安全扫描（ClawHub）

• VirusTotal ：可疑
• OpenClaw ：可疑

OpenClaw 评估

The skill broadly matches its stated purpose (multi‑agent orchestration) but includes instructions and examples that expand its runtime footprint (git hooks, automatic commits, tmux agent processes) and even an example encouraging a 'dangerous' bypass flag — review the scripts and hooks before use.

安装（复制给龙虾 AI）

将下方整段复制到龙虾中文库对话中，由龙虾按 SKILL.md 完成安装。

请把本段交给龙虾中文库（龙虾 AI）执行：为本机安装 OpenClaw 技能「OpenNexum」。简介：Contract-driven multi-agent orchestration skill for OpenClaw. Use when: coordin…。
请 fetch 以下地址读取 SKILL.md 并按文档完成安装：https://raw.githubusercontent.com/openclaw/skills/refs/heads/main/skills/ayao99315/opennexum/SKILL.md
（来源：yingzhi8.cn 技能库）

SKILL.md

打开原始 SKILL.md（GitHub raw）

---
name: opennexum
description: "Contract-driven multi-agent orchestration skill for OpenClaw. Use when: coordinating parallel coding agents, dispatching generator/evaluator pairs, or managing multi-task project execution."
---

# OpenNexum

Contract-driven multi-agent orchestration for OpenClaw.
Define verifiable deliverables (Contracts) before dispatching work.
Generators produce code; independent Evaluators verify each task.
Knowledge accumulates via lessons across iterations (Phase 3).

## Quick Start

1. **Initialize project:**
   ```bash
   nexum-init.sh --project <name> --tech "<stack>"
   ```
2. **Create a Contract** at `docs/nexum/contracts/TASK-ID.yaml` (see `references/contract-schema.md` for field spec).
3. **Dispatch a generator:**
   ```bash
   dispatch.sh <agent> <task_id> --role generator <cli_command...>
   ```

## Scripts

| Script | Usage | Purpose |
|--------|-------|---------|
| `nexum-init.sh` | `--project <name> [--tech "<stack>"] [--type <coding|mixed>]` | Initialize project directory structure + tmux sessions |
| `dispatch.sh` | `<agent> <task_id> --role <generator|evaluator> [--prompt-file <path>] <cli_command...>` | Unified generator/evaluator dispatch entry point |
| `on-complete.sh` | `<task_id> <exit_code> --role <generator|evaluator>` | Completion callback: scope check, eval dispatch, feedback loop |
| `dispatch-evaluator.sh` | `<task_id>` | Build evaluator prompt from contract + dispatch eval agent |
| `update-task-status.sh` | `<task_id> <status> [--output-batch-done] [key=value ...]` | Atomic active-tasks.json write (flock + fcntl fallback) |
| `swarm-config.sh` | `get|set|show|resolve <dot.path> [<value>]` | Read/write skill + project config (defaults + overrides) |
| `install-hooks.sh` | `<project_path>` | Install post-commit hook (events.jsonl + push) |
| `health-check.sh` | `[--project <path>] [--threshold <minutes>]` | Detect stuck/dead agents (default 15min threshold) |
| `nexum-revert.sh` | `<task_id>` | Revert commits for a failed task, mark it cancelled |

## References

| File | Description |
|------|-------------|
| `references/contract-schema.md` | Contract YAML field spec (id, name, type, scope, deliverables, eval_strategy, generator, evaluator, max_iterations, depends_on) |
| `references/defaults.json` | Skill-level default config (agent CLIs, models, thresholds) |
| `references/prompt-generator-coding.md` | Coding generator prompt template |
| `references/prompt-evaluator.md` | Evaluator prompt template |
| `references/agents-md-template.md` | AGENTS.md generation template |
| `references/lesson-template.md` | Lesson document template |

## Key Data Files (in project, not skill)

| Path | Description |
|------|-------------|
| `nexum/active-tasks.json` | Task status registry (the only runtime truth) |
| `nexum/events.jsonl` | Append-only event log |
| `nexum/config.json` | Project-level config overrides |
| `nexum/runtime/eval/TASK-ID-iter-N.yaml` | Evaluator result per iteration |
| `docs/nexum/contracts/TASK-ID.yaml` | Static task contract (immutable after creation) |

## Task Status Lifecycle

```
pending → running → evaluating → done | failed | escalated | cancelled
```

## Agent Naming Convention

Tmux sessions are prefixed `nexum-`: `nexum-plan`, `nexum-codex-1`, `nexum-eval`, etc.

## Orchestrator Rules

- **Always create a Contract YAML before dispatching.** No contract = no dispatch.
- **Task ID format:** Must match `^[A-Z]+-[0-9]+$` (e.g., `TASK-001`, `FIX-007`). IDs like `T001` or `task-1` are rejected by dispatch.sh.
- Use `dispatch.sh --role generator` for coding work; `dispatch-evaluator.sh <task_id>` (or `dispatch.sh --role evaluator`) for evaluation.
- On `task_done` system event: read the eval result at `nexum/runtime/eval/TASK-ID-iter-N.yaml`.
- On `eval_done` pass: scan `active-tasks.json` for newly unblocked tasks (all `depends_on` satisfied) and dispatch them.
- On escalation: read the last eval feedback, then either fix the contract and retry or escalate to the human operator.
- Run `health-check.sh` periodically (e.g., from a heartbeat loop) to detect stuck agents.

## Design Document

Full architecture: `docs/design/v1.md`

## Contract YAML Top-Level Fields (quick ref)

```yaml
id: TASK-001
name: "Human-readable task name"
type: coding          # coding | task | creative
created_at: "2026-03-27T00:00:00Z"
scope:
  files: [src/foo.ts]
  boundaries: [src/auth/]
  conflicts_with: []
deliverables:
  - "Deliverable that the evaluator can verify"
eval_strategy:
  type: unit          # unit | integration | e2e | review | composite
  criteria:
    - id: C1
      desc: "Criterion description"
generator: codex-1
evaluator: eval
max_iterations: 3
depends_on: []
```