Roadmap

Status: Draft v0.1 · Last updated 2026-05-27

Phases are sized in weeks of single-engineer effort, not calendar time. They are sequential — each phase assumes the previous one shipped.

Phase 0 — Foundation (week 1)

Goal: repo is alive, can deploy a hello-world, infra accounts are provisioned.

• Monorepo scaffolding (pnpm + Turborepo)
• Next.js app deployed to Vercel (preview + prod)
• Supabase project provisioned via Vercel Marketplace
• Fly.io app + machine template created (no real logic yet)
• AI Gateway configured, BYOK key set
• Hyperliquid mainnet read-only access verified (read-only endpoints don't require credentials; testnet deprecated in ADR-0015)
• CI: lint + typecheck on PR
• ADRs 0001–0008 ratified

Exit criteria: pnpm dev runs the web app locally with Supabase Auth working end-to-end.

Phase 1 — Skill authoring + Simulation (weeks 2–4)

Goal: an author can describe a strategy and backtest it against historical data.

• packages/skill-schema — zod types, version-on-save
• Skill editor form in web app
• packages/tools registry with first 4 tools (bars, news, portfolio, propose_order)
• packages/execution-engine — validation + risk checks + paper broker
• packages/agent-runtime — runSkill using AI SDK + Gateway
• Historical data ingestion script (Hyperliquid majors, 12 months)
• News ingestion + pgvector embeddings (CryptoPanic free tier to start)
• packages/simulator — replay loop, context assembly, decision snapshots
• CLI: pnpm sim --skill <id> --from <date> --to <date>
• Web UI: trigger sim, view equity curve + trade log + decision snapshots

Exit criteria: an author can create a Skill in the UI, click Backtest, and see a complete report within 2 minutes for 30 days of 5m bars.

Phase 2 — Live deployment + Chat (weeks 5–7)

Goal: an author can deploy a tested Skill live (mainnet, conservative caps) and talk to it.

• Hyperliquid mainnet broker adapter (gated by explicit per-deployment confirmation; testnet deprecated — ADR-0015)
• apps/live-runner — Fly app: control loop + tick loop, command listener via Postgres LISTEN/NOTIFY
• Deploy trigger: web → Fly Machines API
• Agent state writes (agent_state, agent_logs)
• Deployment detail page: state snapshot, decision log, command panel (stop/pause/flatten/kill)
• Chat agent route: streamText with introspection tools
• Conversation persistence (agent_conversations, agent_messages)
• Suggested-prompts row in chat UI

Exit criteria: an author can deploy a Skill on mainnet with conservative caps, watch it tick, ask it "why did you open that position?" and get a grounded answer citing the actual decision snapshot.

Phase 3 — Hardening (weeks 8–10)

Goal: trustworthy enough to point at real money at meaningful size.

• Encrypted API key vault for user exchange credentials
• Daily loss halt + auto-flatten enforcement (engine-side, not agent-side)
• Sentry integration for agent errors
• Cost dashboard per Skill (tokens, $ per decision, $ per day)
• Rate limiting + retry policy on exchange calls
• Sim vs live drift report (compare paper fills to actual fills for the same skill)
• Auth hardening: 2FA, session management, audit log

Exit criteria: a Skill running on mainnet for 7 days without engineer intervention, with complete audit trail.

Phase 4 — Iteration acceleration (weeks 11–14)

Goal: make authoring strategies dramatically faster.

• Skill forking (clone + version)
• Multi-skill comparison view (same date range, side-by-side metrics)
• MCP server attachment (pre-approved community tools)
• Skill diff view (between versions)
• Sim queue with priority (so users running 10 sims don't block each other)
• Optional deterministic helpers (entry.ts, exit.ts) compiled and called from agent loop

Exit criteria: an author can fork an existing Skill, change the model and one risk cap, and have a comparative backtest in under 5 minutes.

Phase 5 — Marketplace foundations (weeks 15+)

Goal: Skills become a shareable, valuable artifact — the platform moat.

• Public Skill profiles (opt-in)
• Standardized performance metrics (Sharpe, max drawdown, profit factor)
• Skill provenance + audit (immutable version log)
• Skill subscription model (consumer pays author per deployment)
• Forking with attribution
• Skill discovery / search
• Reviews and disputes process

Out of scope: trustless on-chain Skill ownership, copy-trading other people's live positions, social features.

Backlog (no phase assigned)

• Multi-exchange support (Binance, Bybit, OKX) — abstract broker interface accommodates this from day one, but adapters are real work
• Multi-agent decomposition (analyst → strategist → risk officer → executor handoffs)
• Team workspaces with RBAC
• Mobile app (probably never; PWA at most)
• Skill plagiarism / copy detection in marketplace
• Realized-PnL tax reports

What we will deliberately not do

• Custody user funds. Always user-keyed exchange access.
• Give financial advice. Skills are user configurations; we execute, we don't recommend.
• Auto-execute on agent-suggested risk overrides. The engine's risk caps are sovereign.
• Promise SLAs we can't back. Until we have ops maturity, no uptime guarantees.