zryph

Steve McManaman and Ian Darke are the most biased commentators in the game. I mean that McManaman guy is so bad he can’t even hide it anymore. @ESPNFC @espn

@steipete agent-scripts is an absolute gold!

@olivercamp_ @sama 🤞🤞🤞

@FCB_Cartel @sama Fahh ran a 32h goal sesh hopefully it’s good

Codex in the last 3 days has been a nightmare. Hallucination after hallucination, writing so much unwanted code, creating things that are not at all intended, taking too much time and burning tokens unnecessarily. Especially /goal. @sama something is wrong. It burned 50% of my Pro plan tokens for the week in 1 night creating a PR of size 104k lines and when asked about the PR : it said it’s useless and cleared the whole PR and closed it.

@EliotGevers @sama What’s the secret?

@FCB_Cartel @sama experiencing the opposite lol, keep going bro

@ProfAdebay @sama Yes it is very good! I use Codex&Claude together with cursor as IDE, Codex has been pretty good until it started behaving oddly since yesterday.

@FCB_Cartel @sama Codex is working well for me. Maybe it’s because I know my way around it. (I didn’t feed it any skill files)

@lionhead_king @sama 🥲🥲

@FCB_Cartel @sama I deleted 10,000 lines of code yesterday. Codex been totally ignoring Agent.md, very weird.

@slopwareindy @sama Will read it! Thanks:)

@FCB_Cartel @sama A bit more upfront effort can go a way in getting decent results. This iterative code design loop once before any request of medium+ complexity will drastically reduce unfavorable results. Also have an improved version I'll post up soon. x.com/slopwareindy/s…

@sheathinkler @sama I used /goal for building an incredibly big project that had to go through like 15 repos and call 11 different MCPs to build flows. It went on and on for 3 full days and built 300 flows precisely. After this I got good confidence to use /goal

@FCB_Cartel @sama Write your own objective mode and recompile the app - keep the focal loop local and don’t get hit by whatever they are screwing with behind /goal. Theirs is omega broken

@HotTakeHawk @sama 🤷‍♂️

@FCB_Cartel @sama Got the same issue , deleted my whole bot trading folder and said sorry after 🫠

@FCB_Cartel @sama This happened to me as well. I gave it a clean set of instructions got stuck in one part of my workflow, looped that that 30 times over the span of 24 hours, came back 24hrs later, nothing useful all garbage.

@catalinbanu @sama

@FCB_Cartel @sama "when asked about the PR : it said it’s useless and cleared the whole PR and closed it." Well that were best invested money! Think if that code was in prod :D

@cyganztrojkatu @sama Good for you! Do you run /goal? What do you mean advanced workflows? Do these workflows happen in one coded session? Or are you talking about multi-agent workflows?

@FCB_Cartel @sama Am I the only one not experiencing any issues? Can anyone relate? I have pretty advanced workflows and I felt no regression at all

@simpsoka @sama After this it went wild. Full night, no stops. Adding sloppy code till I stopped it myself in the morning as I wasn’t expecting that much code at all.

@2034_SWE There was a pre exiting PR, and there were some P1 findings in the PR. I asked to fix them, and gave an answer to an earlier question. After this prompt: it went wild, back to back, whole night.

Idk what the full picture and original user prompt looks like, but based on that reply, I’d imagine: - insufficient/ineffective context engineering - imperfect prompting skills - and bad/no document architecture (for efficient & autonomous progressive disclosure) I have never had an issue like this even once Ever And I’ve had Codex running 24/7 for last 3 days on 50 different things I’ve told it

@OliwierMako @sama Good idea! But this only works when the human is right there waiting for checkpoints, but i usually run /goal when i go to sleep to get some boring tasks done while i'm asleep.

@FCB_Cartel @sama This is why agent workflows need rollback points Not just better models Every 15 min it should leave a clean checkpoint you can inspect or nuke

@andrei_ot @sama I actually switched to Opus 4.7 today, and Opus's time for first byte was 4 mins.

@FCB_Cartel @sama getting the opus 4.7 experience I see, its coming full circle, can't wait for 4.8 to be great for 2-3 weeks

@boneuou @sama Could be /goal related issue in particular. I am not explicitly asking it to use parallel sub-agents rather than doing on its own. Still feels wrong.

@FCB_Cartel @sama He's taking 1h10 to do something easily on fast mode, /goal burnt more than 30% of weekly usage on pro (x20) in 8h ... Everything is broken on it

@FCB_Cartel @sama @thsottiaux I have the same observations.

@FCB_Cartel @sama I had so many issues yesterday that I ended up closing my laptop and stepping away for the entire day, which had never happened since I’ve been using codex. It’s definitely screwy right now

How come the time for the first byte on Claude code is dead slow? Sometimes it’s over 3 mins with 0 response!