Feat/prompt caching

[Unshure]

Summary

• Add CacheConfig with strategy="auto" for automatic prompt caching in BedrockModel
• Cache points are injected at the end of the last assistant message before each model call
• Supports all Claude models on Bedrock that have prompt caching capability

Usage

from strands import Agent
from strands.models import BedrockModel, CacheConfig

model = BedrockModel(
    model_id="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    cache_config=CacheConfig(strategy="auto")
)
agent = Agent(model=model)

Test plan

• Unit tests for cache point injection logic
• Integration test with Claude models on Bedrock confirming cache hits

[Unshure]

/strands review

[github-actions]

Issue: _inject_cache_point mutates the original messages list in-place, which can lead to unexpected side effects on agent.messages.

Suggestion: Make a deep copy of the messages before injecting cache points:

import copy

# In _format_bedrock_messages:
if cache_config and cache_config.strategy == "auto":
    if self._supports_caching:
        messages = copy.deepcopy(messages)  # Work on a copy
        self._inject_cache_point(messages)

This ensures that agent.messages remains unmodified and cache points are only in the request payload sent to Bedrock.

[github-actions]

Suggestion: Consider making this check more robust. Currently it uses string matching which could have edge cases:

1. A model ID containing "claude" in a different context
2. Future Anthropic models with different naming

Consider adding a comment explaining the heuristic, or checking against a known list of supported model patterns.

[github-actions]

Issue: This test uses mock_model which doesn't have the _inject_cache_point method, so it doesn't actually test whether BedrockModel mutates messages.

Suggestion: Move this test to tests/strands/models/test_bedrock.py and use a real BedrockModel instance to verify the mutation behavior. For example:

def test_format_bedrock_messages_does_not_mutate_original(bedrock_client):
    """Test that _format_bedrock_messages doesn't mutate the original messages."""
    model = BedrockModel(
        model_id="us.anthropic.claude-sonnet-4-20250514-v1:0",
        cache_config=CacheConfig(strategy="auto")
    )
    
    messages = [
        {"role": "user", "content": [{"text": "Hello"}]},
        {"role": "assistant", "content": [{"text": "Hi there!"}]},
    ]
    original_content_len = len(messages[1]["content"])
    
    model._format_bedrock_messages(messages)
    
    # Original messages should be unchanged
    assert len(messages[1]["content"]) == original_content_len

[github-actions]

Good docstring! One minor addition - consider documenting that this config is currently only supported by BedrockModel.

[github-actions]

Review Summary

Assessment: Request Changes

Key Themes

The implementation approach is well thought out - automatically moving cache points to the last assistant message is a clean way to maximize cache hits. The code is well-structured and follows the repository's coding patterns.

Critical Issue:

• Message Mutation: The _inject_cache_point method modifies the original messages list in-place. This means agent.messages gets mutated with cache points, which could cause unexpected behavior. A deep copy should be made before injection.

Testing Gaps:

• The test test_cache_config_does_not_mutate_original_messages uses a mock model that doesn't have the injection logic, so it doesn't actually verify the mutation behavior
• No integration test was added for the new CacheConfig feature (existing tests cover the deprecated cache_prompt)

Minor Suggestions:

• The _supports_caching property could benefit from documentation explaining the heuristic
• CacheConfig docstring could mention it's currently only supported by BedrockModel

---

Nice work on the feature design! The cache point auto-injection is a useful addition. Once the mutation issue is addressed, this will be a solid contribution. 🚀