test(integration): look up owned cache keys and clean up sweep models

Three integration tests diffed the whole shared Redis keyspace (scan_iter or dbsize before and after), so
a sibling test's cache write could satisfy or break them. Each now finds its own entry by a value only it
wrote: the Gemini reasoning test by a per-run Gemini responseId, the max_messages test by a per-test
answer text, and the S3 cache-hit test by its request marker

The sweep auth tests created models through raw /model/new on the shared gateway and never deleted them.
They now create them through Scenario.model, which deletes each model on exit, including when setup fails

tests/integration/authorization/test_audit_any_sweep_auth.py

@@ -9,7 +9,7 @@ import httpx
 import openai
 import pytest
 import websockets
-from integration._support.client import Gateway, eventually
+from integration._support.client import Gateway, Scenario, eventually
 from integration._support.database import read_rows
 from integration._support.wire import Reply, Request, Wire, wire_server
 from pydantic import JsonValue
@@ -63,7 +63,8 @@ _RESPONSES_EVENTS: Final = (
 )
 
 
-def _anthropic_model(gateway: Gateway, *, stream: bool) -> str:
+def _anthropic_model(scenario: Scenario, *, stream: bool) -> str:
+    gateway: Final = scenario.gateway
     scenario_id: Final = "audit-anthropic-stream" if stream else "audit-anthropic"
     if stream:
         _register_scenario(gateway, scenario_id, {"content_type": "text/event-stream", "frames": _ANTHROPIC_EVENTS})
@@ -76,23 +77,13 @@ def _anthropic_model(gateway: Gateway, *, stream: bool) -> str:
                 "routes": {"POST /v1/messages": {"content_type": "application/json", "body": _ANTHROPIC_MESSAGE}},
             },
         )
-    name: Final = f"integration-{uuid.uuid4().hex}"
+    return scenario.model(
-    created: Final = gateway.post(
+        model="anthropic/claude-3-5-sonnet-20241022", api_base=f"{gateway.upstream_url}/{scenario_id}"
-        "/model/new",
-        {
-            "model_name": name,
-            "litellm_params": {
-                "model": "anthropic/claude-3-5-sonnet-20241022",
-                "api_key": "integration-provider-key",
-                "api_base": f"{gateway.upstream_url}/{scenario_id}",
-            },
-        },
     )
-    assert isinstance(created["model_info"], dict) and created["model_info"]["id"], created
-    return name
 
 
-def _responses_model(gateway: Gateway, *, stream: bool) -> str:
+def _responses_model(scenario: Scenario, *, stream: bool) -> str:
+    gateway: Final = scenario.gateway
     scenario_id: Final = "audit-resp-stream" if stream else "audit-resp"
     if stream:
         _register_scenario(gateway, scenario_id, {"content_type": "text/event-stream", "frames": _RESPONSES_EVENTS})
@@ -105,20 +96,7 @@ def _responses_model(gateway: Gateway, *, stream: bool) -> str:
                 "routes": {"POST /responses": {"content_type": "application/json", "body": _RESPONSES_BODY}},
             },
         )
-    name: Final = f"integration-{uuid.uuid4().hex}"
+    return scenario.model(api_base=f"{gateway.upstream_url}/{scenario_id}")
-    created: Final = gateway.post(
-        "/model/new",
-        {
-            "model_name": name,
-            "litellm_params": {
-                "model": "openai/gpt-4o-mini",
-                "api_key": "integration-provider-key",
-                "api_base": f"{gateway.upstream_url}/{scenario_id}",
-            },
-        },
-    )
-    assert isinstance(created["model_info"], dict) and created["model_info"]["id"], created
-    return name
 
