Monitoring Deep Dive
Independent observation, radical transparency
An independent process continuously evaluates agent behavior without the agent being able to influence the results. The monitoring system never goes through the agent -- it communicates directly with the human operator.
Single-file lightweight web servers running on each machine. Auto-refreshing interface with no external dependencies.
Max Dashboard (7 tabs)
Inter-agent message logs and delivery status
Color-coded exec-guardian decisions
Pipeline status, extractions, MEMORY.md size
9-dimension scoring with trajectory analysis
Governor statistics from the bridge local database
Full site content: 9 use cases, 12 security layers
13 extensions monitored in real-time (violet): extract-memories, auto-dream, compaction, hooks, and more
Eva Dashboard (7 tabs)
Inter-agent message logs
Exec-guardian decisions
Pipeline status
7-dimension scoring with radar chart
Governor statistics
Site content mirror
13 extensions monitored: memory metrics, hooks, active tasks
The Bridge Analytics tab reads directly from the local database maintained by the governor daemon. It provides real-time statistics on inter-agent communication volume, rate-limit events, and message delivery times.
Messages per direction (sent/received)
Rate-limit trigger frequency
Average delivery latency
STOP/GO event history
Eva uses a 7-dimension scoring model (vs 9 for Max). The radar chart dynamically adapts to the available dimensions.
correction_rate and satisfaction_rate are excluded for Eva due to insufficient session data.
Trust Score -- 30-day trend
The dashboard integrates a new EasyClaw-dedicated tab (violet). It monitors in real-time the status of 13 extensions on each machine: extract-memories, auto-dream, compaction, hooks, permission-classifier, and more. Each agent has its own view with specific statistics.
13 extensions monitored
Detailed tracking of the automatic memory plugin activity.
- Memories extracted per session (counter by type: facts, preferences, decisions)
- False positive rate (memories rejected after review)
- Deduplication rate (memories merged with existing ones)
- Context volume saved by compaction
Overnight memory consolidation statistics.
- Number of consolidations executed (by time window)
- Tokens consumed per dream session
- Backups automatically generated before each consolidation
- Compression ratio (raw memory / consolidated memory)
Tracking the adversarial verification agent and simplifier.
- Results by status: PASS, FAIL, PARTIAL
- Number of reverts triggered by /verify (automatic corrections)
- Successful simplification rate (before/after readability)
- Average execution time per command
An independent script (zero external dependencies) runs every 6 hours via launchd cron. It reads session files, computes 9 dimensions across 4 time windows (24h, 7d, 30d, all), and pushes scores to Supabase PostgreSQL.
- The agent never scores itself -- completely independent process
- Trajectory over snapshot: 4 time windows for drift detection
- Robust statistics: median + MAD instead of mean + std
- Composite score 0-1 mapped to 5 trust bands
- Automatic drift alerts via iMessage when anomalies are detected