Howaboua

Alright folks from the Pi world, my caveman Pi setup and a bit about the workflow. AMA. Extensions: - @howaboua/pi-codex-conversion - I use GPT models, and I want them to have access to their native tools. This gives them that. Whether it's better or worse is up for debate, they work either way. Subjective. - @howaboua/pi-auto-reasoning-tool - lets the agent change it's own reasoning. It's got some predefined templates of when it should be invoked. Resets reasoning to low on every agent_end. - @howaboua/pi-auto-trees - a Theo-stream-twitch-inspired variation on Pi's trees. /marker to set a starting point. /end to summarise conversation and advance the marker, with some changes to summarisation prompt so it's more about what was accomplished and less about details. This way you don't have to go into the trees and track where the heck you should go back to. - @howaboua/pi-subagent-review - /review command that performs a subagent codex-style review in a separate ephemeral, no-context session and outputs the results. Automatically selects what to compare based on the branch. Then the main agent automatically addresses the results. Has some guardrails to prevent suggestions that were architectural decisions, so they don't get automatically treated as fixable. - @howaboua/pi-semantic-grep - adds a tool to look up code by natural language. Does indexing as it goes. Requires an embedder model running on your machine. I use Gemma 300m. - @howaboua/pi-vent - theoretically enables the Clanker to complain about its struggles with your environment. I am yet to see it use it, maybe my environment is perfect. - pi-explore-subagent - local/custom. Very simple two-modes subagent running shallow (spark)/deep (mini) modes. - @howaboua/pi-markdown-workflows - adds /skills & /workflows (GUI), /learn and enables reading nested agents.md files. Bundles a basic skill creator prompt, a tool for the agent to create workflows (which are skills, but on a repo-level and more SOP/procedure focussed than fluff), and /learn instructs the agent to create a workflow and create a tiny agents.md in a relevant place (needs a bit of a prompt-refresh, not using it much, just letting the Clankers cook). Honourable mention goes to @aidenybai's React grab which is absolutely essential when working with React-based apps. Up to the point where if you're working on any thing that can use react, you probably should, no matter if it's an overkill. github.com/aidenybai/reac… Skills: - agent-native-hardening inspired by @nummanali - less godfiles, better structure, DRY, scoring system. Use it after every major feature implementation (TS but if you ask the agent to apply similar principles to whatever language, it also "just works". gist.github.com/IgorWarzocha/b… - anti-ai-copy - some guidelines for anti-slop writing. Kinda works. CBA to refine. And everyone's style differs anyway so not worth posting. - chrome-cdp by @xpasky - customised version the one in the link. Minimally, just to work with my env. Skips the BS of having to use middleware. I don't use it often because Clankers are not good ad judging what good UI/UX is, but works great for browser control. github.com/pasky/chrome-c… - gh-issue-pr-flow - says how I like my GH Clanked up. Asks to post a comment "@codex please review this PR and give me 10-20 issues if any. Categorize findings as required, recommended, or optional." for every PR. This enables to squeeze more juice out of the review bot. Beware, it might get nitpicky so you gotta use your judgement. - impeccable by @pbakaus - self explanatory. banger. github.com/pbakaus/impecc… - make-interfaces-feel better by @jakubkrehel - for these UI/UX microdetails github.com/jakubkrehel/ma… - skill creator skill by yours truly - based on shitloads of web research, my own findings, Anthropic's guide to skills, etc etc. gist.github.com/IgorWarzocha/e… - omarchy-help - completely custom skill (do yourself a favour and remove the default one, it contains bloat that might not apply to your environment) with references for my machine, updated whenever things change and containing reference for issues and fixes. Don't sleep on having a skill like this for computer use and management. I don't touch any Linux settings anymore, my Clankers even tidy up my boot records for me. Running 3 machines like that since July 2025, no issues. Honourable mention to @noizynerd's aka Opencode-Snippets-JosXa's strict typescript rules stack gist.github.com/JosXa/bddbf281… I tend to post any features as issues on GH in a separate session so when I notice things, I can then resolve it in one pass. Most of the time I run 3-4 sessions. Hardly ever work on two things in one project. If something is really small I just ask the agent to create a separate worktree and work in it (skipping /review because it wouldn't work), no fancy tooling around it or folder structure required. The workflow using these is roughly, per session: - I fire up a pi session. "Familarise yourself with the repo, use subagents". - "Let's discuss issues X, Y, Z, grab these from GH" or even "Have a look at open issues let's organise them into groups and see what we can tackle in a single PR" - /marker to set the point in session where the agent has basic context and has read some relevant files. - "Alright, let's tackle issue X" - /review, fix whatever is worthy - /review again. - "Post a PR to GH". - copy paste whatever the Codex bot found. Maybe go back to manual /review if issues are bigger than they seemed. - At some point during this, manual QA and trying to break the feature. - /end - this summarises the implementation so the agent knows what happened, and advances the marker. The agent has basic context and can now work on feature Y. - Rinse & repeat. - After big changes, ask the Clanker to use agent hardening skill on a PR to modularise it and get rid of the slop. I don't care whether the session gets compacted or not. By the magic of Pi's compaction system, it just works. When working on UI from scratch, I give the agent a reference frame. So let's say if I want to create something that looks like Discord, I would tell it that, ask it to control my browser to click through some stuff and take screenshots. Remember every PWA, every Electron app can be launched with browser control enabled. I believe I do it for anything I launch on my computer as a default. Then I feed the agent other bits of interfaces I like and ask it for a mashup (screenshots). At this point I create a mock UI and review it. I am not expecting the Clanker to come up with anything original anyway, so I do not subscribe to the idea of one-shotting a nice frontend. When the mock is done, I start using React grab and screenshots to customise the hell out of it. The final outcome doesn't really resemble anything that it was based on, apart from the basic idea. Derivative work? Yeah, sure. Isn't everything a sidebar and chat window anyway these days? Techstack: bun for bundling, node for the actual apps, react+ts+tailwind. Convex wherever applicable. Think that's it. There you have it. Simple stupid. No loops, no worker subagents. Just raw clanking and understanding what is happening. Read more about the actual process: x.com/Howaboua/statu…

