Tommy D. Rossi

@SIGKITTEN @tszzl young jake gyllenhaal

@tszzl who would u have picked

the dune movies were doomed from the start to be good and not great due to the casting of chalamet as paul. he does not have the gravitas for a child-god and is much better suited for kind of silly coming of age movies

@mitchellh for the state right now I only use markdown files and the email labels (make as read or archive). that's everything I need now maybe you could also wrote to a .csv file for more complex tasks I run everything inside Discord with kimaki kimaki.xyz

@mitchellh I made a cli for gmail. I run scheduled tasks to read specific emails, reply following a prompt.md file and mark emails as read works really well for boring tasks and automations github.com/remorses/zele

I'm dying for pi-mono-style minimal library that handles the hard parts of email (auth, syncing with local state, etc.) and gives me an opinionated way to add agentic loops on top of that. I want to build my own agents and logic and guardrails, I don't trust vendors right now.

@thdxr it's not a change in anthropic but an issue with the original plugin implementation which was using a wrong string enum for the auth callback. latest opencode had a change that makes the plugin no longer work cc @rekram11 the fix is very simple github.com/remorses/openc…

there's some confusion because anthropic may have made changes that broke the plugin today this is unrelated to what we're doing - we haven't released 1.3.0 yet

opencode 1.3.0 will no longer autoload the claude max plugin we did our best to convince anthropic to support developer choice but they sent lawyers it's your right to access services however you wish but it is also their right to block whoever they want we can't maintain an official plugin so it's been removed from github and marked deprecated on npm appreciate our partners at openai, github and gitlab who are going the other direction and supporting developer freedom

I set a kimaki scheduled task each day to see latest sentry errors, debug and create a plan to fix them the benefit is that it can run on your own computer or any machine you install kimaki in the session will be visible in discord and you can follow up there kimaki.xyz

I got this working, here's how 1. Created agent-first clis for viewing logs in (@getsentry and @AxiomFM) 2. Sentry sends a webhook to backend on new error 3. Backend triggers a github action 4. Github action triggers codex (authenticated via chatgpt sub) 5. Github action creates PR when codex is done

today I setup my first automation for kimaki.xyz I basically explain what to do in a markdown file. then create a task to read the .md file and run the task. every day at 5pm 𝚔𝚒𝚖𝚊𝚔𝚒 𝚜𝚎𝚗𝚍 --𝚙𝚛𝚘𝚖𝚙𝚝 '...' --𝚜𝚎𝚗𝚍-𝚊𝚝 '𝟶 𝟸 * * *'

@chubes4 @nayshins but we would still pass the linter error to the agent so what's the difference? auto fix could be an option but it's hard to build

@__morse @nayshins Don't you think we should keep AI and linting separated to avoid more slop and instead do this with a mechanical algorithm?

Has anyone documented all the code slop patterns yet? I want to lint for them and banish them to hades.

@wavefnx beautiful

pure means absolute pure describes what can exist within itself, self-sustained absolute

@ku1deep @duckdb is also built on this assumption. makes things so easy to scale

I worked with a CTO, a decade ago, that generated $400mn alpha on the back of basically this insight.

@yacineMTB matjazperc.com/publications/P…

@yacineMTB Accumulative Law for example investment outcome is normally distributed based on luck/skills. but investment outcome is a multiplication ratio of the input accumulation of normal distributions in a multiplicative process is a Pareto distribution

Is there a tangible intuitive statistical explanation for why the Pareto distribution turns up literally everywhere I look?

I built exactly this in August 2025, using isolated-vm to let an agent run code in a js sandbox instead of bash github.com/remorses/holoc… the result was pretty horrible. the agent tries to create super intricate code to do many steps in a giant spaghetti mess I also did this again with playwriter.dev, it's an MCP and cli that can control the browser using js snippets with the playwright API to make it work well I had to show many examples in the prompt and force the model to only use expressions and not statements, otherwise it would have the same issue this only makes sense for small things that depend on the js API, like playwright. for most things bash is the obvious choice it's pretty rare for LLMs to one shot a js script. it usually has to run it once and update it. using an execute tool the model has to rewrite it each time instead of editing it. this ends up consuming many output tokens and annoy the agent. sometimes it just gives up there is also 100x more training data for bash than for js AI companies do RL training on bash. the models expect a bash tool to be always available. the models thrive for the cli

@codetaur I feel exactly the opposite. using opencode

every single thing I try to do using claude code takes about 5x longer than trying to do it using codex, and makes my entire codebase slowly deteriorate. it feels like going back in time.

@thdxr I hope you have a way to measure the performance regression after this and quickly revert it

we've been experimenting with getting rid of the bash tool agents can write js fine which can do what bash can (though some gaps with things like git) and is more cross platform and then could run that in this

@raythurnvoid the spec defines them as files #where-to-scan" target="_blank" rel="nofollow noopener">agentskills.io/client-impleme…

@__morse They can be, is just a matter of how they are loaded in the system prompt no?

skills should have been urls not local files. brain dead spec

@njpatel @seiflotfy @zeddotdev it's because codex is extremely bad. it may be good at a few things but when creating new code it's just garbage

.@seiflotfy and I were talking about the same thing yesterday - both of us are back in the IDE more and tabbing more (@zeddotdev for me). There's one thing about moving fast but if you do care about the design of interfaces, functions, tests, etc - you want to be closer to the code than an instruction. Been using agents more for hyper localized changes when the surrounding code is already well designed.

@tazsingh @Cloudflare now it's impossible to understand what problem you are referring to. what happened?

Needless to say, I've now deleted all those posts. Nobody asked me to, I just did because I truly hate being that person. I'm a "peace love and community" guy at the end of the day ✌️❤️🙌 However, given this experience, I will still be looking to migrate away from @Cloudflare

I'm so grateful to have @threepointone as a close friend ❤️ Thank you for sorting everything for me dude. I literally do not think it would've been possible without you. Honestly a gift to society and a wonderful human being Still, I'm left absolutely gutted by @Cloudflare

@elie2222 it's in the name: why c? because a and b are not enough

YC decided to invest in this Conductor clone. Because 52 clones wasn't enough