- Introduced Task Attachment Model to clarify how SlashCommand invocations expand workflows by attaching sub-tasks to the current TodoWrite.
- Added Auto-Continue Mechanism details to explain dynamic task management and continuous execution across phases.
- Updated TodoWrite examples to reflect task attachment and collapsing behavior during workflow execution.
- Improved execution flow diagrams for both test-fix-gen and test-gen workflows to illustrate task attachment and phase completion.
- Emphasized critical aspects of continuous execution and task management in the documentation.
- fix(discover-design-files.sh): Fix script errors and expand framework support
* Fix variable quoting and conditional checks
* Remove keyword restrictions for JS file discovery
* Add support for .mjs, .cjs, .vue, .svelte files
* Now correctly discovers all JS/TS framework files recursively
- feat(import-from-code): Add optional gemini/qwen CLI analysis for agents
* Add CLI-Assisted Analysis step to Style/Animation/Layout agents
* Optional usage for large codebases (>20 files) or complex frameworks
* All CLI calls use analysis mode (READ-ONLY)
* Agent can use CLI output to guide file reading and token extraction
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Updated `explore-auto.md`, `imitate-auto.md`, and `import-from-code.md` to replace legacy parameters with a unified `--input` parameter for better clarity and functionality.
- Enhanced input detection logic to support glob patterns, file paths, and text descriptions, allowing for more flexible input combinations.
- Deprecated old parameters (`--images`, `--prompt`, etc.) with warnings and provided clear migration paths.
- Improved error handling for missing inputs and validation of existing directories in `reference-page-generator.md`.
- Streamlined command execution phases to utilize new input structures across workflows.
Refactor style codification workflow into orchestrator pattern with three specialized commands:
Changes:
- Refactor codify-style.md as pure orchestrator coordinating sub-commands
• Delegates to import-from-code for style extraction
• Calls reference-page-generator for packaging
• Creates temporary design run as intermediate storage
• Added --overwrite flag with package protection logic
• Improved component counting using jq with grep fallback
- Create reference-page-generator.md for multi-component reference pages
• Generates interactive preview with all components
• Creates design tokens, style guide, and component patterns
• Package validation to prevent invalid directory overwrites
• Outputs to .workflow/reference_style/{package-name}/
- Create style-skill-memory.md for SKILL memory generation
• Converts style reference packages to SKILL memory
• Progressive loading levels (0-3) for efficient token usage
• Intelligent description generation from metadata
• --regenerate flag support
Improvements based on Gemini analysis:
- Overwrite protection in codify-style prevents accidental data loss
- Reliable component counting via jq JSON parsing (grep fallback)
- Package validation in reference-page-generator ensures data integrity
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Introduce comprehensive test quality assurance framework to prevent "hollow tests"
from masking real issues. Optimize JSON data structures following minimal-but-sufficient principle.
Major Changes:
- Multi-layered test strategy (L0: Static, L1: Unit, L2: Integration, L3: E2E)
- New quality gate task (IMPL-001.5-review) validates tests before fix cycle
- Layer-aware failure diagnosis with test_type field support
- JSON simplification: removed redundant failure_context (~44% size reduction)
File Changes:
- new: cli-planning-agent.md - CLI analysis executor with layer-specific guidance
- mod: test-fix-gen.md - multi-layered test planning and quality gate generation
- mod: test-fix-agent.md - layer-aware test execution and failure classification
- mod: test-cycle-execute.md - 95% pass rate threshold with criticality assessment
Technical Details:
- test_type field tracks test layer (static/unit/integration/e2e)
- IMPL-fix-N.json simplified: removed 350 lines of redundant data
- Single source of truth: iteration-N-analysis.md contains full context
- Quality config: ~/.claude/workflows/test-quality-config.json (not in repo)
Benefits:
- Prevents symptom-level fixes through layer-specific diagnosis
- Ensures test quality with static analysis and coverage validation
- Reduces JSON file size by 44% while maintaining information completeness
- Enforces comprehensive test coverage (happy path + negative + edge cases)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Updated `layout-extract` command to include a refinement mode (`--refine`) for generating single refined layouts, alongside exploration mode for multiple contrasting variants.
- Added detailed reporting and validation for variants count based on the selected mode.
- Implemented logic to load existing layouts for refinement and generate specific refinement options categorized by density, responsiveness, grid specifics, and component arrangement.
- Updated `style-extract` command similarly to support refinement mode, allowing for detailed adjustments to existing design systems.
- Enhanced user interaction phases to accommodate selection of refinements or design directions based on the active mode.
- **generate.md**: Eliminate ambiguity in agent task allocation
- Add explicit "one agent per style" principle with warnings
- Fix distribution strategy: balanced split (12→6+6) instead of fill-first (12→10+2)
- Add distribution formula and concrete examples table
- Simplify redundant constraints and consolidate documentation
- Remove verbose descriptions per minimal information principle
- **explore-auto.md**: Optimize animation extraction conditional logic
- Add user confirmation for animation reuse in code import scenarios
- Implement should_extract_animation flag with comprehensive conditions
- Add skip_animation_extraction for explicit user preference handling
- Update autonomous flow description with conditional animation phase
- Add animation extraction task to TodoWrite tracking
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Remove ambiguous "wait for completion" expressions that could be misinterpreted as waiting for an external command. Replace with clear execution-blocking language.
