Skip to content

Production Readiness

Commander is designed for production from day one. Every component includes observability, safety, and reliability features.

Feature Matrix

CapabilityCommander Status
Type safetyTypeScript strict mode, zero as any / @ts-ignore (ESLint error)
Error handlingZero empty catch blocks across 100+ modules
MetricsUnified MetricsCollector with Prometheus/OpenMetrics counters, gauges, histograms + tenant labels
TracingSpan-based execution with persistent trace store, OpenTelemetry export
Crash safetyAtomic SQLite WAL checkpoints at every step + event sourcing hash chain
Circuit breaker5 failures → 30s open → half-open recovery, per-provider registry
Dead letter queue7 categories, 15 failure modes, persistent storage with replay support
Multi-tenancyPer-tenant rate limits, concurrency quota, storage isolation, cache isolation
Security7-layer EnterpriseSecurityGateway, DLP, capability tokens, Bearer auth, CORS, rate limiting
ObservabilityHealth check, readiness probe, OpenAPI spec, SSE streaming, Grafana dashboards
Event sourcingWAL with SHA-256 hash chain, snapshot recovery, deterministic replay
Plugin sandboxingThird-party plugins restricted via sandboxed load context; permissions never exceed main system

Safety Mechanisms

Circuit Breaker

After 5 consecutive failures, the circuit opens for 30 seconds, then transitions to half-open for recovery. The CircuitBreakerRegistry manages breakers for all active providers.

Dead Letter Queue

Unrecoverable errors are persisted across 7 categories (llm, tool, execution, verification, circuit_breaker, compensation, semantic_drift) with 15 standardized failure modes. Supports replay after root cause is fixed.

Compensation Registry

Failed mutation tools trigger automatic rollback via registered compensation actions. Integrated with Saga coordinator for distributed transactions.

State Checkpointer

Every step saves an atomic checkpoint using SQLite with WAL mode (synchronous=NORMAL, busy_timeout=5000). Resume from any failure without data loss.

Event Sourcing Engine

Write-Ahead Log with SHA-256 hash chain provides tamper-proof event logging. All non-deterministic inputs (timestamps, random values, LLM responses, tool results) are recorded for deterministic replay.

Recovery Bootstrapper

On process startup, scans for zombie runs (EXECUTING/VERIFYING/PAUSED states), acquires fencing lease, and either resumes from checkpoint or aborts with compensation.

Observability

Metrics (Prometheus / OpenMetrics)

typescript
getMetricsCollector().exportOpenMetrics()
// Exports: counters, gauges, histograms with tenant labels

Tracing

Span-based execution traces with persistent storage in TraceStore. OpenTelemetry export with PII redaction (auto-strips gen_ai.prompt, gen_ai.completion, gen_ai.tool.call.arguments from spans).

Health Endpoints

  • /health — Liveness probe
  • /ready — Readiness probe
  • /metrics — Prometheus metrics
  • /health/detailed — Component-level health (circuit breaker, DLQ, compensation, event bus, provider, event sourcing)

Grafana Dashboards

Pre-configured dashboards for two audiences:

  • Developer view: Run success rate, P95 latency, token cost, active runs, tool call success rate
  • Mechanistic view: WAL write latency, WAL file size, DLQ backlog, circuit breaker state, event backlog ratio, SQLite lock contention, compensation execution rate

Testing

Commander maintains zero tolerance for failures:

  • 6700+ tests across unit, integration, chaos, and e2e
  • Chaos-monkey tests for fault injection
  • Multi-tenant isolation tests (28 scenarios)
  • Plugin permission tests (47 scenarios for sandbox enforcement)
  • Stress tests: 10K messages, 50 concurrent calls
  • Coverage thresholds: statements 60%, functions 70%, lines 60%
npx tsx --test tests/*.test.ts   # All green, # fail 0
npx tsc --noEmit                  # Zero type errors

MIT Licensed — Built for multi-agent orchestration.