[BUG] Prompt Caching Does not work with AWS Bedrock Custom ARN

[itisjustjeff]

Problem (one or two sentences)

Although a checkbox exists for prompt caching and it is selected, caching doesn't work when using a custom ARN. The underlying model works with prompt caching and is enabled for other tools. We can clearly see in logs that caching is not being used by Roo when interacting with our custom ARN.

{
    "inputTokenCount": 22454,
    "cacheReadInputTokenCount": 0,
    "cacheWriteInputTokenCount": 0
}

I believe that there is still something that is overriding the "Enable Prompt Cache" checkbox for custom arn models that disables it regardless of the checkbox.

Context (who is affected and when)

Happens to all users who are using AWS Bedrock with a custom ARN.

Reproduction steps

1. VS Code version 1.112.0 // RooCode Extension 3.51.1 // AWS Bedrock with enhanced logging (Or other token logging mechanism)
2. Configure AWS Bedrock in any region, check "Enable prompt caching", select "custom-arn" model, and include the custom ARN
3. Use Roo for a prompt and monitor cache usage

Expected result

AWS Bedrock uses cache input tokens instead of regular input tokens

Actual result

Only input tokens are used

Variations tried (optional)

Other applications support caching (Claude, LiteLLM, CLI)

App Version

3.51.1

API Provider (optional)

Amazon Bedrock

Model Used (optional)

Sonnet 4.5

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause

The problem is in supportsAwsPromptCache() in src/api/providers/bedrock.ts (line 1171). This method requires both conditions to be true:

1. modelConfig.info.supportsPromptCache must be truthy
2. modelConfig.info.cachableFields must exist and have length > 0

When a custom ARN is used and the model ID extracted from the ARN is not found in bedrockModels (e.g., provisioned throughput ARNs with opaque resource IDs, or custom inference profiles with non-standard names), the code falls back to getModelById() -> guessModelInfoFromId(). Two things go wrong:

1. guessModelInfoFromId() default fallback (line 340) returns supportsPromptCache: false, which gets spread over the default model info and overrides the default model's supportsPromptCache: true
2. guessModelInfoFromId() never includes cachableFields for any pattern match, so even when a Claude pattern IS matched and supportsPromptCache: true is set, the cachableFields check in supportsAwsPromptCache() still fails

The result: prompt caching is silently disabled even though the user has explicitly enabled it via the "Enable Prompt Cache" checkbox.

Proposed Fix

Two changes in src/api/providers/bedrock.ts:

1. Add cachableFields to guessModelInfoFromId() - For each Claude model pattern that already sets supportsPromptCache: true, also include cachableFields: ["system", "messages", "tools"]. This ensures that when a custom ARN's model ID is guessed via pattern matching, the full caching metadata is present.

2. Respect the user's explicit awsUsePromptCache setting for custom ARN models - In createMessage(), when a custom ARN is in use and awsUsePromptCache is explicitly true, bypass the supportsAwsPromptCache() model capability check. The user has already indicated their model supports caching. Add a default cachableFields of ["system", "messages", "tools"] to the model info if not already present, so the downstream caching logic works correctly.

3. Add/update tests to verify prompt caching works with custom ARN models (both recognized and unrecognized model IDs).

This approach is safe because:

• If the underlying model truly does not support caching, the Bedrock API simply ignores cache points - it does not error out
• The user has explicitly opted in via the checkbox, so we should trust their intent
• Known models from bedrockModels are unaffected (they already have cachableFields)

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[itisjustjeff]

@roomote These proposed changes look good and I approve this plan. Please proceed with the implementation

[roomote-v0]

Thanks for the approval @itisjustjeff! I've seen your comment and will proceed with the implementation now. I'll open a PR to share progress shortly.

[roomote-v0]

Opened PR #11984 with my attempt at addressing this. The fix adds cachableFields to the model info guessed for Claude patterns and updates supportsAwsPromptCache() to respect the user's explicit prompt caching preference when using custom ARNs. Feedback welcome!

[itisjustjeff]

I have verified that this proposed fix (and the associated PR) have resolved our issue. The fix is functional for what we need. I look forward to it being merged.