Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
JavaScript ai-agentsdashboarddeveloper-toolsdigital-twinquantified-selfsecond-brainself-hostedtailscale
/**
* The one definition of the style a universe contributes to an IMAGE prompt.
*
* A universe carries two different style surfaces and they are NOT
* interchangeable:
*
* - `influences.embrace` / `influences.avoid` — curated, comma-joined visual
* tokens authored in the Universe Builder's chip inputs specifically so a
* diffusion/image model can consume them directly.
* - `styleNotes` — free-text direction written for the WRITING stages: tone,
* staging, pacing, what a culture's scenes must establish. It routinely
* names canon entities that are nowhere near the current shot ("Stage
* <character> as a massive but local physical presence…") and craft notes an
* image model cannot act on ("keep translation slow, funny, imperfect").
* Injected into a render it burns prompt budget and drags unrelated canon
* into the frame, so it never enters an image prompt.
*
* The pipeline's visual stages already worked this way; this module makes the
* rule shared so the FableLoom compiler, character reference/blueprint sheets
* and LoRA dataset renders can't drift back to the free-text field.
*
* `universeCanon.js#buildStyleClause` is the sibling for LLM TEXT prompts —
* there `styleNotes` is exactly the right context and stays included.
*/
import { escapeRegExp } from './textUtils.js';
const tokenList = (values) => (Array.isArray(values) ? values : [])
.map((token) => (typeof token === 'string' ? token.trim() : ''))
.filter(Boolean);
const normalize = (token) => token.toLowerCase().replace(/\s+/g, ' ').trim();
/** Split a comma-joined prompt/negative string back into trimmed tokens. */
export const splitPromptTokens = (text) => (typeof text === 'string' ? text : '')
.split(',').map((token) => token.trim()).filter(Boolean);
/** Curated visual token lists for a universe (`[]` for a missing universe). */
export function universeVisualStyleTokens(universe) {
return {
embrace: tokenList(universe?.influences?.embrace),
avoid: tokenList(universe?.influences?.avoid),
};
}
/**
* Join a chip-token list (embrace OR avoid) into the comma-separated string
* `composeStyledPrompt` consumes. Tokens are trimmed and blank entries
* dropped — the same normalization `universeVisualStyleTokens` applies to a
* whole universe, exposed here for a caller (`universeBuilder/compile.js`'s
* prompt assembly) that already has just the raw array, not the universe.
*/
export function joinInfluenceList(structured) {
return tokenList(structured).join(', ');
}
/**
* Positive style clause for an image prompt.
*
* `override` / `mode` mirror `series.stylePromptOverride` +
* `stylePromptOverrideMode` ('prepend' | 'append' | 'override') so a series can
* lead, trail, or replace the universe's tokens — the same composition the
* client's `universeStylePreset` applies.
*/
export function buildVisualStyleClause(universe, { override = '', mode = 'prepend' } = {}) {
const trimmedOverride = typeof override === 'string' ? override.trim() : '';
const embrace = trimmedOverride && mode === 'override'
? ''
: universeVisualStyleTokens(universe).embrace.join(', ');
const parts = mode === 'append' ? [embrace, trimmedOverride] : [trimmedOverride, embrace];
return parts.filter(Boolean).join('. ');
}
/**
* Remove a style clause the authored prompt already leads with.
*
* The browser composes the same universe preset onto a scene prompt before
* POSTing it (`composeStyledPrompt`), and the compiler prepends the canonical
* clause too — so without this the whole token list lands twice in every
* render. Strip the browser's copy rather than the compiler's: diffusion models
* weight earlier tokens heaviest, and only the compiler's copy sits in front of
* the canon context where that conditioning is worth anything.
*
* Matching is on the JOINED clause, not token by token, which is what keeps a
* token carrying its own punctuation (`M.C. Escher`) intact — a per-token
* split on '.' would never match it and would emit it a second time.
*/
export function stripStyleClause(text, clause) {
const source = typeof text === 'string' ? text : '';
const trimmed = typeof clause === 'string' ? clause.trim() : '';
if (!trimmed || !source) return source;
// Whitespace in the composed prompt need not match the stored token spacing,
// and the clause is followed by the separator composeStyledPrompt inserted.
const pattern = new RegExp(`^\\s*${escapeRegExp(trimmed).replace(/\s+/g, '\\s+')}\\s*[.,;]?\\s*`, 'i');
return source.replace(pattern, '').trim();
}
/**
* Union several comma-joined negative-prompt sources, deduped at TOKEN level.
*
* A plain `unique()` over the sources compares whole strings, so an authored
* negative that is itself a joined token list never matches the individual
* universe tokens and the whole list repeats.
*/
export function mergeNegativePromptTokens(sources) {
const seen = new Set();
const merged = [];
for (const source of Array.isArray(sources) ? sources : []) {
const tokens = Array.isArray(source) ? tokenList(source) : splitPromptTokens(source);
for (const token of tokens) {
const key = normalize(token);
if (seen.has(key)) continue;
seen.add(key);
merged.push(token);
}
}
return merged;
}
/**
* The labeled aesthetic line the canon render prompts (character reference and
* blueprint sheets, LoRA dataset images) have always led with. `''` when the
* universe has no curated tokens, so callers keep their own fallback line.
*/
export function universeAestheticLine(universe, options) {
const clause = buildVisualStyleClause(universe, options);
return clause ? `Universe aesthetic: ${clause}` : '';
}
/**
* The ONE style preset a universe (plus an optional series override) hands to
* `composeStyledPrompt` — `{ prompt, negativePrompt } | null`. Both the browser
* (Universe Builder canon renders, FableLoom's scene composer) and the server
* (the FableLoom visual-canon compiler) must apply the exact same override
* composition, or a series override drifts out of sync between what the
* preview shows and what a render actually receives (#8442) — the browser
* used to carry its own copy (`client/src/lib/universeStylePreset.js`, now a
* re-export of this).
*
* `series` is optional; when omitted the result is the universe-only preset.
* Avoid (negative) tokens always come from the universe — the mode only
* composes the positive embrace clause. See `buildVisualStyleClause` for the
* prepend/append/override semantics of `stylePromptOverrideMode`.
*/
export function universeStylePreset(universe, series = null) {
const prompt = buildVisualStyleClause(universe, {
override: series?.stylePromptOverride,
mode: series?.stylePromptOverrideMode,
});
const negativePrompt = universeVisualStyleTokens(universe).avoid.join(', ');
if (!prompt && !negativePrompt) return null;
return { prompt, negativePrompt };
}