_name,
         ", ".join(unsupported),
     )
@@ -563,9 +564,9 @@ def _streamable_post_call_pipelines(
     Streaming pipelines scan the buffered stream through the endpoint guardrail
     translation of the request route, so every step's guardrail needs the
     unified apply_guardrail interface and the route needs a translation. A
-    pipeline that cannot be run that way yet is left out and the stream is
-    released the way it was before pipelines ran on streams at all, with a
-    warning naming what went ungoverned.
+    pipeline that cannot be run that way yet is left out and its guardrails
+    run on the stream on their own, the way they did before pipelines ran on
+    streams at all, with a warning naming the pipeline.
     """
     post_call_pipelines: Final = _post_call_pipelines(request_data)
     if not post_call_pipelines:
@@ -574,7 +575,8 @@ def _streamable_post_call_pipelines(
     if route and resolve_endpoint_translation(user_api_key_dict, None) is None:
         verbose_proxy_logger.warning(
             "Policies with post_call guardrail pipelines cannot scan streaming responses on route %s yet "
-            "(no endpoint guardrail translation); the stream is released ungoverned by them: %s",
+            "(no endpoint guardrail translation); the stream skips the pipelines and their guardrails run "
+            "on their own: %s",
             route,
             ", ".join(policy_name for policy_name, _pipeline in post_call_pipelines),
         )
@@ -3361,10 +3363,10 @@ class ProxyLogging:
         current_response = response
         stream_needs_translation: Final = ProxyLogging._stream_requires_guardrail_translation(user_api_key_dict)
 
-        pipeline_managed_names: Final = _pipeline_managed_guardrail_names(request_data, "post_call")
+        pipeline_gated_names: Final = _pipeline_step_guardrail_names(post_call_pipelines)
         for resolved_callback, kind in caps.iterator_overrides:
             if isinstance(resolved_callback, CustomGuardrail):
-                if resolved_callback.guardrail_name in pipeline_managed_names:
+                if resolved_callback.guardrail_name in pipeline_gated_names:
                     continue
                 if (
                     resolved_callback.should_run_guardrail(data=request_data, event_type=GuardrailEventHooks.post_call)
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index ad2b4a0efbb..73dd6746b29 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -1620,6 +1620,42 @@ async def test_streaming_iterator_hook_releases_stream_when_pipeline_guardrail_l
     assert any("'response-governance'" in message and "gr-post" in message for message in _warnings(caplog))
 
 
+@pytest.mark.asyncio
+async def test_streaming_iterator_hook_runs_iterator_hook_guardrail_whose_pipeline_cannot_stream(
+    proxy_logging, make_user_api_key_auth, monkeypatch, caplog
+):
+    seen: Dict[str, Any] = {}
+
+    class IteratorHookGuardrail(CustomGuardrail):
+        async def async_post_call_streaming_iterator_hook(self, user_api_key_dict, response, request_data):
+            seen["count"] = seen.get("count", 0) + 1
+            async for item in response:
+                item.choices[0].delta.content = f"[governed] {item.choices[0].delta.content}"
+                yield item
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [IteratorHookGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=True)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+
+    with caplog.at_level(logging.WARNING, logger="LiteLLM Proxy"):
+        delivered = [
+            item
+            async for item in proxy_logging.async_post_call_streaming_iterator_hook(
+                user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+                response=_async_chunk_iter(_stream_chunks()),
+                request_data=data,
+            )
+        ]
+
+    assert seen["count"] == 1
+    assert [item.choices[0].delta.content for item in delivered] == ["[governed] hello ", "[governed] world"]
+    assert any("'response-governance'" in message and "gr-post" in message for message in _warnings(caplog))
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
     "rewrite_attribute, value",

From 91fc1b201027b3d21a7ced292ac19a51661cba0e Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 7 Sep 2026 21:48:19 -0700
Subject: [PATCH 22/25] fix(policy_engine): record a streaming pipeline step
 once and in the applied guardrails header

