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

Project	Deployment	Review	Updated (UTC)
docs	Ready	Preview, Comment	Dec 22, 2025 10:57pm

Greptile Summary

This PR refactors memory management, improves provider type safety, and consolidates cost calculation logic. The changes address three key areas:

Memory Management Improvements

• Changed memory scoping from workflowId to workspaceId for conversation-level persistence
• Wrapped streaming responses to persist assistant messages after stream completion
• Used atomic PostgreSQL array concatenation (||) for race-safe appends
• Migration backfills workspaceId from workflows and deletes orphaned records

Provider Type Safety

• Added explicit TypeScript types from official SDKs (ChatCompletionCreateParamsStreaming, RawMessageStreamEvent, etc.)
• Removed fallback code for legacy response formats (Google's content.function_call)
• Enhanced streaming utilities to track token usage from provider events

Cost Calculation Consolidation

• Moved cost calculation from logger to provider layer for consistency
• Cost now calculated in real-time during streaming using onComplete callbacks
• Both streaming and non-streaming paths now compute costs at the same level

Issues Found

• Memory seeding in agent-handler.ts:714 has race condition potential when concurrent requests use the same conversationId
• Sliding window memory modes fetch-modify-write, which could lose messages during concurrent appends
• Cost tracking happens in streaming callback before response completes - interrupted streams may record costs for incomplete responses

Confidence Score: 4/5

• Safe to merge with minor race condition risks in high-concurrency scenarios
• The PR implements solid architectural improvements with proper type safety and atomic database operations. However, memory seeding and sliding window modes have race condition vulnerabilities when multiple requests use the same conversationId concurrently. These are edge cases that may not surface in typical usage but could cause message loss or duplication under load. The provider type safety improvements and cost calculation consolidation are well-executed.
• apps/sim/executor/handlers/agent/agent-handler.ts (seeding race condition) and apps/sim/executor/handlers/agent/memory.ts (sliding window race condition)

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Client
    participant AgentHandler
    participant Provider
    participant MemoryService
    participant Database

    Client->>AgentHandler: execute(block, inputs)
    AgentHandler->>AgentHandler: buildMessages()
    
    alt Memory Enabled
        AgentHandler->>MemoryService: fetchMemoryMessages()
        MemoryService->>Database: SELECT data FROM memory
        Database-->>MemoryService: existing messages
        
        alt First Run (no existing)
            AgentHandler->>MemoryService: seedMemory(conversationMessages)
            MemoryService->>Database: INSERT memory record
        else Existing Memory
            alt New User Message
                AgentHandler->>MemoryService: appendToMemory(userMessage)
                MemoryService->>Database: UPDATE with array concat
            end
        end
    end

    AgentHandler->>Provider: executeProviderRequest()
    
    alt Streaming Response
        Provider-->>AgentHandler: StreamingExecution
        
        alt Memory Enabled
            AgentHandler->>MemoryService: wrapStreamForPersistence()
            MemoryService-->>AgentHandler: wrapped stream
        end
        
        AgentHandler-->>Client: stream chunks
        
        Note over MemoryService: Stream accumulates content
        
        MemoryService->>MemoryService: flush() callback
        MemoryService->>Database: UPDATE memory (append assistant message)
        MemoryService->>Provider: calculate cost from usage
        
    else Non-Streaming Response
        Provider-->>AgentHandler: BlockOutput
        Provider->>Provider: calculate cost from tokens
        
        alt Memory Enabled
            AgentHandler->>MemoryService: persistResponseToMemory()
            MemoryService->>Database: UPDATE memory (append assistant)
        end
        
        AgentHandler-->>Client: complete response
    end

@greptile

@greptile

@greptile

@greptile