Jakob // theCodeDrift

Chris Griffing (@cmgriffing) had a model build the OAuth consent form for Messijo. Every line item looked like a button, and only the text was clickable. Front-end engineer since 2015. That's the kind of detail he still fixes by hand. Full episode: taskless.io/blog/chris-griff…

CL Kao at @DataRecce: "If you're IC, you are still having a managing job now, congrats." Your team is agents now. Very smart, and in his words, "ADHD without any taste." Full episode: taskless.io/blog/cl-kao-recc…

200 output tokens turned into 7,000 tokens of whitespace. @cmgriffing hit that on a couple of GPT-OSS 120B providers, about 5% of the time. Novita answered his DM and fixed it. Another provider is still switched off. Full episode: taskless.io/blog/chris-griff…

Thank you to everyone who tried Taskless and upvoted on DevHunt today. I'm still answering questions in the comments tonight. devhunt.org/tool/taskless

A few hours left on DevHunt. If you've tried Taskless, an upvote with your GitHub account helps, and so does a question in the comments. devhunt.org/tool/taskless

Setup is one npx command. It finds Claude Code, Cursor, OpenCode or Codex in your repo, lists every permission change it wants to make, and asks once. Then tell your agent to set up Taskless. Questions welcome on DevHunt: devhunt.org/tool/taskless

If your agent wrote the code and you cleaned it up before opening the PR, the correction never shows up in review. It's sitting in your commit history. Taskless reads those commits too. On DevHunt today: devhunt.org/tool/taskless

🚀 Taskless is live on DevHunt today It turns the corrections you keep making to your agent into rules that run locally and in CI. Try it, upvote it, ask me anything in the comments. I'm there all day. devhunt.org/tool/taskless

I asked CL Kao at @DataRecce what breaks when AI agents touch your data. Agents are very confident, and the first result impresses. Keeping that result maintainable is hard. Full episode: taskless.io/blog/cl-kao-recc…

Models will sometimes decide to force push. @cmgriffing built GitBench to grade models on the git commands they want to run before any of them touch a repo. "You don't want a model going oopsie and trying to fix it." Full episode: taskless.io/blog/chris-griff…

NGL, I expected Ruff's "just use index.md" from the llms.txt to work everywhere, including the really useful page that tells you about all the rules Ruff supports docs.astral.sh/ruff/rules/in… @charliermarsh is this a bug or an omission? Agents struggle with the HTML version

CL Kao at @DataRecce lets the agent prototype, then makes it start over with everything it learned. First time: a 7 hour rebuild with the edge cases planned in. Full episode: taskless.io/blog/cl-kao-recc…

I'm giving a video talk Dec 7 at @AICodingSummit Berlin: Static and Runtime Enforcement for Agentic Code. I don't think it's the agent's job to police itself. Register through my badge and you can watch it and a bunch more talks for free: gitnation.com/badges/ai-codi… #AICodingSummit

⛵ seamux 0.5 is out, the web UI for cmux (@cmux) - drive Cursor's agent CLI from the board, next to Claude Code and Codex - the memory pressure warning is live - Fork sends a tangent to its own worktree npx seamux or github.com/thecodedrift/seam…

I was having a nice no-node-drama morning... TYVM nodejs.org/en/blog/vulnerabi…

"People who don't do front end a lot think front end is solved." @cmgriffing on AI output: on a subject you know, you catch every mistake. Flip the page to one you don't, and you take it verbatim. Full episode: taskless.io/blog/chris-griff…