Problem

When using Google Gemini as the primary model via provider: custom + base_url, everything works correctly (confirmed with v0.7.0). However, when trying to use the same Google Gemini endpoint as a fallback provider, resolve_provider_client in auxiliary_client.py fails with:
WARNING agent.auxiliary_client: resolve_provider_client: custom/main requested but no endpoint credentials found
WARNING root: Fallback to custom failed: provider not configured

This happens regardless of whether:

• fallback_model uses provider: custom with base_url + api_key_env
• fallback_model uses provider: custom with base_url + api_key: ${GOOGLE_API_KEY}
• fallback_model uses provider: google
• fallback_providers array entries include base_url/api_key fields
• custom_providers is used to register a named provider
• OPENAI_BASE_URL + OPENAI_API_KEY are set in .env

The root cause appears to be that resolve_provider_client does not resolve custom as a provider (returns None), and google is not in the supported fallback provider list. This is the same underlying issue as #4172, but that fix only addressed the primary model path — the fallback code path in auxiliary_client.py still has the same problem.

Expected Behavior

Google Gemini should work as a fallback provider, either:

1. By adding google as a first-class fallback provider (using GOOGLE_API_KEY from .env), or
2. By making provider: custom actually read base_url/api_key/api_key_env from fallback_model and fallback_providers config entries

Environment

• Hermes Agent v0.7.0 (v2026.4.3)
• Primary: openai-codex/gpt-5.3-codex
• Desired fallback: gemini-3-flash-preview via https://generativelanguage.googleapis.com/v1beta/openai
• GOOGLE_API_KEY set in ~/.hermes/.env
• Google Gemini works perfectly as primary model with provider: custom

Workaround

Using Gemini as the primary model instead and falling back to Codex/OpenRouter.