v2.69.1
20 days ago by Craig Adam
New Features
- Payload Scope Management: AI Agents can now operate on specific portions of parent payloads, enabling focused sub-agent operations with automatic path scoping
- Final Payload Validation: New validation system for AI Agent outputs with configurable retry/fail/warn modes, similar to OutputExample in AI Prompts
- Execution Limits: Min/max execution constraints for both actions and sub-agents within agent runs, providing better control over agent behavior
- Resource Guardrails: Comprehensive limits for agent runs including cost ($), tokens, iterations, and time constraints with automatic termination on breach
- JSON Validation Framework: New lightweight JSON validator with flexible validation rules supporting optional fields, wildcards, array constraints, and type checking
- Validation Syntax Cleaner: Utility to remove validation syntax from JSON responses when AI models mistakenly include validation markers in their output
Improvements
- Enhanced Agent Execution Tracking: Added TotalPromptIterations counter to AI Agent Runs for better performance monitoring
- Cerebras Model Configuration: Fixed Qwen 3 235B model configuration with correct API name
- Documentation Updates: Added comprehensive documentation for the new cleanValidationSyntax method in MJGlobal
- CodeGen Enhancements: Automatic generation of TypeScript validation functions for new CHECK constraints
Bug Fixes
- Model Token Limits: Corrected Cerebras Qwen 3 235B model to use actual provider limits rather than theoretical capabilities
- Migration Ordering: Fixed migration file sequence to prevent order of operations issues during schema updates
- Validation Mode Defaults: Ensured FinalPayloadValidationMode properly defaults to 'Retry' when not specified