dexterm-cli 0.1.1 → 0.1.3

package/dist/packages/voice/src/tts/voices.js

@@ -0,0 +1,209 @@
+// Voice resolution, the novelty ban-list, and speech-text sanitisation (task-1393). All pure and
+// dependency-free so every rule is unit-testable without spawning `say`.
+//
+// Why this exists: SayTts used to be constructed with NO default voice, so macOS fell back to the
+// machine's *system* voice — which differs per machine and can be a novelty/odd voice (a maintainer heard a
+// "terrifying" one). This module makes the default an EXPLICIT deep British male and guarantees a
+// novelty voice (Zarvox, Bad News, …) can never be auto-selected and the system default is never used
+// silently: if nothing safe is installed, voice degrades to text (voice === null), never to "whatever
+// the OS picks".
+/**
+ * Novelty / "scary" voices that must never be auto-selected, normalised to lowercase base names (the
+ * locale qualifier like " (English (UK))" is stripped first — see normalizeVoiceName). The list is the
+ * classic MacinTalk novelty set plus the names the task named explicitly ("and similar"). The resolver
+ * also uses POSITIVE allow-lists, so this is defence in depth; the ban test asserts no resolved voice
+ * is ever on it.
+ */
+export const BANNED_VOICE_NAMES = new Set([
+    'albert', 'bad news', 'bahh', 'bells', 'boing', 'bubbles', 'cellos', 'deranged', 'fred',
+    'good news', 'hysterical', 'jester', 'junior', 'organ', 'pipe organ', 'princess', 'ralph',
+    'superstar', 'trinoids', 'whisper', 'wobble', 'zarvox',
+]);
+/**
+ * Strip the locale qualifier macOS appends to the newer voices ("Eddy (English (UK))" → "Eddy"),
+ * collapse whitespace, lowercase. Used for ban/allow matching so "Eddy (English (US))" and
+ * "Eddy (English (UK))" share one base name.
+ */
+export function normalizeVoiceName(name) {
+    return name.replace(/\s*\(.*\)\s*$/, '').trim().toLowerCase();
+}
+/** True when a voice is on the novelty ban-list (matched on its normalised base name). */
+export function isBannedVoice(name) {
+    return BANNED_VOICE_NAMES.has(normalizeVoiceName(name));
+}
+/**
+ * Parse the output of `say -v '?'`. Each line is `<name>  <locale>  # <sample sentence>`, but the name
+ * column is only padded to a fixed width, so a LONG name ("Eddy (English (UK))") leaves a SINGLE space
+ * before the locale — the old regex required two and silently dropped exactly the en_GB voices we care
+ * about. This parser anchors on the trailing locale token instead, so it never loses a voice.
+ */
+export function parseSayVoices(stdout) {
+    const voices = [];
+    for (const raw of stdout.split('\n')) {
+        if (!raw.trim())
+            continue;
+        const hash = raw.indexOf('#');
+        const left = (hash >= 0 ? raw.slice(0, hash) : raw).trimEnd();
+        // name = everything up to the last whitespace run; locale = the final token (en, en_GB, zh-CN…).
+        const m = left.match(/^(.*\S)\s+([a-z]{2}(?:[_-][A-Za-z0-9]+)?)$/);
+        if (m)
+            voices.push({ name: m[1].trim(), locale: m[2] });
+    }
+    return voices;
+}
+/**
+ * Deep British male voices, best-to-worst. Eddy (the maintainers' pick, 2026-10-15) is the dexterm default; Daniel is the canonical deep British male that
+ * ships with every macOS and is the next fallback. Oliver/Arthur are the newer en_GB male voices (present only if the
+ * user downloaded them). Matched on normalised base name.
+ */
+export const BRITISH_MALE_PREFERENCE = ['eddy', 'daniel', 'oliver', 'arthur'];
+/**
+ * Per-voice delivery defaults, applied only when the user has not saved their own rate/pitch.
+ *
+ * task-1411: the old Eddy override (rate 168 + pbas 30) came from an audition sample the maintainers never
+ * explicitly picked; it was reported as "crazy bad" and "way too slow", and the pitch shift distorted
+ * the voice. It is removed: every voice now uses the engine-wide DEFAULT_RATE (a brisk ~205 wpm) and NO
+ * pitch override unless the user sets `voice.pitch` themselves. Keep this map empty rather than deleting
+ * the hook, so a future per-voice tweak has a home and the public `voiceStyleDefaults` export is stable.
