Release 0.13.0

Added

• GenerateExecutor Enhancements - Advanced Task Tree Generation System

• Dual-mode generation system: single_shot (fast) and multi_phase (high-quality) modes

• Multi-phase CrewAI-based task generation with 4-phase specialization:

• Phase 1: Requirement analysis - understand goals and constraints

• Phase 2: Structure design - design correct task hierarchies with proper parent-child relationships

• Phase 3: Input generation - fill task inputs matching executor schemas exactly

• Phase 4: Review and validation - validate and auto-fix common issues

• Schema formatter module (schema_formatter.py) - provides accurate runtime executor schemas instead of outdated static documentation

• Principles extractor module (principles_extractor.py) - extracts framework principles without code examples to guide LLM generation

• Enhanced validation system addressing three critical issues:

• Schema compliance: validates all task inputs match executor schema definitions (required fields, types, constraints)

• Root task patterns: enforces correct task hierarchy (multi-executor trees must have aggregator root)

• Auto-fix mechanisms: automatically corrects multiple roots, invalid parent IDs, and schema mismatches

• Comprehensive test coverage: 359 tests for GenerateExecutor features, multi-phase crew, schema formatting, and principles extraction

• Documentation: detailed development guide (docs/development/generate-executor-improvements.md) and enhanced examples

• Integration tests for schema-based task tree execution and generation, covering scrape_executor + llm_executor, scrape_executor + crewai_executor, and generate_executor flows (tests/integration/test_schema_based_execution.py).

• Improved demo mode test coverage, including custom demo results, sleep scaling, and fallback logic (tests/core/test_demo_mode.py).

• Enhanced TaskBuilder tests for input/output schema handling and multi-level dependencies (tests/core/test_task_builder.py).

• More comprehensive failure handling tests for TaskManager, including all error types and re-execution (tests/core/execution/test_task_failure_handling.py).

• Tests for hook context modification and persistence in pre_hook and post_hook scenarios (tests/core/storage/test_hook_modify_task_with_context.py).

• Integration tests for TaskExecutor with custom and built-in tools (tests/core/execution/test_task_executor_tools_integration.py).

• New ScrapeExecutor (scrape_executor) for web content extraction and data scraping capabilities

• Extension scanner module (src/apflow/core/extensions/scanner.py) for AST-based scanning of extension files

• Test coverage for extension scanner functionality (tests/core/test_extension_scanner.py)

• New APFLOW_EXTENSIONS_IDS environment variable for fine-grained control over extension ID-based access restrictions

• Documentation section on task data fields to clarify distinction between inputs, params, and result fields

• CrewAI Dependency Data Injection Improvements

• CrewaiExecutor now injects dependency outputs directly into CrewAI task templates and input variables, preventing missing-template-variable errors in complex workflows

• TaskManager executor instantiation path has been extended to pass structured dependency data into CrewAI-based executors for more reliable multi-phase generation

Changed

• GenerateExecutor Improvements

• Refactored task tree generation to use runtime executor schemas instead of static documentation, ensuring schema accuracy

• Moved from single-shot generation to intelligent prompt building with principle extraction and schema formatting

• Enhanced post-processing to ensure correct user_id handling and UUID validation

• Improved error handling and validation with detailed logging for debugging

• Added generation_mode parameter to executor inputs for selecting between fast (single_shot) and high-quality (multi_phase) generation

• Refactored WebSocketExecutor to provide full input/output schema, error handling, and cancellation support (src/apflow/extensions/websocket/websocket_executor.py).

• Refactored and cleaned up JSON-RPC API tests for clarity, conciseness, and modern style (tests/api/a2a/test_http_json_rpc.py).

• Improved dummy executor and related test coverage for TaskBuilder (tests/core/test_task_builder.py).

• Enhanced FailingExecutor and related tests for more robust error simulation (tests/core/execution/test_task_failure_handling.py).

• Improved hook context usage and persistence in hook modification tests (tests/core/storage/test_hook_modify_task_with_context.py).

• Expanded demo mode tests to cover more edge cases and sleep scaling (tests/core/test_demo_mode.py).

• Improved schema-based execution tests for dependency mapping and fallback logic (tests/integration/test_schema_based_execution.py).

• Enhanced extension management system with AST-based scanning and detailed configuration options

• Improved extension loading control with executor ID-based restrictions for enhanced security

• Enhanced documentation for CLI command extensions, overrides, and library usage patterns

• Improved CrewAI-based generation dependency handling in TaskManager and CrewaiExecutor, ensuring dependency data is consistently available for CrewAI tasks and multi-phase generation flows

• Enhanced GenerateExecutor and SchemaFormatter with executor exclusion support and lazy loading of multi-phase CrewAI components, reducing overhead when advanced generation features are not used

Fixed

• Fixed LLM-generated task trees not respecting executor schemas - now validates all inputs match schema definitions
• Fixed poor task tree structure issues - ensures correct root task patterns for multi-executor scenarios
• Fixed outdated documentation in prompts - now uses runtime schemas instead of static markdown docs
• Fixed single-shot LLM quality inconsistencies - added multi-phase mode with specialized crews for complex requirements
• Task failure handling and demo mode functionality improvements with enhanced error tracking and recovery mechanisms
• Improved ExtensionScanner path handling for nested project layouts and ensured multi-phase CrewAI tests are skipped cleanly when CrewAI is not installed, including corrected import paths for MultiPhaseGenerationCrew