Claude installer/updater had a high memory usage problem (Gigabytes of resident memory used during a fresh installation of Clade with it's native install.sh script, or during a claude update). This was reported as early as December 2025, see https://github.com/anthropics/claude-code/issues/12327#issue....
I keep adding selected cases of CLAUDE.md instructions non-compliance reported on claude-code github to that issue https://github.com/anthropics/claude-code/issues/13689. Subjectively the amount of such cases seems lower during the past month. It may be that claude-opus-4-8 (default thinking) is a bit better at instructions following than past models.
I don't get the idea of "trending" for scientific papers. I would like to see the full list of papers that have associated source code from the given research field, like the arXiv cs.SD category. The reason to see all papers is to get an overview of the field.
It's independenent from Hugginface, and uses arXiv API directly.
The code is open source, Apache 2. No AI is involved in parsing the arXiv pdfs, only a pattern match for known code-hosting domains like github.com or gitlab.com. This means there false positives, but that's probably fine, and their patterns could be tuned later.
You can host your own feed, by forking the repo and setting GitHub Actions environment variables. For now I'm hosting cs.AI and cs.SD categories.
The takeaway may be that Anthropic could still profit from selectively labeling the issues as "acknowledged" (suggested in one of the now auto-closed issues https://github.com/anthropics/claude-code/issues/21732), and preventing them from auto-close.
Not all auto-closed issues are slop. A human reviewer can still distinguish them.
As for the impact of auto-closing issues, an example of this happened in the period of January/February 2026, when a misconfigured auto-close bot closed over 500 non-slop, human created issues, despite humans still reporting the problems were not fixed https://github.com/anthropics/claude-code/issues/16497.
It was 30 minutes of daily check by a human to see whether the auto-closed issues were human created or still had humans commenting on them.
Only after over 200 hundred upvotes, the misconfigured bot problem was addressed by Anthropic team by removing the bot, but without reopening any of the incorrectly closed issues.
Today it would be probably a half day for a single product manager familiar with the Claude Code product to decide whether any of the daily auto-closed issues are valuable.
There are about 100 issues closed daily by the https://github.com/anthropics/claude-code/actions/workflows/... workflow.
Reviewing all issues is probably still valuable for Anthropic, because not all problems are discussed in parallel on social media platforms.
In practice, some issues receive maintainer comments, especially the issues associated with activity on social media platforms (like https://news.ycombinator.com/item?id=47660925), but most issues are auto-closed without maintainer comments.
I'm a Linux user and wanted to have a speech-to-text functionality in Claude, so I can talk to it, like Armin Ronacher https://www.youtube.com/watch?v=bpWPEhO7RqE#t=5m37s demonstrates on macOS.
I was not able to find a small codebase doing this, that I can understand.
The project I'm submitting is about 500 lines of Python, and is packaged as Docker, so facilitate the setup.
When creating the project I added some security measures, like running the Docker container as non-root, and performing Whisper output sanitization before passing it to Claude.
Thes setup is Linux-only due to `/dev` device dependencies.
The desire of the models to act at the cost of ignoring user instructions is still noticeable.