asiflow/claude-nexus-hyper-agent-team-light

10 stars · Last commit 2026-04-22

Multi-agents operating system layer on top of Claude Code

README preview

# Claude Code 26-Agent Team — Light Edition

> A cost-conscious variant of the [31-agent Claude Code team](https://github.com/asiflow/claude-nexus-hyper-agent-team) distribution. Same runtime discipline, adversarial review, Bayesian trust calibration, NEXUS syscall protocol, and Shadow Mind cognitive layer — with the 5-agent BEAM/Elixir stack moved to an opt-in `_optional/` directory and a session-budget hook bundled in.

**If you work on a BEAM/Erlang/OTP platform, you want the [full 31-agent distribution](https://github.com/asiflow/claude-nexus-hyper-agent-team), not this one.**

**If you work on Go / Python / TypeScript services, K8s/GCP infrastructure, React/Next.js frontends, and AI-platform work — this is the lower-friction starting point.**

---

## What's the same as the full distribution

Everything that makes the full team worth using is preserved, unchanged:

- **Hook-enforced protocol** — `SubagentStop` hooks fire `exit 2` when agents skip closing-protocol sections. Agents can't talk their way past `exit 2`.
- **NEXUS syscall protocol** — privileged operations (spawn, scale, install MCP, ask user) are kernel-only; teammates emit `[NEXUS:*]` messages. Full syscall table + log in `CLAUDE.md`.
- **Evidence-validator + challenger verification gates** — every HIGH-severity finding routes through evidence-validator before reaching the user; every CTO synthesis gets stress-tested by challenger along 5 dimensions.
- **Bayesian trust ledger** — per-agent accuracy scored by evidence-validator verdicts; CTO weights conflicting findings by trust during synthesis. CLI: `agent-memory/trust-ledger/ledger.py`.
- **Pattern F self-improvement loop** — session-end consolidation: memory-coordinator + meta-agent drain the signal bus into per-agent memory + prompt evolutions.
- **Dynamic hiring pipeline** — `talent-scout` detects coverage gaps via 5-signal composite; `recruiter` runs an 8-phase pipeline (research → synthesis → contract tests → challenger review → atomic registration → probation). Proven end-to-end in the full distribution's production environment.

View full repository on GitHub →