The fleet continual-learning monitor (Phase 177.10.1.2), in TypeScript. Turns a
FleetSnapshot — every robot's per-skill retention vs its certified baseline —
into forgetting alerts, rollback recommendations, a Phase-92 MLOps governance
decision, and a self-contained offline HTML dashboard.
Modules#
types.ts— the data model:SkillRetention,RobotState,FleetSnapshot,ForgettingAlert,RollbackRecommendation, and severity thresholds (safety-critical skills use a tighter bound).alerts.ts—classifySeverity(BWT-banded ok/watch/warn/critical, safety escalation),computeAlerts(sorted most-severe first),recommendRollbacks(safety-critical critical regressions → revert to certified revision), andfleetSummary.governance.ts—governanceGatemapping the fleet state topromote/hold/rollbackwith the robots to revert.dashboard.ts—renderDashboard: a single offline HTML page (per-robot × per-skill retention heatmap coloured by severity, safety-critical markers, ranked alert + rollback panels, governance banner). Inline CSS only; HTML-escaped.
Tests#
npx vitest run (from this dir, bypassing Nx) — 14 tests: severity banding +
safety escalation, alert emission/sorting/omission, rollback triggering,
fleet-summary roll-up, governance promote/hold/rollback, and the HTML renderer
(self-contained, embeds data + governance decision, escapes injection, marks
safety-critical skills). Typechecks under verbatimModuleSyntax.