CustomGuardrail.__init_subclass__ wrapped _StreamRewriteObserver.apply_guardrail in log_guardrail_information, so every streaming step recorded a second standard_logging_guardrail_information entry and span next to the inner guardrail's own. The observer's method now carries the marker that skips the wrapper. The step also adds the guardrail to the applied guardrails header the way the non-streaming unified path does, so streamed spend rows name the guardrail that scanned them
---
 .../proxy/policy_engine/pipeline_executor.py  | 27 +++++++++---
 .../policy_engine/test_pipeline_executor.py   | 44 +++++++++++++++++++
 2 files changed, 65 insertions(+), 6 deletions(-)

diff --git a/litellm/proxy/policy_engine/pipeline_executor.py b/litellm/proxy/policy_engine/pipeline_executor.py
index 970ac20487c..9bc10949e9f 100644
--- a/litellm/proxy/policy_engine/pipeline_executor.py
+++ b/litellm/proxy/policy_engine/pipeline_executor.py
@@ -7,19 +7,21 @@ pass/fail actions (allow, block, next, modify_response) and data forwarding.
 
 import copy
 import time
-from collections.abc import Mapping, Sequence
-from typing import TYPE_CHECKING, Any, Final, Literal
+from collections.abc import Callable, Mapping, Sequence
+from typing import TYPE_CHECKING, Any, Final, Literal, TypeVar
 
 from pydantic import BaseModel
 
 import litellm
 from litellm._logging import verbose_proxy_logger
+from litellm.constants import LOGS_GUARDRAIL_INFORMATION_MARKER
 from litellm.integrations.custom_guardrail import (
     CustomGuardrail,
     ModifyResponseException,
 )
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.core_helpers import independent_snapshot
+from litellm.proxy.common_utils.callback_utils import add_guardrail_to_applied_guardrails_header
 from litellm.proxy.guardrails.guardrail_hooks.unified_guardrail.unified_guardrail import (
     UnifiedLLMGuardrails,
 )
@@ -72,13 +74,23 @@ def _rewrote(sent: tuple[object, ...] | None, returned: tuple[object, ...] | Non
     return sent is not None and returned is not None and returned != sent
 
 
+_GuardrailMethodT = TypeVar("_GuardrailMethodT", bound=Callable[..., object])
+
+
+def _logged_by_inner_guardrail(method: _GuardrailMethodT) -> _GuardrailMethodT:
+    vars(method)[LOGS_GUARDRAIL_INFORMATION_MARKER] = True  # rebind-ok: stamps the method the class body just defined
+    return method
+
+
 class _StreamRewriteObserver(CustomGuardrail):
     """Stand-in handed to the endpoint translation in place of a streaming pipeline step's
     guardrail. It records whether the guardrail returned different output than it was given,
     which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text
     rewrites are deliverable on translations that write them back across the buffered chunks
     (``delivers_ended_stream_text_rewrites``); tool-call rewrites and text rewrites on any
-    other translation are discarded by the executor, which releases the original chunks."""
+    other translation are discarded by the executor, which releases the original chunks.
+    The inner guardrail's ``apply_guardrail`` already records the guardrail information
+    and span, so the observer's stays out of ``log_guardrail_information``."""
 
     def __init__(self, inner: CustomGuardrail) -> None:
         super().__init__(guardrail_name=inner.guardrail_name)
@@ -89,6 +101,7 @@ class _StreamRewriteObserver(CustomGuardrail):
     def structured_messages_cover_full_request(self) -> bool:
         return self.inner.structured_messages_cover_full_request()
 
+    @_logged_by_inner_guardrail
     async def apply_guardrail(
         self,
         inputs: GenericGuardrailAPIInputs,
@@ -305,9 +318,11 @@ class PipelineExecutor:
                 )
         except UndeliverableStreamRewrite:
             _release_original_chunks(step.guardrail, streaming_chunks, originals)
-            return
-        if observer.rewrote_tool_calls or (observer.rewrote_texts and not deliver_rewrites):
-            _release_original_chunks(step.guardrail, streaming_chunks, originals)
+        else:
+            if observer.rewrote_tool_calls or (observer.rewrote_texts and not deliver_rewrites):
+                _release_original_chunks(step.guardrail, streaming_chunks, originals)
+        if not callback.records_own_guardrail_information:
+            add_guardrail_to_applied_guardrails_header(request_data=hook_input, guardrail_name=step.guardrail)
 
     @staticmethod
     async def _run_step(
diff --git a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
index 0685bc6aa1e..54ef1f79f4d 100644
--- a/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
+++ b/tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py
@@ -1099,6 +1099,50 @@ async def test_streaming_step_discards_tool_call_rewrite_and_restores_written_te
     assert chunks == [_chunk()]
 
