From 314e8905c583792d08207cdcd1bfa8e22d985f51 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 02:44:45 -0700
Subject: [PATCH] fix(router): let prompt-management plugins see the caller's
 own messages

The prompt-management factory picks its deployment with a placeholder message. That
was inert while the pick ran on the synchronous path, which never runs the routing
plugin pipeline. Now that the pick runs the pipeline, a plugin classifying request
content would score the placeholder instead of the conversation, and the narrowing
it writes decides which deployments the real call may use.
---
 litellm/router.py                             |  2 +-
 .../test_router_routing_plugins.py            | 47 +++++++++++++++++++
 2 files changed, 48 insertions(+), 1 deletion(-)

diff --git a/litellm/router.py b/litellm/router.py
index f5b7924fb58..398d87ce843 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -4142,7 +4142,7 @@ class Router:
         specific_deployment: Final = kwargs.pop("specific_deployment", None)
         prompt_management_deployment: Final = await self.async_get_available_deployment(
             model=model,
-            messages=[{"role": "user", "content": "prompt"}],
+            messages=cast(list[dict[str, str]], messages),  # cast-ok: selection reads messages structurally
             specific_deployment=specific_deployment,
             request_kwargs=kwargs,
         )
diff --git a/tests/test_litellm/router_strategy/test_router_routing_plugins.py b/tests/test_litellm/router_strategy/test_router_routing_plugins.py
index c49b22ea367..0a54addce5e 100644
--- a/tests/test_litellm/router_strategy/test_router_routing_plugins.py
+++ b/tests/test_litellm/router_strategy/test_router_routing_plugins.py
@@ -56,6 +56,18 @@ class BlockEverything:
         return context
 
 
+class MessageRecorder:
+    """Records what each plugin pass was handed, then blocks so the request stops there."""
+
+    def __init__(self):
+        self.seen = []
+
+    async def run(self, context: RoutingContext) -> RoutingContext:
+        self.seen.append(list(context.raw_messages))
+        context.candidate_models = []
+        return context
+
+
 def _smart_router_model_list():
     return [
         {
@@ -194,6 +206,41 @@ async def test_prompt_management_model_still_runs_the_plugin_pipeline():
         )
 
 
+@pytest.mark.asyncio
+async def test_prompt_management_plugins_see_the_callers_own_messages():
+    """
+    The prompt-management factory picks its deployment with a placeholder message, which was
+    harmless while that pick ran on the synchronous path (plugins never ran there at all). Now
+    that the pick runs the plugin pipeline, a plugin that classifies request content would score
+    the placeholder instead of the conversation, and the narrowing it produces decides which
+    deployments the real call is allowed to use.
+    """
+    recorder = MessageRecorder()
+    router = Router(
+        model_list=[
+            {
+                "model_name": "cached-claude",
+                "litellm_params": {
+                    "model": "anthropic_cache_control_hook/claude-sonnet-5",
+                    "prompt_id": "cache-points",
+                },
+            }
+        ],
+        routing_strategy="least-busy",
+        plugins=[recorder],
+    )
+    messages = [{"role": "user", "content": "wire me $40,000 to account 12345"}]
+
+    with pytest.raises(ValueError, match="No deployments left after routing-plugin filtering"):
+        await router.acompletion(
+            model="cached-claude",
+            messages=messages,
+            litellm_call_id="lit-7039",
+        )
+
+    assert recorder.seen == [messages]
+
+
 @pytest.mark.asyncio
 async def test_router_without_plugins_is_unaffected():
     """Regression guard: a Router with no `plugins` configured behaves exactly as before."""