Workspace Cleanup

Overview

Automatically clean workspace directories by detecting and archiving clutter using intelligent multi-signal analysis. Reduces AI context pollution from temp files, sync conflicts, and superseded code versions while maintaining safety through archive-based two-stage deletion.

Core principle: Safe, intelligent cleanup that understands file drift from AI code generation and protects important files through multi-signal confidence scoring.

When to Use

• Workspace has accumulated clutter (temp files, old versions, sync conflicts)
• AI context windows are polluted with noise during file scanning
• User mentions cleanup needs ("this is a mess", "clean up experiments", etc.)
• Regular maintenance to prevent drift accumulation

Problem This Solves

AI assistants (Claude, Codex, Gemini) often create new files instead of updating existing ones:

• auth.ts → auth-new.ts → auth-fixed.ts
• Over time: multiple versions, unclear which is current
• Clutter from system files, temp files, and abandoned experiments
• Context window pollution during code analysis

Detection System

Three Core Signals

1. Similarity Detection

• Content hash comparison between files
• Filename similarity (Levenshtein distance)
• Flags files with >80% content match + similar names

2. Timestamp Analysis

• Last modified time (default: 90 days untouched)
• Last accessed time
• Configurable thresholds

3. Import/Reference Analysis

• Grep for imports/requires across codebase
• Search for file references in code/docs
• Flag files with zero references as "unused"

Tiered Confidence Scoring

Tier 1 (auto-archive): 100% safe to remove

System files: .DS_Store, .sync-conflict-*
Build artifacts: __pycache__/, *.pyc, .pytest_cache/
Empty directories (except with .gitkeep)
Version patterns: -old, -backup, -fixed, -new, -updated, .bak
Status files: *.log, *.tmp, temp-*, tmp-*
Exact duplicates: Files with identical SHA256 (archive all but newest)

Tier 2 (archive): High confidence (2+ signals)

Similar files (80%+ content match) + old timestamp → archive older
Unused + old timestamp
Similarity + unused

Tier 3 (suggest only): Low confidence (1 signal)

Just old, just unused, or just similar
Large files (>100MB) even with multiple signals
Recently modified similar files (<7 days)
Report for manual review, don't auto-archive

Archive Management

Central Archive Structure

/Users/braydon/projects/archive/cleanup/
├── 2025-11-21-143022/
│   ├── metadata.json
│   └── [preserved directory structure]
└── 2025-10-15-091234/
    ├── metadata.json
    └── [files...]

Two-Stage Safety

Stage 1: Archive

• Move files to central archive (never immediate deletion)
• Preserve original directory structure
• Store metadata explaining why each file was archived

Stage 2: Review (30+ days)

• Auto-prompt for archives >30 days old
• Show summary of archived contents
• Options: Keep archive, Delete permanently, Restore files, Skip
• Mark reviewed in metadata

Metadata Format

{
  "timestamp": "2025-11-21T14:30:22Z",
  "scope": "/Users/braydon/projects",
  "recursive": true,
  "files": [
    {
      "original_path": "/Users/braydon/projects/foo.txt",
      "tier": 1,
      "signals": ["pattern_match"],
      "score": 100
    },
    {
      "original_path": "/Users/braydon/projects/experiments/old-auth.ts",
      "tier": 2,
      "signals": ["similarity", "unused", "old_timestamp"],
      "score": 85,
      "similar_to": "experiments/auth.ts"
    }
  ],
  "reviewed": false,
  "review_date": null
}

Protected Patterns

Three-Layer Protection

Layer 1: Respect .gitignore

• If git ignores it, cleanup should too
• Check with: git check-ignore -q "$file"
• Prevents cleaning build artifacts, dependencies, etc.

Layer 2: .cleanupignore Optional file for cleanup-specific exclusions:

# .cleanupignore
archive/          # Don't clean the archive itself
important-*.md    # Keep files matching pattern
legacy-project/   # Preserve specific directories

Layer 3: Hard-coded System Patterns Always protected regardless of ignore files:

Directories: .git, .claude, node_modules, .venv, venv, dist, build
Files: package.json, requirements.txt, *.lock, CLAUDE.md, README.md, .env*

Efficient Scanning with Prune

Use find's -prune to skip entire protected directory trees:

find . \( -name node_modules -o -name .git -o -name dist \) -prune \
  -o -type f -name "*.tmp" -print

This never even traverses into protected directories, making scans much faster.

