Tooling

Five releases landed while I was writing the last two posts. The one that matters: PixelLab can now edit an existing asset directly, no ComfyUI instance required. Two new generators and a gallery shortcut came along with it.

The Pixelkiln project page said one provider, no characters. Forty-eight releases later, in about three weeks, both claims are wrong, and here’s what actually shipped.

Pixelkiln ships a second set of instructions now, not for you, for whatever’s driving the CLI on your behalf. Here’s what it won’t let an agent do, and where else it lives.

A single 65-badge icon set turned into 350 disconnected objects with no record of what shipped from where. Here’s the lockfile we built to fix that, and the three bugs we found the day we actually put it to the test.

A rename touches more than headings and copy. Fourteen localStorage keys, two cookie names, and a desktop bundle ID that locks in the moment one build ships all had to move together, or not at all. Here’s how the Subpath rename got through that without losing anyone’s saved work.

A year after I built a weekend SVG tool out of frustration, it went quiet for four months, then became a real product: boolean path operations, AI-assisted generation, optional cloud sync, and a native desktop app, all under a new name, Subpath.

Wrapping QLC+ on a Raspberry Pi with the things that make a headless DMX rig pleasant to use: one-command provisioning, a friendly landing page, fixture groups, AI scene generation, and the boring reliability glue (udev rules, WiFi watchdog, doctor command) that turns lights-are-dead calls into a non-event.

Homelab dashboards tend to be either link boards with no real metrics or full monitoring stacks that feel like overkill. Buoy is the middle ground: one container per node that auto-discovers Docker services, shows system vitals, and polls peer nodes for a fleet view. No database, no build step.

A Sunday backlog on coco that should have eaten the whole day took about an hour. Not because the AI wrote my code, but because steering files and specs meant it already knew the project instead of me re-explaining it every session. Here is what is in mine, and what I left out.

Every guide blames a fancy zsh prompt when your AI coding agent cannot read the terminal. So I deleted mine, then deleted bash too, and the commands still were not running. Here is where the trail actually led.

I rebuilt coco’s interface with a TUI design skill, then needed to show it off. Screenshots and demo GIFs turned out to be their own craft, full of subtle traps. So I built a skill for that too.

The regex-based structural diff extractors landed and worked – then hit a ceiling regex was never going to clear. Here’s how coco’s diff pipeline actually moved onto tree-sitter: sharper structural context for commit messages, a layered regex fallback, lazy-loaded grammars – plus the bonus the same parser bought us, syntax-highlighted diffs in the Workstation.

git-scenarios came together fast because the project it serves, Coco’s TUI git workstation, had very clear testing needs. A couple of weeks after the v0.1.0 extraction it’s now stable at v1.0.0: 32 scenarios, 60+ atoms, five framework adapters, 527 tests across three operating systems. The v1 work wasn’t about features, it was about trust: typed error classes, precondition guards, property-tested determinism, cross-platform CI, and shedding the last of its coco branding.

I have written the same 80 lines of git test setup more times than I want to count. @gfargo/git-scenarios v0.5.0 extracts the helper layer that grew up inside Coco: a registry of named scenarios (mid-merge-conflict, submodule-with-history, mid-bisect, and 8 more), a composable atom API for building your own, and a tool-agnostic CLI that drops any tool into a real repo in a known state.

A weekend alone. Four projects. 222 commits in a single day. An honest look at what agentic coding is doing to my attention, what wisdom still costs, and how I am trying to stay balanced at light speed.

localpress is a CLI tool that processes your WordPress media library locally – compression, WebP/AVIF conversion, AI background removal, and round-trip desktop editing – with no cloud SaaS, no recurring credits, and no plugin required on the WordPress side.

Doorman 2.0 ships Cloudflare WAF support alongside Vercel Firewall, a new .doorman.json config format, and 1,100+ tests backing a production-ready multi-provider firewall CLI.

A small filesystem tweak that turned my open source projects into a single workspace my agents can actually reason about. Nested git repos, a centralized orchestrator, and a documentation loop that finally stays in sync.

After years of building terminal apps and a growing inability to justify my GitKraken subscription, I built a Claude Skill for TUI design. Here’s what’s in it, why I made it, and what I’m planning to build with it next.

A client needed help managing a few VPSes, so I wrote a small bash script. Three rewrites and 21 releases later, it’s Strut: an open-source CLI for deploying Docker stacks anywhere there’s SSH, built agent-first for the way I actually code now.

After building Spec Shaver as a CLI for trimming oversized OpenAPI specs, I kept finding myself wanting a visual way to do the same thing. SpecForge is what came out of that itch: a web app for importing, curating, and exporting OpenAPI specs in the browser.

Hitting OpenAI’s 1MB / 30-operation cap on Custom GPT actions when your real API has 150+ endpoints is maddening. Spec Shaver is the CLI I built to reduce a bloated OpenAPI spec down to the operations that actually matter, while keeping the schema valid.

Fargo Flags is the feature-flag toolkit I wished existed for Next.js: zero infrastructure, type-safe, and small enough to drop into any project. Here’s the design call that shaped it and the trade-off I’m still not sure about.

I built Vercel Doorman because clicking through firewall rules in a dashboard is the kind of thing infrastructure-as-code was supposed to fix. This is what it does, why JSON config beat a DSL, and what I’d change if I started over.

Coco is the AI-powered Git assistant I built because I was tired of writing terrible commit messages at 11pm. The story of how it grew from a one-off script into something I now use daily across every project.

A working set of Yup schemas I keep reaching for: emails, phone numbers, git hashes, IP addresses, Ethereum wallets, full user registration. Plus DOMPurify for the output side, since validation alone won’t stop a script tag from rendering.

PixelLab’s Object Creator sat on the project’s own roadmap since the character generator shipped. Two releases now cover it: a new objectPro generator for skeleton-free props and creatures, and gallery buttons that add a new pose or animation without touching the manifest by hand.