 
+class _BlockingStreamGuardrail(CustomGuardrail):
+    def __init__(self):
+        super().__init__(guardrail_name="masker", event_hook="post_call", default_on=True)
+
+    async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+        raise HTTPException(status_code=400, detail={"error": "output blocked"})
+
+
+def _recorded_guardrail_statuses(result):
+    return [
+        entry["guardrail_status"]
+        for entry in result.modified_data["metadata"]["standard_logging_guardrail_information"]
+    ]
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_records_guardrail_information_once_on_mask(monkeypatch):
+    monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(["hello [MASKED]"])])
+
+    result = await _run_streaming_step(_WritingTranslation(), [_chunk()])
+
+    assert result.terminal_action == "allow"
+    assert _recorded_guardrail_statuses(result) == ["success"]
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_records_the_guardrail_in_the_applied_guardrails_header(monkeypatch):
+    monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(["hello [MASKED]"])])
+
+    result = await _run_streaming_step(_WritingTranslation(), [_chunk()])
+
+    assert result.modified_data["metadata"]["applied_guardrails"] == ["masker"]
+
+
+@pytest.mark.asyncio
+async def test_streaming_step_records_guardrail_information_once_on_block(monkeypatch):
+    monkeypatch.setattr(litellm, "callbacks", [_BlockingStreamGuardrail()])
+
+    result = await _run_streaming_step(_WritingTranslation(), [_chunk()])
+
+    assert [step.outcome for step in result.step_results] == ["fail"]
+    assert _recorded_guardrail_statuses(result) == ["guardrail_intervened"]
+
+
 @pytest.mark.asyncio
 async def test_streaming_step_restores_chunks_when_translation_refuses_the_rewrite(monkeypatch, caplog):
     monkeypatch.setattr(litellm, "callbacks", [_TextReturningGuardrail(["hello [MASKED]"])])

From d08a177bc75b45769e5e22efa2bbb4b0ea27ee86 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 7 Sep 2026 21:51:34 -0700
Subject: [PATCH 23/25] fix(policy_engine): keep a policy-added guardrail's
 other stages when a pipeline steps it

A policy that both adds a guardrail and steps it in a post_call pipeline
used to drop the guardrail from the request's guardrail list outright, so
its pre_call stage never ran. The per-hook loops already skip guardrails
by pipeline mode, so the mode-agnostic subtraction only lost coverage
---
 litellm/proxy/litellm_pre_call_utils.py       |  7 +---
 .../proxy/test_litellm_pre_call_utils.py      | 42 +++++++++++++++++++
 2 files changed, 44 insertions(+), 5 deletions(-)

diff --git a/litellm/proxy/litellm_pre_call_utils.py b/litellm/proxy/litellm_pre_call_utils.py
index 56512570448..e4bce4378f0 100644
--- a/litellm/proxy/litellm_pre_call_utils.py
+++ b/litellm/proxy/litellm_pre_call_utils.py
@@ -3079,10 +3079,9 @@ def _apply_resolved_guardrails_to_metadata(
     if metadata_variable_name not in data:
         data[metadata_variable_name] = {}
 
-    # Track pipeline-managed guardrails to exclude from independent execution
-    pipeline_managed_guardrails: set = set()
+    # Record the pipelines and the guardrails they step; the hook loops skip those per pipeline mode
     if pipelines:
-        pipeline_managed_guardrails = PolicyResolver.get_pipeline_managed_guardrails(pipelines)
+        pipeline_managed_guardrails: Final = PolicyResolver.get_pipeline_managed_guardrails(pipelines)
         data[metadata_variable_name]["_guardrail_pipelines"] = pipelines
         data[metadata_variable_name]["_pipeline_managed_guardrails"] = pipeline_managed_guardrails
         verbose_proxy_logger.debug(
@@ -3099,10 +3098,8 @@ def _apply_resolved_guardrails_to_metadata(
         existing_guardrails = []
 
     # Combine existing guardrails with policy-resolved guardrails (no duplicates)
-    # Exclude pipeline-managed guardrails from the flat list
     combined = set(existing_guardrails)
     combined.update(resolved_guardrails)
-    combined -= pipeline_managed_guardrails
     data[metadata_variable_name]["guardrails"] = list(combined)
 
     verbose_proxy_logger.debug("Policy engine: added guardrails to request metadata: %s", list(combined))
diff --git a/tests/test_litellm/proxy/test_litellm_pre_call_utils.py b/tests/test_litellm/proxy/test_litellm_pre_call_utils.py
index 7070617ce3e..78fe2e6df88 100644
--- a/tests/test_litellm/proxy/test_litellm_pre_call_utils.py
+++ b/tests/test_litellm/proxy/test_litellm_pre_call_utils.py
@@ -4148,6 +4148,48 @@ async def test_add_guardrails_from_policy_engine():
     attachment_registry._initialized = False
 
