[BUG] Accurately display reasoning options for OpenRouter models based on their actual capabilities

[hannesrudolph]

Problem (one or two sentences)

OpenRouter models have different reasoning capabilities (binary on/off, effort-based levels, or budget-based tokens), but the current implementation treats all models with "reasoning" in supported_parameters the same way. This causes incorrect UI controls to be displayed - some models show effort level dropdowns when they only support on/off, while others may need budget sliders instead.

Context (who is affected and when)

Users selecting OpenRouter models see reasoning controls that may not match the model's actual capabilities:

• Binary-only models (like grok-4-fast): Show effort dropdown instead of simple on/off checkbox
• Effort-based models: May not show all supported levels (some support none, minimal, xhigh)
• Budget-based models: Need token budget sliders, not effort dropdowns

This creates confusion when users select a reasoning option that has no effect on the model's behavior.

Desired behavior (conceptual, not technical)

The reasoning settings UI should automatically detect and display the correct control type for each OpenRouter model:

1. Binary reasoning models → Simple on/off checkbox (supportsReasoningBinary)
2. Effort-based models → Dropdown with only the supported effort levels (supportsReasoningEffort as array)
3. Budget-based models → Token budget sliders (supportsReasoningBudget)
4. Required reasoning models → No disable option (requiredReasoningBudget / requiredReasoningEffort)
5. No reasoning support → No reasoning controls shown

Constraints / preferences (optional)

• This supersedes/closes [ENHANCEMENT] Support disabling reasoning in grok-4-fast on OpenRouter. #8709 (grok-4-fast reasoning issue) as a comprehensive fix
• Related issues: [ENHANCEMENT] Add "OFF" option to the reasoning effort dropdown #8624 (Add "OFF" option to dropdown), [BUG] OpenRouter DeepSeek V3.1 Terminus forces reasoning; cannot disable #8270 (DeepSeek V3.1 Terminus reasoning)
• The solution may require:
  • Checking if OpenRouter provides more granular supported_parameters (e.g., "reasoning.effort" vs "reasoning.enabled")
  • Maintaining a list of models with specific reasoning types
  • Using heuristics based on model characteristics

Request checklist

• I've searched existing Issues and Discussions for duplicates
• This describes a specific problem with clear context and impact

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Problem Analysis

The root cause is in src/api/providers/fetchers/openrouter.ts where the OpenRouter model fetcher sets:

supportsReasoningEffort: supportedParameters ? supportedParameters.includes("reasoning") : undefined,

This sets supportsReasoningEffort: true (boolean) when a model has "reasoning" in its supported_parameters. However, a boolean true causes the UI to display a dropdown with low/medium/high options for ALL reasoning-capable models, even those like grok-4-fast that only support binary on/off reasoning.

The Codebase Already Supports

1. supportsReasoningBinary: true - Shows a simple on/off checkbox (exactly what binary reasoning models need)
2. supportsReasoningEffort: [...] (array) - Shows only the specified effort levels in a dropdown
3. supportsReasoningBudget: true - Shows a token budget slider

Proposed Solution

1. Add a new constant set in packages/types/src/providers/openrouter.ts for OpenRouter models with binary-only reasoning:

   export const OPEN_ROUTER_BINARY_REASONING_MODELS = new Set([
     "x-ai/grok-4-fast",
     // ... other models with binary reasoning
   ])

2. Add a map for models with specific effort levels (like grok-3-mini which supports only ["low", "high"]):

   export const OPEN_ROUTER_REASONING_EFFORT_LEVELS: Record<string, readonly string[]> = {
     "x-ai/grok-3-mini": ["low", "high"],
     // ... other models with specific levels
   }

3. Update parseOpenRouterModel in src/api/providers/fetchers/openrouter.ts to:

   • Check if the model is in OPEN_ROUTER_BINARY_REASONING_MODELS -> set supportsReasoningBinary: true
   • Check if the model is in OPEN_ROUTER_REASONING_EFFORT_LEVELS -> set supportsReasoningEffort to the specific array
   • Fall back to supportsReasoningEffort: true only for models that actually support effort levels

4. Add tests to verify the correct UI is shown for each reasoning type

Benefits

• Models with binary reasoning will show a simple checkbox (like the existing xAI direct integration)
• Models with specific effort levels will show only those options
• Follows existing patterns in the codebase (similar to OPEN_ROUTER_REASONING_BUDGET_MODELS)
• Maintainable: new models can be easily added to the appropriate set

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.