Claude-Dash Administration Framework

When This Activates

This skill activates for system maintenance:

• Health checks and diagnostics
• Storage cleanup and optimization
• Index management
• Service troubleshooting

System Components

Core Services

Storage Locations

~/.claude-dash/
├── projects/       # Project memory (indexes, summaries)
├── sessions/       # Session data, transcripts
├── learning/       # Corrections, preferences, calibration
├── mlx-tools/      # Local AI tools (Ollama-based)
├── skills/         # Skill definitions
├── logs/           # Service logs
├── indexes/        # HNSW search indexes
└── config.json     # Main configuration

Health Checks

Quick Health Check

memory_health action=status

Full Diagnostic

# Check services
ps aux | grep -E "(watcher|gateway|ollama)"

# Check storage
du -sh ~/.claude-dash/*/

# Check recent errors
tail -20 ~/.claude-dash/logs/watcher-error.log

# Check gateway metrics
gateway_metrics format=summary

Common Maintenance Tasks

1. Clean Up Storage

# Run session archival
python3 ~/.claude-dash/scripts/archive-sessions.py

# Run log rotation
~/.claude-dash/scripts/log-rotation.sh

2. Refresh Indexes

# Trigger HNSW rebuild
hnsw_status action=rebuild project=gyst

# Run freshness check
workers_run worker=freshness project=gyst

3. Consolidate Learning

# Run observation consolidation
workers_run worker=consolidate

4. Restart Services

# Restart watcher
~/.claude-dash/watcher/start-watcher.sh restart

# Restart gateway (via Claude settings or restart Claude)

Troubleshooting

“Memory search returns nothing”

1. Check if project is registered in config.json
2. Verify index files exist: ls ~/.claude-dash/projects/{project}/
3. Check watcher is running
4. Trigger reindex if needed

“Hook is slow”

1. Check hook timeout in settings
2. Review inject-context.sh performance
3. Consolidate to single Python process (inject_all_context.py)

“Watcher crashes”

1. Check watcher-error.log
2. Verify project paths exist
3. Check for permission issues

“Ollama not responding”

1. Check: ollama list
2. Restart: ollama serve
3. Verify model exists: ollama pull gemma3:4b-it-qat

Configuration

Main Config (config.json)

{
  "projects": [...],
  "watcher": {
    "enabled": true,
    "ignorePatterns": [...],
    "scanIntervalMs": 5000
  }
}

Adding a New Project

{
  "id": "project-id",
  "displayName": "Project Name",
  "path": "/full/path/to/project",
  "memoryPath": "projects/project-id"
}

Performance Metrics

Gateway Efficiency

gateway_metrics format=detailed

Shows:

• Cache hit rate (target: >50%)
• Local routing rate (higher = cheaper)
• Token savings

Learning Status

learning_status component=all

Shows:

• Preferences learned
• Corrections recorded
• Confidence calibration

Scheduled Maintenance

These run automatically via cron:

• 3 AM: Log rotation
• Background: Watcher auto-updates

Manual periodic tasks:

• Weekly: Session archival
• Monthly: Review and clean old projects
• As needed: HNSW index rebuild

Emergency Recovery

Reset Project Memory

rm -rf ~/.claude-dash/projects/{project}/*
# Watcher will rebuild on next file change

Reset All Indexes

hnsw_status action=rebuild-all

Clear Corrupted State

# Backup first
cp -r ~/.claude-dash ~/.claude-dash.backup

# Then clear specific problematic data
rm ~/.claude-dash/sessions/observations.json
rm ~/.claude-dash/learning/corrections.json