convez.ai
MCP ServerDevOps & InfraestruturaAvançado
AI-wuji

Wuji Legion for Codex

AI-wuji/wuji-legion-codex

Kernel de execução unificado Codex com roteamento singular e auditoria integrada

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que implementa um kernel de execução em sistema nível para o Codex, consolidando múltiplas capacidades em uma única cadeia principal com ativação esparsa (MoE), portas determinísticas em Go e auditoria independente. Otimizado para reduzir loops, retornos de trabalho e consumo de tokens sem comprometer acurácia ou qualidade de entrega.

Instalação

$npx convez-ai AI-wuji/wuji-legion-codex

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Orquestração de múltiplas habilidades em cadeia principal única
  • Otimização de consumo de tokens em fluxos complexos de IA
  • Execução determinística com gates de verificação em Go
  • Auditoria independente e rastreabilidade de decisões
  • Análise de causa raiz com hipóteses paralelas

Veredito

Solução arquitetural sofisticada para quem precisa consolidar múltiplas modelos de IA em um sistema determinístico de baixo custo. Exige profundo entendimento de orquestração de agentes e token economics; não é para implementação casual.

README

A tradução para o português deste README ainda está na fila.

Wuji Legion for Codex / 无极军团 Codex 版

A Codex system-level meta-skill: it atomically distills the strongest parts of many skills, then routes, mounts, audits, and verifies them through one main chain.

无极军团是 Codex 的系统级总 skill:把众多 skill 的强项拆成原子后优胜劣汰、融合重组,再由唯一主链统一路由、按需挂载、审计验证和交付。

Current version: v11.3 Truth source: kernel-source.json

Why Wuji Legion

无极军团不是“多角色聊天皮肤”,也不是把一堆 skill 叠在一起的合集。它是一套给 Codex 使用的系统级总 skill:把 OpenSquilla、Headroom、Reasonix、PPT、调研、审计、根因修复等来源的有效能力拆成原子,经过查漏补缺、优胜劣汰后,融合进唯一主链。执行时只按任务需要挂载能力,再用 Go 底座做确定性门禁和真实验证。目标很简单:少兜圈、少返工、少烧 token,但不牺牲命中率、证据链和交付质量。

Wuji Legion is not a role-play prompt pack, and it is not a pile of separate skills. It is a Codex system-level meta-skill: useful capabilities from OpenSquilla, Headroom, Reasonix, presentation work, research, audit, root-cause repair, and other sources are broken into atoms, filtered by gap-fill or replacement value, and fused into one main chain. At runtime, only the needed atoms are mounted, then deterministic Go gates and real-run checks verify the result. The goal is simple: fewer loops, less rework, lower total token cost, without weakening accuracy, evidence, or delivery quality.

Core Selling Points / 核心卖点

  • Single main chain / 单一主链: one routing authority, one merge point, no competing agent headquarters.
  • Codex-only execution / Codex 内执行: OpenSquilla-style strengths are distilled into atoms, not kept as an external executor.
  • MoE-style sparse activation / MoE 式稀疏激活: only the owner profile, triggered officers, selected skill, and evidence handles enter context.
  • Low total cost, not fake savings / 优化总成本而非假省 token: concise_execution_contract gates cached, fresh, output, uncached tokens, retries, and tokens per success.
  • Execution budget contract / 执行预算契约: FAST_REPLY, LIGHT_TASK, STRUCTURAL_TASK, and RELEASE_TASK stop small tasks from expanding into full-system scans.
  • Independent officers / 独立审查官: white-hat, guard-office, root-cause officer, audit, quality-inspection, performance benchmark, and compliance stay separate and on-demand.
  • Root-cause over patching / 根因修复优先: root-cause officer, root-cause-radar, parallel-hypothesis-fanout, and patch-debt-root-cure target causes instead of stacking local patches.
  • Deterministic Go gates / Go 确定性门禁: wuji-cli provides repeatable audits, routing reports, context packing, benchmark checks, MCP guards, PPT gates, and closeout checks.
  • Real-run closeout / 真实运行后收口: completion claims require current evidence, not just explanations or plans.

