fix: model-family-aware Mantle base URL resolution for OpenAI models

[Ashutosh0x]

Issue

Fixes #2629

Problem

The _MANTLE_BASE_URL_TEMPLATE hardcodes /v1 for all models, but OpenAI model families require different path prefixes on Bedrock Mantle:

Model Family	Required Path	Example
openai.gpt-5.* (GPT-5.4, GPT-5.5)	/openai/v1	https://bedrock-mantle.{region}.api.aws/openai/v1
openai.gpt-oss-* (GPT-OSS-120B)	/v1	https://bedrock-mantle.{region}.api.aws/v1

As confirmed by @SummerWJLiu in #2629: flipping the template wholesale to /openai/v1 would break gpt-oss models.

Solution

Instead of a single static template, this PR adds model-family-aware URL resolution:

1. New _resolve_mantle_base_url() helper

`python
def _resolve_mantle_base_url(region: str, model_id: str | None = None) -> str:
path_prefix = 'v1'
if model_id:
for prefix in _OPENAI_PATH_MODEL_PREFIXES:
if model_id.startswith(prefix):
path_prefix = 'openai/v1'
break
return _MANTLE_BASE_URL_TEMPLATE.format(region=region, path_prefix=path_prefix)
``n

2. model_id parameter added to

esolve_bedrock_client_args()`nCallers now pass model_id to enable automatic path resolution.

3. base_url override in BedrockMantleConfig`nFor unmapped model families, users can set base_url in the config dict to bypass automatic resolution:

`python
model = OpenAIResponsesModel(
model_id='openai.new-model',
bedrock_mantle_config={'region': 'us-east-1', 'base_url': 'https://bedrock-mantle.us-east-1.api.aws/custom/v1'}
)
``n

Changes

• _openai_bedrock.py: Added _OPENAI_PATH_MODEL_PREFIXES, _resolve_mantle_base_url(), �ase_url field in BedrockMantleConfig, and model_id param to
  esolve_bedrock_client_args()

Note

Callers of
esolve_bedrock_client_args() (in OpenAIModel and OpenAIResponsesModel) need to be updated to pass model_id. The model_id parameter defaults to None (backward compatible — falls back to /v1).

Testing

• gpt-5.4 -> resolves to /openai/v1 path
• gpt-5.5 -> resolves to /openai/v1 path
• gpt-oss-120b -> resolves to /v1 path (unchanged)
• None model_id -> resolves to /v1 path (backward compatible)
• Explicit base_url override -> uses override (escape hatch)

[Ashutosh0x]

Note: This PR currently only modifies _openai_bedrock.py with the model-family-aware resolver. The callers (OpenAIModel and OpenAIResponsesModel) need a one-line update each to pass model_id:

`python

In openai.py (line ~144) and openai_responses.py (line ~198):

Before:

return resolve_bedrock_client_args(self._bedrock_mantle_config, self.client_args)

After:

return resolve_bedrock_client_args(self._bedrock_mantle_config, self.client_args, model_id=self.config.get('model_id'))
``n
The model_id param defaults to None (backward compatible, falls back to /v1), so the core fix is safe to merge independently. Happy to add the caller updates in a follow-up commit if preferred.