Léo Stewart

1st update tomorrow 5pm 🐈

Je vais tester Mistral (Vibe) et quelques automatisations pendant 2-3 semaines on va voir ce que ça donne. Je ferais mes retours sous ce tweet mais je suis hype 👏🏻🇫🇷

@alex_czrl Lent en generation de fichiers, rapport html, copier, déplacer, il m’a perdu des fichiers qui ne le concernait pas etc. En revanche, c’est à mon sens celui qui formate le mieux ses output peut être même mieux que Claude code

@alex_czrl Curieux de voir ce que ça donnera pour toi. Au taff on test différentes solutions, 2 jours ont suffit à calmer deux de mes collègues. Il répondrait souvent à côté de la plaque ou des grosses erreurs et serait très difficile à recadrer. Pour ma part je le trouve super lent en cli.

@housecor Both. Rules or AGENTS.md are not safe enough against AI elucubrations. They take choices, and if WE dont challenge them enough then we definitely need to review their code

We generate code via AI. We review code view AI. Is your team still doing human code reviews? Why or why not?

@benjamincode @romainlanz Aujourd’hui le constat est : niveau models ils se valent, mais pour arriver à un résultat avec Claude code il faut y mettre les tokens, ce qui est de moins en moins acceptable. Codex en revanche a comme défaut de pas être top en design. Tout va dépendre de l’utilisation de l’IA

@LeoWasabee @romainlanz Bah en vrai, j'ai jamais rien eu a lui reprocher a claude, il me fait que des bangers tout le temps, vraiment. Peut être que Codex est mieux mais dans des proportions que je n'aperçois même plus donc j'ai la flemme de changer en vrai !

Ce que tu veux pas du tout voir pendant un hackaton domotique santé de 7 jours chez ta mère qui a Alzheimer... C'est con car j'allais faire un tweet pour dire à quel point Claude Opus 4.8 me carry si fort... A part cette panne, il one shot vraiment tout ce que je lui demande !

@benjamincode @romainlanz Je pense plus qu’il parle context d’ utilisation générale. Mais bon il faudrait demander à Clauclau d’exporter tout ce qui sait du projet dans un fichier qui serait inger’e par codex. Et après ça utiliser un /brain avec Obsidian et /grill-with-docs pour éviter ça a l’avenir :)

@romainlanz Tout le context du projet est sur Claude la. Je vais attendre qu’ils fixent. Ça me fera prendre une pause

@zeeg Ok, thanks! Sorry for the misunderstanding. This area is still quite new to me.

@LeoWasabee this benchmark has nothing to do with that warden.sentry.dev/benchmarking

Here's a baseline quick bench between Anthropic and OpenAI models which I'm now content with. This test runs Warden's stock securitiy-review skill across some known Sentry priors. These are security findings (excluding defense-in-depth) of various levels. I can't explain the changes in Opus 4.7/4.8, but what I can tell you is: its simply not a good product. If a model update breaks usage of a model, I'll use. different model. I'll likely be exploring switching our implementation to run GPT 5.5 (low) for PR review, as that's the right cost tradeoff we're looking for. Generally speaking, the way you can read this is simple: Higher % of known is desirable. Higher total findings is a secondary value add, but cost is the tradeoff. More findings != exclusively more value. It likely means it found more lower sev (or in this case, defense-in-depth-style), that you may not even care to address. In general consider this just an interesting unscientific experiment, one that anecdotally is important but not aimining to be a true benchmark. Read more in the Warden docs: warden.sentry.dev/benchmarking p.s. if you find something wrong in the data please let me know

@conductor_build One feature is missing to conductor ... 🥁 To be able to work on main ! (or at least where the repo is cloned)

How to work with worktrees like a pro

@code Still waiting for codex

📣 Claude Opus 4.8 is now rolling out in @code. Give it a try!

@T_Zahil @ssaulgoodmann69 First of all it uses terminal sessions for your agent, no cli with -p, no sdk. It means all the fuss about Claude code is gone …

@ssaulgoodmann69 How is it better than the others?

Codex VS Conductor VS Claude? I'm currently using Conductor every day, but I have FOMO 😂 What am I missing?

Skills should be: - Concise - Responsible for one thing, not multi-step - Composable - Progressively disclosed - Harness-agnostic What else? Or - what did I get wrong?

@mattpocockuk Si if not responsable for multiple things. Does it means that we should use command for that type on usecase ?

@nikolai_skor @supseset_sh is one of many solutions :)

Constant context switching is a headache for me But if i don't do that, i feel unproductive. I don't know if having 4-6 terminal windows with Claude Code open is actually good for building products. How do you manage that when building?

@FlyaKiet @aboutaaryan @superset_sh Hi @FlyaKiet, I can confirm that my issue about fetching issues and PRs is now working properly on v2 display with @superset_sh v1.8.8 👍

@LeoWasabee @aboutaaryan @superset_sh yes please, I've just released a fix recently as well but once you share error I can reproduce on my side as well

if youre not using this, youre dusted

@FlyaKiet @aboutaaryan @superset_sh It seems that v2 proxies GH issues and PRs lists but it never works. But no problem at all when switching back to v1 I can give you errors tomorrow.

@LeoWasabee @aboutaaryan @superset_sh Hi sorry to hear that. what’s the issue you’re seeing with issues and PRs can fix with some details

@jamesqquick Yeah, more that Pro's limits run out really fast with Claude Code . After a couple of days of real use you're already rate-limited. In practice Max felt like the actual entry point if you wanted to use it seriously. Sorry if I wasn’t able to make myself clear the first time

@LeoWasabee Meaning you think you already need a higher tier?

This is kinda crazy right? What’s the rational?

@zeeg Only for organisations ?

its under capabilities btw

"lets make a help page to enable a feature and not even tell people how to turn on the feature" support.claude.com/en/articles/14…