[BUG] Bedrock cross-region inference profiles get the 8192-token fallback context window

First, thank you to the crewAI maintainers for the native Bedrock provider — the Converse-API integration is genuinely nice to build on. I think I've found a small gap in how the context window is sized for cross-region inference profiles, and I have a fix + tests ready.

Description

AWS recommends invoking newer Bedrock models (Claude 4, etc.) through cross-region inference profiles, whose model ids carry a geographic prefix: us., eu., apac., us-gov. — e.g. bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0.

BedrockCompletion.get_context_window_size() looks the model up in a table keyed on the bare id (anthropic.claude-sonnet-4) using self.model.startswith(model_prefix). When the model carries a region prefix, none of the keys match, so it falls through to the 8192 default. A 200K-token model is therefore treated as ~8K, which triggers premature summarization/truncation of the conversation.

Reproduction

from crewai import LLM

base    = LLM(model="bedrock/anthropic.claude-sonnet-4-20250514-v1:0")
profile = LLM(model="bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0")  # AWS-recommended form

print(base.get_context_window_size())     # 170000  (200000 * 0.85 ratio)
print(profile.get_context_window_size())  # 6963    (8192   * 0.85)  <-- wrong

The same prefix issue affects supports_multimodal(): it hardcodes only us. variants, so eu./apac. Claude/Nova profiles are reported as non-vision.

Root cause

The lookups key on the bare model id but compare against self.model, which (correctly) keeps the cross-region prefix. The sibling validator _normalize_bedrock_fields is already prefix-agnostic — this is one spot where the prefix wasn't accounted for.

Proposed fix

Strip the cross-region prefix before the table lookup via a small shared helper, and reuse it in supports_multimodal. PR linked below.

Possibly related real-world report: #3791 (cross-region inference breakage). Happy to adjust scope if maintainers prefer.

This issue was prepared with the help of an AI agent (Claude Code); a human reviewed the analysis, reproduction, and proposed fix before submitting.

Start at BedrockCompletion.get_context_window_size() and supports_multimodal(), then compare their model-id matching with the prefix-agnostic _normalize_bedrock_fields validator. Run or extend the existing Bedrock tests to cover bare, us., eu., and apac. model IDs; done means context sizes and multimodal support are consistent across those forms.