@thdxr@thsottiaux good call, this is one of the most annoying and intrusive things about CC and one of the reasons I don't use it and think Anthropic is sus
they also made it hard to opt-out, the official docs' way didn't work for me. but maybe it's just sloppiness idk
@thdxr i have this idea that as llms reach their potential limit, they will get commoditized and there will be no clear winner in the provider world, much like airlines. it was lucrative to be the first to market, but now you just compete on price and marginal efficiency gains
@AdmetosL@myhiddenvalue@grok Yeh here it is.
The bottom sign "Tel Aviv" reads "Knor Lon" in Hebrew (what?). The top sign's Hebrew and English text have nothing in common (and Hebrew doesn't make sense, ha-pyokim?) Also there's Hebrew and Arabic on the 2nd line where it's usually just Arabic.
@AdmetosL@myhiddenvalue@grok There was a fuller non-cropped version of this photo circulating with visible street signs in Hebrew and the signs are just nonsense.
@thdxr btw what's going on with context size indicator in codex 5.2? it jumps like from 30k to 140k to 40k during one turn, no compaction. is it some internal thinking artifact? esp. visible with xhigh thinking
Для себя Claude Code за $100 в OpenCode, на работе Cursor и ещё есть Codex который идёт по подписке с ChatGPT (использую оба).
Всё всегда в один поток — пробовал параллельно, не успеваю просматривать код и очень сложно переключаться между задачами, понял что это зло как и любой мультитаскинг.
Из MCP использую только Notion, очень удобно работать с планированием/задачами/документацией — пока один агент пишет код, я могу со вторым прорабатывать какие-то концептуальные моменты, писать планы, разбивать на задачи и тд.
I tried doing something like this, an LSP, but shaped for agents, not IDEs. Basically index tool to get an overview of workspace/directory/file symbols, read_symbol, write_symbol
The problem is that the models are trained to work with a finite set of built-in harness tools (Read, Write, Grep etc.), and unless you ask them to, they don't use external tools.
@FarhadNawab@adamdotdev Are you using it with opencode? When I tried it it just went in rounds for 40m about not being able go use a list tool and planning to edit pyproject.toml file. After 3 compactions I just stopped it.
@thdxr Would be great to have a /plan cmd that will open the "current" plan in .$EDITOR
and also make plan file location configurable? like don't save at all, save in ~/.opencode, save on spec/plans in project etc
@kcosr@parcadei I think the models are finetuned on built-in tools, need a stronger signal to prefer non-standard ones.
A path. forward would be to define the protocol and capabilities that make sense (similar to LSP). I'll experiment with Go or Python and try to see what works.