logan alpaca

[logan alpaca @loganalpaca]

@tianyi https://t.co/mr9vW09YzU
带Papyrus Desktop来凑个热闹~，这是一款AI Agent驱动的知识管理工具

[logan alpaca @loganalpaca]

@thsottiaux Why does Codex Security get forcibly stopped by safety checks every time I scan my own repository? Your company keeps talking about security, yet you’re preventing developers from using GPT for defensive security work.

[logan alpaca @loganalpaca]

@Aditya_181105 Never trust data sent from the frontend on the backend.🤣

[logan alpaca @loganalpaca]

@o4DT4OO2zY9Oyol @show5560 其实早期也是有隔板的，差不多十年前，现在网约车出来就基本没有了

[logan alpaca @loganalpaca]

@Young_AGI 👍，不过额度能不能提上一点

[logan alpaca @loganalpaca]

@ChinuaNwasogwa @Kimi_Moonshot like-sonnet price

[logan alpaca @loganalpaca]

@thsottiaux You can keep a 5-hour counter to help me track my quota, but please don't actually enforce the 5-hour limit.

[logan alpaca @loganalpaca]

@MPirrongelli @jasondoesstuff yeah

[logan alpaca @loganalpaca]

@jasondoesstuff Actually, Ultra mode is unnecessary most of the time. It can even lead to worse coding decisions because of overthinking. For everyday use, Medium to XHigh is more than enough.

[logan alpaca @loganalpaca]

@philipnft_ @rezoundous Go to Settings → Reasoning Effort and select Ultra.

[logan alpaca @loganalpaca]

阿迪王回家告诉用户：
token涨价了
用户问：
那你会少耗费token吗
阿迪王：
不，你以后会少额度

[logan alpaca @loganalpaca]

@YuG98s 对，而且还有缓存低的原因，两者叠一起消耗还更大了

[logan alpaca @loganalpaca]

核实了一下codex的subagent不使用指定模型而是父模型的问题。

codex的subagent不��使用你指定的模型，而是直接继承你使用的父模型，也就是说，你如果用sol medium，即使你在agent.toml里声明了model，唤起subagent时依然会使用你的父模型。

[logan alpaca @loganalpaca]

Today I used Sol Medium with Luna for the sub-agents, then started coding and watched my quota disappear almost instantly. Later I found out that because the sub-agent model field isn’t exposed, all of my sub-agents had actually been using Sol too
my quota！！！@thsottiaux

[logan alpaca @loganalpaca]

reset!!!@thsottiaux

[logan alpaca @loganalpaca]

workaround方法是将你使用subagents的模型effort和modelid写进agents.md，强制模型写这两个字段

[logan alpaca @loganalpaca]

https://t.co/E0xJhP9wXN
排查提示词参考，你可以把它输入进你的codex来排查子代理问题

[logan alpaca @loganalpaca]

这是脱敏后适配字数的简短报告，原对话��敏感信息发不出来

[logan alpaca @loganalpaca]

实测 Codex 子代理路由异常：explorer.toml 已指定 gpt-5.6-luna / medium，但通过 spawn_agent(task_name="explorer") 启动后，子会话实际仍是父线程的 gpt-5.6-terra / low。说明 task_name 不会自动加载同名 agent.toml，多模型分工与成本优化可能失效，应以子会话 turn_context 为准。

[logan alpaca @loganalpaca]

我的排查主模型配置是terra low，subagents配置是luna的几个推理版本

[logan alpaca @loganalpaca]

因为x不允许过多字数，因此��给重点