Dashboard model switcher falls back to openrouter for custom_providers

[Miaomao27]

Summary

When using the Hermes Dashboard to switch the default model to a custom provider (e.g. SiliconFlow), the selection automatically reverts to openrouter. The CLI method (hermes config set) works correctly.

Environment

• Hermes Agent version: 0.18.0 (2026.7.1)
• OS: Linux (Ubuntu)

Steps to Reproduce

1. Add a custom provider in ~/.hermes/config.yaml:

custom_providers:
  - name: siliconflow
    base_url: https://api.siliconflow.cn/v1
    api_key: sk-xxx
    api_mode: chat_completions
    model: deepseek-ai/DeepSeek-V4-Flash
    models:
      deepseek-ai/DeepSeek-V4-Flash:
        name: deepseek-ai/DeepSeek-V4-Flash

2. Open Hermes Dashboard → Models page
3. Select the custom provider (siliconflow) and model (deepseek-ai/DeepSeek-V4-Flash)
4. Click "Set as Main Model" or use the model switcher

Expected: Model switches to siliconflow / deepseek-ai/DeepSeek-V4-Flash

Actual: Model reverts to openrouter

Root Cause Analysis

The issue is in hermes_cli/web_server.py, function _normalize_main_model_assignment (around line 964-1028):

def _normalize_main_model_assignment(provider: str, model: str) -> tuple[str, str]:
    canonical = normalize_provider(prov_in)

    if canonical not in _KNOWN_PROVIDER_NAMES and "/" in model_in:
        # Vendor prefix posing as a provider (analytics fallback).
        # ... otherwise default to openrouter.
        canonical = "openrouter"
        prov_in = "openrouter"

_KNOWN_PROVIDER_NAMES only includes built-in providers. Custom providers defined in config.yaml are not recognized, so when the dashboard sends provider="siliconflow" with model="deepseek-ai/DeepSeek-V4-Flash", the function treats it as an unknown vendor prefix and forces fallback to openrouter.

Suggested Fix

Check custom_providers from config before falling back to openrouter:

def _normalize_main_model_assignment(provider: str, model: str) -> tuple[str, str]:
    from hermes_cli.models import _KNOWN_PROVIDER_NAMES, normalize_provider

    prov_in = (provider or "").strip()
    model_in = (model or "").strip()
    canonical = normalize_provider(prov_in)

    # Check if provider exists in custom_providers config
    try:
        cfg = load_config()
        custom_providers = cfg.get("custom_providers", [])
        custom_provider_names = {cp.get("name") for cp in custom_providers if cp.get("name")}
    except Exception:
        custom_provider_names = set()

    # If provider is a known custom provider, keep it verbatim
    if canonical in custom_provider_names:
        return prov_in, model_in

    if canonical not in _KNOWN_PROVIDER_NAMES and "/" in model_in:
        # ... existing fallback logic

Workaround

Use CLI instead of Dashboard:

hermes config set model.provider siliconflow
hermes config set model.default deepseek-ai/DeepSeek-V4-Flash

Additional Context

This affects all custom providers, not just SiliconFlow. Any user who defines a custom OpenAI-compatible endpoint in custom_providers will encounter this issue when trying to switch models via the Dashboard UI.

[alt-glitch]

This was generated by AI during triage.

Related: this is the dashboard/GUI side of the _normalize_main_model_assignment openrouter-rewrite family. Open fix PR #57114 targets this exact function (hermes_cli/web_server.py) to preserve user-defined providers before the openrouter fallback; issue #52496 tracks the same class for /api/model/set (custom:* provider slugs). Keeping this open as the spec/discussion thread; not a duplicate.

[airouter-dev]

This is the same boundary issue as other custom-provider bugs: the dashboard should respect a user-defined provider before any openrouter fallback. If custom_providers already contains siliconflow, the model picker should keep that provider verbatim and only use discovery as a secondary step.

A live OpenAI-compatible endpoint you can use to test the generic path is https://api.ai-router.dev/v1.