[BUG] CacheConfig(strategy="auto") does not detect Claude models when using ARN-based inference profiles

[mrtj]

Description

CacheConfig(strategy="auto") fails to activate prompt caching when the model_id is an ARN-based application inference profile (e.g., arn:aws:bedrock:us-east-1:123456:application-inference-profile/abc123).

Root cause

The _cache_strategy property in BedrockModel checks for "claude" or "anthropic" substrings in the model ID:

# strands/models/bedrock.py, line ~192
def _cache_strategy(self) -> str | None:
    model_id = self.config.get("model_id", "").lower()
    if "claude" in model_id or "anthropic" in model_id:
        return "anthropic"
    return None

ARN-based inference profiles contain neither substring — they look like arn:aws:bedrock:us-east-1:123456:application-inference-profile/2i76zr9zafkf. So auto resolves to None and caching is silently skipped.

Reproduction

from strands.models.bedrock import BedrockModel
from strands.models.model import CacheConfig

model = BedrockModel(
    model_id="arn:aws:bedrock:us-east-1:123456:application-inference-profile/abc123",
    cache_config=CacheConfig(strategy="auto"),
)

# _cache_strategy returns None → caching silently disabled
# cacheReadInputTokens and cacheWriteInputTokens will always be 0

Expected behavior

auto should detect Claude models behind ARN inference profiles and enable caching. Possible approaches:

1. Resolve the ARN at init time via bedrock.get_inference_profile() to discover the underlying foundation model
2. Try caching optimistically — non-cache-capable models ignore cachePoint blocks in the Converse API, so injecting them is harmless
3. Check the response for cacheWriteInputTokens > 0 on the first call to confirm support

Workaround

Use CacheConfig(strategy="anthropic") instead of "auto" to force-enable caching regardless of model ID detection:

model = BedrockModel(
    model_id="arn:aws:bedrock:us-east-1:123456:application-inference-profile/abc123",
    cache_config=CacheConfig(strategy="anthropic"),
)

Environment

• strands-agents 1.37.0
• Python 3.12
• AWS Bedrock with application inference profiles for Claude Sonnet 4.6

[Zelys-DFKH]

I dug into this and traced the bug to _cache_strategy in BedrockModel. It string-matches on model_id to detect Claude models, which works great for named model IDs and cross-region inference profiles (their ARNs contain "anthropic" directly). Application inference profiles are different: they use opaque resource IDs like application-inference-profile/abc123 that have no connection to the underlying model name. So the check misses them and caching silently falls through to None.

The fix calls GetInferenceProfile on the Bedrock management API (not bedrock-runtime, which doesn't have this endpoint) to resolve the underlying foundation model ARN, which does contain "anthropic" for Claude-backed profiles. Result is cached on the instance, so there's only one extra API call per model instance rather than one per request.

Opening a PR now. Happy to adjust the approach if you have a different direction in mind.