The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Feb 7, 2026 5:00am

Greptile Overview

Greptile Summary

Updated model configurations and improved error handling across multiple AI providers. Key changes include:

Model Configuration Updates

• Simplified maxOutputTokens from object structure to single number across all models
• Updated context windows (o3: 128k→200k, o4-mini: 128k→200k, o1-preview: 400k→128k)
• Changed default thinking levels from 'medium' to 'high' for Claude Sonnet models
• Added 'auto' option for reasoning effort and verbosity parameters
• Added 'none' option for thinking level to allow explicit disabling
• Moved reasoning effort, verbosity, thinking level, temperature, and max tokens to advanced mode in agent block UI
• Updated pricing for various models (Azure o3, Claude 3 Haiku cached input)

Anthropic Provider Improvements

• Added createMessage() helper that automatically uses streaming internally when max_tokens exceeds 21,333 (SDK non-streaming threshold) to prevent HTTP timeouts
• Enhanced thinking configuration with proper constraint handling (budget_tokens must be less than max_tokens, adjusted automatically)
• Added compatibility checks: thinking mode disables temperature and limits tool_choice to 'auto'/'none' only
• Preserved thinking blocks during tool use loops to maintain reasoning continuity
• Improved type safety with proper Anthropic SDK types throughout
• Error handling now propagates errors with enhanced timing metadata

OpenAI/Azure Provider Improvements

• Added checks to skip 'auto' values for reasoningEffort and verbosity to prevent sending invalid parameter values to API
• Improved type safety by replacing any with proper OpenAI SDK types and Record<string, unknown>
• Enhanced schema validation with stricter typing in enforceStrictSchema function

Other Provider Updates

• Bedrock: added support for 'any' tool choice type, included pricing in cost objects
• Gemini: changed to only activate thinking when user explicitly selects level (no longer uses model defaults)

Type Safety & Code Quality

• Replaced numerous any types with proper typed interfaces and Records
• Removed redundant comments and improved code clarity
• Fixed OpenAI model pattern regex from /^o1/ to /^o\d/ to properly match o-series models

Confidence Score: 4/5

• This PR is safe to merge with low risk - primarily configuration updates and type safety improvements
• Score reflects well-structured refactoring with improved type safety and error handling. The changes address real API constraints (Anthropic SDK streaming threshold, thinking mode compatibility rules). Minor concern: temperature silently overridden when thinking enabled without warning user
• Pay attention to apps/sim/providers/anthropic/core.ts for the temperature override behavior when thinking is enabled

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant User
    participant AgentBlock
    participant AgentHandler
    participant ProviderRegistry
    participant AnthropicProvider
    participant OpenAIProvider
    participant ModelConfig

    User->>AgentBlock: Configure agent (model, thinkingLevel, reasoningEffort, etc)
    AgentBlock->>AgentHandler: Execute with inputs
    
    AgentHandler->>AgentHandler: Build provider request
    Note over AgentHandler: Add thinkingLevel parameter
    
    AgentHandler->>ProviderRegistry: Route request to provider
    
    alt Anthropic Provider
        ProviderRegistry->>AnthropicProvider: executeRequest()
        AnthropicProvider->>ModelConfig: getMaxOutputTokensForModel()
        ModelConfig-->>AnthropicProvider: maxTokens (simplified from object to number)
        
        AnthropicProvider->>AnthropicProvider: Build thinking config (if thinkingLevel !== 'none')
        Note over AnthropicProvider: Check compatibility:<br/>- Adjust max_tokens for budget_tokens<br/>- Set temperature=undefined<br/>- Limit tool_choice to 'auto'/'none'
        
        AnthropicProvider->>AnthropicProvider: createMessage() helper
        Note over AnthropicProvider: Auto-switch to streaming<br/>if max_tokens > 21333
        
        AnthropicProvider->>AnthropicProvider: Preserve thinking blocks in tool loops
        AnthropicProvider-->>AgentHandler: Response with enhanced error handling
    else OpenAI/Azure Provider
        ProviderRegistry->>OpenAIProvider: executeRequest()
        OpenAIProvider->>OpenAIProvider: Skip 'auto' for reasoningEffort/verbosity
        Note over OpenAIProvider: Prevent invalid API values
        OpenAIProvider-->>AgentHandler: Response with improved typing
    end
    
    AgentHandler-->>User: Final response with tokens, cost, timing

@greptile

@cursor review