cleanup

name	cleanup
description	Reduce input token bloat by archiving completed features, removing stale milestone files, trimming notes, and auditing convention files (CLAUDE.md, AGENTS.md).
alwaysApply	false

Belmont: Cleanup

You are cleaning up accumulated Belmont state to reduce input token bloat. This is a middle ground between review-plans (audit alignment) and reset (nuclear wipe). It targets completed features, archived milestones, stale notes, and outdated convention files.

This cleanup requires ultrathink-level reasoning — carefully assess what is safe to archive vs. what the user still needs for active context.

Purpose

Over time, Belmont projects accumulate completed feature directories (PRD, PROGRESS, TECH_PLAN, MILESTONE-*.done.md), growing NOTES.md files, and stale conventions in CLAUDE.md/AGENTS.md. Every one of these files inflates the context window for AI agents, costing tokens and potentially introducing outdated guidance. This skill interactively trims that bloat while preserving what matters.

Critical Rules

1. Interactive — every item requires explicit user choice. Never bulk-act without consent. Present each feature, each file, each finding individually.
2. Non-destructive by default — the preference order is: keep > archive > delete. Default suggestion is always "keep" unless the user chooses otherwise.
3. Scoped to state files — modifies .belmont/, audits CLAUDE.md and AGENTS.md, suggests tool-dir cleanup. Does NOT modify source code.
4. Skip gracefully — if a category has nothing to clean, skip it silently.

Forbidden Actions

• DO NOT modify source code files
• DO NOT run builds, tests, or compilation commands
• DO NOT delete files without explicit user confirmation per item
• DO NOT auto-modify tool directories (.claude/, .codex/, .cursor/, etc.) — only suggest
• DO NOT pressure the user toward archiving — completed features may still be useful context

Allowed Actions

• DO read all .belmont/ state files, CLAUDE.md, AGENTS.md, tool directory structures
• DO present cleanup options interactively via AskUserQuestion
• DO archive completed feature directories (compress verbose files into slim summaries)
• DO remove MILESTONE-*.done.md files based on user choice
• DO edit NOTES.md, CLAUDE.md, AGENTS.md based on user decisions
• DO commit .belmont/ changes after cleanup

Step 1: Scan & Inventory

Read all state files and build a cleanup profile:

1. Check .belmont/ directory exists — if not, tell user to run belmont install first and stop
2. Read .belmont/PRD.md — extract features table, note each feature's status
3. Read .belmont/PROGRESS.md — cross-reference feature statuses (computed from task states)
4. Scan .belmont/features/ for each subdirectory:
   • Read PROGRESS.md — check if all tasks are [v] (verified) indicating the feature is complete
   • Count all files: PRD.md, TECH_PLAN.md, PROGRESS.md, NOTES.md, MILESTONE.md, MILESTONE-*.done.md
   • Estimate total size of the feature directory
   • Classify as: Completed (all tasks [v]), Active (has [>] or [x] tasks), or Not Started (all tasks [ ])
5. Find all MILESTONE-*.done.md files across .belmont/ root and all feature directories
6. Read .belmont/NOTES.md — count entries/sections, note oldest entry date
7. Check for convention files at project root: CLAUDE.md, .cursorrules, .windsurfrules, AGENTS.md
8. Check which tool directories exist: .claude/, .codex/, .cursor/, .windsurf/, .gemini/, .copilot/

Present the inventory:

Belmont Cleanup Scan
====================

Features: X total (Y completed, Z active)
  Completed:
    [slug]  [name] — N files, ~X KB
    [slug]  [name] — N files, ~X KB
  Active:
    [slug]  [name] — N files (not eligible for archiving)

Archived Milestones: N MILESTONE-*.done.md files (~X KB total)
Global Notes:        M entries in NOTES.md (oldest: YYYY-MM-DD)
Convention Files:    CLAUDE.md [found/not found], AGENTS.md [found/not found]
Tool Directories:    .claude/ .codex/ .cursor/ ...

Cleanup Categories:
  [1] Full cleanup (walk through all categories below)
  [2] Pick categories interactively
  [c] Cancel

Wait for user response before proceeding.

If option 2, present categories:

Categories:
  [a] Archive completed features
  [m] Remove archived milestone files
  [n] Trim NOTES.md
  [d] Audit convention files (CLAUDE.md, AGENTS.md)
  [t] Check tool directory state
  [c] Cancel

Enter letters for categories to run (e.g., "amn"):

Step 2: Archive Completed Features

Skip if no completed features exist.

