fix: Handle "prompt is too long" from Anthropic (#1137)

PR#1078 mentioned that context overflows were not handled, but I wasn't able to reproduce using the code changes in it.  However, in testing (using @DEA's suggested test) I was able to reproduce and consistently got a "prompt is too long:" error

Co-authored-by: Mackenzie Zastrow <[email protected]>

Author: zastrowm

src/strands/models/anthropic.py

@@ -39,6 +39,7 @@ class AnthropicModel(Model):
     }
 
     OVERFLOW_MESSAGES = {
+        "prompt is too long:",
         "input is too long",
         "input length exceeds context window",
         "input and output tokens exceed your context limit",

tests_integ/models/test_model_anthropic.py

@@ -5,7 +5,10 @@
 
 import strands
 from strands import Agent
+from strands.agent import NullConversationManager
 from strands.models.anthropic import AnthropicModel
+from strands.types.content import ContentBlock, Message
+from strands.types.exceptions import ContextWindowOverflowException
 
 """
 These tests only run if we have the anthropic api key
@@ -152,3 +155,30 @@ def test_structured_output_multi_modal_input(agent, yellow_img, yellow_color):
     tru_color = agent.structured_output(type(yellow_color), content)
     exp_color = yellow_color
     assert tru_color == exp_color
+
+
+@pytest.mark.asyncio
+def test_input_and_max_tokens_exceed_context_limit():
+    """Test that triggers 'input length and max_tokens exceed context limit' error."""
+
+    # Note that this test is written specifically in a style that allows us to swap out conversation_manager and
+    # verify behavior
+
+    model = AnthropicModel(
+        model_id="claude-sonnet-4-20250514",
+        max_tokens=64000,
+    )
+
+    large_message = "This is a very long text. " * 10000
+
+    messages = [
+        Message(role="user", content=[ContentBlock(text=large_message)]),
+        Message(role="assistant", content=[ContentBlock(text=large_message)]),
+        Message(role="user", content=[ContentBlock(text=large_message)]),
+    ]
+
+    # NullConversationManager will propagate ContextWindowOverflowException directly instead of handling it
+    agent = Agent(model=model, conversation_manager=NullConversationManager())
+
+    with pytest.raises(ContextWindowOverflowException):
+        agent(messages)