pi-fusion: Local inference-time scaling for coding agents

OpenRouter's Fusion feature got my attention because it matched results from Binks, an internal code-review project I've been working on. In those evals, panels of model calls kept doing better than I expected. The first surprise was that a handful of fast, "dumber" models could match or beat parts of our benchmark. The second surprise was that a panel of frontier models could beat any single model in the panel. The mix mattered: cheap planners plus a strong synthesizer behaved differently from a frontier-only panel, and both were useful in different places. I wanted that shape in my daily pi loop, but I didn't want every fused turn to go through OpenRouter. I wanted to swap models, change worker count, restrict tools, cap output, and inspect exactly what got handed to the main agent. So I built an extension for pi-coding-agent called pi-fusion that recreates it, plus a few more bells and whistles. It runs a local planning panel before the normal actor turn. It can load shared…