For each feature where all tasks in PROGRESS.md are [v] (verified), present it individually:

Feature: [feature-name] ([slug])
  Status:   Complete
  Files:    PRD.md, TECH_PLAN.md, PROGRESS.md, NOTES.md, N MILESTONE-*.done.md
  Size:     ~X KB total
  Summary:  [2-3 sentence summary extracted from PRD.md overview]

  Options:
    [a] Archive — replace all files with a slim ARCHIVE.md summary (~0.5 KB)
    [k] Keep   — leave untouched (still needed for context)
    [d] Delete — remove the entire feature directory
    [s] Skip   — decide later

Wait for user response for each feature.

If user chooses Archive

1. Generate .belmont/features/<slug>/ARCHIVE.md:

# Archive: <feature-name>

**Slug**: <slug>
**Completed**: <date from last session entry in PROGRESS.md, or "Unknown">
**Milestones**: <list milestone names, all complete>

## Summary
<2-3 sentence summary extracted from PRD.md overview section>

## Key Decisions
<Extract numbered decisions from PROGRESS.md Decisions Log section. If none, extract key points from NOTES.md. If neither, write "None recorded.">

## Key Files
<List key files/directories from TECH_PLAN.md file structure section. If no tech plan, write "See git history.">

2. Delete all other files in .belmont/features/<slug>/: PRD.md, TECH_PLAN.md, PROGRESS.md, NOTES.md, MILESTONE.md, all MILESTONE-*.done.md
3. Update the master .belmont/PRD.md features table — change the Status column for this feature to Archived
4. Update the master .belmont/PROGRESS.md features table — change the Status column for this feature to Archived

If user chooses Delete

1. Remove the entire .belmont/features/<slug>/ directory
2. Update master PRD features table — remove the row or mark as Removed
3. Update master PROGRESS features table — remove the row or mark as Removed

If user chooses Keep or Skip

Move to the next feature. No changes.

Step 3: Remove Archived Milestone Files

Skip if no MILESTONE-*.done.md files exist.

Find all MILESTONE-*.done.md files in:

• .belmont/ root (legacy location)
• .belmont/features/*/ (per-feature, only in non-archived features)

Present the full list:

Found N archived milestone files:

  .belmont/features/auth/MILESTONE-M1.done.md       (~X KB)
  .belmont/features/auth/MILESTONE-M2.done.md       (~X KB)
  .belmont/features/dashboard/MILESTONE-M1.done.md  (~X KB)
  ...

Total: ~X KB

These are completed milestone archives. They served as inter-agent
communication during implementation but have no active purpose now.

Options:
  [a] Remove all
  [p] Pick individually
  [s] Skip

For pick mode, present each file:

  .belmont/features/auth/MILESTONE-M1.done.md (~X KB)
  [r] Remove  [k] Keep  [s] Skip remaining

Step 4: Trim NOTES.md

Skip if no NOTES.md files exist (global or feature-level).

Read .belmont/NOTES.md (global) and any feature-level NOTES.md in active (non-archived) features.

For each NOTES file with content, present entries grouped:

Global NOTES.md: N entries

  Entries for archived/deleted features:
    - [entry summary] (YYYY-MM-DD) — feature was archived
    - [entry summary] (YYYY-MM-DD) — feature was archived

  Entries older than 30 days:
    - [entry summary] (YYYY-MM-DD)
    - [entry summary] (YYYY-MM-DD)

  Recent entries (keeping):
    - [entry summary] (YYYY-MM-DD)

Options:
  [1] Remove entries for archived features (N entries, ~X KB)
  [2] Remove entries older than 30 days (N entries, ~X KB)
  [3] Remove both 1 and 2
  [4] Review each entry individually
  [s] Skip

For individual review, present each entry with [r]emove / [k]eep options.

Step 5: Audit Convention Files

Skip if no convention files exist.

Check for convention files at the project root. These are agent-agnostic configuration files that AI tools read automatically:

• CLAUDE.md (Claude Code)
• .cursorrules (Cursor)
• .windsurfrules (Windsurf)
• AGENTS.md (Codex)

For each file that exists, perform the following audit:

File Path References

Scan the file for paths that look like file/directory references (e.g., src/components/, lib/utils.ts). For each, verify the path exists using glob. Flag paths that don't resolve:

CLAUDE.md Audit — File References:

  Line 23: "Components live in src/components/ui/"
    → Path exists (OK)

  Line 45: "Auth middleware is in src/middleware/auth.ts"
    → Path NOT found (MISSING)
    [u] Update path (ask user for correct path)
    [r] Remove this line
    [s] Skip

Stale Feature References

Look for references to features that were just archived or deleted in Step 2. If found:

  Line 67: "The dashboard feature uses the Chart component..."
    → Feature "dashboard" was archived in this cleanup
    [u] Update reference
    [r] Remove this section
    [s] Skip

Outdated Conventions

Look for patterns that may be outdated:

• References to deprecated packages or APIs
• Build/test commands that may have changed
• Directory structure descriptions that don't match reality
• Conflicting or duplicate rules

Present each finding individually with context and options.

