Changelog

v0.2.1-pre (Unreleased)

Features

• SQLite persistence: SqliteWarRoomStore with WAL mode, 9 indexes, transaction-safe writes. Enable via WARROOM_STORAGE=sqlite; createWarRoomStore() factory provides transparent switching
• Event sourcing engine: Write-Ahead Log with SHA-256 hash chain tamper protection, snapshot-based recovery, deterministic event replay (IF-05)
• Recovery bootstrapper: Process startup scans for zombie runs (EXECUTING/VERIFYING/PAUSED), acquires fencing lease, and either resumes from checkpoint or aborts with compensation
• Petri net scheduler: Formal resource allocation for sandbox execution slots (v8_slots, seccomp_slots, wasm_slots, tee_slots) with deadlock detection
• Enterprise security gateway: 7-layer defense-in-depth (zero-trust signatures, auth, rate limiting, input scanning, cost pre-check, request processing, DLP output scanning)
• Data Loss Prevention (DLP): 12+ sensitive data pattern detections (API keys, JWT, PEM keys, credit cards, SSN, etc.) with 4 redaction strategies (REDACT/MASK/HASH/ALLOW)
• Plugin security hardening: Sandboxed load context prevents privilege escalation; withTimeout() enforces per-plugin maxExecutionTimeMs; updateConfig() routes through sandbox
• RAG knowledge base plugin: Built-in optional plugin (builtin-rag) with HNSW vector search, OpenAI/local embedding fallback, auto-inject context mode
• Unified audit log: Cross-source audit aggregation (security events, approvals, execution traces, user actions, config changes) with unified query, statistics, and export
• Web onboarding wizard: 4-step guided setup (welcome → provider config → first task → complete)
• Grafana dashboards: Developer view (business metrics) and Mechanistic view (ops metrics) with pre-configured Prometheus datasource
• Agent degradation detection: QualityGater using Agent Capsules method — starts with compound execution, escalates to per-agent when quality drops

Improvements

• 260+ new unit tests across 15 test files (sandbox, deliberation, memory, plugin, orchestration, runtime, API security)
• ESLint no-explicit-any upgraded from warn to error
• Coverage thresholds raised: statements 40→60%, functions 55→70%, lines 40→60%
• Prettier + EditorConfig formatting enforced in CI
• Staging environment CD pipeline

v0.2.0 (2026-05-19)

Features

• Agent War Room: Web GUI dashboard with live execution feed, topology visualization, and agent roster
• OpenAPI 3.0: Full REST API specification with Swagger UI
• Readiness probe: /ready endpoint for Kubernetes deployment
• Prometheus metrics: /metrics endpoint with OpenMetrics-compatible format
• Multi-tenant isolation: Per-tenant rate limits, concurrency, storage, memory, cache isolation
• Circuit breakers: Per-provider circuit breaker with 3-state (CLOSED/OPEN/HALF-OPEN) and registry
• Dead letter queue: Persistent failure storage with replay support, 7 categories, 15 failure modes
• Compensation registry: Automatic rollback for mutation tools via registered compensation actions
• Graceful shutdown: Connection draining with configurable timeout
• HTTP security: Default localhost binding, Bearer token auth, configurable CORS
• Structured logging: All console.* replaced with structured logger; zero empty catch blocks
• Docker build: 6-stage multi-architecture build with Nginx reverse proxy and tini init
• Canonical topologies: 5 canonical topologies (SINGLE/CHAIN/DISPATCH/ORCHESTRATOR/REVIEW) aligned with Anthropic's agent ontology, with 9 legacy aliases for backward compatibility
• Artifact-based communication: Agents communicate via references instead of raw text to prevent information loss
• A2A protocol: Agent-to-Agent HTTP server with Agent Card discovery and mandatory auth tokens

CLI

• commander company — Enterprise pipeline with quality gates and memory
• commander swarm — Recursive decomposition with parallel execution
• commander drive — Autonomous step-by-step execution
• commander goal — Multi-round convergence loop
• commander review — AI code review (P0-P3 findings)
• commander saga — Saga transaction operations
• commander budget — Token budget management
• commander checkpoint — Checkpoint operations
• commander cost — Cost analysis

v0.1.0 (2026-05-17)

Features

• 8 orchestration topologies: SINGLE, SEQUENTIAL, PARALLEL, HIERARCHICAL, HYBRID, DEBATE, ENSEMBLE, EVALUATOR-OPT
• 18 LLM providers: OpenAI, Anthropic, Google, DeepSeek, GLM, MiMo, Groq, Together, Ollama, vLLM, Bedrock, xAI, OpenRouter, Mistral, Cohere, Replicate, Fireworks, Perplexity
• 25+ built-in tools: Filesystem, LSP, web research, AST transformation, sub-agent delegation
• Multi-tenant architecture: Per-tenant rate limits, concurrency, storage, memory
• Production readiness: Circuit breakers, dead letter queues, compensation registry, atomic checkpoints
• Observability: Prometheus metrics, span-based tracing, SSE stream events
• Plugin system: 19 hook points (before/after LLM, tool, run, etc.)
• Meta-learning: Thompson Sampling + Reflexion for self-evolution

CLI

• commander <task> — Quick task analysis
• commander run <task> — Full execution
• commander plan <task> — Deliberation plan
• commander watch <task> — SSE streaming
• commander gui — Agent War Room dashboard
• commander tui — Terminal dashboard
• 15 total commands

Benchmarks

• GAIA: 69.7% (115/165) — +48.5 pp over bare MiMo
• BFCL: 60.0% tool / 91.4% parameter (35-scenario); 91.7%/91.7% (12-core)
• PinchBench: 97.7% (42/43) — vs OpenClaw 89.5%
• HumanEval+: 91.5%

Deployment

• Docker Compose (6-stage multi-arch build)
• Production overlay with CPU/memory limits, health checks, JSON logging
• One-click VM deploy script
• GitHub Actions CI/CD