Issue:
- In Bash version, new_install_manifest creates file immediately
- save_install_manifest calls remove_old_manifests_for_path
- This deletes ALL manifests for the path, including the new one
- Result: "WARNING: Failed to save installation manifest"
Solution:
- Add current_manifest_file parameter to remove_old_manifests_for_path
- Skip deletion if file matches current manifest
- Pass manifest_file to exclude it from deletion
Note: PowerShell version does not have this issue because it creates
the manifest file AFTER deleting old ones (hashtable → file conversion)
- Add Remove-OldManifestsForPath function to automatically clean old manifests
- Modify Save-InstallManifest to remove old manifests before saving new one
- Update Get-AllInstallManifests to return only latest manifest per installation path
- Apply same strategy to both Install-Claude.ps1 (PowerShell) and Install-Claude.sh (Bash)
Benefits:
- Each installation location registers only once
- Only latest version manifest is retained
- Uninstall UI shows only latest version per location
- Prevents manifest file accumulation
Restructure lite-plan.md (844→668 lines, -176) and lite-execute.md (597→569 lines, -28) following agent document patterns. Move Data Structures sections to end as reference, simplify repeated content, improve hierarchical organization. All original content preserved.
Key changes:
- Data Structures moved to end (from beginning)
- Simplified Execution Process to avoid duplication
- Improved section hierarchy and flow
- Consistent structure across both documents
Restructure cli-lite-planning-agent.md and cli-planning-agent.md following action-planning-agent.md's clear hierarchical pattern. Merge duplicate content, consolidate sections, and improve readability while preserving all original information.
Issue found by Gemini analysis:
- executionContext.planObject.tasks was incorrectly documented as string[]
- Actual runtime structure is array of structured task objects (7 fields)
- This caused documentation mismatch between lite-plan and lite-execute
Fix:
- Update executionContext definition to show full task object structure
- Add comment clarifying it's an array of structured objects (7 fields each)
- Aligns with actual implementation and lite-execute consumption logic
Impact: Documentation only (no code changes, runtime behavior was already correct)
Major changes:
- Add cli-lite-planning-agent.md for generating structured task objects
- Upgrade planObject.tasks from string[] to structured objects with 7 fields:
- title, file, action, description (what to do)
- implementation (3-7 steps on how to do it)
- reference (pattern, files, examples to follow)
- acceptance (verification criteria)
- Update lite-execute.md to format structured tasks for Agent/Codex execution
- Clean up agent files: remove "how to call me" sections (cli-planning-agent, cli-explore-agent)
- Update lite-plan.md to use cli-lite-planning-agent for Medium/High complexity tasks
Benefits:
- Execution agents receive complete "how to do" guidance instead of vague descriptions
- Each task includes specific file paths, implementation steps, and verification criteria
- Clear separation of concerns: agents only describe what they do, not how they are called
- Architecture validated by Gemini: 100% consistency, no responsibility leakage
Breaking changes: None (backward compatible via task.title || task fallback in lite-execute)
Changes:
1. Initialize previousExecutionResults array in Step 5.1
2. Add id field to executionCalls objects (format: [Method-Index])
3. Add execution loop structure in Step 5.2 showing sequential processing
4. Clarify executionId comes from executionCalls[currentIndex].id
5. Add comments explaining ID storage for result collection
Benefits:
- Clear definition of where executionId comes from
- Explicit initialization of tracking variables
- Better understanding of execution flow and result collection
- Proper context continuity across multiple execution calls
Improvements:
1. Add plan summary in confirmation question for quick review
2. Add previousExecutionResults tracking for multi-execution flows
3. Include execution result collection mechanism after each call
4. Update both Agent and Codex execution prompts with context continuity
Benefits:
- Subsequent executions can see what previous calls completed
- Avoid duplicate work across multiple execution calls
- Better dependency management and task flow
- Clear context propagation: executionId, status, tasks, outputs, notes
Change Phase 4 confirmation flow from single-step to two-step process:
- Step 4.1: Display complete plan as readable text output
- Step 4.2: Collect three-dimensional input via AskUserQuestion
Benefits:
- Clearer plan presentation (not embedded in question)
- Simpler question interface focused on user decisions
- Better user experience with logical separation
- Deleted the `list` command for design runs.
- Removed the `update` command and its associated documentation.
- Introduced `design-sync` command to synchronize finalized design system references to brainstorming artifacts.
- Updated command references in `COMMAND_REFERENCE.md`, `GETTING_STARTED.md`, and `GETTING_STARTED_CN.md` to reflect the new command structure.
- Ensured all related documentation and output styles are consistent with the new command naming and functionality.
Changes:
- Switch from full directory replacement to incremental merge for all directories (.claude, .codex, .gemini, .qwen)
- Preserve user's custom files and folders during installation
- Add priority backup for critical config files:
- .codex/AGENTS.md
- .gemini/GEMINI.md, CLAUDE.md
- .qwen/QWEN.md
- Only overwrite files present in installation package
- Auto-cleanup empty backup folders
- Update both PowerShell and Bash installers with consistent behavior
Benefits:
- Non-destructive updates that preserve user customizations
- Safer upgrade path with automatic backup
- Better user experience during reinstallation
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