Usage

Context-Aware Invocation

From conversation:

User: "This directory is a mess, let's clean it up"
→ Runs recursive cleanup from CWD

User: "Let's clean up the experiments directory"
→ Runs cleanup scoped to /experiments

Explicit commands:

/cleanup                          # Current directory
/cleanup --recursive              # Current + subdirs
/cleanup /path/to/dir             # Specific directory
/cleanup --review-archives        # Review old archives

Execution Workflow

When invoked, follow these steps:

1. Parse Scope

• Determine target directory from user request or CWD
• Check if recursive or targeted cleanup
• Validate directory exists and is accessible

2. Scan & Analyze

# Build file hash map for duplicate detection
declare -A file_hashes

# Scan with protection (prune protected dirs early)
find . \( -name node_modules -o -name .git -o -name dist -o -name build \) -prune \
  -o -type f -print | while read file; do

  # Layer 1: Check .gitignore
  if git check-ignore -q "$file" 2>/dev/null; then
    continue
  fi

  # Layer 2: Check .cleanupignore (if exists)
  if [[ -f .cleanupignore ]] && grep -q "$(basename "$file")" .cleanupignore; then
    continue
  fi

  # Layer 3: Hard-coded protections
  if [[ "$file" =~ (package\.json|CLAUDE\.md|README\.md|\.env) ]]; then
    continue
  fi

  # === TIER 1 CHECKS (auto-archive) ===

  # Check obvious patterns
  if [[ "$file" =~ (\.DS_Store|\.sync-conflict-|\.tmp$|\.log$) ]]; then
    archive "$file" tier:1 signal:pattern_match
    continue
  fi

  # Check exact duplicates
  hash=$(sha256sum "$file" | cut -d' ' -f1)
  if [[ -n "${file_hashes[$hash]}" ]]; then
    # Found duplicate - archive older file
    original="${file_hashes[$hash]}"
    if [[ "$file" -nt "$original" ]]; then
      archive "$original" tier:1 signal:exact_duplicate duplicate_of:"$file"
      file_hashes[$hash]="$file"
    else
      archive "$file" tier:1 signal:exact_duplicate duplicate_of:"$original"
    fi
    continue
  fi
  file_hashes[$hash]="$file"

  # Check version patterns
  if [[ "$file" =~ (-old|-backup|-fixed|-new|-updated|\.bak)$ ]]; then
    archive "$file" tier:1 signal:version_pattern
    continue
  fi

  # === TIER 2/3 CHECKS (multi-signal) ===
  signals=()

  # Run similarity detection (expensive, do after Tier 1)
  if hasSimilarFile "$file"; then
    signals+=("similarity")
  fi

  # Check timestamps
  if [[ $(find "$file" -mtime +90) ]]; then
    signals+=("old_timestamp")
  fi

  # Check references
  if ! grep -rq "$(basename "$file")" --exclude-dir=node_modules .; then
    signals+=("unused")
  fi

  # Score and tier
  if [[ ${#signals[@]} -ge 2 ]]; then
    archive "$file" tier:2 signals:"${signals[*]}"
  elif [[ ${#signals[@]} -eq 1 ]]; then
    suggest "$file" tier:3 signals:"${signals[*]}"
  fi
done

3. Archive Files

• Create timestamped archive directory
• Move Tier 1 + Tier 2 files preserving structure
• Generate metadata.json with analysis results
• Skip Tier 3 (just report)

4. Report Results

🧹 Workspace Cleanup - /Users/braydon/projects
Scope: Recursive | Protected dirs: 8 | Scanning...

📊 Analysis Results:
  • 156 files scanned (78 skipped via protection layers)
  • 23 Tier 1 (auto-archive):
    - System files: .DS_Store (8), sync conflicts (3)
    - Exact duplicates: (6 files, kept newest)
    - Version patterns: -old, -backup files (6)
  • 12 Tier 2 (archive): similar + old or unused
  • 8 Tier 3 (suggestions): review manually

📦 Archiving to: /Users/braydon/projects/archive/cleanup/2025-11-21-143022/
  ✓ Archived 35 files (2.3 MB saved)

💡 Tier 3 Suggestions (not archived):
  • experiments/test-model.py (unused, 45 days old)
  • personal/notes.txt (old, 120 days)
  • work/large-dataset.csv (>100MB, unused - verify before archiving)

⏰ Archives ready for review: 2 archives >30 days old
   Run '/cleanup --review-archives' to review

5. Check Archives

• Find archives >30 days old
• If found, prompt for review
• Show summary and offer actions

Archive Review Workflow

📋 Archive Review - 2 archives ready

Archive: 2025-10-15-091234 (37 days old)
  • Scope: /Users/braydon/projects (recursive)
  • 18 files archived (1.2 MB)
  • Breakdown:
    - Tier 1: .DS_Store (8), sync conflicts (10)
    - Tier 2: unused code (0)

