@thdxr Solved this for my app by making one MCP running in runtime manager level and giving remote MCPs to sessions/runtimes, but it was kinda annoying to deal with.
@pkqs91@mattpocockuk Tbh, I prefer the other way - description of feature + AC + ways to verify, then make agent implement it. After I get first pass with messy implementation, I cleanup/upgrade it in next few iterations. I find this converging to better results compared to making overdesigned plans
@MCfromNC88@ScorpiusEnt Omens are expensive, just don't use the expensive tablet (extra rerolls) and it's easy profit. Whittling/erasure go for 3-5 divs each, one tablet for omen chance, one for deferral cost reduction + head + freedom cost like 250-300ex per rite.
@SigmaNymous47@media_express_ Out of people I know only stupid ones did the conscription service. It's a sort of IQ test in a way, any self-respecting person would find a way to dodge this waste of time lol.
@thsottiaux@RhysSullivan Any chance we will get refreshed/better spark? It's quite nice to use when the task is simple enough, but it's not smart enough for any more complex tasks :(
@michael_chomsky I can tell how this happened. They started to fuck around with interface, making it basically impossible to gradually control which parts of edit to accept. This was sort of a selling point for cursor - deep integration and control over code, it just stopped working.
@cmuratori "here's why it's not a self-ownage that all these self-proclaimed critics(some are actual artists btw) who think that ai is inferior at core who just made fun of themselves by searching (and finding!) reasons why a real Monet painting is soulless AI gen" type of post lmao
@tabularasa31 I made 3 company subdivision-wide seminars on "how to vibe code correctly, good and bad practices, efficient tools" etc, from what I see that helped some people to start using these tools better.
@theo Mostly for quick adjustments for small things. Like quite often I do main work in codex app, while I have OpenCode in terminal to fast check configs, do sanity checks, cd into some place and spin OC for a quick job there.
@LukeParkerDev Hey, a stupid question: I see that you're showcasing this in chrome here, is this a fair way to test? I'm asking because in my previous OC desktop experience (before and after the electron move) web OC in browser was always way more performant compared to the desktop version.
Crazy that vLLM still has bugs for tool call parsing with streaming. Not just not steaming the tool call argument deltas, but straight up breaking the empty arguments!
Had to use my LLM proxy for parsing instead of relying on vLLM implementation. https://t.co/QqRt57MFen
@RyanLeeMiniMax@MiniMax_AI Love the deep dives like these from labs, always cool to see when researchers do in-detail explanations :) thanks for the post!
Honestly, the openclaw user experience is straight up dogshit. The telegram connect is bugged (I keep getting the access issue like daily), I set up reminder with cron and it never fired, and just now it fully ignored my message. Idk how it got so hyped, just not usable.