Support model aliases with built-in thinking-effort mapping, e.g. gpt-5.4-low → gpt-5.4 + reasoning_effort=low

[ZeaoZhang]

Summary

Please support configuring model aliases that automatically map to a base model plus a predefined thinking/reasoning level.

Example:

• gpt-5.4-low -> upstream model gpt-5.4 with reasoning_effort=low
• gpt-5.4-medium -> upstream model gpt-5.4 with reasoning_effort=medium
• gpt-5.4-high -> upstream model gpt-5.4 with reasoning_effort=high

Current behavior

CLIProxyAPI already supports:

1. dynamic thinking suffix parsing such as:

   • gpt-5.4(low)
   • gpt-5.4(medium)

2. model aliases in some provider configs (such as openai-compatibility, vertex, claude)

However, there is currently no native config mechanism to declare aliases like:

• gpt-5.4-low
• gpt-5.4-medium
• gpt-5.4-high

and have them automatically inject the corresponding thinking/reasoning config.

For codex-api-key, there is also no models/alias section, so this is not currently possible in a clean built-in way.

Why this is useful

Some clients only work well with plain model names and do not conveniently support the model(low) syntax.

Model aliases such as gpt-5.4-low are easier to:

• expose in /v1/models
• select in clients like Cherry Studio
• distinguish for different latency/cost profiles
• use as stable presets for end users

Expected behavior

Allow configuration like this (example only):

codex-api-key:
  - api-key: "..."
    base-url: "..."
    models:
      - name: "gpt-5.4"
        alias: "gpt-5.4-low"
        reasoning-effort: "low"
      - name: "gpt-5.4"
        alias: "gpt-5.4-medium"
        reasoning-effort: "medium"
      - name: "gpt-5.4"
        alias: "gpt-5.4-high"
        reasoning-effort: "high"

or a generic alias mapping mechanism such as:

model-aliases:
  - alias: "gpt-5.4-low"
    target: "gpt-5.4"
    reasoning-effort: "low"
  - alias: "gpt-5.4-medium"
    target: "gpt-5.4"
    reasoning-effort: "medium"
  - alias: "gpt-5.4-high"
    target: "gpt-5.4"
    reasoning-effort: "high"

## Expected result

- these aliases should appear in /v1/models
- requests using these aliases should route to the base model
- the configured thinking/reasoning setting should be injected automatically
- this should work without requiring the client to use gpt-5.4(low) syntax

## Notes

At the moment this can only be approximated by combining:

- alias registration in some compatibility providers
- payload override rules

But that is not a native first-class solution, especially for Codex/OpenAI model routing.

A built-in alias + reasoning mapping feature would make this much easier to use.