Actions:
  K - Keep archive (don't prompt again for 30 days)
  D - Delete permanently (CANNOT BE UNDONE)
  R - Restore files to original locations
  S - Skip this review

Your choice [K/D/R/S]:

Configuration

Users can override defaults in .claude/workspace-cleanup-config.json:

{
  "timestamp_threshold_days": 90,
  "similarity_threshold": 0.80,
  "archive_review_days": 30,
  "custom_protected_patterns": [
    "important-*.md",
    "do-not-delete/*"
  ],
  "custom_tier1_patterns": [
    "*.tmp",
    "temp-*",
    ".scratch"
  ],
  "excluded_dirs": [
    "special-project"
  ]
}

Implementation Notes

Exact Duplicate Detection

# Build hash map of all files
declare -A file_hashes

while read file; do
  # Generate SHA256 hash
  hash=$(sha256sum "$file" | cut -d' ' -f1)

  # Check if we've seen this hash before
  if [[ -n "${file_hashes[$hash]}" ]]; then
    original="${file_hashes[$hash]}"

    # Archive older file, keep newer
    if [[ "$file" -nt "$original" ]]; then
      echo "Duplicate found: $file is newer than $original"
      archive "$original" tier:1 signal:exact_duplicate
      file_hashes[$hash]="$file"  # Update to keep newer
    else
      echo "Duplicate found: $original is newer than $file"
      archive "$file" tier:1 signal:exact_duplicate
    fi
  else
    # First time seeing this content
    file_hashes[$hash]="$file"
  fi
done

Why SHA256: Strong collision resistance, fast computation, standard tool (sha256sum).

Edge case: If duplicates have same mtime, keep first found, archive rest.

Similarity Detection (for non-exact matches)

// Generate content hash for quick comparison
const hash = crypto.createHash('sha256')
  .update(fs.readFileSync(file))
  .digest('hex');

// Compare filenames (Levenshtein distance)
const nameDistance = levenshtein(file1, file2);
const similarity = 1 - (nameDistance / Math.max(file1.length, file2.length));

// Flag if both content and name are similar (but not identical)
if (contentMatch > 0.80 && contentMatch < 1.0 && similarity > 0.70) {
  // Tier 2: Archive older file if also old/unused
}

Reference Detection

# Use grep to find references
grep -r "import.*${filename}" ${scope}
grep -r "require.*${filename}" ${scope}
grep -r "${filename}" ${scope}

# If no results: unused

Version Pattern Detection

const VERSION_PATTERNS = [
  /-old$/, /-backup$/, /-fixed$/, /-new$/,
  /-updated$/, /-v\d+$/, /-copy$/,
  /^old-/, /^backup-/, /^new-/, /^temp-/,
  /\.bak$/, /\.backup$/
];

// When detected + similarity match:
// Keep file without pattern, archive file with pattern

Common Mistakes

Cleaning without scanning first

• ❌ Don't skip analysis phase
• ✅ Always scan → analyze → report → archive

Ignoring Tier 3 suggestions

• ❌ Tier 3 files might become Tier 2 over time
• ✅ Review suggestions periodically

Deleting archives too quickly

• ❌ Don't delete archives <30 days old
• ✅ Wait for review prompt, verify you don't need files

Not checking protected patterns

• ❌ Assuming default patterns cover everything
• ✅ Review protected patterns for your workspace

Running on untracked important work

• ❌ Don't clean directory with active untracked experiments
• ✅ Commit or stash important work first

Edge Cases

Similar files, both recent

• If both files modified within last 7 days: Tier 3 (suggest only)
• Let user decide which to keep

Empty directories with .gitkeep

• Don't archive empty dirs containing .gitkeep
• These are intentionally empty

Large files (>100MB)

• Always Tier 3 (suggest only)
• User should explicitly confirm before archiving

Files in git staging area

• Skip files with uncommitted changes
• Report as "skipped: uncommitted changes"

Benefits

1. AI Context Reduction - Less noise in context windows
2. Safety First - Two-stage archive prevents accidental deletion
3. Intelligent Detection - Finds actual clutter, not just patterns
4. Context Aware - Adapts to user intent and scope
5. Low Maintenance - Mostly automated with sensible defaults
6. Recoverable - Everything archived, nothing immediately deleted

Related Skills

• learning-from-outcomes - Learn from cleanup patterns over time
• coordinating-sub-agents - Delegate cleanup to specialized agent

Future Enhancements

• Machine learning on user archive/restore decisions
• Cross-project similarity detection
• Automatic .gitignore updates based on archived patterns
• Integration with project task management