 
 def _openai_client(gateway: Gateway, key: str) -> openai.OpenAI:
@@ -219,8 +197,8 @@ def test_allowed_key_chat_completions_restamps_alias(
 @pytest.mark.parametrize("client_kind", ("httpx", "anthropic", "anthropic_async"))
 @pytest.mark.parametrize("stream", (False, True))
 def test_allowed_key_messages(gateway: Gateway, client_kind: str, stream: bool) -> None:
-    model: Final = _anthropic_model(gateway, stream=stream)
     with gateway.scenario() as scenario:
+        model: Final = _anthropic_model(scenario, stream=stream)
         key: Final = scenario.key(models=[model])
         body: Final = {
             "model": model,
@@ -264,8 +242,8 @@ def test_allowed_key_messages(gateway: Gateway, client_kind: str, stream: bool)
 @pytest.mark.parametrize("client_kind", ("httpx", "openai", "openai_async"))
 @pytest.mark.parametrize("stream", (False, True))
 def test_allowed_key_responses(gateway: Gateway, client_kind: str, stream: bool) -> None:
-    model: Final = _responses_model(gateway, stream=stream)
     with gateway.scenario() as scenario:
+        model: Final = _responses_model(scenario, stream=stream)
         key: Final = scenario.key(models=[model])
         body: Final = {
             "model": model,
@@ -334,20 +312,10 @@ async def test_realtime_websocket_allowed_and_denied(gateway: Gateway) -> None:
             ),
         },
     )
-    name: Final = f"integration-{uuid.uuid4().hex}"
-    created: Final = gateway.post(
-        "/model/new",
-        {
-            "model_name": name,
-            "litellm_params": {
-                "model": "openai/gpt-4o-realtime-preview",
-                "api_key": rt_scenario,
-                "api_base": gateway.upstream_url,
-            },
-        },
-    )
-    assert isinstance(created["model_info"], dict) and created["model_info"]["id"], created
     with gateway.scenario() as scenario:
+        name: Final = scenario.model(
+            model="openai/gpt-4o-realtime-preview", api_key=rt_scenario, api_base=gateway.upstream_url
+        )
         key: Final = scenario.key(models=[name])
         other_model: Final = scenario.model()
         denied_key: Final = scenario.key(models=[other_model])

tests/integration/caching/response_cache_case.py

@@ -282,16 +282,19 @@ def redis_store() -> Redis:
     return Redis(host=os.environ["REDIS_HOST"], port=int(os.environ["REDIS_PORT"]))
 
 
-def entry_key(store: Redis, needle: str) -> str | None:
+def entry_keys(store: Redis, needle: str) -> Iterator[str]:
-    """The response cache key (a bare sha256) whose stored value names ``needle``, or ``None``."""
     for raw_key in store.scan_iter(count=1000):
         key: Final = raw_key.decode()
         if not CACHE_KEY.match(key) or store.type(key) != b"string":
             continue
         value: Final = store.get(key)
         if value is not None and needle.encode() in value:
-            return key
+            yield key
-    return None
+
+
+def entry_key(store: Redis, needle: str) -> str | None:
+    """The response cache key (a bare sha256) whose stored value names ``needle``, or ``None``."""
+    return next(entry_keys(store, needle), None)
 
 
 def await_entry(store: Redis, needle: str) -> str:

tests/integration/caching/test_cache_max_messages.py

@@ -15,6 +15,7 @@ from integration._support.openai_wire import chat_reply, responses_reply
 from integration._support.process import owned_proxy
 from integration._support.provider import SharedProvider
 from integration._support.wire import Reply, Request, wire_server
+from integration.caching.response_cache_case import await_entry, redis_store
 from pydantic import JsonValue, TypeAdapter
 from redis import Redis
 from redis.exceptions import ResponseError
@@ -133,11 +134,6 @@ def _first(value: JsonValue) -> dict[str, JsonValue]:
     return object_value(value[0])
 
