🦞🦞🦞🦞🦞
Posts by Chris Banes
This is opposed to Gradle + KMP builds (with iOS) which takes 15 mins, at a _minimum_.
Starting hacking on an AI Agent orchestration tool (for myself).
Decided to build it in Rust (because why not), and holy cow: the entire CI run takes ~40 seconds, and most of that is setting up the toolchain.
github.com/chrisbanes/e...
✨ ChromaDial is a new #ComposeMultiplatform library that enables you to create beautiful Dials for your #JetpackCompose app
Its API is similar to the Slider, but instead works in degrees
Just pass in and set the degree state
#buildinpublic
Do you have any (rough) numbers of how much these cost at runtime?
The $20 sub for Codex is really generous for the price. I wouldn’t be surprised if it’s close to the CC $100 plan in terms of allowance
GPT-5.2-Codex is listed at 272k but I’m not sure if that is actually true. I can’t find anything to confirm that anyway.
I like Copilot (it’s basically a more basic version of Cursor), but beware that a lot of the models have a reduced context window size.
models.dev?search=copil...
OpenAI mostly (it uses oAuth so you can use your CGPT sub), and Antigravity (mostly for the free Opus/Sonnet allowance).
After becoming infuriated by Codex (the tool) lately, I've switched to @mariozechner.at's `Pi` and it's like a breathe of fresh air 🧘
All of the stuff I actually use, it's super fast, and most importantly it doesn't constantly ask permission to run `git`
shittycodingagent.ai
In some ways I really like GitLab, in other ways it’s infuriating. It’s CI system being top of my list…
+1 to this, although I personally prefer a mix of Codex 5.2 and Sonnet for coding (depending on task complexity). Codex is also my favourite for code reviews.
Opus is too expensive to be a pure coder imo.
code when? 😆
Thanks Don! 🙌
Definitely try other models. I know that Gemini 3 scores highly on benchmarks but I still find the Claude models the best in real world coding (by a good margin).
I have a follow up blog post in draft which goes more into this.
Just get AI to review it? (Only half joking as I do actually do this)
Totally agreed, AI isn’t the only reason this happens.
If your engineering team is 'senior-only' + AI-assisted… where do tomorrow's seniors come from?
New blog post on the 'disappearing middle':
chrisbanes.me/posts/disapp...
Not sure. It doesn’t seem to run any builds anyway, as the commits are full of compile errors
just found the limitation of Copilot. Since it can't run builds/lints/etc as part of the Agent run, you end up doing a lot of bouncing around via CI.
If your CI is slow (GitHub Actions are not exactly the fastest) then you get into a bit of loop.
Yep, but I did this solely from my phone and the GitHub app.
The ecosystem moat that GH have is significant.
And that’s where we’re at. Still need to finalise a few things, but I’m sure that this PR will be landing later today without me touching a physical keyboard.
5/
I wasn’t quite happy with the final solution, so added a review with some comments…
github.com/chrisbanes/h...
4/
About a hour later, the Agent was in full swing. It had fully fleshed out the plan, and began implementing…
github.com/chrisbanes/h...
3/
Started an ‘Agent Chat’ via the GitHub app. I gave it the URL of the GitHub issue, and we iterated on the high-level proposed plan to fix.
Once happy, I asked it to “Upload a Draft PR”
It created a placeholder PR immediately:
github.com/chrisbanes/h...
2/
Tried out @github Copilot to write a small feature for Haze, all from my phone. It’s the ultimate in lazy programming, but I’m pretty mind blown.
🧵…
Yep. I have a Hugo partial which takes the embed URL and calls the microlink REST API. It returns a JSON dump containing various bits of data (title, image, etc).
Here’s the partial: gist.github.com/chrisbanes/a...
True. I guess it depends how critical the unfurled content is to your product.
Localization can be largely fixed by passing through the locale info from the client.
Anything geo related is obviously trickier, but can be improved by deploying the cloud function to lots of regions.
Is this not incredibly more efficient to do this on a backend somewhere?
I personally use microlink.io as part of my Hugo blog build for exactly this purpose.
If you want to write it yourself, it seems like something which is perfect for a cloud function.
Nice post!