tryingEveryThing

@MaximeRivest md.collab?

Markco is for the Markdown collaboration With AI and people and sound like a name so the idea is that the AI collaborator is named Markco. But I dont love either too. I tried: Atelier. For a while, workshop in French or feuille (sheet in French) But naming this one has been hard for me. Mrmd was a temp name maximerivest's markdown. But I don't find myself finding what to move to.

I can't decide between mrmd or markco. do you have a strong preferences? why?

@trashpandaemoji in the tools we use there is always a progression. You just find the right tool for the job. and Pi is definitively the tool for the exploration and prototyping phase. But when you figured it out, and worked quite well with all the things nico put out. Then it is time to dspy it.

I used it a bit for GEPA, but I never really put 2 and 2 together for our specific use case. This definitely makes way more sense for something more controlled and deterministic like report generation. It would be much easier to implement in our pipeline too. Thanks for the tip!!

I wonder if you could take Pi, package it with skills and prompts, then make it an executable binary with Bun and distribute it? Thinking of doing this at work for a specific agentic workflow with a well defined input and output. Provider is solved via our intranet, and it’s better than just distributing a skill since we might want to run this programmatically in a pipeline.

@trashpandaemoji if you dont like python, there are multiple different reimplementations in ts, go, rs ;) For getting started: dspy.ai/community/comm… dspy.ai/community/comm…

@tryingET Ah i see what you're saying now! I'm not that familiar with dspy, will check that out.

@trashpandaemoji I get that, but I still dont get why you just do the systemd or cron job with a run command of dspy. If you want something established it should be an established program with clear signatures. So the structure is 100% repeatable. Use the right tool for the job.

Yeah that would be us iterating on the prompt which we could theoretically do. For distribution, its mainly so we can run this thing hands-off and get a report. I don't want to have to do this on my laptop weekly and distribute the report. Ideally I'd stick it in our existing data pipeline or cron system, it would run, spit out the report into S3 and then get distributed to necessary parties. The hard part is that this is an agentic process, something akin to: Look at this data, extract trends, here are the trends from last week, how are things changing? Anything new?

@trashpandaemoji True. But don't you want to have a weekly job to be reliable and improve over time? I don't get the nned of distributions. But that is probably a me thing

Isn't GEPA mainly to optimize the prompt? I'm mainly concerned with distribution. For my use case, I built a flexible agentic workflow that outputs a report. Distributing a skill is cool for ad-hoc runs that people can do. But what if we wanted to run this weekly? I haven't seen too many ways to encapsulate a unit of agentic work. 🤔

@VictorTaelin Just let pi call codex exec search?

To elaborate - I never liked the "Linux way": > ship a half assed software, and let the user modify it Instead, I always bought the Apple way: > pay us and we'll give you the best possible defaults This worked for me, because I wanted to spend my time writing compilers, NOT fixing driver issues. So, when people told me "Pi comes with just the bare essentials and you can add what you want", that definitely did NOT paint a good picture. But it is different. The time to modify is minimal. "Pi, extend yourself so that I can spawn sub agents in a specific way that works Bend2's prelude" One or two prompts later, and it is done. It modifies itself for what I need and I suddenly have a new tool to help me get things done. It just works. You can't do that with Claude Code and Codex. That said, I'm still not sure that'll always work. How the hell do I make my Pi browse the web, for example? Seems like the author doesn't want it, it is definitely important, and there's no easy / satisfactory way to add 🤔

Ok so I thought that was a dumb gimmick but now I'm completely sold on how pi is a self-modifiable software. It literally knows how to modify itself very cleanly and that's extremely useful in practice I'm not using Codex / Claude Code anymore Bend2 should definitely be like this! I mean, constructed in a way that AI's can easily navigate it and know how to modify it to add any feature the user wants. Perhaps we're past the era of open source software and into the era of forkable software, where the most hackable project wins?

@unclebobmartin than change your mutation testing before hand based on the new behavior you want to see.

Mutation testing has a dark side. Not only does it consume rather large amounts of CPU and wall time; but it makes is much more difficult to remove old behavior and replace it with new, "better", behavior. Those extra tests do their job of stabilizing the behavior very well -- perhaps a bit too well. Caveat Emptor.

Maybe something for you @badlogic

@VictorTaelin Finally arriving. Was counting the days you will switch.

@doodlestein This might work with this inflow you have currently. But when you receive 100s a day it will get harder

I know people get really bent out of shape by the onslaught of low-quality "slop PRs" and GitHub issues in open-source projects, but when you automate the entire process, it doesn't bother you at all. Note the 2 "Spam/nonsensical issues" that Claude closed with extreme prejudice:

@sawyerhood Make a pi extension instead please

does codex have plugins yet? i'm working on a revamped version of dev-browser and want to make sure that i support them (if they exist)

@mkurman88 no. have a look at all what @doodlestein build. he is the agents guy. his complete guide: agent-flywheel.com/complete-guide He has about 60 as intense or even more that will change the game. completely

@tryingET Is it your project?! nice!

Okay, after 4 hours it produced tui; not good, not bad either :D But that was a test. I've made 11 key observations, and now Codex is working on improvements to the entire goal-oriented stack. I used GLM-5-TEE via Chutes. Unfortunately, the agent was idle for about 55 minutes due to rate limits.

@mkurman88 Glad you like it. If you have any issues just post an issue there. The ai will take care of it in 1 to 3 days. Usually this is the turnaround for me for his projects.

@tryingET Bro, I had no idea this existed! Frankentui is like a honeycomb for my FrankQwenStein soul 😅

@itsyourcode Different parts of a system to build the system for teaching kids based on their individual needs. Middle school teacher

@tryingET Then VSTO/HTN is your "verifier" step. My process conceptually the same, but my verifier is reviewing the code. It is great that you have found a method that is working for you. What are you building?

Agentic coding replaces typing, not thinking. Happy Wednesday!

@LLMJunky Why not just use niri? I don't get it

I got so tired of everyone raving about how great cmux is. Panes this. Browser that. EXHAUSTING. And that's because I'm on Linux, where we get none of the coolest toys. So...I built it myself. And my God. You were right. It's amazing. Introducing Limux, a a GPU-accelerated terminal workspace manager for Linux, powered by Ghostty's rendering engine, with split panes, tabbed workspaces, and a built-in browser. Think cmux, but native Linux. If you're interested in something like this, be sure to leave a comment and I'll release it. Special thanks to @manaflowai and @mitchellh for making this possible.

@itsyourcode True, but this was just an example. My way of doing is: envision the correct version of your intent. Drill down on what you want to built exactly. The end goal. Is it a cli, core plus adapter or whatever. And then you go in an vsto / htn approach until you are done.

@tryingET Yes, that is a narrowly scoped example. Large systems exceed not only the largest context windows many times over, but also the attention capacity of the model within that window (which is considerably less than the window itself).

@itsyourcode Also we live now in the time where we will see faster and faster iterations of cheaper and better llms come out. Just habe a look at qwen3.5 and minimax 2.7 We are off to the races. Self improving llms based on auto research and improved versions from karparthy

@tryingET That which you cannot verify cannot be reasoned about as the system grows larger. Eventually the LLM suffers this too, and its ability to express your meta-intent degrades. Then who will fix it?