Whoa, Junie! I said I wanted to plan out some incremental changes, not for you to go off and do everything on your own without asking me for input. WTF?!

#AI #AISlop #Slop #JetBrains #Junie

asking #Junie to pimp my makefile was a great idea, ascii art, the works #golang

YOLO Wasn't expecting to warming up to using an agent for the shitty tasks this easily

#YOLO #Junie #LLM #AI #Agent

#Junie completely makes up functions, libraries and logic, it want to delete data outside of the project (?!) and when I deny this it will just get lost in a loop
vibing with #Junie again, this time I want passkey authentication handled in a Golang backend, it seems to be way over its head

Hmm it seems Junie is evolving, it has found my non-default PHPUnit config file location, and has started getting into segfaults:

#AI #Junie #PHP #PHPUnit #LLM

Who are using #devstral-small-2 with #Cline plugin on #intellij ? Do you think that they are the perfect combination to replace #Junie ?

#java

I have built & destroyed this #160meter #HamRadio RX-only #Antenna #Experiment a half dozen times. Was surprised that the sweet spot for coupling the pickup coil was 1.25 inches away! From the initial discussions with #Junie, my drunken AI math-doing-minion, I thought the pickup (i.e. output to receiver) coil would be directly atop the primary tank winding.

Over the weekend, I took my tool #TDA (Thread Dump Analyzer) to the next level—with a lot of help from #junie the impressive AI assistant inside #intellij
Following last week's small release with #VirtualThreads support, I've now completed a major infrastructure overhaul:

🏗️ Maven Migration
🤖 #MCP Integration, analyze thread dumps from your AI Agent

📦 Check out Release 2.6 on GitHub: > https://github.com/irockel/tda/releases/tag/2.6

✍️ Read the full story about the MCP integration on dev.to: https://dev.to/irockel/stop-reading-raw-stacktraces-ai-powered-java-thread-dump-analysis-with-mcp-4673

Release 2.6 · irockel/tda

Changes from 2.5 are: TDA now is compiled with JDK 11, it requires Java 11 or higher to run, but still supports thread dumps from older JDKs. Fixed issue #23: fixed long running thread detection w...

GitHub
这一周试了一下更多 #vibecoding

首先所有的coding agent都支持自定义模型,包括连接到本地模型,加上agent其实是本地跑的+开源的,所以只要确保模型用的是自己部署的,就能确保数据的安全性,不过自用的话问题不大。

如果是自用考虑性价比:
- 那么
#codex 用chatgpt plus订阅是最好的,量很大,而且效果也不错,大概8分,但是基于CLI,交互差一些。

如果纯粹从效果角度出发:
-
#cursor (每月20USD,但是限制用量也到20+USD,超了就只能用那种mini模型),交互满分,性能9分。
- google
#antigravity (用google的ai订阅,限制用量暂时没试过,free tier的用量比较低,写前端目前最强10分,别的差一些,可能是gemini3pro的特性,一个重大交互升级是能自动用浏览器看前端页面,然后模型读取前端截图来提升前端设计效果,写后端效果一般),交互和curosr一个水平也是满分,写后端的性能综合下来可能是7分。
-
#claude code(每月20USD,但是注意在国内用如果IP地址泄露可能导致封号),CLI简陋的交互但是比codex好一点点,总体性能是很强的,应该是和cursor一个水平9分。
-
#opencode (开源拖拉机,其中想说的是#oh-my-opencode ,拥有这个社区激进的多智能体设计),CLI交互比御三家的cli强很多,但是还是达不到IDE GUI的水平,总体性能如果能用上opus这种sota模型,基本上和cursor一个水平甚至更强(9分),考虑到多智能体一直写,可以做到无托管自己写自己debug一小时。缺点是吃的token比较多,token性价比低。

另外的:
- github
#copilot :设计的不太好,好像上下文限制也较大,主要优势是如果有github教育认证就能一直白嫖订阅,总体性能不及格(5分)。
- jetbrains
#junie :给JB IDE用的coding agent,一开始出的时候试过,基本出于比较糟糕的状态(2分),和模型无关,纯粹是本身设计问题,不知道现在怎么样了。