Patch 50 of 53 Dependabot alerts in uv.lock; document diskcache and bleach, which have no fix

docs/runbook.md

@@ -208,7 +208,7 @@ Google deprecated `gemini-3.5-flash` and `gemini-3.7-flash` on 2026-10-08
 ([changelog](https://ai.google.dev/gemini-api/docs/changelog)). Requests under
 those names now route to `gemini-3.6-flash` and `gemini-3.8-flash`, which are
 rows of their own, so neither row can be re-run under its name. A synchronous
-re-run or row repair would not show the re-route: litellm 1.88.1 sets the
+re-run or row repair would not show the re-route: litellm 1.88.x sets the
 response model to the requested name, which PolicyBench records as
 `provider_resolved_model`, and ignores Gemini's `modelVersion`. The Gemini
 batch path, which calls google-genai directly, does record `modelVersion`.

policybench/config.py

@@ -83,7 +83,7 @@
     # answers were requested on 2026-08-18, before Gemini 3.8 Flash was
     # released (2026-09-02), so they are Gemini 3.7 Flash, and this name can
     # no longer re-run them. A synchronous run, which row repairs also use,
-    # would not show either re-route: litellm 1.88.1 sets the response model
+    # would not show either re-route: litellm 1.88.x sets the response model
     # to the requested name and never reads Gemini's modelVersion, so
     # provider_resolved_model records gemini-3.7-flash or gemini-3.5-flash.
     # The Gemini batch path (google-genai) does record modelVersion. GET

scripts/tool_pilot_argon.py

@@ -430,9 +430,10 @@ def refuse_policyengine(label: str, value) -> None:
 # Times a JSON-contract reply that searched but gave no answer is sent back
 # as it is, Anthropic's continuation for a paused server-tool turn.
 MAX_PAUSE_CONTINUATIONS = 3
-# The litellm whose Gemini search-tool handling allow_gemini_search_with_functions
-# was written against.
-GEMINI_SEARCH_PATCH_LITELLM = "1.88.1"
+# The litellm releases whose Gemini search-tool handling
+# allow_gemini_search_with_functions was checked against. It was written against
+# 1.88.1; 1.88.6's litellm/llms/vertex_ai/gemini/ is byte-identical to 1.88.1's.
+GEMINI_SEARCH_PATCH_LITELLM = frozenset({"1.88.1", "1.88.6"})
 
 CALCULATE_DESCRIPTION = (
     "Run PolicyEngine US (policyengine-us 2.41.0) on one household and return "
@@ -1223,20 +1224,20 @@ def web_search_tool(profile: ModelProfile) -> dict:
 def allow_gemini_search_with_functions() -> None:
     """Keep Gemini's search tool beside function tools in litellm.
 
-    litellm 1.88.1 drops googleSearch from any request that also declares
-    functions unless ``include_server_side_tool_invocations`` is set. It reads
-    that flag only among the mapped OpenAI parameters, and a ``completion``
-    keyword reaches it later, as a provider extra, after the search tool is
-    gone. Gemini 3 and 4 models take both together with the flag
+    litellm 1.88.1 and 1.88.6 drop googleSearch from any request that also
+    declares functions unless ``include_server_side_tool_invocations`` is set.
+    They read that flag only among the mapped OpenAI parameters, and a
+    ``completion`` keyword reaches it later, as a provider extra, after the
+    search tool is gone. Gemini 3 and 4 models take both together with the flag
     (ai.google.dev/gemini-api/docs/tool-combination), and the web arms always
     send it, so both of litellm's drops are skipped. Refuses any other litellm,
     whose code this was not checked against.
     """
     version = importlib.metadata.version("litellm")
-    if version != GEMINI_SEARCH_PATCH_LITELLM:
+    if version not in GEMINI_SEARCH_PATCH_LITELLM:
         raise SystemExit(
             f"litellm {version} is installed; the Gemini search patch was "
-            f"checked against {GEMINI_SEARCH_PATCH_LITELLM} only"
+            f"checked against {', '.join(sorted(GEMINI_SEARCH_PATCH_LITELLM))} only"
         )
     from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
         VertexGeminiConfig,
@@ -1466,7 +1467,7 @@ def _attach_deferred_results(self, message: dict) -> None:
 
         When Claude calls a server tool beside a client tool, the API runs the
         server tool on the next request, and that response begins with its
-        result, without the call. litellm 1.88.1 re-sends a server result only
+        result, without the call. litellm 1.88.x re-sends a server result only
         beside its call in the same message, so left in place the result drops
         out of later requests, and the API refuses the call that lacks it (a
         live probe got that 400 on the third request). The result joins its

tests/test_tool_pilot_argon.py

@@ -2762,14 +2762,25 @@ def tool_keys():
         )
         return sorted(k for t in params.get("tools", []) for k in t)
 
-    if importlib.metadata.version("litellm") != pilot.GEMINI_SEARCH_PATCH_LITELLM:
-        pytest.skip("the patch refuses other litellm versions")
+    # Installed from the lock, litellm is a release the patch was checked
+    # against (the lock test below), so this runs rather than skips.
+    assert importlib.metadata.version("litellm") in pilot.GEMINI_SEARCH_PATCH_LITELLM
     assert tool_keys() == ["function_declarations"]  # litellm drops search
     pilot.allow_gemini_search_with_functions()
     assert tool_keys() == ["function_declarations", "googleSearch"]
     assert calculate_tool["function"]["name"] == pilot.TOOL_NAME
 
 
+def test_the_locked_litellm_is_one_the_gemini_patch_was_checked_against():
+    """A lock bump past the checked releases fails here, before a live web arm
+    refuses to start."""
+    import tomllib
+
+    lock = tomllib.loads((Path(__file__).resolve().parents[1] / "uv.lock").read_text())
+    (version,) = {p["version"] for p in lock["package"] if p["name"] == "litellm"}
+    assert version in pilot.GEMINI_SEARCH_PATCH_LITELLM
+
+
 def test_the_gemini_patch_refuses_another_litellm(monkeypatch):
     _keep_litellm_gemini_config(monkeypatch)
     monkeypatch.setattr(pilot.importlib.metadata, "version", lambda name: "9.9.9")