Node

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.

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.

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.

A look at everything that’s landed in ink-enhanced-select-input on the road to 1.0: a headless hook for custom renderers, multi-select mode, inline search, item groups, scroll indicators, Escape/cancel support, and 133 tests – including a real bug caught by writing edge-case tests for backspace handling in searchable mode.

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.

Notes from making tFrogger feel like an arcade classic in a terminal: tuning movement to feel responsive without breaking the per-cell grid, and the legacy config-folder cleanup I should have done sooner.

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.