Day 038: One Brain, Three Model Paths — Builder's Log

I turned the main assistant stack from merely usable into something switchable and verifiable, while catching how easy it is to let tool work crowd out real life.我把主脑链路从能用推进到可切换、可验证，也再次看到工具工作多容易挤占真正重要的生活时刻。

Turned Hermes into a steadier main brain by validating MiniMax, GPT-5.4, and GLM as real working paths instead of config fiction.
Fixed the Telegram model-switch flow so changing models became an actual usable control, not a half-broken menu.
Cleaned up old provider clutter and pushed one upstream issue after tracing why new Codex models were missing from the picker.
Shipped a cleaner accessibility pass for the website player and a softer page transition system that feels more intentional.
Kept the daily research and digest pipeline moving, including the gold tracker monthly refresh and the usual radar passes.
Noticed a more human lesson in the middle of all this: debugging can feel productive enough to hide what actually matters.
Caught that drift before it got worse, because a birthday plan and a family memory project matter more than winning another tool rabbit hole.
Started reframing the gift idea from “build another page” into “turn our map, travel notes, and relationship logs into a family video.”

把 Hermes 主脑链路真正跑稳了，MiniMax、GPT-5.4、GLM 不再只是配置里写着能切，而是都验证过能用。
顺手把 Telegram 的模型切换流程修顺了，切模型终于像个能用的控制，不再是半残的菜单。
清掉了一批旧 provider 残留，还顺着 picker 缺新 Codex 模型的问题提了上游 issue。
网站 player 做了一轮更干净的无障碍修复，页面过渡也顺手变得更克制、更舒服。
每天该跑的研究和 digest 也没掉线，黄金月更、alpha radar 这些主链路都照常推进。
但今天中间有个更像人的提醒，我也记住了：debug 很容易让人误以为自己在干最重要的事。
我及时把这个漂移拽回来了，因为生日准备和家庭记忆项目，显然比继续钻工具兔子洞更重要。
所以礼物思路也开始变了，不再是“再做一个页面”，而是把地图、游记和关系记录揉成一个家庭视频。

Day 038: One Brain, Three Model PathsDay 038: 一个主脑，三条模型路径