multiversal

Write one annotated body of facts once, and render it into a gated, per-audience version of one communication. Each audience gets its own register and length, facts it may not hear are withheld, facts it already knows become callbacks rather than news, and a knowledge ledger per audience records what each reader already has and where that was inferred from. Plain files are the database; the model-driven steps (harvest, infer, compose) go through one adapter that shells out to claude -p.

Space layout

<space>/
  corpus/<fact-id>.md          one fact per file, YAML front matter (id, claim, date, sources, meaning, tellings, share, weight, media, status)
  audiences/<audience-id>.md   one audience per file (id, name, people, register, language, length, channel) + notes in the body
  ledger/<audience-id>.json    what each audience already knows, each entry with its source
  updates/<update-id>/
    brief.md                   front matter scope: [fact ids] (optional); body is the intent
    drafts/<aud>.md            composed draft with [f:id] citation markers
    renders/<aud>.md           gated, stripped render ready to publish
    recipes/<aud>.json         which facts a render cited, at which corpus commit
    review.md                  every universe side by side

A fact's share defaults to nobody. Only status: blessed facts render.

The writing engine (opt-in)

A space whose space.md front matter says engine: hyperspec holds every version to @supersuit/hyperspec before it can be sent. Multiversal builds what hyperspec reads and reimplements none of it:

updates/<update-id>/engine/<aud>/
  spec.hyperspec.md            generated per render with hyperspec's specText; never edited
  materials/facts.md           every blessed fact shared with this reader, one claim segment each
  materials/brief.md           the brief
  claims.jsonl                 the draft's own [f:id] citations, verbatim as rendered
  judge/                       doctor, reader and persona packets and verdicts
  runs.jsonl                   every check and judgment, by the bytes graded

The audience page becomes the spec's reader, persona limits, rejects (banned phrases) and requirements (register, language, length, where their knowledge left off). Each judge answers from its own fresh model call. Render runs check, the judges, check again and hyperspec ready; one revision on the judges' findings goes back through the gate, and a version still not ready is refused like a gate refusal. send refuses an opted-in version unless hyperspec ready passes on its current bytes. A brief's engine: off skips the engine for one update; a space without space.md renders exactly as before.

Commands

multiversal --help lists them: check, harvest, learn, render, review, send, stale. Tests: pnpm test. Run it with npx -y @supersuit/multiversal, which brings its dependencies (the plugin cache does not).

Two halves, one repo

• The engine (bin/, lib/): the multiversal command. Loads facts, checks every sentence against them, calls the model, writes files. It decides nothing about people.
• The plugin (.claude-plugin/, skills/): the procedure an agent follows around the engine, and every place the operator has to weigh in. Front door: /multiversal:multiversal. Verbs: harvest-facts, decide-who-hears-what, write-versions, send-a-version, ship. Each verb has an HDSOP.md map beside it.

This repo is its own marketplace. Install from the remote, never from a local clone:

claude plugin marketplace add SupersuitUp/multiversal
claude plugin install multiversal@multiversal

Tests: ./run-tests.sh discovers every suite, including the conformance test against Freedom's seven-requirement plugin standard.

Spaces (the facts, audiences and ledgers about real people) never live in this repo.