Architecture / 架构

flowchart LR
    A["User task / 用户任务"] --> B["task-routing"]
    B --> C["capability-mount"]
    C --> D["deterministic-execution"]
    C --> E["On-demand officers / 按需独立官"]
    E --> C
    D --> F["Verified delivery / 验证后交付"]

The fused kernel has exactly three layers:

  1. task-routing
  2. capability-mount
  3. deterministic-execution

无极军团只有这三层主链。它不允许第二套路由、第二个用户入口、第二套外部执行系统,或绕过主链的并行总控。

Wuji Legion has only these three main-chain layers. It forbids a second router, a second user entry, an external execution system, or a parallel command authority.

What It Distills / 它吸收了什么

Wuji Legion keeps useful ideas as atoms and rejects project-shell stacking.

无极军团只吸收原子能力,不叠加外部系统外壳。

  • OpenSquilla-style task detection, capability gap detection, skill filtering, MCP awareness, and memory/session routines.
  • Headroom/Reasonix-style token and hit-rate discipline.
  • Prior-art search before invention.
  • Root-cause analysis, patch-debt cleanup, and parallel hypothesis fanout.
  • Context compression with evidence handles instead of replaying whole logs.
  • Guarded MCP/plugin intake and source/security/compliance checks.

Marker: distilled_atom_kernel.

Independent Officers / 独立官

Independent officers are explicit review seats, not simulated tones:

独立官是真正的审查席位,不是语气模拟:

  • white-hat: challenge assumptions, catch fake fusion, check weak evidence.
  • guard-office: screen external pages, repos, MCP manifests, plugins, scripts, dependencies, and assets.
  • root-cause-officer: judge failure causes, low efficiency, rework, and patch debt.
  • audit: check process truth, gates, and evidence chain.
  • quality-inspection: final acceptance and real-run verification.
  • performance-benchmark-on-demand: token, hit-rate, speed, cost, and resource measurement.
  • compliance-on-demand: license, source, privacy, attribution, and release boundary.

They can reject, return, or set release conditions. They do not directly edit files or become a second commander.

Token And Context System / Token 与上下文系统

Wuji Legion is built around a sparse context governor:

无极军团的省 token 不是“少看证据”,而是稀疏装配:

  • hotpath-manifest.json: resident, on-demand, cold-ledger, and forbidden-resident surfaces.
  • context-bloat-audit: rejects high cache-hit runs when cached volume, fresh input, output, loaded bytes, or activated roles exceed budget.
  • runtime-context-audit: required only for token, cost, cache, backend usage, or outer-context claims.
  • bench-report: binds the measured workspace, command, timestamp, log reference, and usage hash.
  • Raw prompts, messages, content, secrets, sessions, keys, cookies, and tokens never enter resident reports.

Optimization markers: fusion-audit, optimization-audit, context-bloat-audit, runtime-context-audit, execution_budget_contract, analysis_completeness_contract, complete-materials-before-architecture-analysis, reversible-evidence-handle.

Typical Use Cases / 适合场景

  • Codex workflows that keep getting slow, bloated, or over-routed.
  • AI coding tasks that need root-cause repair instead of another patch.
  • Research tasks that need GitHub/source evidence without unsafe external execution.
  • PPT, HTML, UI, visual, and document delivery that needs real QA gates.
  • MCP/plugin adoption where security, privacy, source, and license boundaries matter.
  • Teams that want AI agents to be auditable, sparse, and disciplined.

适合这些场景:

  • Codex 经常兜圈、跑偏、上下文膨胀的项目。
  • 修 bug 不想继续打补丁,而是要快速定位根因。
  • 调研需要 GitHub、源码、证据句柄,但不想接入外部执行系统。
  • PPT、HTML、UI、视觉、文档交付需要真实质检门禁。
  • MCP/插件接入需要安全、隐私、来源、许可审查。
  • 希望 AI agent 更可审计、更克制、更稳定的团队。

Quick Start / 快速开始

Clone and verify / 克隆并验证

git clone https://github.com/AI-wuji/wuji-legion-codex.git
cd wuji-legion-codex
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\ensure-wuji-cli.ps1 -RepoRoot .
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\test-wuji-cli.ps1

Expected result:

RESULT: PASS - wuji-cli deterministic gates

Install as a Codex skill / 安装为 Codex skill

For supply-chain safety, the installer requires a pinned 40-character commit SHA.

