boathackathon.com/#featured-…
my shardCI is selected to demo at @BoatHackathon. Fastest shard CI for your repo. ( I will try google neon model tonight and see it can keep improving the performance)
CI was the bottleneck, so we built shardci. We put everything we learned about speeding up CI into a skills playbook and used @boatdotdev to run CI shards on a bunch of machines at once.
Built aboard the Klamath for #boathackathon ⛵
chaorders.com/boat
13
Give this prompt to your fable5.2 high: measure the last 30 days of CICD baseline, then continue improve and climb the performance mountains. General CI p50 should be 2 mins and CD p50 is 3mins. Keep improving. Do not buy any new runners.
@BoatHackatho
29
Everyone needs his own Claude code/mode and codex and linear and more.
最近发现 Raft 有一些深度编程的问题。
在对话过程中,尤其是使用时间比较长了,它可能是因为构造上下文的问题,导致 Agent 注意力越来越倾向于对话,而不是进行深度编程。
比如,当你有两个 Agent 进行协作时,它们总会倾向于以另一个 Agent 的话作为原点进行 case by case 的优化,而不是进行思辨的、系统性的思考。
而如果用 Claude Code 直接在里面去编码的话,其实是没有问题的。
我的推测是因为上下文构造的问题,它在构造上下文时会带上不同的 channel 和一些历史记录,这些东西会导致它的行为偏向于对话而不是深度编程。
我现在的解决方法,就是让 Raft 的 Agent 使用 Herdr 来进行编程,某种程度上是更加把 Agent 当人了
这个方案同时也带来了一个非常大的好处,就是我可以直接在 herdr 里面查看任何一个 Agent 实时编码的状况,并且我也可以进行微管理。
注意力管理也是门大学问。
18
Build this search.chaorders.com/ in one day and did not get a prize in a hackathon. Very very interesting.
15
I helped a builder to reserve a time slot for city inspection by asking codex to spin a check every 20ms and monitor the backend api. He thought he can use Muse or ChatGPT to do that, what he got is every 20mins, he got an email said their is no spot.
AI will make difference of us greater.
1
62
Won 3rd place for obvious.ai hackathon. And they have a great discount only for today @obvious good product and recommend to try
35








