📰 Dev.to · Brad Kinnard

Semgrep and ESLint flag risky code. This flags what they miss: tests weakened, errors swallowed, renames left half-done.

An AI agent can make CI go green without making the code right. Here's how that happens, and how far you can get catching it.

Agent transcripts on disk aren't evidence. Here are the cryptographic primitives needed to make a Claude Code or Codex CLI session verifiable off-host.

An offline-first multimodal evidence capture system built on Gemma 4, designed for environments where cloud access and chain-of-custody assumptions fail.

v8.0.2 is out now and it cleans up several rough edges that kept showing up under heavy tournament...

An independent verification layer that wraps any AI coding agent CLI and runs five checks on every patch before allowing it to merge.

I ran skillcheck v1.2.0 against 500 random skills from a 1,436-skill corpus. Here's what the SKILL.md ecosystem actually looks like in production.

Your notebook works because your kernel has hidden state. It dies for everyone who runs it clean. Static linter for that class of bug.

An arXiv paper from March 2026 proposes a specification-first approach with deterministic verification and bounded AI review. A practical open-source implementa

Copilot and Claude Code have shipped self-verification. The gaps that remain are where Swarm Orchestrator v4.2.0 sits.

Two AI agents scored 70% adherence on the same CLAUDE.md rules. They failed on entirely different ones. RuleProbe is the tool that found that.