为了供应链安全,安装脚本要求使用固定的 40 位 commit SHA,而不是浮动分支或 tag。

powershell -NoProfile -ExecutionPolicy Bypass -File scripts\wuji-install.ps1 -Ref <40-char-commit-sha> -Bootstrap -InstallAgents

Install target:

%USERPROFILE%\.agents\skills\wuji-legion

Go Execution Base / Go 执行底座

wuji-cli is the deterministic base of Wuji Legion. It is intentionally boring: gates, audits, routing reports, benchmark reports, context packs, MCP guards, PPT checks, and closeout checks.

wuji-cli 是无极军团的确定性底座。它不抢主链,只负责可重复的门禁、审计、路由报告、基准报告、上下文装配、MCP 检查、PPT 检查和收口检查。

Common commands:

.\.wuji-tools\wuji-cli.ps1 canon-report --report outputs\canon-report.json
.\.wuji-tools\wuji-cli.ps1 route-task --config config.json --query "fix bug" --report outputs\route-report.json
.\.wuji-tools\wuji-cli.ps1 fusion-audit --workspace . --report outputs\fusion-audit-report.json
.\.wuji-tools\wuji-cli.ps1 optimization-audit --workspace . --report outputs\optimization-audit-report.json
.\.wuji-tools\wuji-cli.ps1 context-bloat-audit --workspace . --report outputs\context-bloat-audit-report.json

Closeout Sound: non-fast tasks may schedule scripts\beep.ps1 complete -SpawnDelayed -DelayMs 1200 before the final response. Sound failure is non-blocking.

Repository Map / 仓库结构

kernel-source.json          structured truth source
SKILL.md                    Codex skill mirror
GLOBAL_AGENTS.md            global runtime mirror
config.json                 routing and provider overlay, no secrets
tools/wuji_cli.go           Go deterministic execution base
scripts/test-wuji-cli.ps1   full deterministic gate test suite
units/                      unit mirrors
experts/                    owner and officer cards
hotpath-manifest.json       sparse context and hot-path budget policy
fusion-matrix.json          distilled atom decision ledger

What Wuji Legion Is Not / 它不是什么

  • Not a second AI agent platform.
  • Not an external OpenSquilla runner.
  • Not a prompt pack that relies on long resident context.
  • Not a swarm where every role speaks all the time.
  • Not a tool that stores your accounts, keys, tokens, cookies, addresses, or private sessions.

它不是:

  • 第二套 AI agent 平台。
  • 外部 OpenSquilla 执行器。
  • 靠超长常驻提示词维持质量的 prompt 包。
  • 每个角色都常驻发言的群聊式 swarm。
  • 保存账号、key、token、cookie、地址或私有会话的工具。

Release Discipline / 发布纪律

A completion claim is valid only after current gates pass:

发布或完成声明必须有当前门禁结果:

  • fusion-audit
  • optimization-audit
  • context-bloat-audit

Token, cost, cache, backend usage, and outer-context claims also require runtime-context-audit with numeric-only usage evidence.

Contributing / 参与贡献

Useful contributions are welcome, especially:

  • smaller hot-path context
  • better deterministic gates
  • safer MCP/plugin intake
  • stronger root-cause workflows
  • better bilingual docs
  • reproducible benchmark evidence

欢迎贡献,尤其是:

  • 更小的常驻上下文
  • 更强的确定性门禁
  • 更安全的 MCP/插件接入
  • 更好的根因修复流程
  • 更好的中英双语文档
  • 可复现的基准证据

License / 许可证

See repository license if present. Third-party source pools and distilled ideas must keep their own license, source, privacy, and attribution boundaries.

如仓库包含许可证,请以仓库许可证为准。外部 source pool 和被蒸馏能力必须保留各自的许可、来源、隐私和归属边界。

Support / 支持一下

If Wuji Legion helps your Codex workflow, a small donation keeps the project moving.

如果无极军团帮到了你,欢迎随手赞赏,支持继续打磨。

Ver o original no GitHub

Bom para

Relacionados