fix(llm_translation): extract session_id from JSON-encoded user_id metadata for prompt_cache_key

litellm/llms/anthropic/experimental_pass_through/utils.py

@@ -1,3 +1,4 @@
+import json
 import os
 from collections.abc import Mapping
 from types import MappingProxyType
@@ -21,6 +22,23 @@ _THINKING_OFF: Final = "none"
 def prompt_cache_key_from_user_id(user_id: object) -> str | None:
     if user_id is None:
         return None
+    if isinstance(user_id, dict):
+        if session_id := user_id.get("session_id"):
+            return str(session_id)[:OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH] or None
+        if user_key := user_id.get("user_id"):
+            return str(user_key)[:OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH] or None
+    elif isinstance(user_id, str):
+        trimmed = user_id.strip()
+        if trimmed.startswith("{") and trimmed.endswith("}"):
+            try:
+                parsed = json.loads(trimmed)
+                if isinstance(parsed, dict):
+                    if session_id := parsed.get("session_id"):
+                        return str(session_id)[:OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH] or None
+                    if user_key := parsed.get("user_id"):
+                        return str(user_key)[:OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH] or None
+            except Exception:
+                pass
     return str(user_id)[:OPENAI_MAX_PROMPT_CACHE_KEY_LENGTH] or None
 
 

tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_handler.py

@@ -28,6 +28,40 @@ def test_build_responses_kwargs_derives_prompt_cache_key_from_user_id():
     assert responses_kwargs["prompt_cache_key"] == "session-abc"
 
 
+def test_build_responses_kwargs_derives_prompt_cache_key_from_json_user_id_metadata():
+    json_user_id = json.dumps(
+        {
+            "device_id": "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
+            "account_uuid": "",
+            "session_id": "12345678-abcd-ef01-2345-6789abcdef01",
+        }
+    )
+    responses_kwargs = _build_responses_kwargs(
+        max_tokens=1024,
+        messages=MESSAGES,
+        model="openai/gpt-5.6-luna",
+        metadata={"user_id": json_user_id},
+        extra_kwargs={"custom_llm_provider": "openai"},
+    )
+    assert responses_kwargs["user"] == json_user_id[:64]
+    assert responses_kwargs["prompt_cache_key"] == "12345678-abcd-ef01-2345-6789abcdef01"
+
+
+def test_build_responses_kwargs_derives_prompt_cache_key_from_dict_user_id():
+    dict_user_id = {
+        "device_id": "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
+        "session_id": "sess-guid-999",
+    }
+    responses_kwargs = _build_responses_kwargs(
+        max_tokens=1024,
+        messages=MESSAGES,
+        model="openai/gpt-5.6-luna",
+        metadata={"user_id": dict_user_id},
+        extra_kwargs={"custom_llm_provider": "openai"},
+    )
+    assert responses_kwargs["prompt_cache_key"] == "sess-guid-999"
+
+
 def test_build_responses_kwargs_prefers_explicit_prompt_cache_key_over_derived():
     responses_kwargs = _build_responses_kwargs(
         max_tokens=1024,