Cadence

Coding agents have one input channel: text. Your prompt doesn't say you're tired but still shipping, that you just got back from a long break, or that you want the agent to stop asking permission and make the call. Cadence is the second channel — an ambient context layer I'm building and shipping in the open as a Cullum & Co project.

It reads the room around the prompt: time, git state, prompt rhythm, music, Focus mode, and a self-reported state become visible context, distilled into four independent dials — pace, tone, posture, and proactivity. Every prompt then carries an inspectable state block plus a lens for reading what you meant. It's a lens, not a leash: it never rewrites your words or constrains the agent, and it always defers to what you actually typed.

The result is the same prompt landing differently in different rooms. In a shipping cadence, "how should I structure the retry logic?" gets the call made — backoff with jitter, here's the diff. In a thinking cadence, it gets the trade-offs explored patiently, no pressure to pick. The alpha ships as a Claude Code plugin, macOS-first because the signals read the Mac around you — but Claude Code is just the alpha surface. The product is the context layer underneath it.