Tighten Responses bad request fallback

Keep fallback for endpoint-specific and Venice-style Responses payload rejections, but stop treating generic Bad Request validation text as proof that Responses is unsupported.

Add a regression that generic pre-output Responses 400s raise without trying Chat Completions.

helpers/litellm_transport.py

@@ -1574,10 +1574,9 @@ def _looks_like_responses_request_rejected(text: str) -> bool:
             "expected object, received string",
             "expected string, received array",
             "zod",
-            "invalid request",
-            "invalid type",
-            "failed to deserialize",
-            "validation error",
+            "failed to deserialize input",
+            "failed to deserialize response",
+            "failed to deserialize responses",
         )
     )
 

helpers/litellm_transport.py.dox.md

@@ -26,7 +26,7 @@
 - Strip Agent Zero internal kwargs before sending requests to LiteLLM.
 - Do not send orphan tool controls when no tools are present; strict OpenAI-compatible servers can reject empty `tools` arrays.
 - Prefer Responses API when configured, but fallback to Chat Completions when the provider does not support Responses.
-- Fall back to Chat Completions when a Responses request is rejected before any output by a Bad Request/validation error that indicates the provider cannot parse the Responses request shape.
+- Fall back to Chat Completions when a Responses request is rejected before any output by an endpoint-specific or shape-specific Bad Request indicating the provider cannot parse Responses payloads.
 - Preserve provider-state metadata when Responses API calls succeed, and fall back to local replay when provider state is unsupported.
 - Keep prompt-cache markers only for providers that accept them.
 

tests/test_stream_tool_early_stop.py

@@ -611,6 +611,48 @@ async def test_unified_call_falls_back_when_responses_bad_request_rejects_shape(
     assert calls == ["responses", "chat"]
 
 
+@pytest.mark.asyncio
+async def test_unified_call_raises_generic_responses_bad_request(monkeypatch):
+    class BadRequestError(Exception):
+        status_code = 400
+
+    calls: list[str] = []
+
+    async def fake_aresponses(*args, **kwargs):
+        calls.append("responses")
+        raise BadRequestError(
+            "BadRequestError: validation error: invalid request: max_tokens is too high"
+        )
+
+    async def fake_acompletion(*args, **kwargs):
+        calls.append("chat")
+        raise AssertionError("generic 400 should not fallback to chat")
+
+    async def fake_rate_limiter(*args, **kwargs):
+        return None
+
+    monkeypatch.setattr(litellm_transport, "aresponses", fake_aresponses)
+    monkeypatch.setattr(litellm_transport, "acompletion", fake_acompletion)
+    monkeypatch.setattr(models, "apply_rate_limiter", fake_rate_limiter)
+
+    wrapper = models.LiteLLMChatWrapper(
+        model="test-model",
+        provider="openai",
+        model_config=None,
+    )
+
+    async def response_callback(chunk: str, full: str):
+        return None
+
+    with pytest.raises(BadRequestError):
+        await wrapper.unified_call(
+            messages=[],
+            response_callback=response_callback,
+        )
+
+    assert calls == ["responses"]
+
+
 @pytest.mark.asyncio
 async def test_unified_call_preserves_cache_control_with_chat_for_non_native_responses(
     monkeypatch,