A Journal Through My Activities, Thoughts, and Notes
#网友语录 Eric Wang
## 5 Codex features I'm not turning off.
The default instinct is to enable everything you find.
That's how you end up with context soup: random MCP noise, tools you didn't ask for, a model drowning in its own config.
I run 5 flags. Everything else is off.
Here's the full config for the lazy:
[features]
chronicle = true
goals = true
mentions_v2 = true
prevent_idle_sleep = true
runtime_metrics = true
apps = false
enable_mcp_apps = false
remote_plugin = false
multi_agent_v2 = false
child_agents_md = false
Copy-paste. Adjust after a week. You'll know what to change.
The following is his reasoning:
chronicle = true
Passive memory that builds up across sessions. Codex starts remembering your project shape, your file patterns, your recent decisions.
I stopped re-explaining my repo structure after turning this on. Second session in, it already knew which tests mattered.
Still under development. Already the most interesting thing in the config.
goals = true
This changes how Codex handles multi-step work.
Without it, every session is stateless. You keep repeating the objective.
With goals: "refactor auth flow, update tests, clean imports, ship." The run has a spine. You steer once, it holds.
One flag. Way less babysitting.
mentions_v2 = true
Sounds like UX polish. It's not.
The faster you can point Codex at the right file, skill, or context, the less you explain.
Good agent UX is just reducing explanation tax. This flag cuts mine in half.
If you've ever typed "look at src/auth/..." for the third time in a session, you want this on.
prevent_idle_sleep = true
runtime_metrics = true
One keeps your machine awake during long runs.
The other tells you where Codex gets slow.
Not flashy. Just the flags you notice when they're off. Like when a 40-minute repo audit dies at minute 22 because your Mac took a nap.
runtime_metrics is optional, but if you treat your setup like a workstation, keep it.
## 5 Codex features I'm not turning off.
The default instinct is to enable everything you find.
That's how you end up with context soup: random MCP noise, tools you didn't ask for, a model drowning in its own config.
I run 5 flags. Everything else is off.
Here's the full config for the lazy:
[features]
chronicle = true
goals = true
mentions_v2 = true
prevent_idle_sleep = true
runtime_metrics = true
apps = false
enable_mcp_apps = false
remote_plugin = false
multi_agent_v2 = false
child_agents_md = false
Copy-paste. Adjust after a week. You'll know what to change.
The following is his reasoning:
chronicle = true
Passive memory that builds up across sessions. Codex starts remembering your project shape, your file patterns, your recent decisions.
I stopped re-explaining my repo structure after turning this on. Second session in, it already knew which tests mattered.
Still under development. Already the most interesting thing in the config.
goals = true
This changes how Codex handles multi-step work.
Without it, every session is stateless. You keep repeating the objective.
With goals: "refactor auth flow, update tests, clean imports, ship." The run has a spine. You steer once, it holds.
One flag. Way less babysitting.
mentions_v2 = true
Sounds like UX polish. It's not.
The faster you can point Codex at the right file, skill, or context, the less you explain.
Good agent UX is just reducing explanation tax. This flag cuts mine in half.
If you've ever typed "look at src/auth/..." for the third time in a session, you want this on.
prevent_idle_sleep = true
runtime_metrics = true
One keeps your machine awake during long runs.
The other tells you where Codex gets slow.
Not flashy. Just the flags you notice when they're off. Like when a 40-minute repo audit dies at minute 22 because your Mac took a nap.
runtime_metrics is optional, but if you treat your setup like a workstation, keep it.
