- Don't state the obvious: I wouldn't hand a senior human dev a copy of "Clean Code" before every ticket and expect them to work faster.
- File vs. Prompt is a false dichotomy: The paper treats "Context Files" as a separate entity, but technically, an AGENTS.md is just a system prompt injection. The mechanism is identical. The study isn't proving that "files are bad," it's proving that "context stuffing" is bad. Whether I paste the rules manually or load them via a file, the transformer sees the same tokens.
- Latent vs. Inferable Knowledge: This is the key missing variable. If I remove context files, my agents fail at tasks requiring specific process knowledge - like enforcing strict TDD or using internal wrapper APIs that aren't obvious from public docs. The agent can't "guess" our security protocols or architectural constraints. That's not a performance drag; it's a requirement. The paper seems to conflate "adding noise" with "adding constraints."
Really interesting Project, thanks for sharing. Reminded me a bit of my time coding assembly on the C64 (yeah, I'm old). For 3D (wire-frame) we also needed to find creative ways around hardware limitations, especially the lack of a multiply instruction.
Posted almost at the same time about Kata. I'm trying to use Kata as replacement for the standard docker runtime (since I already have a tool based on docker).
The idea is to simply use the runtime flag (after kata install):
docker run -d --runtime=kata -p 8080:8080 codercom/code-server:latest
Hope this works, with this I could keep my existing docker setup.
I'm currently building a Docker dev environment for VSCode (github.com/dg1001/xaresaicoder) usable in a browser and hit the same issue. Without docker-in-docker it works well - I even was able to add transparent proxy in the Docker network to restrict outbound traffic and log all LLM calls (pretty nice in order to document your project).
For docker-in-docker development and better security isolation, I'm considering Kata Containers instead of Vagrant. Which gives me real VM-level isolation with minimum perf overhead, while still be able to use my docker stuff. Still on my TODO list though.
Has anyone actually run Kata with vs code server? Curious about real-world quirks - I've read that storage snapshot performance can be rough.
Does anybody also see this with claude code and haiku 4.5 (tried to set the env var, no change): "API Error: Claude's response exceeded the 8192 output token maximum. To configure this behavior, set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable."?
Nice UI btw, plus good use of AI for scoring/summarization.
I wrote something kinda similar that scrapes HN (using Firebase) for particular keywords I'm interested in. It gathers all hyperlinks mentioned in comments and uses NLTK to summarize. Kind of a curated HN reading list.
I'm currently working on using an LLM for the summaries.
Your project has given me a few ideas for mine. Thanks!
Glad to hear it. Your NLTK approach sounds interesting — would love to hear more about it. BTW, I’m planning to improve my project’s documentation. Funny enough, under “consideration” it flagged itself: "Documentation quality is not explicitly high. The effectiveness of the LLM’s scoring criteria is subjective and not deeply explained." Sadly, that’s true.
Thanks for the feedback — definitely not meant as spam. It’s open source, not a commercial blog. Just wanted to surface some fresh ideas that might otherwise get missed. Really appreciate the tip!
To make it clear I'm not accusing you of spamming, but I do think the supply vs demand for AI articles right now is out of whack.
What I was trying to say about spam is that your karma doesn't count so much as being a participant in the community counts. Like, I post a lot of articles to web sites that aren't my blog so if I post a few links to my blog it is OK. Now I'm an extreme case because I post a lot of links, but somebody who has like 1/10,000 the karma I have or less could post what they like so long as they aren't posting all links to the same web site. That's the place where your participation matters -- my posts don't do any better than anyone else's.
Got it, thanks for clarifying. Makes sense that participation matters more than raw karma. I’ll keep that in mind and try to contribute more broadly, posting less about my own projects.
Qwen3-Coder-30B-A3B-Instruct-FP8 is a good choice ('qwen3-coder:30b' when you use ollama). I have also had good experiences with https://mistral.ai/news/devstral (built under a collaboration between Mistral AI and All Hands AI)
- Don't state the obvious: I wouldn't hand a senior human dev a copy of "Clean Code" before every ticket and expect them to work faster.
- File vs. Prompt is a false dichotomy: The paper treats "Context Files" as a separate entity, but technically, an AGENTS.md is just a system prompt injection. The mechanism is identical. The study isn't proving that "files are bad," it's proving that "context stuffing" is bad. Whether I paste the rules manually or load them via a file, the transformer sees the same tokens.
- Latent vs. Inferable Knowledge: This is the key missing variable. If I remove context files, my agents fail at tasks requiring specific process knowledge - like enforcing strict TDD or using internal wrapper APIs that aren't obvious from public docs. The agent can't "guess" our security protocols or architectural constraints. That's not a performance drag; it's a requirement. The paper seems to conflate "adding noise" with "adding constraints."