 
+@pytest.mark.asyncio
+async def test_add_guardrails_from_policy_engine_keeps_a_policy_added_guardrail_its_pipeline_also_steps():
+    from litellm.proxy.policy_engine.attachment_registry import get_attachment_registry
+    from litellm.proxy.policy_engine.policy_registry import get_policy_registry
+    from litellm.types.proxy.policy_engine import (
+        GuardrailPipeline,
+        PipelineStep,
+        Policy,
+        PolicyAttachment,
+        PolicyGuardrails,
+    )
+
+    data = {"model": "gpt-4", "messages": [{"role": "user", "content": "Hello"}], "metadata": {}}
+    policy_registry = get_policy_registry()
+    policy_registry._policies = {
+        "response-governance": Policy(
+            guardrails=PolicyGuardrails(add=["pii_blocker"]),
+            pipeline=GuardrailPipeline(mode="post_call", steps=[PipelineStep(guardrail="pii_blocker")]),
+        ),
+    }
+    policy_registry._initialized = True
+    attachment_registry = get_attachment_registry()
+    attachment_registry._attachments = [PolicyAttachment(policy="response-governance", scope="*")]
+    attachment_registry._initialized = True
+
+    try:
+        await add_guardrails_from_policy_engine(
+            data=data,
+            metadata_variable_name="metadata",
+            user_api_key_dict=UserAPIKeyAuth(api_key="test-key"),
+        )
+    finally:
+        policy_registry._policies = {}
+        policy_registry._initialized = False
+        attachment_registry._attachments = []
+        attachment_registry._initialized = False
+
+    assert data["metadata"]["guardrails"] == ["pii_blocker"]
+    assert data["metadata"]["_pipeline_managed_guardrails"] == {"pii_blocker"}
+    assert [pipeline.mode for _policy_name, pipeline in data["metadata"]["_guardrail_pipelines"]] == ["post_call"]
+
+
 @pytest.mark.asyncio
 async def test_add_guardrails_from_policy_engine_accepts_dynamic_policies_and_pops_from_data():
     """

From 08b60c409a0b556efb6bbe6470402a4530666870 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 7 Sep 2026 21:53:26 -0700
Subject: [PATCH 24/25] refactor(guardrails): drop the unused
 rewrites_streamed_output hook

Nothing calls it since the streaming pipeline detects rewrites at run time
through the stream observer, so the base method and the content filter's
override were dead code with dead tests
---
 litellm/integrations/custom_guardrail.py      |  3 -
 .../litellm_content_filter/content_filter.py  |  9 ---
 .../content_filter/test_content_filter.py     | 56 -------------------
 3 files changed, 68 deletions(-)

diff --git a/litellm/integrations/custom_guardrail.py b/litellm/integrations/custom_guardrail.py
index 883329c9fa8..2d66a280663 100644
--- a/litellm/integrations/custom_guardrail.py
+++ b/litellm/integrations/custom_guardrail.py
@@ -773,9 +773,6 @@ class CustomGuardrail(CustomLogger):
     def uses_apply_guardrail_interface(self) -> bool:
         return type(self).apply_guardrail is not CustomGuardrail.apply_guardrail
 
-    def rewrites_streamed_output(self) -> bool:
-        return self.mask_response_content
-
     def _deployment_pre_call_target(self) -> "CustomLogger":
         if not self.uses_apply_guardrail_interface() or self.use_native_lifecycle_hooks:
             return self
diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
index 85eb50c78e7..722f96ef814 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
@@ -1947,15 +1947,6 @@ class ContentFilterGuardrail(CustomGuardrail):
                 exception_str=exception_str,
             )
 