 
-def _cached_responses(redis: Redis[bytes]) -> frozenset[bytes]:
-    digests: Final = tuple(key for key in redis.scan_iter() if len(key) == 64)
-    return frozenset(key for key in digests if b'"response"' in (redis.get(key) or b""))
-
-
 _SCENARIOS: Final = (
     pytest.param("/v1/messages", "anthropic/claude-sonnet-5-5", _claude_code_turns, id="messages"),
     pytest.param("/v1/chat/completions", "openai/gpt-5.4", _agent_turns, id="chat-completions"),
@@ -149,22 +145,22 @@ _SCENARIOS: Final = (
 def test_cache_serves_a_turn_under_max_messages_and_skips_one_past_it(
     gateway: Gateway, provider: SharedProvider, path: str, model: str, turns: _Turns
 ) -> None:
-    under_cap, past_cap = turns(f"update the config {uuid.uuid4().hex}")
+    task: Final = uuid.uuid4().hex
-    provider.expect(_reply(path, "first answer"), _reply(path, "second answer"), _reply(path, "third answer"))
+    under_cap, past_cap = turns(f"update the config {task}")
+    answers: Final = tuple(f"{ordinal} answer {task}" for ordinal in ("first", "second", "third"))
+    provider.expect(*(_reply(path, answer) for answer in answers))
 
-    redis_client: Final[Redis[bytes]] = Redis(host=os.environ["REDIS_HOST"], port=int(os.environ["REDIS_PORT"]))
+    with redis_store() as store:
-    with redis_client as redis:
-        cached_before: Final = _cached_responses(redis)
         gateway.post(path, _body(path, model, under_cap))
-        eventually(lambda: _cached_responses(redis) - cached_before, lambda written: len(written) == 1)
+        await_entry(store, answers[0])
     repeated: Final = gateway.post(path, _body(path, model, under_cap))
     past_cap_twice: Final = (
         gateway.post(path, _body(path, model, past_cap)),
         gateway.post(path, _body(path, model, past_cap)),
     )
 
-    assert _answer(path, repeated) == "first answer", "a repeated turn under max_messages was not served from the cache"
+    assert _answer(path, repeated) == answers[0], "a repeated turn under max_messages was not served from the cache"
-    assert tuple(_answer(path, answer) for answer in past_cap_twice) == ("second answer", "third answer"), (
+    assert tuple(_answer(path, answer) for answer in past_cap_twice) == answers[1:], (
         "a turn past max_messages was served from the cache"
     )
     assert len(provider.received()) == 3

tests/integration/observability/test_s3_v2_flush_surfaces.py

@@ -1,11 +1,9 @@
-import os
 import re
 import uuid
 from pathlib import Path
 from typing import Final
 
 import pytest
-from redis import Redis
 from _s3_v2_support import (
     BUCKET,
     PREFIX,
@@ -21,6 +19,7 @@ from _s3_v2_support import (
 from integration._support.client import Gateway, eventually
 from integration._support.process import owned_proxy
 from integration._support.wire import wire_server
+from integration.caching.response_cache_case import entry_keys, redis_store
 
 pytestmark: Final = pytest.mark.shares_worker_state
 
@@ -117,13 +116,14 @@ def test_s3_v2_cache_hit_twins_log_one_object_per_request(gateway: Gateway, tmp_
                 model="anthropic/claude-sonnet-4-5-20250929", api_base=provider.url, api_key="synthetic-provider-key"
             )
             key: Final = scenario.key(models=[openai_model, anthropic_model])
-            cache: Final = Redis(host=os.environ["REDIS_HOST"], port=int(os.environ["REDIS_PORT"]))
-            keys_before: Final = cache.dbsize()
             warmed: Final = tuple(
                 call_surface(candidate, surface, openai_model, anthropic_model, key, f"{marker}-{surface}")
                 for surface in SURFACES
             )
-            eventually(cache.dbsize, lambda size: size >= keys_before + len(SURFACES), seconds=30)
+            with redis_store() as cache:
+                eventually(
+                    lambda: frozenset(entry_keys(cache, marker)), lambda owned: len(owned) >= len(SURFACES), seconds=30
+                )
             repeated: Final = tuple(
                 call_surface(candidate, surface, openai_model, anthropic_model, key, f"{marker}-{surface}", False)
                 for surface in SURFACES

tests/integration/translation/chat_completions/reasoning/test_chat_completions_reasoning_gemini.py

