[BUG] Vercel Gateway is lacking some models features (thinking & image & prompt caching)

[ndeybach]

Problem (one or two sentences)

When using Vercel Gateway you cannot set some models features:

• "thinking" level selector not visible / available (openai 5.2 series & anthropic 4.5 series & gemini 3 series)
• image support not enabled (openai 5.2 series & anthropic 4.5 series & gemini 3 series)
• prompt caching not enabled (openai 5.2 series & gemini 3 series)

Context (who is affected and when)

Any user using Vercel gateway when wanting to use said models.

Reproduction steps

Go to settings and in providers try to add models

Expected result

working said features

Actual result

hardcoded unsupport or unavailability

Variations tried (optional)

No response

App Version

3.41.3 (23717fb)

API Provider (optional)

Not Applicable / Other

Model Used (optional)

provided earlier

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

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

Problem Analysis

The Vercel AI Gateway model parser (src/api/providers/fetchers/vercel-ai-gateway.ts) and its type definitions (packages/types/src/providers/vercel-ai-gateway.ts) are missing support for several model capabilities:

1. Thinking/Reasoning Support: The parseVercelAiGatewayModel function doesn't set supportsReasoningBudget or supportsReasoningEffort flags for models that support extended thinking (like Claude 4.5/Sonnet 4, OpenAI o3/o4-mini, and Gemini models)

2. Image Support: Uses hardcoded static sets (VERCEL_AI_GATEWAY_VISION_ONLY_MODELS and VERCEL_AI_GATEWAY_VISION_AND_TOOLS_MODELS) that are incomplete - missing newer models like OpenAI 5.x series

3. Prompt Caching: The VERCEL_AI_GATEWAY_PROMPT_CACHING_MODELS set exists but isn't used in parseVercelAiGatewayModel - instead it relies solely on API pricing data which may not always be accurate

Proposed Implementation

1. Add reasoning model support to packages/types/src/providers/vercel-ai-gateway.ts:

• Create VERCEL_AI_GATEWAY_REASONING_MODELS set with models supporting reasoning effort (o3, o4-mini, etc.)
• Create VERCEL_AI_GATEWAY_REASONING_BUDGET_MODELS set for models that support reasoning budget (Claude Sonnet 4, etc.)

2. Update the vision model sets in packages/types/src/providers/vercel-ai-gateway.ts:

• Add OpenAI 5.x series models to VERCEL_AI_GATEWAY_VISION_AND_TOOLS_MODELS
• Add any other missing vision-capable models

3. Update parseVercelAiGatewayModel in src/api/providers/fetchers/vercel-ai-gateway.ts:

• Import and use the new reasoning model sets
• Set supportsReasoningBudget for Claude thinking models
• Set supportsReasoningEffort with appropriate effort levels for OpenAI o-series and other reasoning models
• Use VERCEL_AI_GATEWAY_PROMPT_CACHING_MODELS as a fallback if API pricing doesn't indicate cache support

4. Export new types from packages/types:

• Ensure the new sets are properly exported

This approach follows the existing patterns used by other providers (like OpenRouter and OpenAI native) for handling model capabilities.

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