+ */
+export const VOICE_STYLE_DEFAULTS = {};
+/** The delivery defaults for a resolved voice name, or undefined to use the engine-wide defaults. */
+export function voiceStyleDefaults(name) {
+    return name ? VOICE_STYLE_DEFAULTS[normalizeVoiceName(name)] : undefined;
+}
+/**
+ * en_GB voices that read masculine / neutral and are acceptable British fallbacks when no first-class
+ * British male is installed (Grandpa is explicitly a male voice; Eddy/Reed/Rocko are the modern
+ * expressive voices under audition). Matched on normalised base name, en_GB locale only.
+ */
+export const BRITISH_FALLBACK = ['grandpa', 'eddy', 'reed', 'rocko'];
+/** Safe English male voices in any English locale, used only when nothing British is installed. */
+export const ENGLISH_MALE_ALLOW = ['daniel', 'oliver', 'arthur', 'alex', 'tom', 'aaron', 'rishi'];
+const isEnglish = (locale) => /^en([_-]|$)/i.test(locale);
+const isBritish = (locale) => /^en[_-]gb$/i.test(locale);
+/** Pick the best installed match for a base name, preferring an Enhanced/Premium variant and the given locale. */
+function bestMatch(installed, base, preferLocale) {
+    const matches = installed.filter((v) => normalizeVoiceName(v.name) === base && !isBannedVoice(v.name));
+    if (!matches.length)
+        return null;
+    const score = (v) => {
+        let s = 0;
+        if (/\((enhanced|premium)\)/i.test(v.name))
+            s += 4; // higher-quality download, if present
+        if (preferLocale?.(v.locale))
+            s += 2;
+        return s;
+    };
+    return matches.slice().sort((a, b) => score(b) - score(a))[0];
+}
+/**
+ * Find the installed voice a saved name refers to. The `,` key saves the FULL name ("Eddy (English (UK))"),
+ * and Eddy/Reed/Rocko each exist in many languages, so an exact full-name match must win. Only when the
+ * saved name is a bare base name ("Eddy") do we choose among same-base voices: en_GB first, then any
+ * English, then anything else (a verifier found "Reed (English (UK))" resolving to the German Reed).
+ */
+function findConfiguredVoice(installed, configured) {
+    const exact = installed.find((v) => v.name === configured && !isBannedVoice(v.name));
+    if (exact)
+        return exact;
+    const folded = configured.toLowerCase();
+    const ci = installed.find((v) => v.name.toLowerCase() === folded && !isBannedVoice(v.name));
+    if (ci)
+        return ci;
+    const base = normalizeVoiceName(configured);
+    const rank = (v) => (isBritish(v.locale) ? 0 : isEnglish(v.locale) ? 1 : 2);
+    return installed
+        .filter((v) => normalizeVoiceName(v.name) === base && !isBannedVoice(v.name))
+        .sort((a, b) => rank(a) - rank(b))[0];
+}
+/**
+ * Resolve the voice `say` should use. The fallback chain, in order:
+ *   1. the user's configured voice (if installed and not banned)
+ *   2. the chosen default: a British male (Eddy → Daniel → Oliver → Arthur), Enhanced/Premium preferred
+ *   3. the next British voice (en_GB Grandpa/Eddy/Reed/Rocko)
+ *   4. any safe English male voice (Alex/Tom/…)
+ *   5. null → replies stay text-only
+ * A banned (novelty) voice is NEVER returned, and the OS system default is NEVER used silently.
+ */
+export function resolveVoice(opts) {
+    const installed = opts.installed ?? [];
+    const configured = opts.configured?.trim();
+    // 1. Configured voice wins when it is actually installed and safe.
+    if (configured) {
+        const hit = findConfiguredVoice(installed, configured);
+        if (hit && !isBannedVoice(hit.name)) {
+            return { voice: hit.name, source: 'config', reason: `your configured voice (${hit.name})` };
+        }
+        // Configured but missing or banned: fall through to the safe default rather than honour a bad value.
+    }
+    // 2. Chosen default: British male (Eddy → Daniel → …). en_GB only: Eddy also exists in other locales.
+    const britishInstalled = installed.filter((v) => isBritish(v.locale));
+    for (const base of BRITISH_MALE_PREFERENCE) {
+        const v = bestMatch(britishInstalled, base, isBritish);
+        if (v) {
+            const isDefault = base === BRITISH_MALE_PREFERENCE[0];
+            return {
+                voice: v.name,
+                source: 'default',
+                reason: isDefault ? `${v.name} — deep British male (dexterm default)` : `${v.name} — British male voice`,
+            };
+        }
+    }
+    // 3. Next British voice (en_GB, masculine/neutral).
