Cyrene_v2/frontend/web/src/lib/parseContent.ts

// Structured parser for persona action/narration markup (design/07 §1).
// A single message `content` may mix plain text with an action/narration block in
// any of three equivalent forms, so a model that drifts from the canonical XML
// still peels into the same muted note instead of leaking markup into a bubble:
//   1. <action>…</action> / <narration>…</narration>  — the canonical form;
//   2. ＊…＊ full-width star  (＊动作：弯起眼睛＊);
//   3. *…* half-width star, only when it clearly carries a role-play keyword
//      (*动作：弯起眼睛*) — plain *emphasis* stays as text.
// This splits `content` into an ordered parts array so the UI can position each
// block (inline, or as an attachment note via `splitBubble`) while the structured
// data stays available for downstream feeds (mind-map / state) via
// `summarizeContent`.

export type ContentPartType = "text" | "action" | "narration";

export interface ContentPart {
  type: ContentPartType;
  text: string;
}

/** Matches one action/narration block in any of its three accepted forms:
 *  `<action>…</action>` / `<narration>…</narration>`, a `＊…＊` full-width star
 *  block, or a `*…*` half-width star block that begins with a role-play keyword.
 *  The alternation is ordered left-to-right (the leftmost token wins), and each
 *  body is non-greedy (`[\s\S]*?`) so a block may span newlines while a malformed
 *  or unclosed tag/star simply falls through as plain text. Capture groups:
 *  1 = XML tag name, 2 = XML body, 3 = full-star body, 4 = half-star body. */
const ROLE_RE =
  /<\s*(action|narration)\s*>([\s\S]*?)<\/\s*\1\s*>|＊([\s\S]*?)＊|\*((?:动作|旁白|内心|神态|轻声)\s*[:：]?[\s\S]*?)\*/g;

/** Leading role-play keyword of a star block, stripped before it becomes a note. */
const ROLE_KEYWORD_RE = /^(动作|旁白|内心|神态|轻声)\s*[:：]?\s*/;

/** Caps how long a keyword-less full-width star block may be and still count as a
 *  role-play remark (the drift shape ＊弯起眼睛＊). A long keyword-less ＊…＊
 *  paragraph stays as plain text rather than being mislabelled as an action note. */
const STAR_SHORT_RUNES = 40;
const NEWLINE_RE = /[\n\r]/;

/**
 * Split `content` into an ordered parts array preserving block position.
 * Pure text with no block yields a single `{type:"text"}` part.
 */
export function parseContent(content: string): ContentPart[] {
  const parts: ContentPart[] = [];
  let last = 0;
  ROLE_RE.lastIndex = 0;

  let m: RegExpExecArray | null;
  while ((m = ROLE_RE.exec(content)) !== null) {
    const before = content.slice(last, m.index);
    if (before) parts.push({ type: "text", text: before });
    parts.push(blockToPart(m));
    last = m.index! + m[0].length;
  }

  const tail = content.slice(last);
  if (tail) parts.push({ type: "text", text: tail });
  if (parts.length === 0) parts.push({ type: "text", text: content });

  return parts;
}

/** Turn a single matching block into its content part, per which form it took. */
function blockToPart(m: RegExpExecArray): ContentPart {
  if (m[1]) {
    // Canonical <action>/<narration>: body is used as-is (it has no keyword prefix).
    return { type: m[1] === "action" ? "action" : "narration", text: m[2] };
  }
  if (m[3]) return classifyFullStar(m[3]) ?? { type: "text", text: m[0] };
  return classifyHalfStar(m[4]);
}

/** A `＊…＊` block is role-play (returned as a part) whenever it carries a role-play
 *  keyword or is a short stand-alone fragment; a long keyword-less paragraph wrapped
 *  in stars returns `null`, so the caller keeps it as plain text star-for-star. */
function classifyFullStar(body: string): ContentPart | null {
  const kw = matchRoleKeyword(body);
  if (kw) return { type: kw.kind, text: kw.rest };
  const t = body.trim();
  if (t && t.length <= STAR_SHORT_RUNES && !NEWLINE_RE.test(t)) {
    return { type: "action", text: t };
  }
  return null;
}