Changes:
- "WAIT for completion" → "Execute phase (blocks until finished)"
- "Upon each phase completion" → "When each phase finishes executing"
- "execution pauses until completion" → "execution pauses until the command finishes"
Affected files (6):
- brainstorm/auto-parallel.md: 3 changes
- plan.md: 3 changes
- tdd-plan.md: 1 change
- test-gen.md: 1 change
- ui-design/explore-auto.md: 12 changes
- ui-design/imitate-auto.md: 10 changes
Impact: Clarifies that SlashCommand is blocking execution, not waiting for separate completion task
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Optimized agent grouping to process multiple layouts per agent:
- Changed from one-layout-per-agent to max-10-layouts-per-agent
- Grouped by target × style for better efficiency
- Reduced agent count from O(T×S×L) to O(T×S×⌈L/10⌉)
- Shared token loading within agent (read once, use for all layouts)
- Enforced target isolation (different targets use different agents)
Benefits:
- Reduced agent overhead by ~83% in typical scenarios
- Minimized redundant token reads
- Maintained parallel execution capability (max 6 concurrent agents)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add universal question interaction rules to UI design commands:
- Use batch text format (1a 2b) when questions > 4 OR options > 3
- Otherwise use AskUserQuestion tool
- Support multi-selection: [N][key1,key2] format for layout/style commands
- Variable-based templates with option descriptions for clarity
Updated commands:
- animation-extract.md: Single selection per question
- layout-extract.md: Multi-selection per target support
- style-extract.md: Multi-selection for variants
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Update default interactive_mode from true to false
- Add --interactive parameter documentation
- Non-interactive batch generation is now the default behavior
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add --interactive flag to style-extract and layout-extract commands
- Enhance animation-extract --mode interactive with selection storage
- Implement unified user-selections storage in .intermediates/user-selections/
- Add parameter passthrough in explore-auto (default: enabled) and imitate-auto (always enabled)
- Support fallback to generate all variants when no selection file exists
- Fix explore-auto.md bug: duplicate --base-path in import-from-code call (line 313)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Updated `layout-extract` command to support user multi-selection of layout concepts, allowing for multiple layout templates to be generated based on user preferences.
- Revised argument hints and output structures to reflect changes in selection handling and output file formats.
- Enhanced user interaction by enabling multi-select options in both layout and style extraction phases.
- Streamlined the generation of layout templates and design systems to accommodate multiple selected variants, improving flexibility and user experience.
- Adjusted validation and output verification processes to ensure consistency with new multi-selection features.
- Refactor test-task-generate.md to use action-planning-agent
- Add two-phase execution flow (Discovery → Agent Execution)
- Integrate Memory-First principle and MCP tool enhancements
- Support both agent-mode (default) and cli-execute-mode
- Add test-specific context package with TEST_ANALYSIS_RESULTS.md
- Align with task-generate-agent.md architecture
- Remove 556 lines of redundant old content (Phase 1-4 old structure)
- Update test-gen.md and test-fix-gen.md to reflect agent-driven changes
Changes include:
- Core Philosophy with agent-driven principles
- Agent Context Package structure for test sessions
- Discovery Actions for test context loading
- Agent Invocation with test-specific requirements
- Test Task Structure Reference
- Updated Integration & Usage sections
- Enhanced Related Commands documentation
Now test task generation uses autonomous agent planning with MCP enhancements for better test coverage analysis and generation.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Change --agent flag to --cli-execute for consistency
- Make Agent Mode the default execution mode
- Update CLI Mode to use --cli-execute flag
- Align with agent-driven task generation architecture
- Update Phase 5 command documentation
- Update Related Commands section
This completes the migration to agent-driven architecture for both /workflow:plan and /workflow:tdd-plan commands.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Refactor task-generate-tdd.md to use action-planning-agent
- Add two-phase execution flow (Discovery → Agent Execution)
- Integrate Memory-First principle and MCP tool enhancements
- Support both agent-mode (default) and cli-execute-mode
- Change --agent flag to --cli-execute for consistency
- Add TDD-specific context package with test coverage integration
- Align with task-generate-agent.md architecture
Now both /workflow:plan and /workflow:tdd-plan use agent-driven task generation for autonomous planning and execution.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Remove 150+ lines of duplicate content from task-generate-agent.md
- Implement reference-based design following Content Uniqueness Rules
- Simplify plan.md to use task-generate-agent exclusively
- Remove --agent parameter (agent mode is now default)
- Improve separation of concerns between command and agent layers
Changes:
- task-generate-agent.md: Replace detailed specs with references to action-planning-agent.md
- plan.md: Remove task-generate command, unify on agent-driven approach
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Updated explore-auto.md to include a new phase for intelligent path detection and source selection, allowing the system to identify existing file paths from prompts and images.
- Introduced conditional code import and completeness assessment based on detected design sources (code only, visual only, hybrid).
- Modified phase execution flow to accommodate new checks for style, animation, and layout completeness based on the design source.
- Added error handling for missing design elements and user prompts for visual supplementation.
- Enhanced imitate-auto.md with intelligent path detection and a new phase for code import and completeness assessment, ensuring a hybrid approach when applicable.
- Implemented detailed reporting for each phase, including missing elements and recommendations for improvement.
- Created a comprehensive import-from-code.md file outlining the workflow for importing design systems from code files, detailing the execution process and error handling.