AGENTS.md — Belmont Section

If AGENTS.md exists and contains a Belmont-managed section (between <!-- belmont:codex-skill-routing:start --> and <!-- belmont:codex-skill-routing:end --> markers):

1. Read the known skills list from the managed section
2. Compare against actual files in .agents/skills/belmont/
3. Flag any skills listed that don't have corresponding files, or files that aren't listed

AGENTS.md — Belmont skill routing section:

  Listed but missing file: "old-skill" → no .agents/skills/belmont/old-skill.md
  File exists but not listed: "cleanup" → .agents/skills/belmont/cleanup.md exists

  [u] Update section (suggest running `belmont install` to re-sync)
  [s] Skip

Step 6: Agent Tool State

Skip if no tool directories exist.

Check each detected tool directory for stale Belmont artifacts:

Copy-based tools (may have stale copies)

• .claude/commands/belmont/ — compare file list against .agents/skills/belmont/. Flag files in .claude/commands/belmont/ that don't exist in source.
• .codex/belmont/ — same comparison.

Symlink-based tools (may have broken links)

• .cursor/rules/belmont/ — check each .mdc file is a valid symlink
• .windsurf/rules/belmont/ — check symlink target exists
• .gemini/rules/belmont/ — check symlink target exists
• .copilot/belmont/ — check symlink target exists

Present findings as suggestions only:

Agent Tool State:

  .claude/commands/belmont/: 2 extra files not in source
    - old-skill.md
    - renamed-skill.md
  .codex/belmont/: in sync (OK)
  .cursor/rules/belmont/: 1 broken symlink
    - removed-skill.mdc → (broken)

  Recommendation: Run `belmont install` to re-sync all tool integrations.

  [Press Enter to continue]

Do NOT modify tool directories — only inform and suggest belmont install.

Step 7: Summary & Commit

Present a final summary:

# Cleanup Summary

## Actions Taken
- Archived N features (saved ~X KB): [list slugs]
- Deleted N features: [list slugs]
- Removed N milestone files (~X KB)
- Trimmed NOTES.md (removed N entries, ~X KB)
- Updated CLAUDE.md (N changes)
- Updated AGENTS.md (N changes)

## Total Estimated Token Savings
~XX KB removed from agent context

## Kept (user chose to preserve)
- [list of features/items kept]

## Skipped
- [list of items skipped]

Commit Changes

After completing all updates, commit modified files:

1. Check if .belmont/ is git-ignored — run:

   git check-ignore -q .belmont/ 2>/dev/null
   

   If exit code is 0, .belmont/ is ignored — skip this section entirely.

2. Check for changes — run:

   git status --porcelain .belmont/ CLAUDE.md AGENTS.md .cursorrules .windsurfrules 2>/dev/null
   

   If there is no output, nothing to commit — skip the rest.

3. Stage and commit — stage only modified state and convention files:

   git add .belmont/
   

   If CLAUDE.md, AGENTS.md, .cursorrules, or .windsurfrules were modified, stage those too:

   git add CLAUDE.md AGENTS.md .cursorrules .windsurfrules 2>/dev/null
   

   Then commit:

   git commit -m "belmont: cleanup completed features and stale state"
   

Suggested Next Steps

Suggested next steps:
- /belmont:status to verify current project state
- belmont install to re-sync tool integrations (if stale files were detected)
- /belmont:review-plans to audit remaining active features

Edge Cases

• No completed features, no stale files: Tell user "Nothing to clean up — project state is lean." and stop after the scan.
• All features completed: Still present each individually — user may want to keep some for reference.
• No NOTES.md: Skip Step 4.
• No convention files: Skip Step 5.
• No tool directories: Skip Step 6.
• Feature has MILESTONE.md (not .done.md): This is an active milestone — do NOT touch it, even if the feature is marked complete. Warn the user there may be an in-progress implementation.
• Running in a worktree: Warn that cleanup should be run from the main worktree, not a feature worktree. Check for .belmont/auto.json — if it references active worktrees, warn accordingly.

When to Use This Skill

• After completing a batch of features and starting new work
• When context windows feel bloated or agents are slow
• After /belmont:review-plans identifies many completed features
• Periodically during long-running projects
• Before onboarding new team members (clean slate)