Soul-driven AI agent with permission-hardened tools, token budgets, and multi-channel access. Runs 24/7 from CLI, Telegram or More.
import { create } from "zustand";
import { ApiError } from "~/api/client";
import { speakApi } from "~/api/models";
/**
* Read a reply aloud with Mercury's speech service. One reply at a time
* across the app: starting another stops the first.
*
* The reply is turned into plain speakable text (code, tables and link
* addresses aren't read out), split into parts at sentence boundaries, and
* each part is fetched while the one before plays, so the first words come
* quickly even for a long answer.
*/
export type ReadPhase = "loading" | "playing" | "needs-setup" | "error";
interface ReadAloudState {
/** The message being read (or that needs attention), and how it's going. */
id: string | null;
phase: ReadPhase | null;
error: string | null;
play(id: string, markdown: string): void;
stop(): void;
dismiss(): void;
}
/** Each part stays well under the runtime's 4000-character limit; small first parts start sooner. */
const FIRST_PART = 280;
const PART = 1500;
/** Markdown → what a person would say out loud. */
export function speakableText(markdown: string): string {
return markdown
.replace(/```[\s\S]*?(```|$)/g, "\n(code omitted)\n")
.replace(/^[ \t]*\|?[ \t]*:?-{3,}[-|: \t]*\n?/gm, "")
.replace(/^[ \t]*\|.*\|[ \t]*$/gm, (row) => row.split("|").map((c) => c.trim()).filter(Boolean).join(", ") + ".")
.replace(/!\[([^\]]*)\]\([^)]*\)/g, "$1")
.replace(/\[([^\]]+)\]\([^)]*\)/g, "$1")
.replace(/<?https?:\/\/[^\s>)]*[^\s>).,;:!?'"]>?/g, "a link")
.replace(/`([^`]+)`/g, "$1")
.replace(/^[ \t]{0,3}#{1,6}[ \t]+(.*)$/gm, "$1.")
.replace(/^[ \t]*>[ \t]?/gm, "")
.replace(/^[ \t]*[-*+][ \t]+\[[ xX]\][ \t]+/gm, "")
.replace(/^[ \t]*(?:[-*+]|\d+[.)])[ \t]+/gm, "")
.replace(/(\*\*|__)(.+?)\1/g, "$2")
.replace(/(^|[^\w*])[*_]([^*_\n]+)[*_](?=[^\w*]|$)/g, "$1$2")
.replace(/~~(.+?)~~/g, "$1")
.replace(/^[ \t]*(?:---|\*\*\*|___)[ \t]*$/gm, "")
.replace(/<\/?[a-z][^>]*>/gi, "")
.replace(/[ \t]+/g, " ")
.replace(/\.(\s*\.)+/g, ".")
.replace(/\n{3,}/g, "\n\n")
.trim();
}
/** Split into parts at paragraph, then sentence, then word boundaries. */
export function speechParts(text: string, first = FIRST_PART, size = PART): string[] {
const sentences = text.split(/(?<=[.!?…:;])\s+|\n+/).map((s) => s.trim()).filter(Boolean);
const parts: string[] = [];
let current = "";
const limit = () => (parts.length === 0 ? first : size);
const push = () => { if (current) parts.push(current); current = ""; };
for (let sentence of sentences) {
while (sentence.length > size) {
// One very long sentence: cut at a word boundary.
push();
const cut = sentence.lastIndexOf(" ", size) > size / 2 ? sentence.lastIndexOf(" ", size) : size;
parts.push(sentence.slice(0, cut).trim());
sentence = sentence.slice(cut).trim();
}
if (current && current.length + 1 + sentence.length > limit()) push();
current = current ? `${current} ${sentence}` : sentence;
}
push();
return parts;
}
let session: { abort: AbortController; audio: HTMLAudioElement | null; urls: string[] } | null = null;
function teardown() {
if (!session) return;
session.abort.abort();
session.audio?.pause();
session.urls.forEach((u) => URL.revokeObjectURL(u));
session = null;
}
function playBlob(blob: Blob, s: NonNullable<typeof session>): Promise<void> {
return new Promise((resolve, reject) => {
const url = URL.createObjectURL(blob);
s.urls.push(url);
const audio = new Audio(url);
s.audio = audio;
const done = () => { s.abort.signal.removeEventListener("abort", done); resolve(); };
audio.onended = done;
audio.onerror = () => reject(new Error("This audio couldn't be played."));
s.abort.signal.addEventListener("abort", done);
audio.play().catch(reject);
});
}
export const useReadAloud = create<ReadAloudState>((set) => ({
id: null,
phase: null,
error: null,
stop: () => { teardown(); set({ id: null, phase: null, error: null }); },
dismiss: () => set({ id: null, phase: null, error: null }),
play: (id, markdown) => {
teardown();
const parts = speechParts(speakableText(markdown));
if (!parts.length) return;
const s = { abort: new AbortController(), audio: null, urls: [] as string[] };
session = s;
const mine = () => session === s;
set({ id, phase: "loading", error: null });
void (async () => {
try {
let next: Promise<Blob> = speakApi.audio(parts[0], s.abort.signal);
for (let i = 0; i < parts.length; i++) {
const blob = await next;
if (!mine()) return;
// Fetch the next part while this one plays.
if (i + 1 < parts.length) {
next = speakApi.audio(parts[i + 1], s.abort.signal);
next.catch(() => {}); // handled when awaited
}
set({ phase: "playing" });
await playBlob(blob, s);
if (!mine()) return;
}
if (mine()) { teardown(); set({ id: null, phase: null }); }
} catch (err) {
if (!mine()) return;
teardown();
if (err instanceof ApiError && err.data.needs === "speech") set({ id, phase: "needs-setup", error: null });
else if (err instanceof ApiError && err.status === 404) set({ id, phase: "error", error: "This Mercury can't read aloud yet. Update Mercury to use it." });
else set({ id, phase: "error", error: err instanceof Error ? err.message : String(err) });
}
})();
},
}));