[FEATURE] Support for per-model request parameters

[shekohex]

Feature Request

Add support for passing model-specific request parameters through configuration.

Use Case

OpenAI GPT-5.1 models support additional request parameters that control reasoning behavior:

• reasoningEffort: "low" | "medium" | "high"
• reasoningSummary: "auto" | "detailed"
• textVerbosity: "low" | "medium" | "high"
• include: ["reasoning.encrypted_content"]
• store: boolean

Currently, these cannot be configured per model mapping. The actual OpenAI models are just:

• gpt-5.1
• gpt-5.1-codex
• gpt-5.1-codex-mini

But with different parameters, they behave as distinct models (e.g., gpt-5.1-low, gpt-5.1-medium, gpt-5.1-high).

Proposed Solution

Extend model mapping configuration to support optional parameters:

[[models]]
name = "gpt-5.1-high"

[[models.mappings]]
actual_model = "gpt-5.1"
priority = 1
provider = "openai-codex"
options = { reasoningEffort = "high", reasoningSummary = "detailed", textVerbosity = "high", include = ["reasoning.encrypted_content"], store = false }

[[models]]
name = "gpt-5.1-low"

[[models.mappings]]
actual_model = "gpt-5.1"
priority = 1
provider = "openai-codex"
options = { reasoningEffort = "low", reasoningSummary = "auto", textVerbosity = "medium", include = ["reasoning.encrypted_content"], store = false }

Real-World Example

OpenCode uses this pattern:

"gpt-5.1-codex-high": {
  "name": "GPT 5.1 Codex High (OAuth)",
  "limit": {
    "context": 272000,
    "output": 128000
  },
  "options": {
    "reasoningEffort": "high",
    "reasoningSummary": "detailed",
    "textVerbosity": "medium",
    "include": ["reasoning.encrypted_content"],
    "store": false
  }
}

Implementation Notes

Would require:

1. Add options: Option<HashMap<String, serde_json::Value>> to ModelMapping struct
2. Extend provider request structs (e.g., OpenAIRequest) with optional parameter fields
3. Merge model-specific options into request body when constructing provider requests
4. Update config parsing to handle nested options in TOML

References

• OpenAI Responses API: https://platform.openai.com/docs/api-reference/chat/create
• Similar feature in other proxies: LiteLLM supports per-model parameters via config

[9j]

Brilliant! That’s a fantastic idea. I think this will allow for much deeper customization. I’ll do some codebase optimizations and give it a try. Thank you!