Implement Refactor

This command continues from /plugin-creator:assessor. After planning completes, use this to execute the refactoring tasks.

<refactor_input> $ARGUMENTS </refactor_input>

---

Resolve Task File

If $1 is:

• A .md path: Use directly
• A slug (e.g., python3-development): GLOB for .claude/plan/tasks-refactor-{slug}.md

---

Load and Analyze

1. Read Task File

READ the task file completely. Extract:

• All tasks with ID, name, status, dependencies, priority, Agent
• Design spec path (from header or linked file)
• Parallelization information

IMPORTANT: Each task specifies its assigned Agent field. Use this to route to the correct specialized agent during execution.

2. Read Design Spec

The task file links to its design spec (e.g., refactor-design-{slug}.md). READ it to understand the overall refactoring plan.

3. Build Dependency Graph

Identify:

• Ready tasks: Status ❌ NOT STARTED with all dependencies ✅ COMPLETE or "None"
• Blocked tasks: Have incomplete dependencies
• Parallel groups: Tasks that can run together (from "Can Parallelize With" field)

4. Create Progress Todos

TodoWrite(todos=[
    {"content": "Task {ID}: {Name}", "status": "pending", "activeForm": "Implementing {Name}"},
    ... for each task ...
    {"content": "Final: Verify refactoring complete", "status": "pending", "activeForm": "Verifying completion"}
])

---

Execute Tasks

Agent Routing Strategy

Route each task to the appropriate specialized agent based on the Agent field in the task:

Issue Type	Agent	When to Use
SKILL_SPLIT	plugin-creator:refactor-skill	Tasks splitting large skills into smaller focused ones
AGENT_OPTIMIZE	subagent-refactorer	Tasks improving agent prompts and descriptions
DOC_IMPROVE	claude-context-optimizer	Tasks improving skill/agent documentation quality
ORPHAN_RESOLVE	claude-context-optimizer	Tasks integrating orphaned reference files
STRUCTURE_FIX	claude-context-optimizer	Tasks fixing broken links or structural issues
Validation	plugin-assessor	Post-refactoring validation tasks
Documentation	plugin-docs-writer	README and documentation generation tasks

Launch Strategy

For each ready task, READ the Agent field from the task and launch that agent:

Task(
    subagent_type="{task.agent}",  # From task's **Agent** field
    description="Task {ID}: {Name}",
    prompt="/start-refactor-task {task_file_path} --task {task_id}"
)

Parallel execution: If multiple tasks can parallelize, launch them in a SINGLE message with multiple Task calls.

Example parallel launch:

# Launch skill split tasks in parallel (no shared files)
Task(
    subagent_type="plugin-creator:refactor-skill",
    description="Task 1: Split python3 core skill",
    prompt="/start-refactor-task .claude/plan/tasks-refactor-python3-development.md --task 1"
)
Task(
    subagent_type="subagent-refactorer",
    description="Task 2: Optimize python-cli-architect agent",
    prompt="/start-refactor-task .claude/plan/tasks-refactor-python3-development.md --task 2"
)

On Task Completion

When a sub-agent completes:

1. Verify task status changed to ✅ COMPLETE in task file
2. Mark TodoWrite item as completed
3. Recalculate ready tasks (dependencies may be satisfied now)
4. Launch newly-ready tasks

Progress Loop

WHILE tasks remain incomplete:
    ready = tasks where status=❌ and dependencies satisfied
    IF ready is empty AND incomplete tasks exist:
        → Deadlock. Report blocked tasks and their dependencies.

    parallel_groups = group ready tasks by "Can Parallelize With"
    FOR each group:
        Launch all tasks in group (single message if multiple)
        Wait for completion
        Update status

---

Plugin Validation Requirements

After completing refactoring tasks, you MUST validate:

1. Plugin.json Schema Validation

Validate against authoritative plugin.json schema from claude-plugins-reference-2026:

Required validation steps:

# Validate plugin structure
claude plugin validate {plugin-directory}

Common plugin.json issues after refactoring:

