test(bedrock): repair a vacuous test, drop a dead condition, pin the hardening

Round 3 of review. The production logic held up -- 6 of 7 mutations killed, no
breakage across the full bedrock directory (1089 passed; the 5 errors are a
missing pytest-mock in this env and reproduce at base) -- but one round-2 test
asserted nothing.

test_response_format_injected_tools_still_honour_parallel_tool_calls picked
"anthropic.claude-sonnet-4-5-20250929-v1:0", which advertises
supports_native_structured_output, so _translate_response_format_param takes the
outputConfig branch and injects NO synthetic tool. optional_params had no tools,
the list was empty, and the whole assertion sat behind `if listed:` and never ran
-- it passed with the round-2 fix reverted, which made it the only coverage for
half that fix and worthless. Of the 59 models carrying
supports_parallel_tool_use_config exactly one lacks native structured output
("claude-sonnet-4-5-20250929-v1:0", bare id), so the test now uses that and
asserts the exact payload unconditionally:

    toolConfig.tools = [json_tool_call]
    additionalModelRequestFields.tool_choice = {"type": "tool",
        "name": "json_tool_call", "disable_parallel_tool_use": True}

Verified it now dies under the mutation it exists to catch.

Dropped `and "type" in passthrough_tool_choice` from
_apply_parallel_tool_use_config: the builder sets type on all three branches, so
it could never be false -- the same reasoning round 1 used to delete the
unreachable merge branch two lines above.

Softened the guard comment, which claimed more than the guard delivers. A
non-empty optional_params["tools"] does not imply a non-empty toolConfig.tools:
_bedrock_tools_pt later drops tools carrying neither "function" nor
"input_schema" (the Responses built-ins, e.g. web_search), so a request whose
tools are all of that kind still reaches the provider with a tool_choice and no
toolConfig. Measured, and identical before this feature, so it is not a
regression -- but the comment should not assert the payload is prevented.

Added the one test that pins the non-dict native["tool"] hardening, which no
mutation was catching.

Confirmed by review that the tool_choice="none" early return is in the right
place: Anthropic's ToolChoiceNone carries only `type`, and
disable_parallel_tool_use exists on auto/any/tool only -- so emitting the bare
flag is the "missing field type" 400 this PR fixes, and {"type":"none",
disable_parallel_tool_use} would be rejected as an extra input. Emitting nothing
is the only correct option, which is also what litellm's own Anthropic transform
does.

174 pass.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: pseudotensor

litellm/llms/bedrock/chat/converse_transformation.py

@@ -123,8 +123,10 @@ def _apply_parallel_tool_use_config(
     # hybrid such as {"type": "auto", "name": <user>, "disable_parallel_tool_use": true}.
     additional_request_params.update(parallel_tool_use_config)
 
-    passthrough_tool_choice = parallel_tool_use_config.get("tool_choice")
-    if isinstance(passthrough_tool_choice, dict) and "type" in passthrough_tool_choice:
+    # No `"type" in ...` check: _map_parallel_tool_use_config sets type on every
+    # branch, so that condition could never be false. (Same reasoning that removed
+    # the unreachable merge branch above.)
+    if isinstance(parallel_tool_use_config.get("tool_choice"), dict):
         inference_params.pop("tool_choice", None)
 
 
@@ -1014,10 +1016,17 @@ def _map_parallel_tool_use_config(
         # A truthy non-list ``tools`` (e.g. a bare dict) passes a
         # non_default_params check but is skipped by the mapping loop's
         # isinstance(value, list) guard, so the payload came out with
-        # additionalModelRequestFields.tool_choice and NO toolConfig at all --
-        # the exact 400 this guard exists to prevent. In the other direction a
-        # json_schema response_format injects a synthetic tool into
-        # optional_params with nothing in non_default_params.
+        # additionalModelRequestFields.tool_choice and NO toolConfig at all. In the
+        # other direction a json_schema response_format injects a synthetic tool
+        # into optional_params with nothing in non_default_params.
+        #
+        # This narrows that payload but does not eliminate it: _bedrock_tools_pt
+        # later drops tools carrying neither "function" nor "input_schema" (the
+        # Responses built-ins such as web_search), so a request whose tools are ALL
+        # of that kind still reaches the provider with a tool_choice and no
+        # toolConfig. Measured, and unchanged from before this feature -- the only
+        # place the answer is known is _transform_request_helper, where
+        # bedrock_tools has already been computed.
         if not optional_params.get("tools"):
             return
 

tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py

@@ -6107,8 +6107,15 @@ def test_response_format_injected_tools_still_honour_parallel_tool_calls():
     with nothing in non_default_params, so reading the raw value silently ignored
     parallel_tool_calls for it.
     """
+    # The bare id, NOT "anthropic.<id>": of the 59 models carrying
+    # supports_parallel_tool_use_config this is the only one WITHOUT
+    # supports_native_structured_output, so it is the only one that takes the
+    # synthetic-tool branch instead of outputConfig. Picking any other model makes
+    # this test vacuous -- no tool is injected, `toolConfig.tools` is empty, and a
+    # conditional assertion never runs. (An earlier version of this test did
+    # exactly that and passed with the fix reverted.)
     data = _converse_request(
-        "anthropic.claude-sonnet-4-5-20250929-v1:0",
+        "claude-sonnet-4-5-20250929-v1:0",
         {
             "response_format": {
                 "type": "json_schema",
@@ -6120,10 +6127,14 @@ def test_response_format_injected_tools_still_honour_parallel_tool_calls():
             "parallel_tool_calls": False,
         },
     )
-    listed = (data.get("toolConfig") or {}).get("tools") or []
-    if listed:  # only meaningful if litellm injected the synthetic tool
-        passthrough = data["additionalModelRequestFields"]["tool_choice"]
-        assert passthrough["disable_parallel_tool_use"] is True
+    listed = [t["toolSpec"]["name"] for t in (data.get("toolConfig") or {}).get("tools") or []]
+    assert listed == ["json_tool_call"], f"expected the synthetic tool, got {listed}"
+    passthrough = data["additionalModelRequestFields"]["tool_choice"]
+    assert passthrough == {
+        "type": "tool",
+        "name": "json_tool_call",
+        "disable_parallel_tool_use": True,
+    }, passthrough
 
 
 def test_tool_choice_none_is_not_reinvented_as_auto():
@@ -6143,3 +6154,18 @@ def test_tool_choice_none_is_not_reinvented_as_auto():
         drop_params=True,
     )
     assert "tool_choice" not in data.get("additionalModelRequestFields", {})
+
+
+def test_a_malformed_native_tool_choice_does_not_raise():
+    """Defensive: unreachable through get_optional_params (every writer of
+    optional_params["tool_choice"] produces a well-formed block), but the builder
+    reads native["tool"].get("name") and would raise AttributeError on a string."""
+    config = AmazonConverseConfig()
+    optional_params = {"tools": _TOOL_PARAM, "tool_choice": {"tool": "get_weather"}}
+    config._map_parallel_tool_use_config(
+        non_default_params={"tools": _TOOL_PARAM, "parallel_tool_calls": False},
+        optional_params=optional_params,
+    )
+    tool_choice = optional_params["_parallel_tool_use_config"]["tool_choice"]
+    assert tool_choice["type"] == "tool"
+    assert "name" not in tool_choice  # nothing usable to forward