hermes - 💡(How to fix) Fix [Bug] GCP Vertex AI connection fails with 404 using 'gcp' and 'rest' drivers, while direct curl works

Describe the bug
Hermes is unable to connect to a custom provider for Google Cloud Vertex AI using Anthropic's Claude 3 Sonnet model. It consistently fails with an HTTP 404 Not Found error, even though a direct curl command to the same endpoint with the same credentials works perfectly. This occurs with both the gcp driver and the more generic rest driver.

To Reproduce

1.  Configure a custom provider in config.yaml as follows:

    yaml
    # ~/.hermes/config.yaml
    model:
      default: claude-sonnet-4-6
      provider: custom
      subprovider: vertex-claude

    providers:
      custom_providers:
      - name: vertex-claude
        driver: rest
        auth:
          type: gcp
        base_url: https://aiplatform.googleapis.com/v1/projects/rtp-client-agents/locations/global/publishers/anthropic/models/{model}:streamRawPredict
        extra_body:
          anthropic_version: "vertex-2023-10-16"
        models:
        - name: claude-sonnet-4-6
          context_window: 200000
          supports_tools: true
          supports_vision: true
          default_params:
            max_tokens: 4096


2.  Ensure you are authenticated with GCP via gcloud auth application-default login.

3.  Run a simple prompt command:
    bash
    hermes -z "This is a test."


Expected behavior
Hermes should connect to the Vertex AI API and return a response from the model, just like the direct curl command does.

Actual behavior
Hermes fails after 3 retries with a 404 error:

API call failed after 3 retries: HTTP 404: Error code: 404 - [{'error': {'code': 404, 'message': 'models/claude-sonnet-4-6 is not found for API version v1main, or is not supported for generateContent. Call ModelService.ListModels to see the list of available models and their supported methods.', 'status': 'NOT_FOUND'}}]


Crucial Evidence: Working curl command

The following curl command, run from the same terminal session, works perfectly and returns a valid response from the model. This proves that GCP permissions, project ID, region, model name, and authentication are all correct.

bash
curl -X POST \
  -H "Authorization: Bearer *** auth print-access-token)" \
  -H "Content-Type: application/json; charset=utf-8" \
  -d '{
    "anthropic_version": "vertex-2023-10-16",
    "messages": [
      {
        "role": "user",
        "content": "This is a test, reply with the word: confirmed."
      }
    ],
    "max_tokens": 256
  }' \
  "https://aiplatform.googleapis.com/v1/projects/rtp-client-agents/locations/global/publishers/anthropic/models/claude-sonnet-4-6:streamRawPredict"


Successful curl Output:
json
{"model":"claude-sonnet-4-6","id":"...","type":"message","role":"assistant","content":[{"type":"text","text":"Confirmed."}],"stop_reason":"end_turn", ...}


Conclusion
Since the direct API call via curl works, the issue must lie in how Hermes is constructing or sending the HTTP request via its rest (and gcp) driver, causing the same valid endpoint to return a 404.