-    def rewrites_streamed_output(self) -> bool:
-        return (
-            super().rewrites_streamed_output()
-            or any(entry["action"] == ContentFilterAction.MASK for entry in self.compiled_patterns)
-            or any(action == ContentFilterAction.MASK for action, _ in self.blocked_words.values())
-            or any(action == ContentFilterAction.MASK for _, _, action in self.category_keywords.values())
-            or any(action == ContentFilterAction.MASK for _, _, action in self.always_block_category_keywords.values())
-        )
-
     async def async_post_call_streaming_iterator_hook(
         self,
         user_api_key_dict: UserAPIKeyAuth,
diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
index 73020fe3e6f..be55ac47bde 100644
--- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
+++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/content_filter/test_content_filter.py
@@ -3068,59 +3068,3 @@ class TestContentFilterToolCallArguments:
                 request_data={},
                 input_type="response",
             )
-
-
-class TestRewritesStreamedOutput:
-    def test_block_only_rules_do_not_rewrite(self):
-        guardrail = ContentFilterGuardrail(
-            guardrail_name="cf",
-            patterns=[ContentFilterPattern(pattern_type="prebuilt", pattern_name="us_ssn", action=ContentFilterAction.BLOCK)],
-            blocked_words=[BlockedWord(keyword="kumquat", action=ContentFilterAction.BLOCK)],
-        )
-
-        assert guardrail.rewrites_streamed_output() is False
-
-    def test_mask_blocked_word_rewrites(self):
-        guardrail = ContentFilterGuardrail(
-            guardrail_name="cf",
-            blocked_words=[BlockedWord(keyword="persimmon", action=ContentFilterAction.MASK)],
-        )
-
-        assert guardrail.rewrites_streamed_output() is True
-
-    def test_mask_pattern_rewrites(self):
-        guardrail = ContentFilterGuardrail(
-            guardrail_name="cf",
-            patterns=[ContentFilterPattern(pattern_type="prebuilt", pattern_name="us_ssn", action=ContentFilterAction.MASK)],
-        )
-
-        assert guardrail.rewrites_streamed_output() is True
-
-    def test_mask_response_content_rewrites(self):
-        guardrail = ContentFilterGuardrail(
-            guardrail_name="cf",
-            blocked_words=[BlockedWord(keyword="kumquat", action=ContentFilterAction.BLOCK)],
-            mask_response_content=True,
-        )
-
-        assert guardrail.rewrites_streamed_output() is True
-
-    @pytest.mark.parametrize("action, expected", [("MASK", True), ("BLOCK", False)])
-    def test_category_keywords_follow_the_category_action(self, action, expected):
-        guardrail = ContentFilterGuardrail(
-            guardrail_name="cf",
-            categories=[{"category": "bias_gender", "enabled": True, "action": action}],
-        )
-
-        assert guardrail.category_keywords and not guardrail.always_block_category_keywords
-        assert guardrail.rewrites_streamed_output() is expected
-
-    @pytest.mark.parametrize("action, expected", [("MASK", True), ("BLOCK", False)])
-    def test_always_block_category_keywords_follow_the_category_action(self, action, expected):
-        guardrail = ContentFilterGuardrail(
-            guardrail_name="cf",
-            categories=[{"category": "age_discrimination", "enabled": True, "action": action}],
-        )
-
-        assert guardrail.always_block_category_keywords and not guardrail.category_keywords
-        assert guardrail.rewrites_streamed_output() is expected

From 6475443efbfecb52ef693bfd4e7149908b4283a2 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 8 Sep 2026 15:00:25 -0700
Subject: [PATCH 25/25] fix(policy_engine): run per-chunk hook guardrails whose
 post_call pipeline cannot stream

The per-chunk streaming hook skipped every guardrail stepped by a post_call pipeline, even when the pipeline is left out of the stream for lacking the unified apply_guardrail interface, so a default_on guardrail that only implements async_post_call_streaming_hook stopped governing streams it governed on the merge base. The skip set now comes from the pipelines that gate the stream, the same way the iterator hook already computes it
---
 litellm/proxy/utils.py                        | 29 ++++++++++++---
 .../proxy_logging/test_guardrail_pipeline.py  | 37 ++++++++++++++++++-
 2 files changed, 59 insertions(+), 7 deletions(-)

diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index 8b36061c6be..d40cc34476a 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -555,6 +555,24 @@ def _pipeline_is_streamable(policy_name: str, pipeline: "GuardrailPipeline") ->
     return False
 
