feat: Add Amazon Bedrock provider

[unexge]

Adding initial Amazon Bedrock provider using @aws-sdk/client-bedrock-runtime with the ConverseStream API.

[Copilot]

Pull request overview

This PR adds support for Amazon Bedrock as a new AI provider, leveraging the AWS SDK's ConverseStream API for streaming responses. The implementation includes comprehensive model definitions, tool calling support, and image handling capabilities while maintaining consistency with existing provider patterns.

Key changes:

• New Amazon Bedrock provider implementation using @aws-sdk/client-bedrock-runtime
• Support for 60+ Bedrock models including Claude, Llama, DeepSeek, and Amazon Nova variants
• Integration of cross-region inference profiles for select Claude models
• Authentication via AWS credentials (environment variables or profile) instead of API keys

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
packages/ai/src/providers/amazon-bedrock.ts	Core provider implementation handling streaming, message conversion, tool calling, and image processing for Bedrock's ConverseStream API
packages/ai/src/stream.ts	Integration of Bedrock provider into the unified streaming interface with AWS credential handling
packages/ai/src/types.ts	Type definitions for Bedrock API and options including region and profile configuration
packages/ai/src/models.generated.ts	Generated model definitions for 60+ Bedrock models with pricing, context windows, and capability metadata
packages/ai/scripts/generate-models.ts	Model generation logic including cross-region inference profile handling for specific Claude models
packages/coding-agent/src/core/model-resolver.ts	Default model configuration for amazon-bedrock provider
packages/ai/test/stream.test.ts	E2E test suite covering basic generation, tool calling, streaming, and image input
packages/ai/package.json	Added @aws-sdk/client-bedrock-runtime dependency and bedrock keyword
packages/ai/README.md	Documentation update listing Amazon Bedrock as a supported provider
packages/ai/src/providers/anthropic.ts	Fixed typo in error message (corrected "unkown" to "unknown" and "ocurred" to "occurred")

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[badlogic]

Cool! Configuring thinking etc. should be part of this PR, please. Also need to add bedrock to all other ai package tests that have multiple providers to be tested in them.

Any way you can disable Copilot for PRs send to this repo? It's never really helpful.

[unexge]

Yup I'll convert this PR into a draft until I finish adding remaining bits.

Yeah sorry about the Copilot, I was testing it on my personal project and I think I messed up with configurations to review all PRs I create :)

[unexge]

Created a PR on models.dev to add cross-region inference profiles for Claude 4.x family: anomalyco/models.dev#607

[unexge]

@badlogic provider related changes should be ready to review now. Would you also prefer having coding agent changes in this PR as well (haven't checked the necessary changes and not sure if those would be small or big tbh)?

Also, I'm running tests on my local currently, but do you also plan to set up an AWS account for CI as well?

[badlogic]

CI currently doesn't do any LLM endpoint tests, I let them run locally before a release, or when the ai package changes. Too costly otherwise.

The coding-agent shouldn't really need any changes and work out of the box if the ai changes are complete. Iiuc Bedrock doesn't need pi-side login, that's handled by the SDK which looks for env vars/files on disk.

[unexge]

Added it to the coding agent as well - it wasn't much effort as you mentioned. I think now it should be ready for review.

Due to cross-region inference, some model identifiers needs to be tweaked. I've added a test to ensure all models are usable and added a TODO to upstream correct model identifiers to models.dev repository if they accept my initial PR. But for now all models work.

Currently, only AWS_PROFILE is supported to provide credentials, but we can add more if people want to use other mechanisms as well. AWS_PROFILE should be flexible enough to support most credential types.

[unexge]

Rebased and also added a CHANGELOG entry. @badlogic appreciate your review once you have some time!