fusion-harness

Fuse frontier models instead of racing them. AND, not OR. A standalone Pi coding agent extension harness for two-model agentic engineering.

📺 Watch this video to get the full breakdown of this codebase: GPT-5.6 Sol vs Fable 5 Is the Wrong Question (Fusion) on YouTube

"Which model is best" is a benchmark question, not an engineering question. One model plans, another builds, and the results fuse: you combine compute instead of selecting it. Aider called the pattern architect/editor (the original fusion); Devin calls it fusion; OpenRouter calls it model fusion. This repo makes the pattern a first-class, on-camera-clear workflow: two hard-labeled agents, a gate-first validation loop, and attributed fusion. You don't have to pick a winner when you can hire both.

---

Install

Agentic Install

claude "/install"   # runs the /install slash command in Claude Code (or Pi, or your favorite agentic coding tool)

The /install command lives at .claude/commands/install.md and handles toolchain checks, key checks, and project-specific setup.

Manual Install

Prereqs: pi, just, jq, uv.

npm install -g @earendil-works/pi-coding-agent    # the Pi coding agent
brew install just jq uv                           # task runner, json inspection, gate runner
printf 'ANTHROPIC_API_KEY=sk-ant-...\nOPENAI_API_KEY=sk-...\n' >> .env   # architect + builder keys
just fh-workhorse                                 # launch on the cheap test pair

---

Why this exists

Every frontier release restarts the same argument: GPT-5.6 Sol or Fable 5? Benchmarks crown a winner, teams switch, and next quarter the crown moves. Meanwhile the two models are genuinely different animals: one plans and critiques with more depth, the other ships working diffs faster. Picking one means losing the other's edge on every single task.

The deeper problem is review, the second constraint of agentic engineering. A single unchecked agent drif