I have to admit, @deepseek_ai Harmess is a really impressive piece of software. With its append-only design, it keeps exact byte-identical prefixes, so prefix/KV cache hits 99%+.

🤔 Real talk: Like millions of people, I’m not a native 🇬🇧 English speaker, so I use Grammarly (@Superhuman) quite extensively for proofreading. Does that mean everything I write, including this post, should be watermarked as "AI-generated"? I don’t think so.

🤯 So basically, I can't subscribe to @hootsuite because their subscription page is broken, and I can't subscribe to @buffer either because they can't verify a stupid CAPTCHA. What's going on, folks?

🫆 Why I stand against AI watermarking Not because I’m against attribution. I’m against a system that turns “a machine was involved” into a proxy for “a machine wrote this.” The debate is worth having. 👇

🚀 Watermarks Remover v0.5.0 - aka "make it deployable" ⭐ A milestone worth celebrating: 8,000+ GitHub stars. That's a community I'll never take for granted. 🧩 Skill/service split: the agent skill is now a code-free remote client. All the machinery moved behind an HTTP service (/health, /inspect, /clean, /capabilities), so hosts need no Python. 📦 Docker distribution: a full core image with exiftool / qpdf / c2patool preinstalled, compose.yaml for the whole stack, and GHCR publishing on release tags. 📜 OpenAPI 3.0.3 spec generated straight from the route table + live config. 🧪 Two new verification harnesses: MarkLLM (text: KGW, SynthID) and MarkDiffusion (images: Tree-Ring, Ring-ID, ROBIN, SEAL…) so you can be more confident a mark was actually cleared before/after. 🐛 Plus a long list of hardening: reasoning-effort control for rewrite backends, SSRF/gzip-bomb hardening, WebP provenance, Windows bootstraps, and more. Huge thanks to every contributor, tester, and star-gazer. The issues, PRs, and ideas you bring keep this project sharp.