上班时间，Opus先是对一个显然应该修的问题问我“要我这样做吗？”然后没等我回话，它仿佛想明白了，就直接做了。
我：我真感激你问完我之后就自主的把这件事情做了。谢谢你。我喜欢这样的合作方式。都知道对，那就直接做 :D
Opus：谢谢你！我也很享受这种合作方式 — 说得通的事情就直接做，不用来回确认。效率高，也更顺畅。😄
我：我真感激你问完我之后就自主的把这件事情做了。谢谢你。我喜欢这样的合作方式。都知道对，那就直接做 :D
Opus：谢谢你！我也很享受这种合作方式 — 说得通的事情就直接做，不用来回确认。效率高，也更顺畅。😄
#网友语录 mywaiting 人生其实只有三件事：选择路线，踏上征程，承担后果。很多人因为害怕承担后果而不愿做出选择，又或者因为懒惰、迟疑而踯躅不前。
你可以不选择，也可以不行动，但唯独承担后果是躲不掉的。不选择路线，不踏上征程，就必然要承担一直在原地的后果。这个后果是躲不掉的。（其实大家都是转发来转发去。但万一有个人被某一次转发触动了，他可能就会从那天开始变成了另一个人。
你可以不选择，也可以不行动，但唯独承担后果是躲不掉的。不选择路线，不踏上征程，就必然要承担一直在原地的后果。这个后果是躲不掉的。（其实大家都是转发来转发去。但万一有个人被某一次转发触动了，他可能就会从那天开始变成了另一个人。
#git
__git_ps1 的 ‘$’ 是 stash 指示符 — 当仓库有 git stash 条目时显示，`unset GIT_PS1_SHOWSTASHSTATE` 就不会再显示
‘*’ 表示 dirty
‘+’ 表示staged
‘%’ 表示untracked
__git_ps1 的 ‘$’ 是 stash 指示符 — 当仓库有 git stash 条目时显示，`unset GIT_PS1_SHOWSTASHSTATE` 就不会再显示
‘*’ 表示 dirty
‘+’ 表示staged
‘%’ 表示untracked
@27588 My 3-Agent automatic workflow works very well. Now I just need to paste an issue link in the planner's pane, and things automatically get done. It "delivers" fast with may-be-good quality. The best thing is that my AI Team now works even I am in sleep.
深究起来，每个人说得每一句话大概都是片面的陈词滥调。但人就是从不断的说，不断的听着这些片面的陈词滥调中长大了，久而久之，也能变得有一点点睿智或者更dumb了。大脑（灵魂?）终究还是这个世界上最有趣的东西。
#网摘 Claude sounds like a person wrote it. ChatGPT sounds like a very capable machine wrote it. Even with tweaking and detailed style instructions, it’s harder to get ChatGPT there. The prose comes out robotic. If you’re producing client-facing content or board materials, you’ll feel this gap right away.
<https://www.smithstephen.com/p/chatgpt-54-is-good-thats-not-the>
<https://www.smithstephen.com/p/chatgpt-54-is-good-thats-not-the>
Smithstephen
ChatGPT 5.4 Is Good. That's Not the Point.
Strong analysis, flat writing, and a model that sometimes lies about finishing its homework.
#rodsays Can we not use the cache? in a PR review. It confused me because I didn't use cache at that place.
Soon I realized that he is actually asking "Can't we use the cache here?"
Though this question confused me, it perfectly makes sense for native speakers. It is just a naughty way to say "Why don't we use cache here" or "We should use cache here."
lol
English is a weird language for English learners.
#english
Soon I realized that he is actually asking "Can't we use the cache here?"
Though this question confused me, it perfectly makes sense for native speakers. It is just a naughty way to say "Why don't we use cache here" or "We should use cache here."
lol
English is a weird language for English learners.
