← Lab / 04

Motion Workbench

Turn formulas into motion. Edit recipes, draw curves, follow program execution, and diagnose a blocked job.

Built with GPT-6.1 Sol through OpenCode. The C++ application runs in your browser through WebAssembly; its “About the code” explorer is also AI-authored.

Start with Guided Demo → From formula to motion. Best explored on a desktop with a keyboard and mouse.

Open standalone ↗
About this experiment

Seven editable recipes combine parameters, loops, branching, and mathematical expressions. Guided lessons connect formulas, geometry, and motion; searchable Docs explain the controls and language.

This is an illustrative 2D pen simulation, not hardware control or production CNC planning. Continuous contours model tangential speed and acceleration, not physical cornering, jerk, or servo dynamics. Fixture checking covers 2D pen paths rather than complete machine collisions.

The application runs locally in your browser, without a simulation backend. Export a recipe to keep your work. The application has its own light/dark preference.

From idea to finished demo

I used GPT6-Astra at Medium reasoning through my OpenAI Pro subscription to help with high-level design decisions and to turn my ideas and feedback into prompts for GPT 6.1 Sol, also at Medium reasoning.

GPT 6.1 Sol did the implementation through the OpenCode harness, using OpenRouter at API prices: initial project and tooling setup, build pipeline installation, programming, helper scripts, tests, debugging, and visual debugging. I supplied the idea, reviewed the results, and steered the iterations. The code explorer is an AI-authored explanation of the implementation, rather than a claim that I personally understand every part of the code.

Part of the experiment was seeing how far a bounded API budget could go. I used GPT 6.1 Sol exclusively for the main development work. Going from idea to finished demo took several hours, with $27.81 in API usage reported by my OpenRouter dashboard and the Pro subscription used separately for design and prompt guidance. The conversation grew large enough to require automatic compaction.

The OpenCode project report records one session, six prompts, 329 model steps, and 464 tool calls, with no subagents. All recorded model steps used GPT 6.1 Sol at Medium reasoning through OpenRouter. OpenCode reports $27.59 for this project; that is slightly below the dashboard total.

Token counters as reported by OpenCode: 984 input, 228,832 output, 266,809 reasoning, 96,052,390 cache reads, and 933,897 cache writes. These are cumulative usage counters, not the size of a single conversation: cached context can be counted again across many requests. The categories are shown separately to avoid assuming whether reasoning is already included in another counter.