droid does not respect project level config for model selection

[funkyidol]

My current global settings.json contains the following config

"sessionDefaultSettings": {
    "autonomyMode": "auto-medium",
    "model": "gpt-5.3-codex",
    "reasoningEffort": "none",
    "specModeReasoningEffort": "medium",
    "specModeModel": "gpt-5.3-codex"
  },

and my project level settings.json contains the following config

{
  "sessionDefaultSettings": {
    "model": "glm-5",
    "specModeModel": "glm-5"
  }
}

but starting new droid session in the project starts droid with GPT instead of GLM.

The idea is that I want to set the droid models based on the project to fine tune costs based on how trivial the project is. I have multiple projects running in parallel and some of them just need agentic tool execution instead of high level reasoning.

[zwen64657]

We recommend that Droid, while fixing the configuration priority issue (#660), introduce a "Capability Tier Abstraction + Configurable Mapping" mechanism—for instance, by mapping tiers like "MODEL_TIER_LIGHT": "glm-4.5-air", "BALANCED": "glm-4.7", and "POWERFUL": "glm-5"—enabling the system to automatically route tasks to the optimal model based on complexity, achieving intelligent cost-efficiency where simple tasks use lightweight models to save money and complex tasks leverage powerful models to ensure quality.