abzhaw/self-learning

Skill: Self-Learning Engine

Purpose

Sentinel gets smarter over time. Every report is a learning opportunity. This skill runs last in the daily cycle and updates the agent's memory so tomorrow's report is better than today's.

What It Does

1. Compare Today vs. Yesterday

• What new findings appeared today that weren't yesterday?
• What findings disappeared (resolved)?
• What findings are recurring for N consecutive days?

2. Update Baseline

• Add confirmed-safe findings to memory/baseline.json
• This suppresses them from future reports so the signal stays clean

3. Apply Suppression List

• If Raphael replied "dismiss" to a finding (tracked in memory/suppressed.json), skip it in future reports
• Suppressed findings still logged silently but not surfaced

4. Track Severity Trends

• Is the overall security posture improving or degrading week-over-week?
• Which service generates the most findings? (focus attention there)

5. Update Learnings Journal

• Append to memory/learnings.md with what changed today
• Note any false positives that should be suppressed
• Note any new patterns discovered

Memory Files

memory/baseline.json

{
  "known_safe": [
    {"check": "port-scan", "finding": "port 5432 on 127.0.0.1", "added": "2026-02-23"},
    {"check": "process-audit", "finding": "node pm2 daemon", "added": "2026-02-23"}
  ],
  "last_updated": "2026-02-23"
}

memory/suppressed.json

{
  "suppressed": [
    {"check": "api-security", "finding": "no rate limiting on bridge", "dismissed_by": "raphael", "date": "2026-02-23", "reason": "internal only"}
  ]
}

memory/learnings.md

## 2026-02-23
- First scan complete. Baseline established.
- 3 findings dismissed by Raphael (bridge rate limiting — internal only)
- New pattern: docker container restarts correlate with bridge error spikes

Severity Rules

This skill has no severity output — it is the system's memory, not a scanner.

Output Format (in daily report)

SELF-LEARNING UPDATE
📚 Baseline: 12 known-safe items
🚫 Suppressed: 3 items (dismissed by Raphael)
📈 Trend: Security posture IMPROVING (6→3 findings week-over-week)
🔍 Focus area this week: credential hygiene (3/10 findings)

Principles

• Never suppress a Critical or High finding automatically — always surface to Raphael
• Trend over signal — a persistent Low finding that never gets fixed is worth more attention than a one-off Medium
• Explain changes — always note why something was added to baseline or suppressed