clean

Act on findings from `/moltbloat:audit`. Walk through each finding interactively, explain what it is, and offer to fix it — with explicit user confirmation before any destructive action. Supports dry-run mode for previewing changes without applying them.

<Use_When>

• User wants to clean up their Claude Code ecosystem
• User says "clean", "cleanup", "remove bloat", "fix findings"
• User says "dry run", "preview", "what would clean do"
• After running /moltbloat:audit and seeing findings they want to address </Use_When>

<Do_Not_Use_When>

• User just wants to see what's wrong — use /moltbloat:audit
• User only wants token cost info — use /moltbloat:token-budget </Do_Not_Use_When>

- NEVER delete or modify anything without explicit user confirmation (in interactive mode) - In dry-run mode, NEVER modify anything — only report what would be done - NEVER auto-fix multiple findings at once — present each individually - ALWAYS explain what will change and what the impact is - ALWAYS offer to skip any finding - For plugin uninstalls, prefer `claude plugin remove` over manual deletion - Back up configs before modifying settings files

1. Parse the command

   Check if the user specified --dry-run or dry-run:

   • /moltbloat:clean — interactive mode (default)
   • /moltbloat:clean --dry-run or /moltbloat:clean dry-run — preview mode only

   In dry-run mode, set a flag and announce:

   Running in DRY-RUN mode — no changes will be made. Showing what would be cleaned.

2. Run the audit first

   If no recent audit results are available in the conversation, run the /moltbloat:audit skill first to collect findings. If audit results are already present from an earlier invocation in this session, use those.

3. Present the cleanup menu

   Show the user a numbered list of all findings from the audit, grouped by severity:

   # Moltbloat Cleanup
   
   Found X issues to review. I'll walk through each one — you decide what to fix.
   
   ## CRITICAL (fix these)
   1. [Skills] 30+ skill name collisions between <plugin-a> and <plugin-b>
   2. ...
   
   ## HIGH (act on these)
   3. [Agents] 25 local agents duplicate <plugin> agents
   4. ...
   
   ## MEDIUM (worth considering)
   5. [Plugin] <name> — 0 skills, 0 MCP servers, 0 agents
   6. [Cache] Stale versions consuming X MB
   7. ...
   
   ## LOW (optional cleanup)
   8. [Plugin] <name> — disabled for 60+ days
   9. [Project] Orphaned config — X MB
   10. ...
   

   In interactive mode, prompt:

   Reply with numbers to fix (e.g., "1, 3, 5") or "all high", "all", or "skip".

   In dry-run mode, continue to step 4 with all findings selected automatically for preview.

4. Process each selected finding

   In dry-run mode, for each finding, output:

   ## [DRY-RUN] Finding N: <description>
   
   **Action that would be taken**: <specific command or removal>
   **Estimated token savings**: ~X tokens/session
   **Estimated disk savings**: X MB
   **Risk**: <low/medium/high> — <explanation>
   

   In interactive mode, for each finding the user selects, present the specific action and ask for confirmation.

   For each finding type, use the appropriate template. Replace all placeholders with actual values from the audit.

   Skill collision fix

   ## Finding N: <count> skill collisions between <plugin-a> and <plugin-b>
   
   **What**: Both plugins provide skills with the same names: <list names>
   **Why fix**: Only one version runs — you may not be getting the one you expect.
   **Action**: Uninstall whichever plugin provides less unique value. Run
   `/moltbloat:why <plugin>` on each to compare.
   **Risk**: Losing skills unique to the removed plugin
   
   Which plugin to remove? (<plugin-a> / <plugin-b> / skip)
   

   If user chooses, run the uninstall command. If it fails, report the error and move on.

   Agent dedup

   ## Finding N: Duplicate agent — <name>.md
   
   **What**: `<name>.md` exists in both ~/.claude/agents/ (local) and
   <plugin> (plugin-provided).
   **Comparison**: <read first 5 lines of each and summarize differences>
   **Action**: Remove local ~/.claude/agents/<name>.md
   **Risk**: Low if versions are identical. If the local version has custom
   modifications, those will be lost.
   
   Proceed? (yes/no/skip)
   

   Before removing a local agent, read both versions. If the local version has custom modifications, warn the user and recommend keeping it.

   Zero-skill plugin removal

   ## Finding N: <plugin> — no skills, no MCP, no agents
   
   **What**: Plugin is installed but provides nothing detectable.
   **Action**: Uninstall plugin
   **Command**: `claude plugin remove <plugin>`
   **Risk**: Plugin may provide hooks or rules not detected by this scan
   
   Proceed? (yes/no/skip)
   

   Disabled plugin cleanup

   ## Finding N: <plugin> — disabled for <X> days
   
   **What**: Plugin is installed but disabled. Last updated <date>.
   **Action**: Uninstall to reduce clutter
   **Risk**: If you need it later, you can reinstall
   
   Proceed? (yes/no/skip)
   

   Stale cache cleanup

   ## Finding N: Old plugin cache — X MB recoverable
   
   **What**: <count> old version cache directories consuming X MB total
   **Action**: Remove old version caches (keeps current active versions)
   **Risk**: Cannot roll back to previous plugin versions after removal
   
   Proceed? (yes/no/skip)
   

   Orphaned project config

   ## Finding N: Orphaned project config — X MB
   
   **What**: Project config at <path> doesn't map to an existing project directory
   **Action**: Delete the orphaned config directory
   **Risk**: Loss of project-specific memory and settings for a project that no longer exists
   
   Proceed? (yes/no/skip)
   

5. Generate cleanup report

   In dry-run mode, output:

   # Dry-Run Summary
   
   ## Findings Reviewed
   - X findings would be addressed
   - Y findings would be skipped (user preference or safety)
   
   ## Estimated Impact (if applied)
   - Disk space that would be recovered: ~X MB
   - Token overhead that would be removed: ~X tokens/session
   - Estimated cost savings: ~$<calculated> per message
   
   ## Commands That Would Run
   <list each command that would be executed>
   
   To apply these changes, run: `/moltbloat:clean` (without --dry-run)
   

   In interactive mode, after processing all selected findings:

   # Cleanup Complete
   
   ## Actions Taken
   - <list each action performed with specific names>
   - ...
   
   ## Skipped
   - <list each finding the user chose to skip>
   - ...
   
   ## Disk Space Recovered
   ~X MB
   
   ## Estimated Token Savings
   ~X tokens per session
   
   ## Next Steps
   - Run `/moltbloat:audit` again to verify clean state
   - Run `/moltbloat:token-budget` to see updated context costs
   - Run `/moltbloat:snapshot` to save this clean state as a baseline
   

6. Done