wbmnfktr

We might have a minor misunderstanding here. Building modules, tools, apps, anything with little to a lot of logic and functions is not the issue here. My problem is not that there are functions or config fields (Admin UI) missing or not working in a way or form I imagined them to work. The problem is that the Agents and LLMs are not ready yet to build module backends (Admin UI) - even when outlined in high detail. Sometimes they are almost there or at least it looks like it, sometimes they invent something totally new and won't even use existing field types, wrappers, and classes. Having an existing module, adding new functionality, new config settings, or even pages works incredibly well. That's because the Agent/LLMs can re-use existing structure of backend code. Besides that I totally support your step-by-step process and workflow. You can't just throw a vague idea towards whatever tool and expect a wonder. That's for sure. Little side note: I asked Kimi and Opus to build a module that's only purpose is to showcase module pages, all sorts of inputs and outputs. So I have a working base for interfaces I can point at later on. The prompt for those interested: The result: https://github.com/webmanufaktur/ProcessShowcase

So far I tried with: Z.AI GLM 5.1 Kimi K2.6 MiniMax 2.7 Opus 4.7 Codex 5.3 DeepSeek Pro & Flash V4 Results don't vary that much here in this case. I generated a few variations now and will pick the best parts from each variation. Huge context windows do help when I tell the agent to inspect a module that has a certain type of block i want to use/copy. Like tables or listing, like button groups or dropdown buttons. Outside of ProcessWire modules absolutely. They do a great first draft you can fine-tune then. I played a bit with ui.sh and impeccable.style to achieve quite good results. Even landing pages, newsletter templates, and everything non-dashboard type of design. Shadcn/TailwindCSS are great for fast prototyping of good looking interfaces. Even when using models like the older MiniMax 2.5. I did a small comparison a while back using older models and no skills or design frameworks. It was fine. https://log.nerd.to/log/ai-frontend-design-comparison/ Pointing in 2 directions helps: pointing at existing modules and how they did it, or just copy and paste the code (which isn't that easy sometimes) pointing at static examples and design systems for the fine-tuning - in this case the one from @maximus. I'm complaining on a high level here. 😂 The main part, developing the module, took an hour. So there is room left to design the UI a bit. The module is up and running, does what it should do. I am happy for now.

So... the very first iteration looks like this. Used a "Look there, apply to that"-prompt. We wiill get there. Now I have to dig a little bit deeper to fine-tune. Big help already! Thanks @maximus!

🤯 this might actually be the solution to everything wrong in my modules. I am not sure if I already read about it here somewhere but couldn't find it when looking up module theming/styling. I will give this a try. Might have to try a few variations as most projects, even new ones, still use the old styling.

How do you prompt and guide Agents (or any AI tool) to get a good understanding of module pages? Below are screenshots of what I currently have (finally). Yet both could be much more beautiful or at least organised and looking more clean. Sure I could go fully custom here but i want it to work with AdminThemeUiKit (and all existing flavours of it). Do you add other modules as examples to the context or do you go this step as we did 5 years ago and code it yourself?

I recently started to use more desktop apps, in this case OpenCode Desktop, which is quite similar to OpenAI Codex and the Claude Desktop app, but actually available for Linux. One great feature is Git worktrees support. They just work with NextJS, AstroJS, HonoJS but not with ProcessWire in my DDEV setup. I'm not sure if anyone else already found a great solution or maybe even built a module, script or anything like that (and I couldn't find anything) to support Git worktrees in DDEV environments with almost no manual intervention, so... I asked my agent to help me. I just pushed the inital/experimental version to Github. I tested it quite a bit with two projects locally in my setup and it works on my machine. https://github.com/webmanufaktur/processwire-ddev-worktree This is more like a proof-of-concept than a stable release. In case you want to test it, maybe don't. Or at least not with important projects. There might be plenty of edge-cases it won't handle right now or even break things on your machine. But it just solved a pretty big problem (for me) I ran into today. Look into it, feel free to fork it or send a PR.

Had a bit of fun with the latest DEV version and AgentTools version... yet ran into some issues. 1. I asked the Engineer to get a good grasp of the project, explained what it was for and what the Engineer should look at. Got this response: Refreshed the page, waited a bit and got what I was looking for. The result was on point and the next task was executed without issues, it wrote a migration and what i was looking for. 2. A bit later I asked the Engineer to answer a question I found here in the forums and got: Every request was using GLM 5.1 which can run long tasks and has tons of tools calls it can process. Just some findings I wanted to report.

That answers my question perfectly fine. You understood me here.

I might repeat myself here but: 🤯 The idea and concept of collections in addition to the interface. That's a great one!

I feel that! But it became way better than it was a few years back when the world was going nuts. The things I changed in the last 12-18 months: getting all the regular vitamins and supplements for someone my age getting additional supplements that help my brain to keep working checking in on gut health and eating less/none heavily processed foods checking calories and protein in-take - fixing both and removing empty carbs and calories getting back to the gym - first only cardio, now additional strength training All the above helped me to not being stressed out everytime and all the time. I couldn't care less about things thrown at me these days. A full day with meetings, a workshop, and a coffee with friend somewhere in-between, and of course the usual time at the desk don't affect me in a negative way anymore. I don't know how the math works here, but these days I have enough hours each day to have plenty of time, even time to balance everything out. PLUS the weekends for side-projects, fun, friends, family, and just things.

I'm blown away by this. 🤯

ProcessWire is more alive than ever.

Fix is confirmed. Installed the latest version from Github and it's working as before. 🎉 GLM is pretty nice. Feel free to test it a bit more using that key.

This is the agent export dialog's result: glm-5.1 | <key_removed> | https://api.z.ai/api/coding/paas/v4 | Z.AI GLM 5.1 config->advanced=true; is set in config.php spaces before and after the pipe "|" look weird but are PW-style, so... no. Nothing incorrect. I created a full export, including the db, and put it up for download - I send you the link via PM.

AgentTools version 0.1.1 in ProcessWire 3.0.259 and 3.0.260 broke all agents in my current setups. Asking the Engineer only returns: ProcessAgentTools: API error (404): Not Found Added new agents (OpenAI-compatible and Anthropic/Claude): still no success. Only difference is, the ones with Anthropic (Claude) won't return the error message, yet no output or anything as far as I can see and tell. Update: Used the backup of version 0.0.9 - works in both instances with all agents.