feat: enable openai provider use aws profile

[JackYPCOnline]

Description

Amazon Bedrock's OpenAI-compatible endpoint (Mantle) at
https://bedrock-mantle.<region>.api.aws/v1 serves the Chat Completions and
Responses APIs plus Responses-only features like stateful conversations and
reasoning controls — capabilities Converse doesn't expose.

Reaching Mantle today means hand-rolling the plumbing: mint a bearer token
with aws-bedrock-token-generator
(or sign with SigV4), compute the regional base URL, and pass both into
OpenAIModel / OpenAIResponsesModel via client_args. This PR adds a
first-class bedrock_mantle_config kwarg to both classes so the Bedrock routing stops
leaking into application code.

Public API

from strands.models.openai import OpenAIModel
from strands.models.openai_responses import OpenAIResponsesModel

OpenAIModel(model_id="openai.gpt-oss-120b", bedrock_mantle_config={"region": "us-east-1"})

OpenAIResponsesModel(
    model_id="openai.gpt-oss-120b",
    stateful=True,
    params={"reasoning": {"effort": "medium"}},
    bedrock_mantle_config={"region": "us-east-1"},
)

Related Issues

Documentation PR

In-repo docs update covers the feature for Python. TypeScript tab in the
new section notes the pathway isn't yet supported on that SDK.

Type of Change

New feature

Testing

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[codecov]

Codecov Report

❌ Patch coverage is 97.14286% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/strands/models/_openai_bedrock.py	95.23%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

Issue: This PR introduces a new public API surface (aws_config kwarg on two model classes, AwsConfig TypedDict) that customers will directly use. Per the API Bar Raising guidelines, this qualifies as a "moderate change" (new parameter on frequently-used classes with a new TypedDict customers interact with).

Suggestion: Add the needs-api-review label to this PR so an API reviewer can evaluate the public surface from a customer perspective. The PR description is already well-prepared for this review.

[github-actions]

Assessment: Request Changes

The feature design is well thought out and the PR description is excellent. However, there's a critical import issue that would break existing users of the sagemaker and litellm extras.

Review Categories

• Critical: Top-level import of aws_bedrock_token_generator in _openai_bedrock.py breaks sagemaker and litellm users at import time since those extras don't include this dependency. Use a lazy import inside resolve_bedrock_client_args() instead.
• API Design: AwsConfig TypedDict should use Required[str] for region to surface misconfiguration at type-check time, matching existing patterns (AnthropicConfig, GeminiConfig). PR should also have the needs-api-review label.
• Robustness: provide_token failures will surface as opaque tracebacks — wrapping with context will improve debuggability. Duplicate region validation between __init__ and resolve_bedrock_client_args should be consolidated.

The integration test rewrite is a nice simplification and the per-request token minting is the right design for long-lived agents.

[github-actions]

Assessment: Approve (with minor suggestion)

All critical and important issues from the previous review have been addressed well. The lazy import, Required[str] typing, consolidated validation, error wrapping, and new test coverage all look good.

Remaining minor item

• Docstring accuracy: Both __init__ docstrings still list ValueError: If aws_config is missing a region under Raises, but region validation now happens at resolve time (in _resolve_client_args), not at construction time. Minor, non-blocking.

Clean implementation with solid test coverage — nice work addressing the feedback.

[github-actions]

Assessment: Approve

All feedback from prior reviews has been fully addressed. The code is clean, well-documented, and ready to merge.

Verification of fixes

• ✅ Lazy import in resolve_bedrock_client_args() — sagemaker/litellm safe
• ✅ Required[str] for region in AwsConfig
• ✅ Docstrings accurate — no misleading Raises sections
• ✅ RuntimeError wrapping for provide_token failures
• ✅ ImportError with actionable install guidance
• ✅ Consolidated validation (single source in helper)
• ✅ Test coverage for error paths
• ✅ PR description accurately describes lazy import behavior

[github-actions]

Issue: The PR description's "Public API" section shows aws_config={"region": "us-east-1"} but the actual kwarg name in the code is bedrock_mantle_config. This will confuse reviewers and readers.

Suggestion: Update the PR description code snippets to use the actual parameter name:

OpenAIModel(model_id="openai.gpt-oss-120b", bedrock_mantle_config={"region": "us-east-1"})

[github-actions]

Assessment: Approve (with minor suggestions)

Good evolution from aws_config to bedrock_mantle_config — the name is more specific and the region resolution chain (explicit → boto_session → default boto3) mirrors BedrockModel's behavior, which is consistent for AWS users.

Minor items (non-blocking)

• PR description: The "Public API" code snippet still shows aws_config instead of bedrock_mantle_config
• Fail-fast validation: The base_url/api_key conflict check in client_args could be moved to __init__ for earlier feedback
• Test coverage: test_explicit_region_wins_over_boto_session exists for OpenAIModel but not OpenAIResponsesModel (shared code, so low risk)

The rename, stricter validation, and region resolution chain are all improvements. Implementation is clean and well-tested.

[github-actions]

Assessment: Approve

The fail-fast validation for client_args conflicts is now at init time as suggested — good improvement. The PR description still shows aws_config in the first code snippet (OpenAIModel example) which should be bedrock_mantle_config, but this was flagged in the prior review and is non-blocking.

No new issues found. The implementation is clean and ready to merge.