+    for (const base of BRITISH_FALLBACK) {
+        const v = installed.find((x) => normalizeVoiceName(x.name) === base && isBritish(x.locale) && !isBannedVoice(x.name));
+        if (v)
+            return { voice: v.name, source: 'british', reason: `${v.name} — British voice (no Daniel installed)` };
+    }
+    // 4. Any safe English male voice, any English locale.
+    for (const base of ENGLISH_MALE_ALLOW) {
+        const v = installed.find((x) => normalizeVoiceName(x.name) === base && isEnglish(x.locale) && !isBannedVoice(x.name));
+        if (v)
+            return { voice: v.name, source: 'english', reason: `${v.name} — English voice (no British male installed)` };
+    }
+    // 5. Nothing safe: text only. Never the system default.
+    return { voice: null, source: 'none', reason: 'no safe voice installed — spoken replies are off (text still works)' };
+}
+// ------------------------------------------------------------------ speech-text sanitisation
+/** Default cap on spoken length. Long analyst answers are summarised upstream; this is a hard backstop. */
+export const MAX_SPEAK_CHARS = 600;
+const BASE58 = '1-9A-HJ-NP-Za-km-z';
+/** A Solana mint / base58 address (32-44 chars) or a 0x… EVM address — never read 44 characters aloud. */
+const ADDRESS_RE = new RegExp(`\\b(?:0x[0-9a-fA-F]{40}|[${BASE58}]{32,44})\\b`, 'g');
+/** Short, speakable form of an address: "7xKX…Dgkp" reads as a few characters, not forty-four. */
+function shortAddress(addr) {
+    const body = addr.startsWith('0x') ? addr.slice(2) : addr;
+    return `${body.slice(0, 4)}…${body.slice(-4)}`;
+}
+/**
+ * Make arbitrary text safe to hand to `say`. Untrusted text (token names, socials, analyst output) can
+ * contain `say`'s own `[[…]]` embedded speech commands — `[[slnc 5000]]`, `[[rate 10]]`, even a voice
+ * switch — which `say` INTERPRETS whether the text arrives on argv or stdin. We also never want a
+ * leading `-` parsed as a flag. This:
+ *   - defuses `[[` / `]]` so no embedded command is ever executed (kept readable, not deleted)
+ *   - neutralises a leading `-`/`+` (defence in depth; we also deliver via stdin, not argv)
+ *   - collapses control characters and whitespace
+ *   - shortens contract addresses to a 4…4 form instead of reading 44 characters
+ *   - caps the length on a word boundary
+ */
+export function sanitizeSpeech(text, opts = {}) {
+    const maxChars = opts.maxChars ?? MAX_SPEAK_CHARS;
+    let s = String(text ?? '');
+    // Defuse `say` embedded commands: a split `[ [` / `] ]` is no longer a command token.
+    s = s.replace(/\[\[/g, '[ [').replace(/\]\]/g, '] ]');
+    // Strip control characters (NUL, ESC, etc.) that could confuse the child or the terminal.
+    s = s.replace(/[\u0000-\u0008\u000B\u000C\u000E-\u001F\u007F]/g, ' ');
+    // Shorten addresses so we don't read 44 characters aloud.
+    s = s.replace(ADDRESS_RE, (m) => shortAddress(m));
+    // Collapse whitespace.
+    s = s.replace(/\s+/g, ' ').trim();
+    // A leading dash/plus could be read as a flag if it ever reached argv; neutralise it.
+    s = s.replace(/^[-+]+\s*/, '');
+    // Cap length on a word boundary.
+    if (s.length > maxChars) {
+        const cut = s.slice(0, maxChars);
+        const sp = cut.lastIndexOf(' ');
+        s = (sp > maxChars * 0.6 ? cut.slice(0, sp) : cut).trim();
+    }
+    return s;
+}

package/package.json

@@ -1,9 +1,13 @@
 {
   "name": "dexterm-cli",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "dexterm: terminal DEX screener with a local Claude CLI analyst",
   "author": "The dexterm authors",
   "license": "AGPL-3.0-only",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/dexterm-xyz/dexterm.git"
+  },
   "type": "module",
   "bin": {
     "dexterm": "./dist/packages/cli/bin/dexterm.mjs"
@@ -32,13 +36,5 @@
     "@dexterm/recorder": "0.0.1",
     "@dexterm/tui": "0.0.1",
     "@dexterm/voice": "0.0.1"
-  },
-  "repository": {
-    "type": "git",
-    "url": "git+https://github.com/dexterm-xyz/dexterm.git"
-  },
-  "homepage": "https://dexterm.xyz",
-  "bugs": {
-    "url": "https://github.com/dexterm-xyz/dexterm/issues"
   }
 }