Search

My coding agent almost wiped my git repo. Here's what I built to stop it. - DEV Community

After a coding agent looped through edits, produced noisy logs, invented a dependency, and attempted a hard Git reset, the author added safeguards to git-oops. The tool blocks destructive commands, stops repeated edits, trims and redacts logs, and checks patches for syntax errors and nonexistent imports.

Claude Security continuously scans the code and traces the corrections - Artificial Intelligence

Translated title. Original: Claude Security scansiona in continuo il codice e traccia le correzioni - Intelligenza Artificiale

Claude Security now performs continuous monitoring of repositories with fix tracking and automatic re-opening of issues, moving from periodic scans to risk-based, ongoing evaluation. The beta includes integrations with GitHub and GitLab, a configurable threat model, and role-based access, aimed at reducing noise and manual verification in security workflows.

If you bundle a project with tar or zip and hand it to AI, it will include .env and also .git history: with git archive you only include tracked files (comparing on real devices) #Git - Qiita

Translated title. Original: プロジェクトを tar や zip で固めてAIに渡すと .env も .git の履歴も入る：git archive なら追跡ファイルだけ（実機で比較） #Git - Qiita

Warning about packaging code for AI workflows, noting risks of including sensitive files and history; mentions deletion consequences and draft deletion.

GitHub - formstep/claude-mod-comment-diff: A Claude Code mod to comment on lines of Claude's diffs, like a GitHub review · GitHub

A Claude Code mod enables line-by-line commenting on diffs, allows batching comments into reviews, and provides UI to track, edit, and resolve comments. It includes installation steps and in-chat workflows for commenting, submitting, and resolving reviews within a terminal or desktop app.

I was fed up with drowning in VS Code windows, so I built my own development environment to run Claude Code in parallel, one worktree at a time.

Translated title. Original: VS Codeのウィンドウだらけで限界だったから、Claude Codeをworktreeごとに並列で回せる開発環境を自作した話

Worktabs is a Windows and macOS app that groups VS Code worktrees into tabs and tracks parallel Claude Code sessions with status badges and notifications. Its author explains the workflow it addresses, its trade-offs, and why it keeps the standard VS Code editor and Claude Code extension.

Ten million AI-signed commits a month, almost all in projects nobody watches - MACCUS

A single AI coding tool, Claude Code, now contributes millions of public commits to GitHub, overwhelmingly in small, underwatch projects. The open web is quietly accepting agent-written code, with limited ability to review it, raising concerns about provenance and oversight in the long tail of open-source repositories.

What Are Human Code Reviews For? - DEV Community

Coding agents can handle routine checks before a pull request, but teams need shared practices to make those checks visible and consistent. Human reviewers can then focus on system-level judgment, domain knowledge, and precise feedback.

AI Makes Linux 7.4 Merge Window the Largest in Kernel History - Slashdot

The Linux 7.4 merge window set a record for commits, with Phoronix noting 19,671 since 7.3 and an uptick tied to AI/LLM activity; final timing suggests some additional commits before the 7.4-rc1 release, while Linus Torvalds cautions that AI patches are not the only reason for larger releases.

Why Git Needs Help: Moving from Files to Systems for Source Control – O’Reilly

The author argues that agent-focused source control should move beyond file diffs to server-hosted syntax graphs, targeted code edits, and saved agent reasoning. The proposed changes aim to reduce prompt overhead, catch dependency impacts earlier, and preserve diagnostic history between agent sessions.

How I'm fighting the AI Brain Fry — Sergey Kuhne

The author describes a Git-based queue for managing AI coding work, with agents handling approved tasks and a single inbox collecting decisions that need human input. Closing each ticket in the queue helps the author focus on one item at a time, though better context summaries and ways to handle blocked tickets are still needed.