npx github:soniccyclone/runbox init

Detection ranks the list of agents and never gates it, so installing runbox before your editor is not a dead end. Pass --agent claude --agent cursor to skip the prompt, or --global to install once for every project.

agent	project	global
Claude Code	.claude/skills/	~/.claude/skills/
Cursor	.cursor/skills/	~/.cursor/skills/
GitHub Copilot	.github/skills/	~/.copilot/skills/
opencode	.opencode/skills/	~/.config/opencode/skills/

Rerunning is also the update path: it always writes the current skills.

That step puts the skill where an agent will read it. A second step wires the harness into the project whose runs you want to compare:

cd <your-game>
.\.claude\skills\runbox\install.ps1 -Launch "godot --path {game}"
.\harness\serve.ps1

install.ps1 writes a harness\serve.ps1 wrapper, adds run output to .gitignore, and prints what the project still owes. It takes its target from your working directory rather than from where the skill sits, so a --global install works the same as a local one.

Windows, for now. The service and gallery are portable; install.ps1 and the generated harness are not.

The .NET SDK 10 or newer, because the service ships as a file-based app: a single .cs with no project file. There is nothing to build.

ffmpeg is optional. Without it runs index and compare fine and simply have no video.

• A carousel of every run, newest first, with each run's clip as its cover.
• A diff bay holding two runs, played from one transport, seeking together.
• Event markers on the scrubber, per slot, clickable to jump both runs to that frame.
• Deltas between the two runs, signed B minus A, over whatever your project records.
• Drift: what a prototype claims about itself against what its runs measured.
• Verdicts: mark the run you preferred; it outlives restarts and regenerated output.
• Launch: start a prototype from the gallery, via a command template.
• Export: bake the current view into one self-contained file with the clips inline.

The live demo is an export, produced by the real tool from three real runs.

They are not, and the evidence is written down. FINDINGS.md records ten findings, each tied to a rule that a reasonable person would otherwise simplify away:

• A closed-form constant disagreed with the running game by 9.5%.
• The drift is a fixed offset, not a ratio, so a fitted calibration constant is wrong everywhere except where it was fitted.
• A respawn is indistinguishable from a landing, and pairing across one reported 6.7px of jump reach.
• GIF is four times larger than the frames it encodes.
• Sparse captures desynchronised video from their own markers, and the failure looked like working software.
• The first human verdict could not have been measured.