subagent_middleware

SubAgent Middleware with Session Access Example

This example demonstrates how SubAgent middleware can access session history for post-execution processing.

Features Demonstrated

Pre-execution Processing

• Injecting context information (timestamps, user data, environment info)
• Modifying arguments before template rendering

Post-execution Processing

• Accessing session history via result.Session
• Extracting metrics (message count, execution duration, token usage)
• Memory extraction from conversation history
• Modifying result data based on session analysis

Prerequisites

Set your OpenAI API key:

export OPENAI_API_KEY="your-api-key-here"

Running the Example

go run main.go

Code Structure

The middleware function demonstrates both phases:

func createMiddleware() func(gollem.SubAgentHandler) gollem.SubAgentHandler {
    return func(next gollem.SubAgentHandler) gollem.SubAgentHandler {
        return func(ctx context.Context, args map[string]any) (gollem.SubAgentResult, error) {
            // Pre-execution: Inject context and start timer
            startTime := time.Now()
            args["_execution_time"] = startTime.Format(time.RFC3339)
            args["_user_context"] = "Example user"

            // Execute
            result, err := next(ctx, args)
            if err != nil {
                return gollem.SubAgentResult{}, err
            }

            // Post-execution: Access session and extract insights
            executionDuration := time.Since(startTime)
            history, _ := result.Session.History()
            if history != nil {
                result.Data["metrics"] = map[string]any{
                    "message_count":      len(history.Messages),
                    "execution_duration": executionDuration.String(),
                }
            }

            return result, nil
        }
    }
}

Use Cases

This pattern is particularly useful for:

1. Memory-aware Agents: Extract and save learned information from conversation history
2. Record Extraction: Parse structured data from session responses
3. Metrics Collection: Track token usage, execution time, and performance
4. Context Injection: Add dynamic context based on user, environment, or system state
5. Audit Logging: Record detailed execution traces for compliance

Key Concepts

• SubAgentResult: Contains both Data (result for parent) and Session (for middleware)
• Session Access: Middleware can call result.Session.History() to access conversation
• Read-only: Session is provided for analysis, not modification
• Error Handling: Session access errors should be handled gracefully (log and continue)

Related Documentation

• See CHANGELOG.md for migration guide from older versions
• See docs/architecture.md for SubAgent architecture details