/** A `*…*` block is role-play only because the regex already required a keyword,
 *  so it is peeled; a keyword-less `*emphasis*` never makes it here and stays text. */
function classifyHalfStar(body: string): ContentPart {
  const kw = matchRoleKeyword(body);
  if (kw) return { type: kw.kind, text: kw.rest };
  return { type: "action", text: body.trim() };
}

/** Matches the leading `动作/旁白/内心/神态/轻声` keyword of a star block, mapping
 *  it to an action/narration kind and returning the bare text after it. */
function matchRoleKeyword(body: string): { kind: "action" | "narration"; rest: string } | null {
  const m = ROLE_KEYWORD_RE.exec(body);
  if (!m) return null;
  const kind: "action" | "narration" =
    m[1] === "旁白" || m[1] === "内心" ? "narration" : "action";
  return { kind, rest: body.slice(m[0].length).trim() };
}

/** One extracted tag value. */
export interface ContentTagItem {
  text: string;
}

/** Flat summary derived from the parts: the plain body string plus the collected
 *  action / narration values. Useful for mind-map / state feeds that want the
 *  structured payload rather than the inline render. */
export interface ContentSummary {
  text: string;
  actions: ContentTagItem[];
  narrations: ContentTagItem[];
}

/** Collapse a message into `{text, actions, narrations}`. */
export function summarizeContent(content: string): ContentSummary {
  const summary: ContentSummary = { text: "", actions: [], narrations: [] };
  for (const part of parseContent(content)) {
    if (part.type === "text") summary.text += part.text;
    else if (part.type === "action") summary.actions.push({ text: part.text });
    else summary.narrations.push({ text: part.text });
  }
  return summary;
}

/** Result of `splitBubble`: a renderable plain-text `body` plus the annotation
 *  blocks that belong above (`head`) or below (`tail`) the bubble. `head`/`tail`
 *  hold the leading / trailing action & narration parts in order; `body` is the
 *  concatenated text with any annotation blocks removed. */
export interface BubbleSplit {
  body: string;
  head: ContentPart[];
  tail: ContentPart[];
}

/**
 * Peel the `<action>` / `<narration>` blocks out of `content` so the bubble
 * body is pure text while the annotations attach as notes outside it.
 *
 *  - `head` = the leading run of action/narration before the FIRST text part
 *    (rendered above the bubble).
 *  - `tail` = the trailing run of action/narration after the LAST text part
 *    (rendered below the bubble).
 *  - `body` = every text part concatenated and outer-trimmed.
 *  - Annotation blocks interleaved BETWEEN text paragraphs have no stable text
 *    anchor once the body collapses, so they lift to `head` — the only hard
 *    requirement is that they never render inside the bubble.
 *  - A content with no text part at all yields an empty `body` with everything
 *    in `head`, so the notes alone carry the message.
 */
export function splitBubble(content: string): BubbleSplit {
  const parts = parseContent(content);
  let firstTextIdx = -1;
  let lastTextIdx = -1;
  parts.forEach((part, i) => {
    if (part.type === "text") {
      if (firstTextIdx < 0) firstTextIdx = i;
      lastTextIdx = i;
    }
  });

  const head: ContentPart[] = [];
  const tail: ContentPart[] = [];
  let bodyText = "";

  parts.forEach((part, i) => {
    if (part.type === "text") {
      bodyText += part.text;
      return;
    }
    if (firstTextIdx < 0 || i < firstTextIdx || (i > firstTextIdx && i < lastTextIdx)) {
      head.push(part);
    } else {
      tail.push(part);
    }
  });

  return { body: bodyText.trim(), head, tail };
}