Roadmap
What to build first, in order. Resist over-splitting and over-building.
MVP — build these first
- Nexus Core in Rust with axum.
- MongoDB collections:
agents,skills,tasks,runs,memories. - Nexus UI: Agents + Skills + Board + Runs.
- Taiga adapter.
- Telegram command gateway.
- Kubernetes job launcher.
- One generic agent-runner container.
First useful command:
/nexus create goal "Build authentication system for my app"
Response:
Goal created.
Drafted 8 tasks.
Waiting for approval in Nexus UI.
Then in the UI: approve plan → create Taiga user stories → assign planner/implementer/tester agents → run Kubernetes jobs.
Repos for MVP
Create only:
nexus-platform
nexus-agent-images
nexus-gitops
Add nexus-examples and split nexus-docs public later.
After MVP
- Memory retrieval upgrade — add a vector store (Qdrant / Atlas Vector Search) once text + tags hits its limits.
- Blueprints — templates to stop repeating agent config.
- Auto-approval for low-risk memory and actions.
- Jira adapter — second
BoardProvider. - More agent images — devops, solana, and language-specific toolchains.
- Executable skills — tool / workflow / verification skill kinds beyond instruction skills.
- SSE event stream wiring across the UI for live run logs.
Capability target — feature parity with Hermes
A deep analysis of NousResearch/hermes-agent (agent isolation, communication, and the self-improving learning loop) and the concrete capabilities Nexus should add to be as robust — grouped into proposed tiers. Timing is undecided; this is the target spec.
→ Capability parity (Hermes gap analysis)
Two focused target specs drill into the highest-value areas:
- → Goal lifecycle engine — turn the one-shot
decompose into a staged pipeline (clarify → classify → decompose → replan →
close) in a new
nexus-orchestratorcrate. - → Kanban board parity (Hermes) — the board feature surface specifically: task fields, live updates, dispatcher self-healing, and the command surface.
HARVEST (Solana accumulation) — live
The on-chain treasury module is built and running (armed since
2026-08-29): laddered SOL/cbBTC accumulation, lending yield while waiting,
JitoSOL staking and LP positions decided by an AI desk and executed by the
nexus-solana venue service within mechanical caps.
→ HARVEST — as built — what runs, its contract, API, and what is still missing (start here)
→ HARVEST (design spec) — the target spec it was built from, annotated where the build diverged
→ Accumulation desk — repurposing the perp trading-desk + dream mobs into an intelligent accumulator that trades rebounds inside a bounded sleeve beside the ladder.
→ Measuring success — the USDT-NAV objective, return and attribution accounting, the benchmark suite and the scenario table the 2026-09-05 audit asks for (specified, not implemented).
Audit status registers live under Audits.
And an operational target spec that spans the whole fleet:
→ Monitoring & alerting — the UI redesigned around a fleet health grid, with rule-based Telegram alerting for every component.
Explicitly not yet
- GraphQL API (REST suffices).
- Cursor-based pagination (defaults are generous).
- Per-agent custom runner binaries (the generic runner is the whole point).