[BUG] _supports_caching doesn't recognize Bedrock application inference profile ARNs

[maud-mcevoy]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.26.0

Python Version

3.12

Operating System

macOS 26.2

Installation Method

pip

Steps to Reproduce

1. Create a Bedrock application inference profile that copies from a Claude system inference profile (e.g., us.anthropic.claude-haiku-4-5-20251001-v1:0)
2. Create a BedrockModel using the application inference profile ARN as the model_id
3. Enable prompt caching with cache_config=CacheConfig(strategy="auto")
4. Make any API call through the model

Expected Behavior

Message caching is enabled and cache points are injected into messages

Actual Behavior

1. Warning logged on every API call: "model_id=<...> | cache_config is enabled but this model does not support caching"
2. _inject_cache_point is never called
3. Message caching is silently disabled

Additional Context

The _supports_caching property only checks if "claude" or "anthropic" is in the model_id string:

def _supports_caching(self) -> bool:
    model_id = self.config.get("model_id", "").lower()
    return "claude" in model_id or "anthropic" in model_id

This works for system inference profile IDs (e.g., us.anthropic.claude-haiku-4-5-20251001-v1:0) but fails for application inference profile ARNs which don't contain these strings.

Possible Solution

No response

Related Issues

No response

[maud-mcevoy]

This is how our instance profiles are constructed:
arn:aws:bedrock:<region>:<account-id>:application-inference-profile/<random-id>

i.e. arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abcdefghijk2

[kevmyung]

cc @Unshure

Proposed Solution

The issue is that application inference profile ARNs don't contain "claude" or "anthropic" strings, so _supports_caching returns False.

Option 1: Detect Model via GetInferenceProfile API

Check the actual model by calling Bedrock API during initialization.

Major issue: This is a breaking change. It requires a new IAM permission (bedrock:GetInferenceProfile) that users might not have, causing existing code to break in production.

def _check_profile_supports_caching(self, profile_id: str) -> bool:
    try:
        import boto3
        client = boto3.client('bedrock')
        response = client.get_inference_profile(
            inferenceProfileIdentifier=profile_id
        )
        model_arn = response['models'][0]['modelArn']
        return 'anthropic.claude' in model_arn.lower()
    except Exception:
        return False  # Fails silently if no permission

• Pros: 100% accurate
• Cons: Breaking change, adds latency, requires new IAM permission

Option 2: Add explicit strategy parameter

Let users specify the caching strategy in CacheConfig:

model = BedrockModel(
    model_id='arn:aws:bedrock:...:application-inference-profile/xxx',
    cache_config=CacheConfig(strategy="anthropic")
)

• Pros: No breaking changes, zero latency, explicit control
• Cons: Requires manual specification for application inference profiles

Recommendation

Option 2 avoids breaking changes and doesn't require new IAM permissions. Users with application inference profiles can simply add strategy="anthropic" to their config.