hermes - 💡(How to fix) Fix Support per-model reasoning_effort in custom_providers[].models [1 participants]

custom_providers:
  - name: my-provider
    base_url: https://api.example.com/v1
    api_key: ${API_KEY}
    models:
      deepseek-v4-pro:
        context_length: 1000000
        reasoning_effort: high      # NEW: per-model reasoning
      deepseek-v4-flash:
        context_length: 1000000
        reasoning_effort: medium
      qwen3-6-plus:
        context_length: 1000000
        reasoning_effort: medium    # Reasoning model, medium triggers CoT
      glm-5:
        context_length: 200000
        reasoning_effort: none      # Explicitly disable for unsupported models
      minimax-m2-7:
        context_length: 204800
        reasoning_effort: none

---

# Around line 2139 in config.py
reasoning_effort = entry.get("reasoning_effort")
if isinstance(reasoning_effort, str) and reasoning_effort.strip():
    # Validate against known levels
    from hermes_constants import VALID_REASONING_EFFORTS
    if reasoning_effort.strip().lower() in VALID_REASONING_EFFORTS:
        normalized["reasoning_effort"] = reasoning_effort.strip().lower()

---

def _get_model_specific_reasoning(model: str, config: dict) -> str | None:
    """Look up per-model reasoning_effort from custom_providers."""
    for provider in config.get("custom_providers", []):
        models = provider.get("models", {})
        if model in models:
            return models[model].get("reasoning_effort")
    return None

# Priority: per-model > global agent.reasoning_effort > default (medium)
per_model_reasoning = _get_model_specific_reasoning(self.model, config)
if per_model_reasoning:
    self.reasoning_effort = per_model_reasoning
elif self.reasoning_effort:  # Already set from agent.reasoning_effort
    pass
else:
    self.reasoning_effort = "medium"  # Default

---

# In process_command() for /model
new_model = args.strip()
new_reasoning = _get_model_specific_reasoning(new_model, config)
if new_reasoning:
    self.reasoning_config = parse_reasoning_effort(new_reasoning)
    # Optionally notify user
    print(f"Model switched to {new_model}, reasoning set to {new_reasoning}")

---

# User's config.yaml
custom_providers:
  - name: my-custom-provider
    base_url: https://my-api-server.example.com/v1
    models:
      deepseek-v4-pro-high:
        context_length: 1000000
        reasoning_effort: high
      glm-5:
        context_length: 128000
        reasoning_effort: none

# In session
> /model deepseek-v4-pro-high
Model: deepseek-v4-pro-high
Reasoning: high (auto-applied from config)

> /model glm-5  
Model: glm-5
Reasoning: none (auto-applied from config, no API overhead)