Project Kaze — Planning Roadmap
Phases
1. Technology Selection
Status: Done See decisions.md — D16-D26, Section 2.
2. First Vertical & MVP Scope
Status: Done See mvp.md — verticals, platform components, parallel build plan.
3. Detailed Technical Design
Status: Done See technical-design.md — TypeScript interfaces, SQL schemas, API contracts.
Components specified:
- [x] Agent runtime contract — YAML+TS skill/agent definitions, AgentRuntime engine, task lifecycle, inter-agent messaging
- [x] Knowledge schema — 4 memory types, tri-factor retrieval, git-inspired versioning, ABAC access control
- [x] LLM Gateway API spec — complete/embed interface, model hints, key routing, budget enforcement, provider adapters
- [x] Tool Integration Framework — typed definitions, registry, Vault auth, retry policies, MVP tools (GitHub, Calendar, SEMrush, Toddle DB)
- [x] Observation Logger — 18 event types, structured payloads, batched writes, query/trace/metrics interface
- [x] Task Scheduler — cron + event triggers, idempotency, polling with SKIP LOCKED for HA
4. Implementation
Status: In progress
Completed
| Component | Repo | What's built |
|---|---|---|
| Agent Runtime | kaze-runtime | Two-layer agent model (VerticalAgent → SubAgent), YAML+TS skill definitions, HTTP dispatch API, supervision ramp, actor-based lifecycle. Port 4100. |
| LLM Gateway | kaze-gateway | Vercel AI SDK, multi-provider (Gemini/Claude), model hints (fast/balanced/best), tool execution with credential injection, Langfuse observability. Port 4200. |
| Knowledge Service | kaze-knowledge | Mem0 OSS for per-agent episodic memory, fact extraction via LLM, Google embeddings (gemini-embedding-001), PostgreSQL + pgvector vector store. Port 4300. |
| Agent Ops (V0) | kaze-agent-ops | Internal Ops vertical with GitHub skill (issues, PRs, labels via github_api tool). |
| CI/CD | All repos | GitHub Actions → Tailscale SSH deploy to EC2. Dockerized services. |
Architecture (as implemented)
┌─ kaze-gateway (port 4200) ──────────────────────────────┐
│ POST /llm/generate → Vercel AI SDK → Gemini/Claude │
│ POST /tools/execute → credential injection → APIs │
│ GET /tools/catalog → registered tools │
│ Secrets: LLM keys, GitHub token │
│ Observability: Langfuse tracing │
└──────────────────────────▲──────────────────────────────┘
│
┌──────────────────────────┴──────────────────────────────┐
│ kaze-runtime (port 4100) │
│ VerticalAgent → SubAgent (per-task, per-skill) │
│ Memory: search before LLM, store after LLM │
│ Zero secrets — calls gateway + knowledge via HTTP │
└──────┬───────────────────────────────────▲──────────────┘
│ │
┌──────▼───────────────────────────────────┴──────────────┐
│ kaze-knowledge (port 4300) │
│ POST /memory/search → vector similarity │
│ POST /memory/add → Mem0 fact extraction + store │
│ Own LLM key (Gemini) for fact extraction + embeddings │
│ Storage: PostgreSQL + pgvector │
└─────────────────────────────────────────────────────────┘Not yet built
- Task Scheduler (cron + event triggers)
- Observation Logger (standalone — currently using Langfuse)
- Shared knowledge tier (pgvector direct, quality gates, ABAC)
- Additional tools (Calendar, SEMrush, Toddle DB)
- V1 SEO and V2 Toddle verticals
- NATS messaging (using direct HTTP calls)
- Budget tracking / token usage enforcement