 
+def _route_supports_streaming_pipelines(user_api_key_dict: UserAPIKeyAuth) -> bool:
+    return not user_api_key_dict.request_route or resolve_endpoint_translation(user_api_key_dict, None) is not None
+
+
+def _stream_gated_guardrail_names(
+    request_data: Mapping[str, object], user_api_key_dict: UserAPIKeyAuth
+) -> frozenset[str]:
+    if not _route_supports_streaming_pipelines(user_api_key_dict):
+        return frozenset()
+    return _pipeline_step_guardrail_names(
+        tuple(
+            (policy_name, pipeline)
+            for policy_name, pipeline in _post_call_pipelines(request_data)
+            if all(_pipeline_step_supports_unified_streaming(step.guardrail) for step in pipeline.steps)
+        )
+    )
+
+
 def _streamable_post_call_pipelines(
     request_data: Mapping[str, object], user_api_key_dict: UserAPIKeyAuth
 ) -> tuple[tuple[str, "GuardrailPipeline"], ...]:
@@ -571,13 +589,12 @@ def _streamable_post_call_pipelines(
     post_call_pipelines: Final = _post_call_pipelines(request_data)
     if not post_call_pipelines:
         return ()
-    route: Final = user_api_key_dict.request_route
-    if route and resolve_endpoint_translation(user_api_key_dict, None) is None:
+    if not _route_supports_streaming_pipelines(user_api_key_dict):
         verbose_proxy_logger.warning(
             "Policies with post_call guardrail pipelines cannot scan streaming responses on route %s yet "
             "(no endpoint guardrail translation); the stream skips the pipelines and their guardrails run "
             "on their own: %s",
-            route,
+            user_api_key_dict.request_route,
             ", ".join(policy_name for policy_name, _pipeline in post_call_pipelines),
         )
         return ()
@@ -3271,15 +3288,15 @@ class ProxyLogging:
             # dict lookups + llm_router.get_deployment() per callback per chunk.
             _cached_guardrail_data: dict | None = None
             _guardrail_data_computed = False
-            pipeline_managed: Final = (
-                _pipeline_managed_guardrail_names(data, "post_call") if caps.has_guardrail else frozenset()
+            pipeline_gated: Final = (
+                _stream_gated_guardrail_names(data, user_api_key_dict) if caps.has_guardrail else frozenset()
             )
 
             for callback in litellm.callbacks:
                 try:
                     _callback: CustomLogger | None = None
                     if isinstance(callback, CustomGuardrail):
-                        if callback.guardrail_name in pipeline_managed:
+                        if callback.guardrail_name in pipeline_gated:
                             continue
                         # Main - V2 Guardrails implementation
                         from litellm.types.guardrails import GuardrailEventHooks
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
index 73dd6746b29..5cb595840fc 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_guardrail_pipeline.py
@@ -2128,7 +2128,11 @@ async def test_per_chunk_streaming_hook_skips_pipeline_managed_guardrail(
             seen[self.guardrail_name] = seen.get(self.guardrail_name, 0) + 1
             return None
 
-    managed = RecordingGuardrail(
+    class UnifiedRecordingGuardrail(RecordingGuardrail):
+        async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+            return inputs
+
+    managed = UnifiedRecordingGuardrail(
         guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=True
     )
     free = RecordingGuardrail(
@@ -2147,3 +2151,34 @@ async def test_per_chunk_streaming_hook_skips_pipeline_managed_guardrail(
     assert result is not None
     assert seen.get("gr-post") is None
     assert seen["gr-free"] == 1
+
+
+@pytest.mark.asyncio
+async def test_per_chunk_streaming_hook_runs_guardrail_whose_pipeline_cannot_stream(
+    proxy_logging, make_user_api_key_auth, monkeypatch
+):
+    seen: Dict[str, Any] = {}
+
+    class ChunkHookGuardrail(CustomGuardrail):
+        async def async_post_call_streaming_hook(self, user_api_key_dict, response):
+            seen["count"] = seen.get("count", 0) + 1
+            seen["response"] = response
+            return None
+
+    monkeypatch.setattr(
+        litellm,
+        "callbacks",
+        [ChunkHookGuardrail(guardrail_name="gr-post", event_hook=GuardrailEventHooks.post_call, default_on=True)],
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None, raising=False)
+    data = _post_call_pipeline_data(stream=True)
+
+    result = await proxy_logging.async_post_call_streaming_hook(
+        data=data,
+        response=_stream_chunks()[0],
+        user_api_key_dict=make_user_api_key_auth(request_route="/v1/chat/completions"),
+    )
+
+    assert result is not None
+    assert seen["count"] == 1
+    assert seen["response"] == "hello "