Issue	Cause	Fix
agents: Invalid input	Used "./agents/" directory string instead of array	Change to array of file paths: ["./agents/file1.md", "./agents/file2.md"]
name: Required	Missing required name field	Add "name": "plugin-name" in kebab-case
Invalid path format	Absolute paths or missing ./ prefix	All paths must be relative and start with ./
Referenced file doesn't exist	Path in plugin.json points to moved/deleted file	Update paths to match new file locations after refactoring

SOURCE: Lines 25-92 of claude-plugins-reference-2026/SKILL.md

2. Hook Configuration Validation

If plugin includes hooks, validate hook configuration:

Hook validation checklist:

• Hook config file exists at path specified in plugin.json
• Hook matchers reference valid tool names (Read, Write, Edit, etc.)
• Hook script paths use ${CLAUDE_PLUGIN_ROOT} variable
• Hook scripts are executable (chmod +x script.sh)
• Hook event types are valid (PreToolUse, PostToolUse, SessionStart, etc.)

Valid hook events:

• PreToolUse, PostToolUse, PostToolUseFailure
• PermissionRequest, UserPromptSubmit, Notification
• Stop, SubagentStart, SubagentStop
• Setup, SessionStart, SessionEnd, PreCompact

SOURCE: Lines 186-227 of claude-plugins-reference-2026/SKILL.md

3. MCP Server Validation

If plugin bundles MCP servers, validate MCP configuration:

MCP validation checklist:

• MCP config file exists (.mcp.json or inline in plugin.json)
• Server commands use ${CLAUDE_PLUGIN_ROOT} for plugin-relative paths
• Server binaries are executable or installed as dependencies
• Server args arrays are properly formatted
• Environment variables are properly defined

SOURCE: Lines 235-270 of claude-plugins-reference-2026/SKILL.md

4. LSP Server Validation

If plugin provides LSP servers, validate LSP configuration:

LSP validation checklist:

• LSP config file exists (.lsp.json or inline in plugin.json)
• LSP server binary is documented as separate installation requirement
• extensionToLanguage mapping is defined for all supported file types
• command references binary in PATH or uses absolute path with ${CLAUDE_PLUGIN_ROOT}

IMPORTANT: LSP servers require separate binary installation. Plugin only configures connection, doesn't bundle the server.

Example LSP validation error:

LSP server 'gopls' not found in $PATH
→ User must install separately: go install golang.org/x/tools/gopls@latest

SOURCE: Lines 271-338 of claude-plugins-reference-2026/SKILL.md

5. Plugin Caching Path Resolution

CRITICAL: Plugins are copied to cache directory during installation. Validate path resolution:

Path resolution warnings to check:

• No ../ parent directory references (will break after caching)
• All paths relative to plugin root with ./ prefix
• External dependencies documented (symlinks or restructure required)
• ${CLAUDE_PLUGIN_ROOT} used in all hook/MCP/LSP commands

Common caching issues:

Issue	Problem	Solution
../shared-utils reference	Parent directory not copied to cache	Use symlink or restructure marketplace to include shared dir
Hook script uses relative path without ./	Ambiguous path resolution	Change to ./scripts/hook.sh or use ${CLAUDE_PLUGIN_ROOT}
MCP server references user home directory	Won't work for other users	Use plugin-relative paths or environment variables

SOURCE: Lines 349-398 of claude-plugins-reference-2026/SKILL.md

6. Agent Dependencies

Agents included in plugin-creator:

• subagent-refactorer - Used for AGENT_OPTIMIZE tasks (✅ included)
• claude-context-optimizer - Used for DOC_IMPROVE and ORPHAN_RESOLVE tasks (✅ included)
• plugin-assessor - Used for validation tasks (✅ included)

Known external agent dependencies:

• plugin-docs-writer - Used for documentation generation (not in plugin-creator)

Action if external agent missing:

1. Check if agent exists in user's ~/.claude/agents/ or project .claude/agents/
2. If missing, create follow-up task to install required agent plugin
3. OR modify task routing to use included agents only

Completion and Verification Loop

When all tasks show ✅ COMPLETE:

Invoke Complete Refactor

AUTOMATICALLY invoke the complete-refactor command to trigger verification:

Skill(skill="complete-refactor", args="{task_file_path}")

This runs 4 phases:

1. Plugin Validation - Re-assess plugin structure, verify improvements
2. Code Review - Validates refactored code against project standards
3. Documentation Audit - Checks for documentation drift
4. Gap Identification - Creates follow-up tasks if issues found

Check for Follow-up Tasks

After complete-refactor finishes, CHECK if follow-up tasks were created:

GLOB for: .claude/plan/tasks-refactor-{plugin-slug}-followup*.md

IF follow-up tasks exist:

1. DISPLAY:

================================================================================
                    FOLLOW-UP TASKS IDENTIFIED
================================================================================

The review found issues that need resolution:
- {list of follow-up task files}

Continuing recursive refactoring...
================================================================================

2. RECURSIVELY call implement-refactor on each follow-up task:

Skill(skill="implement-refactor", args="{followup_task_file_path}")

3. REPEAT until no more follow-up tasks are generated

IF no follow-up tasks:

1. UPDATE REFACTOR-PLAN.md: Move entry from Active to Completed with scores
2. DISPLAY final summary:

================================================================================
                    PLUGIN REFACTORING COMPLETE
================================================================================

Plugin: {plugin_name}
Task File: {task_file_path}

COMPLETED TASKS:
✅ Task {ID}: {Name}
✅ Task {ID}: {Name}
...

VERIFICATION PASSED:
✅ Plugin Validation: Score improved from X to Y
✅ Code Review: No issues found
✅ Documentation: Synced with implementation

All quality gates passed. Plugin refactoring is complete.
================================================================================

---

Recursive Development Cycle

This command implements a recursive refactoring loop:

┌─────────────────────────────────────────────────────────────────┐
│                    IMPLEMENT-REFACTOR                            │
│                                                                  │
│  ┌──────────────┐    ┌───────────────────┐    ┌──────────────┐ │
│  │ Execute      │───▶│ Complete          │───▶│ Follow-up    │ │
│  │ All Tasks    │    │ Refactor          │    │ Tasks?       │ │
│  └──────────────┘    └───────────────────┘    └──────┬───────┘ │
│                                                       │         │
│                              ┌────────────────────────┴───┐     │
│                              │                            │     │
│                              ▼                            ▼     │
│                         ┌────────┐                  ┌─────────┐ │
│                         │ YES    │                  │ NO      │ │
│                         └───┬────┘                  └────┬────┘ │
│                             │                            │      │
│                             ▼                            ▼      │
│                    ┌─────────────────┐          ┌────────────┐  │
│                    │ Recurse:        │          │ DONE       │  │
│                    │ implement-      │          │ Plugin     │  │
│                    │ refactor on     │          │ Refactored │  │
│                    │ follow-up tasks │          └────────────┘  │
│                    └────────┬────────┘                          │
│                             │                                   │
│                             └───────────────────────────────────┘
└─────────────────────────────────────────────────────────────────┘

The cycle continues until the review finds no more issues.

---

Error Handling

Task Failure

If a sub-agent reports failure:

1. Keep task as 🔄 IN PROGRESS
2. Display error details
3. Ask user: "(1) Retry, (2) Skip, (3) Abort"

Dependency Deadlock

If no tasks are ready but tasks remain:

1. Display blocked tasks and their unmet dependencies
2. Ask user to resolve manually

Design Conflicts

If sub-agent reports design spec conflicts with actual skill structure:

1. STOP implementation
2. Report the conflict
3. The design spec may need revision before continuing

---

Orchestrator Responsibilities

You coordinate. Sub-agents implement.

• You read task files and identify what's ready
• You launch sub-agents with /start-refactor-task
• You track overall progress with TodoWrite
• Sub-agents do the actual refactoring work
• Sub-agents run verification steps
• Sub-agents report completion or blocking issues

If a sub-agent is blocked by concurrent edits from another agent, that's expected in parallel execution. Help them understand the changes and continue.