@@ -8,11 +8,11 @@ from unittest.mock import ANY
 from uuid import uuid4
 
 import pytest
-from integration._support.client import Gateway, eventually, list_value, object_value, string_value
+from integration._support.client import Gateway, list_value, object_value, string_value
 from integration._support.process import graceful_stop_seconds
 from integration._support.provider import SharedProvider
 from integration._support.wire import Reply
-from integration.caching.response_cache_case import CACHE_KEY, redis_store
+from integration.caching.response_cache_case import await_entry, redis_store
 from integration.translation.case import SSEFrame, TranslationTestCase
 from integration.translation.chat_completions.bases.gemini import GEMINI_3_8_FLASH_TEST_CASE
 from integration.translation.chat_completions.bases.vertex_ai import (
@@ -840,7 +840,7 @@ def without_closed_schema(tool: JsonValue) -> JsonValue:
     return {**object_value(tool), "function": {**function, "parameters": open_parameters}}
 
 
-def cacheable_run(case: TranslationTestCase, marker: str) -> TranslationTestCase:
+def cacheable_run(case: TranslationTestCase, marker: str, reply_id: str) -> TranslationTestCase:
     messages: Final = list_value(case.litellm_request["messages"])
     system: Final = object_value(messages[0])
     instruction: Final = object_value(case.expected_provider_request["system_instruction"])
@@ -853,6 +853,7 @@ def cacheable_run(case: TranslationTestCase, marker: str) -> TranslationTestCase
             "messages": [{**system, "content": text}, messages[1]],
             "tools": [without_closed_schema(tool) for tool in list_value(case.litellm_request["tools"])],
         },
+        mock_provider_response={**case.mock_provider_response, "responseId": reply_id},
         expected_provider_request={
             **case.expected_provider_request,
             "contents": [list_value(case.expected_provider_request["contents"])[0]],
@@ -861,14 +862,10 @@ def cacheable_run(case: TranslationTestCase, marker: str) -> TranslationTestCase
     )
 
 
-def response_cache_keys() -> frozenset[str]:
+def stored_reply_key(case: TranslationTestCase, gateway: Gateway, provider: SharedProvider) -> str:
-    keys: Final = (raw.decode() for raw in redis_store().scan_iter())
+    reply: Final = assert_translation(case, gateway, provider)
-    return frozenset(key for key in keys if CACHE_KEY.match(key))
+    with redis_store() as store:
-
+        return await_entry(store, string_value(reply["id"]))
-
-def stored_reply_key(before: frozenset[str]) -> str:
-    (key,) = eventually(lambda: response_cache_keys() - before, lambda added: len(added) == 1)
-    return key
 
 
 GEMINI_3_8_FLASH_SUMMARY_ALIAS_TEST_CASE: Final = with_request(
@@ -888,12 +885,11 @@ def test_a_summary_alias_ask_and_a_plain_request_never_share_a_cached_reply(
     first: TranslationTestCase, second: TranslationTestCase, thoughts_off: Gateway, provider: SharedProvider
 ) -> None:
     marker: Final = uuid4().hex
-    before: Final = response_cache_keys()
+    second_run: Final = cacheable_run(second, marker, uuid4().hex)
-    assert_translation(cacheable_run(first, marker), thoughts_off, provider)
+    first_key: Final = stored_reply_key(cacheable_run(first, marker, uuid4().hex), thoughts_off, provider)
-    first_key: Final = stored_reply_key(before)
+    second_key: Final = stored_reply_key(second_run, thoughts_off, provider)
-    assert_translation(cacheable_run(second, marker), thoughts_off, provider)
+    assert second_key != first_key
-    second_key: Final = stored_reply_key(before | {first_key})
+    repeat: Final = thoughts_off.request("POST", second.litellm_endpoint, second_run.litellm_request)
-    repeat: Final = thoughts_off.request("POST", second.litellm_endpoint, cacheable_run(second, marker).litellm_request)
     assert repeat.status_code == 200, repeat.text
     assert repeat.headers.get("x-litellm-cache-key") == second_key, sorted(repeat.headers)
     assert provider.received() == ()