[BUG] Roo doesn't correctly infer details for claude 4.5 when using a custom ARN in a specific region with bedrock. (+ regions not being handled well generally)

[XL-Lewis]

Problem (one or two sentences)

i.e. region ap-southeast-2 has inference ID au.anthropic.claude-sonnet-4-5-20250929-v1:0.

When I specify that custom ID, along with a region, I get the following incorrect parameters:

Context Window: 128,000 tokens
Max output: 5,000 tokens
Supports images
Does not support computer use
Does not support prompt caching

NOTE: Using the arn provided in the roo dropdown returns an error, which is why I have to submit a custom arn anyway. I believe this is due to some issue with cross-region-inference and the specific regions allowed to use each inference id.

Context (who is affected and when)

• Bedrock users
• Custom ARNs
• Claude 4.5 users
• Local model for inference

Reproduction steps

1. Using amazon bedrock
2. Select claude 4.5 via a local region inference ID
3. Select cross region inference
4. Attempt to prompt

Expected result

The model displays a series of settings for the model, including caching and reasoning

Actual result

I get an incorrect settings set.

Usage pricing is significantly increased due to lack of caching in settings. Reasoning not available etc.

Variations tried (optional)

No response

App Version

3.28.15

API Provider (optional)

AWS Bedrock

Model Used (optional)

Claude 4.5 sonnet

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I'm exploring a potential fix for the Bedrock Claude 3.5 Sonnet model capability detection issue with custom ARNs and cross-region inference. I'll investigate why the model parameters are being incorrectly inferred and work on ensuring the correct capabilities are recognized.

[roomote]

I've opened PR #8562 to address this issue: #8562

The fix improves the model detection logic for AWS Bedrock when using custom ARNs with regional inference IDs. It now correctly identifies Claude 4.5 Sonnet capabilities including:

• ✅ Context Window: 200,000 tokens
• ✅ Max output: 8,192 tokens
• ✅ Computer use support
• ✅ Prompt caching support
• ✅ Reasoning support

The implementation also adds support for additional regional prefixes like au. that weren't previously recognized.

All CI checks have passed successfully. Feedback welcome!

[hannesrudolph]

Bedrock doesn’t currently expose full runtime capability details through its APIs. You can get basic metadata via ListFoundationModels or GetFoundationModel (model ID, provider, modalities, region support, etc.), but fields like reasoning support, prompt caching, or fine-grained token limits aren’t part of those responses.

Closing this because there’s nothing further to fix on our side for this specific issue.