How did I not know this deepvoodoo.com

man i thought they were in openclaw outfits at first i need to get out of the bubble

Instead of counting "please" and swear words, count how many times you used "just". "just do it", "just fix it". In my case: - User messages containing “just”: 1,628 - Total user messages scanned: 11,724 - Session files scanned: 901 One one hand it might show lazy prompting, but also the model not carrying out the task and resorting to lecturing you about stuff. (I clean up sessions somewhat regularly)

Free the political prisoners in Britain!

🚨BREAKING: Keir Starmer has told friends that he intends to stand down as Prime Minister and set out an orderly timetable for his departure [@DailyMail]

In thread search, comin thru.

Oh yeah. However... At the risk of sounding like I'm high on my own AI farts: I am wondering if we're optimising for the wrong thing. I have a half-cooked project where you're just talking to an orb appearing as a widget, and it talks back to you. No GUI, nothing. Because we don't really need it. We're not chatting to humans via GUIs (IRL at least, lol). My point being, exploring automation and optimising UIs might make no sense in half a year. Sorry, rambling.

@Howaboua correct, I agree with you there. Just saying there is a lot of optimization potential for automation that is under explored currently.

What vibecoders can never do, but agentic software engineers will: I am adding keybinds to Howcode. While sidebar needs a bit of rethink, one of the keybinds is to scroll through threads in projects. It was PAINFULLY SLOW. 1. Asked the Clanker to fix. It kept adding longer debouncing timeouts, thinking that was the reason. It kept looping in that direction. 2. I obvs tested and none of this worked. My diagnosis was that while scrolling, we were calling the "thread load" function no matter if it was scrolled past quickly or not. 3. Clanker agreed, we were calling "open thread" whenever a thread was highlighted. The fix didn't work though. "READ MORE FILES". And guess, what it worked. The Clankers don't know UI/UX. They can't tell what is slow and what is fast. They are not constrained by performance of what they're coding. How did I arrive at my diagnosis? Intuition? Knowing how models ignore performance being an issue? READING what the agent wrote to me, asking him about function names/calls etc, and thinking logically? You choose. Whatever it was, it now works. PS. These sessions load FAST.

I believe what you're thinking of would help with seeing what's slowing things down, correct. But telling the clanker to "just optimise performance" will not make it suddenly work how it's supposed to be in the app itself. Maybe it would figure out at some point that what it's doing is stupid, but it clearly couldn't for a few turns until I told it "what it looks like is happening"... Dunno

@Howaboua it is flame graphs, isnt it? and you can just tell it to use bonfire to just fuzz all over the place?! But I have no clue, and I am just guestimating.

@tryingET Can't time ui/ux actions. Can't test them all using a Clanker either. And then I would have to define what I find acceptable timing per every action. This is not feasible. Clankers don't feel pain & time.

@Howaboua just give your agents the ability to time stuff. than it can optimize what you wanna achieve otherwise it is blind.

@fitchmultz Yes I have. Pi by default uses your composer model. Crisis averted though, it turns out I WAS using gpt mini - I am assuming spark was erroring out on me when I hit it with too much context.

@Howaboua Did you modify this intentionally? What does pi use by default?

I've just realised I've been using Spark for compaction in Pi. This might not be the best idea. Yes, it uses up a separate quota and is super fast, but I am worried that it might be too dumb. Thoughts? Should I switch to mini?

@steipete I always thought of you as an agentic software engineering researcher. Openclaw as a framework is merely a side effect of this ;)

People freaking out over my AI spend. What nobody sees: Part of what excites me so much about working on OpenClaw is that I'm trying to answer the question: How would we build software in the future if tokens don't matter? We constant run ~100 codex in the cloud, reviewing every PR, every issue. If a fix on main lands, @clawsweeper will eventually find that 6 month old issue and close it with an exact reference. We run codex on every commit to review for security issues (as it's far too easy to miss). We run codex to de-duplicate issues and find clusters and send reports for the most pressing issues. We have agents that can recreate complex setups, spin up ephemeral crabbox.sh machines, log into e.g. Telegram, make a video and post before/after fix on the PR. There's codex that watch new issues and - if it fits our documented vision well, automatically create a PR of it. (that then another codex reviews) We have codex running that scans comments for spam and blocks people. We have codex instances running that verify performance benchmarks and report regressions into Discord. We have agents that listen on our meetings and proactively start work, e.g. create PRs when we discuss new features while we discuss them. We build clawpatch.ai to split all our projects into functional units to review and find bugs and regresssions. We do the same split for security with Vercel's deepsec and Codex Security to find regressions and vulnerabilities. All that automation allows us to run this project extremely lean.

This went hard. Not particularly on local models, but a really good watch :D Nice to see techtubers go deep on tech-adjacent stuff, when we have more popular youtube scientists spreading misinformed shit about AI. Props. @0xSero youtube.com/watch?v=ibt52G…

@KEisuke62350514 @steipete glad you like it ;) it has its moments haha any feedback please let me know or submit an issue!

@Howaboua @steipete Howcode最高です👍

Howcode + Agent Pages FTW <3 This whole idea about giving your agents a canvas (credit where credit is due, @steipete) is growing on me :D igorwarzocha.github.io/howcode/ github.com/IgorWarzocha/a…

ngl I am not going to attempt a fix :P

It's trying...