为什么值得关注

它引入了一种基于每 Token 质量的系统化路由策略，支持显式模型感知发送、独立子智能体分类，且不降级到旧模型。

适合谁使用

使用 Codex CLI 或 GPT Work 的开发者
希望自动化复杂代码生成和调试流程的团队
需要对子任务进行精细模型选择的 GPT Codex 用户
优化 AI 辅助开发成本和吞吐量的工程师

典型使用场景

将机械性任务路由到 Luna Low，将微妙的跨模块修复路由到 Terra High
即使父任务使用 Luna，也将安全关键的子任务委托给 Sol
在继续之前显式验证当前模型和推理强度是否与所选路由匹配
使用显式 model 和 thinking 参数 fork 一个 Codex 任务，以确保使用预期的模型

项目优势

每个子任务独立分类，并获得显式的 model 和 reasoning_effort 值
从不降级到 GPT-5.5 或其他非 5.6 模型；仅在 Luna、Terra、Sol 之间路由
当无法原地热切换时，使用最接近的原生 Codex API 等价方式，并如实报告路由
Ultra 推理从不自动选择，且隐藏的 max 从不使用

使用前须知

无法对正在运行的 Codex 主任务进行原地热切换；需要可见的 fork
Ultra 推理仅在用户显式请求时激活，不会自动启用
安装需要手动克隆并运行脚本，没有包管理器集成

README 快速开始

GPT5.6-SOLTELU Model Inverter

Quality-per-token model routing for GPT Codex, GPT Work, and Codex CLI.

English | 한국어

English

What It Does

GPT5.6-SOLTELU Model Inverter selects the most appropriate GPT-5.6 model and reasoning effort for the main task and every delegated subtask:

Work type	Default route
Clear, mechanical, inexpensive to retry	Luna Low or Medium
Normal implementation and bounded debugging	Terra Medium
Subtle cross-module implementation	Terra High or Extra High
Architecture, conflicting evidence, security, data, or production risk	Sol High or Extra High

The objective is not minimum token use. It is the best balance between answer quality, reasoning cost, handoff cost, verification, and avoided rework.

Routing is limited to gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6-sol. It never falls back to GPT-5.5 or another non-5.6 model; it selects another appropriate GPT-5.6 profile or reports the unavailable gate.

Codex-First Routing

Codex is the primary workflow. For nontrivial work, the skill stays in place only when the current model and effort are positively verified to exactly match the selected route. Unknown settings count as a mismatch: when native controls are available, it must create one visible same-directory Codex fork, send the active request with explicit model and thinking, and open that task so the user continues working directly in the selected model.

The current Codex API does not expose an in-place hot swap for a running host task. GPT5.6-SOLTELU Model Inverter uses the closest native equivalent and reports the route honestly. A profile label alone never counts as a model switch; the explicit model-aware send must succeed first.

Subagent Optimization

Every delegated subtask is classified independently. A Sol parent does not force Sol children, and a Luna parent does not prevent a security-critical child from using Sol. Each spawn receives explicit model and reasoning_effort values.

Ultra is never selected automatically. It is available only when the user explicitly requests Ultra in the current task. Hidden max is never used.

Install

git clone https://github.com/AlexAI-MCP/GPT5.6-SOLTELU-Model-Inverter.git
cd GPT5.6-SOLTELU-Model-Inverter
./scripts/install.sh

Use a custom home or preview the destinati

项目描述

Quality-per-token GPT-5.6 Sol, Terra, and Luna model router for Codex, Work, CLI, and subagents.

GPT5.6-SOLTELU-Model-Inverter