Changes:
- Removed model="haiku" from @test-fix-agent parallel invocation
- Use default model selection for flexibility
- Updated comment to remove model reference
Rationale:
- Let system auto-select appropriate model
- Avoid hardcoded configuration
- Maintain flexibility for future model changes
Based on Gemini architectural analysis (verdict: "File necessary but simplify"):
Removed Redundant Fields:
- session_id → Derive from file path or workflow-session.json
- current_iteration → Calculate as iterations.length + 1
- max_iterations → Read from task.meta.max_iterations (single source)
- stuck_tests → Calculate by analyzing iterations[].failed_tests history
- regression_detected → Calculate from consecutive pass_rate values
Retained Essential Fields:
- current_task: Active task pointer (Resume critical)
- selected_strategy: Current iteration strategy (runtime state)
- next_action: State machine next step (Resume critical)
- iterations[]: Historical log (source of truth, cannot rebuild)
Enhanced iterations[] Array:
- Added failed_tests[] to each entry (was stuck_tests at top-level)
- Removed redundant regression_detected (calculated from pass_rate)
Benefits:
- Single source of truth (no data duplication)
- Less state to maintain and synchronize
- Clearer data ownership and calculation points
- Adheres to data normalization principles
Documentation Updates:
- Added "Purpose" and "Field Descriptions" sections
- Added "Removed Redundant Fields" with derivation logic
- Updated orchestrator responsibilities (Runtime Calculations)
- Updated agent template (Orchestrator-Calculated metadata)
Impact: File remains essential for Resume capability, but now leaner and normalized.
Enhancement:
- Add dynamic prompt template for @test-fix-agent
- Supports 3 task types: test-gen, test-fix, test-fix-iteration
- Auto-adapts based on task.meta.type via configuration maps
Template Features:
- taskTypeObjective: Task-specific goals
- taskTypeSpecificReads: Required file reads per type
- taskTypeGuidance: Execution instructions per type
- taskTypeDeliverables: Expected outputs per type
- taskTypeSuccessCriteria: Completion validation per type
Benefits:
- Orchestrator uses single template for all task types
- Claude dynamically fills in type-specific details
- Reduces duplication, maintains consistency
- Clear guidance for criticality assessment (high/medium/low)
- Progressive testing integration (affected_only vs full_suite)
Lines: +86 (429 → 515, still -35% from original 791)
Structure Changes:
- Quick Start → What & Why → How It Works → Enhanced Features → Reference
- User-facing content first, technical details later
- 540 → 429 lines (-21%, total -46% from original 791)
Improvements:
- Usage examples at top (immediate access)
- Simplified "How It Works" section (tree diagram)
- Agent roles as table (quick reference)
- Technical details moved to Reference section
- All core features preserved
Benefits:
- New users find usage commands immediately
- Conceptual understanding before technical details
- Reference section for implementation details
- Maintains all functionality and enhancements
- Implemented the `docs-related-cli` command for context-aware documentation generation and update for changed modules using CLI execution with tool fallback.
- Introduced the `lite-fix` command for lightweight bug diagnosis and fix workflow, featuring intelligent severity assessment and optional hotfix mode for production incidents.
- Update action-planning-agent and task-generate-agent to load context
via file paths instead of embedded context packages
- Streamline test-cycle-execute workflow execution
- Remove redundant content from conflict-resolution and context-gather
- Update SKILL.md and tdd-plan documentation
- Add English README.md based on Chinese version
- Update version badges to v5.9.2 in both README files
- Add Lite-Fix workflow documentation and examples
- Document session artifact structure and management
- Include lite-fix quick start examples (standard and hotfix modes)
- Align with latest lite-fix session management features
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add CLI mode example to implementation_approach showing optional command field
- Document two execution modes:
1. Default Mode (agent execution): No command field, agent interprets steps
2. CLI Mode (command execution): With command field, uses CLI tools (codex/gemini/qwen)
- Add Implementation Approach Execution Modes section with:
- Mode descriptions and use cases
- Required fields for each mode
- Command patterns for CLI mode (codex, gemini)
- Mode selection strategy
- Simplify implementation_approach examples to use generic placeholders
- Emphasize command field is optional, agent decides based on task complexity
Benefits:
- Clear documentation of both execution patterns
- Agents understand when to use CLI tools vs direct execution
- Pattern examples show structure without over-specifying content
- Supports both autonomous agent work and CLI tool delegation
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Remove context_signature field from meta object
- Simplify meta object to focus on essential fields
- Keep execution_group for parallelization support
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add explicit note in Phase 1 validation to prevent main Claude from
incorrectly looking for manifest.json in .workflow/active/WFS-xxx/.
manifest.json only exists in .workflow/archives/ (for archived sessions).
Active sessions only have workflow-session.json (metadata).
This prevents confusion during plan execution before agent invocation.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add note in Phase 2 to clarify that historical archive analysis
(querying .workflow/archives/manifest.json) is optional, addressing
discrepancy between context-gather.md (4 tracks) and agent implementation
(3 tracks).
This prevents confusion about manifest.json location and makes the
optional nature of archive querying explicit.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add Phase 6 to both synthesis.md and design-sync.md to automatically
update context-package.json when artifacts are modified, preventing
stale cache issues when transitioning to /workflow:plan.
Changes:
- synthesis.md: Add Phase 6 to sync updated role analyses to context-package
- design-sync.md: Add Phase 5 to sync design system refs + role analyses
This ensures:
- synthesis → plan: context-package reflects synthesis enhancements
- ui-design → plan: context-package includes design system references
- No manual context-gather re-run needed after artifact updates
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Implemented `/memory:docs-full-cli` for comprehensive project documentation generation using CLI execution with batched agents and tool fallback.
- Introduced `/memory:docs-related-cli` to generate/update documentation for git-changed modules, optimizing for efficiency with direct parallel execution for fewer modules and agent batch processing for larger sets.
- Defined execution flows, strategies, and error handling for both commands, ensuring robust documentation processes.
Generated system design and usage examples documentation based on project
context and provided templates.
ARCHITECTURE.md provides an overview of the system's architecture,
including core principles, structure, module map, interactions, design patterns,
API overview, data flow, security, and scalability.
EXAMPLES.md offers end-to-end usage examples, covering quick start, core use cases,
advanced scenarios, testing examples, and best practices.
Explicitly set run_in_background: false for all Bash commands in update-full and update-related workflows to prevent unwanted background execution. This ensures synchronous execution without requiring BashOutput polling.
Changes:
- Phase 1-4: Add run_in_background: false to all Bash calls
- Simplify code examples and remove redundant explanations
- Update both direct execution and agent batch execution patterns
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Added optional quiet mode for backup notifications to reduce output clutter.
- Improved file merging with progress reporting and summary of backed up files.
- Implemented a cleanup function to move old installations to a backup directory before new installations.
- Enhanced manifest handling to distinguish between Global and Path installations.
- Updated uninstallation process to preserve global files if a Global installation exists.
- Improved error handling and user feedback during file operations.
- Create dedicated session folder (.workflow/.lite-plan/{task-slug}-{timestamp}/)
for each lite-plan execution to organize all planning artifacts
- Always export task.json (removed optional export question from Phase 4)
- Save exploration.json, plan.json, and task.json to session folder
- Add session artifact paths to executionContext for lite-execute delegation
- Update lite-execute to use artifact file paths for CLI/agent context
- Enable CLI tools (Gemini/Qwen/Codex) and agents to access detailed
planning context via file references
Benefits:
- Clean separation between different task executions
- All artifacts automatically saved for reusability
- Enhanced context available for execution phase
- Natural audit trail of planning sessions