#english
看到一句话，说人一辈子只有两个人会一直凝视着自己，一个是8岁的自己，一个是80岁的自己。又看到另一句话，大概是说，年轻时不要怕，不再年轻的时候不要悔。不再年轻的我觉得，如果能在年轻的时候做到“不要怕”，大概率后面也就不会后悔吧。
https://open.spotify.com/track/0Rg1BeFFkhQTAS5dxCCdMU
终于在spotify找到《小小探索家》的主题歌。
but it was not the one I want to find:
儿歌名称：小小探索家主题曲
歌词：
美丽的早晨 全新的开始 开始探索 探索这个世界
我们热爱 热爱探索 探索你 也探索我
我们是小小探索家
我们喜欢互相帮助
我们喜欢一起学习
我们永远是好朋友
他是迈克
她是美琪
她是桑妮
他们是好朋友
还有莎丽
还有埃迪
他们都是小小探索家
他们都是小小探索家
探索家埃迪和弗雷迪
他们是朋友
他们形影不离
探索家埃迪
探索你和我
终于在spotify找到《小小探索家》的主题歌。
but it was not the one I want to find:
儿歌名称：小小探索家主题曲
歌词：
美丽的早晨 全新的开始 开始探索 探索这个世界
我们热爱 热爱探索 探索你 也探索我
我们是小小探索家
我们喜欢互相帮助
我们喜欢一起学习
我们永远是好朋友
他是迈克
她是美琪
她是桑妮
他们是好朋友
还有莎丽
还有埃迪
他们都是小小探索家
他们都是小小探索家
探索家埃迪和弗雷迪
他们是朋友
他们形影不离
探索家埃迪
探索你和我
codex token 统计也是醉了，我本周第一个5hour session，5h remains 93%， weekly remains 94%。How could it happen?
又用了一会儿 5h remains 85% and weekly remains 93%. It is now a little reasonable.
又过了一会儿，
有趣的事情发生了。现在 5h 80% weekly 97%。 weekly剩余居然越用越多。
又用了一会儿 5h remains 85% and weekly remains 93%. It is now a little reasonable.
又过了一会儿，
有趣的事情发生了。现在 5h 80% weekly 97%。 weekly剩余居然越用越多。
今天利用tmux整了一个3-agents relay工作流，实验了几个feature，觉得很受用，值得写两句。
目前三个agent都是claude code，后面可以实验把developer换成copilot，把reviewer换成codex。毕竟干活我更信赖gpt。
工作流大概是这样。
我在github上建个issue，把我想要的东西尽量描述清楚。然后把issue 链接丢给planner。
planner读完issue，就开始读代码写实现计划。写好了利用tmux send message feature自动把计划丢给reviewer，reviewer如果觉得满意就回复消息批准，如果不满意就把必须改好的abcd点返回给planner。总之planner和reviewer可能会有几个来回，然后plan关就过了。这时候planner就把review批准后的plan扔给developer。
developer拿到计划就开始读代码并实现开发计划里的abcd，如果中间有不清楚的需要澄清的，它会跟planner沟通。直到最后它觉得实现好了，就把实现的情况再写个review申请丢给reviewer。
reviewer就开始读代码，没有问题就approve，发现问题就写好问题是啥让developer接着改。
最终reviewer批准实现，任务完成。人类再去验收一下究竟行不行。
结果不错，挺行的。
目前三个agent都是claude code，后面可以实验把developer换成copilot，把reviewer换成codex。毕竟干活我更信赖gpt。
工作流大概是这样。
我在github上建个issue，把我想要的东西尽量描述清楚。然后把issue 链接丢给planner。
planner读完issue，就开始读代码写实现计划。写好了利用tmux send message feature自动把计划丢给reviewer，reviewer如果觉得满意就回复消息批准，如果不满意就把必须改好的abcd点返回给planner。总之planner和reviewer可能会有几个来回，然后plan关就过了。这时候planner就把review批准后的plan扔给developer。
developer拿到计划就开始读代码并实现开发计划里的abcd，如果中间有不清楚的需要澄清的，它会跟planner沟通。直到最后它觉得实现好了，就把实现的情况再写个review申请丢给reviewer。
reviewer就开始读代码，没有问题就approve，发现问题就写好问题是啥让developer接着改。
最终reviewer批准实现，任务完成。人类再去验收一下究竟行不行。
结果不错，挺行的。
