fix(agent): support AWS Bedrock provider for inference and auxiliary tasks

[zlyuan9]

What changed and why

Three bugs prevented Hermes from working when configured with AWS Bedrock as the inference provider (e.g. provider: bedrock, model: us.anthropic.claude-sonnet-4-6).

1. agent/model_metadata.py — context-length probe hits unsupported Bedrock endpoint

_is_known_provider_base_url() didn't recognize Bedrock URLs (bedrock-runtime.*.amazonaws.com), so the context-length detection fell through to a GET /models probe. Bedrock does not expose this endpoint → HTTP 400 error before any real inference request was ever attempted.

Fix: Added "bedrock-runtime" and "bedrock-mantle" to _URL_TO_PROVIDER, so Bedrock URLs are recognized as a known provider and the code falls straight to the static context-length table (200k for Claude Sonnet 4.6).

2. run_agent.py — regional inference profile IDs mangled by normalize_model_name()

Bedrock cross-region inference profile IDs use dot notation: us.anthropic.claude-sonnet-4-6. _anthropic_preserve_dots() returned False for Bedrock, so normalize_model_name() converted all dots to hyphens → us-anthropic-claude-sonnet-4-6. Bedrock rejects this as an invalid model identifier.

Fix: Added "bedrock" to the provider set in _anthropic_preserve_dots(), and added Bedrock URL substrings to the base-URL fallback check.

3. agent/auxiliary_client.py — aws_sdk auth type unhandled

resolve_provider_client() handled api_key, external_process, and oauth_device_code auth types but not aws_sdk (Bedrock's auth type). It fell through to an unknown provider warning and returned (None, None) — a confusing log message that looked like a bug rather than an expected fallback.

Fix: Added an explicit aws_sdk branch that returns (None, None) with a debug-level log, so callers fall back to an OpenAI-compatible auxiliary provider as intended.

How to test

Reproduction steps (before fix):

1. Set model.default: us.anthropic.claude-sonnet-4-6, model.provider: bedrock in ~/.hermes/config.yaml
2. Set AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY in ~/.hermes/.env
3. Run hermes chat -q "Hello" → HTTP 400 "The provided model identifier is invalid"

Verification (after fix):

1. Same config as above
2. Run hermes chat -q "Hello" → agent responds normally
3. Enable debug logging and confirm no unhandled auth_type aws_sdk warning appears
4. Confirm model name sent to Bedrock is us.anthropic.claude-sonnet-4-6 (dots preserved)

Platforms tested

• macOS (darwin)

Notes

The existing debug request dump in _dump_api_request_debug() always reconstructs the URL as /chat/completions and auth as Bearer {key} for Bedrock requests regardless of actual api_mode — this is a pre-existing display artifact and not addressed in this PR.

[limitMe]

This fix works for me

[alt-glitch]

Related to #11132, #11612, #12365, and #12979 — multiple open PRs addressing Bedrock support. Consider coordinating to avoid merge conflicts.

[zlyuan9]

Seems that the issue's been resolved already, I think we can close this PR