L2 Planner Operational Runbook

Monitoring

Dashboard

See monitoring/dashboards/l2-planner-dashboard.json for the dashboard definition.

Key panels:

• Escalation Rate: Should trend downward over time
• Success Rate: Target > 80%
• Confidence Distribution: Healthy system has most plans > 0.7
• Learned Rules Count: Should grow steadily

Alerts

Condition	Severity	Action
Escalation rate > 50% (1h window)	Warning	Check L2 planner health
Escalation rate > 80% (1h window)	Critical	Check A2A connectivity and action graph
P95 latency > 30s	Warning	Check A2A response times
Learned rule failure rate > 30%	Warning	Trigger rule audit

Common Operations

Adjusting Confidence Thresholds

Edit src/config/routing-config.yaml:

routing:
  l2_confidence_threshold: 0.5  # Lower = more autonomous, higher = more escalations

Inspecting Learned Rules

const registry = dispatcher.getRuleRegistry();
const stats = registry.getStats();
const rules = registry.getAll();

Rolling Back a Learned Rule

const migration = new RuleMigration(registry, versioning, auditor);
migration.rollback("rule-id");

Checking Escalation Trends

const tracker = dispatcher.getEscalationTracker();
const trend = tracker.getTrend(3600_000); // 1-hour buckets
const reasons = tracker.getTopReasons(10);

Troubleshooting

High Escalation Rate

1. Check A2A client connectivity (is Ava responding?)
2. Verify action graph has sufficient actions for the goal types
3. Check if confidence threshold is too high
4. Review top escalation reasons: tracker.getTopReasons(10)

Learned Rule Causing Failures

1. Identify the rule: registry.findByGoal(goalPattern)
2. Check failure count: rule.failureCount
3. Rollback if needed: migration.rollback(rule.id)
4. Audit trail: auditor.getForRule(rule.id)

L2 Planner Not Learning

1. Check minimum learning confidence: plans below 0.8 are not extracted
2. Check promotion threshold: rules need 3 successes before promotion
3. Verify registry is not at max capacity (500 default)
4. Check if plans are too long (max 10 actions for extraction)