Local agent interface

Public documentation, preset JSON, the data catalog, and validation records are readable without WebGPU or user authentication. Private configurations and results exist only in the user's browser and local files. There is no hosted simulation API, MCP server, OAuth flow, or remote calculation endpoint.

Discovery

robots.txt, sitemap.xml, canonical metadata, HTTP Link headers, Markdown content negotiation, llms.txt, and /.well-known/agent-skills/index.json describe the public surface. The skill is a standard SKILL.md with name and description front matter, describing the real browser interface. No capability is advertised merely to pass an unrelated commerce or authentication check.

The catalog follows the Agent Skills Discovery RFC v0.2; the skill follows the Agent Skills format. Catalog URLs and SHA-256 digests identify the actual published skill. The deployment record states the live checker result and its selected scope.

Requests with Accept: text/markdown to /, /docs/, and the documentation pages receive Markdown. JSON and binary assets are ordinary read-only static resources. Content-Signal: search=yes, ai-input=yes, ai-train=no in robots.txt states the public-content preference. Crawler permission does not authorize access to local user files or exported simulation results.

User-authorized browser control

Ask the user to authorize a specific local task. They enable Local agent control in Diagnostics for the current tab. This consent is not stored across reloads. With an authorized browser automation connection, use window.mc in that tab:

window.mc.version; // 1
window.mc.presets(); // Included preset identifiers
window.mc.configuration(); // Clone of the current local configuration
window.mc.selectPreset('fixed'); // Consent required; wait for reinitialization
window.mc.setConfiguration(config); // Validates; clears earlier results
window.mc.statistics(); // Null while initializing, otherwise progress
window.mc.run(); // Consent required; no remote execution
window.mc.pause(); // Pauses between bounded GPU batches
await window.mc.reset(); // Clears current results
await window.mc.exportResults(); // Pauses; explicitly reads local GPU field

After changing a configuration, wait until statistics() returns a new ready configuration before running. Do not run against a stale engine during initialization. Imported CE data must already be loaded locally. The local API does not load paths, read arbitrary files, access credentials, or upload results. User authorization for local simulation is not authorization to transmit configuration or result data elsewhere.

exportResults() returns the configuration, scalar progress, and voxel moment array. Array order is flux/capture/fission, then energy group, then x-fast Cartesian cell. The second half contains second moments; the last float is active-batch count. Divide sums by that count for means. Flux and rates are normalized per external source history and cm³. Do not call export repeatedly to animate the grid; the renderer already consumes GPU buffers directly.

Review the preset assumptions and physics limits before launching. Treat results as preliminary and report uncertainty, active/inactive generations, balance, and adapter. Changes invalidate previous results. Disable local agent control when the task finishes.