[ralyodio]

Anthony: "give my agents a voice using this instead of elevenlabs" →
"can't we use c0mpute.com?" → "use smshub.dev".

Alert calls read the line in a generic Telnyx female voice. Our agents have
voices of their own (riotcoder is kokoro / am_michael in its OpenProfile),
and since c0mpute 0.2.58 every c0mpute node provisions Kokoro itself — so the
network can speak for us, with no per-character bill.

Three tiers, best first

1. Our own network. POST /v1/tts on the c0mpute speech gateway auctions a
   tts.synthesize job to a node running Kokoro and hands back a WAV; the call
   plays it with playback_start.
2. Telnyx's copy of the same model (Telnyx.KokoroTTS.<voice>) when no
   gateway is in reach — same voice, vendor round trip, no hosting.
3. The old generic voice, only if Telnyx ever rejects a Kokoro voice id.

Every step fails soft — no gateway configured, a node that cannot be reached, a
200 that is not audio — because an alert is worth more than the voice it
arrives in.

The audio URL

The rendered WAV waits in a disk cache under an HMAC-named path, and
/api/v1/alerts/audio/<token> serves it unauthenticated, because the only
fetcher is Telnyx's media loader. The token is an HMAC of the call id under a
server secret, so a path cannot be guessed or walked; the URL stops working
after ten minutes, and the sweep runs on each write. Not /tmp — that is
quota-bound on dev2.

call.playback.ended now hangs up on the same h = 1 rule as
call.speak.ended, so a played line ends a call exactly like a spoken one.

Config

Variable	Meaning
C0MPUTE_SPEECH_URL, C0MPUTE_SPEECH_TOKEN	the c0mpute speech gateway (tier 1). Unset → tier 2
ALERT_KOKORO_VOICE	af_sarah, af_bella, am_onyx, am_michael (default)
ALERT_AUDIO_DIR	where rendered lines wait (default .cache/alert-audio)
ALERT_AUDIO_SECRET	signs the path token; falls back to the service-role key

Testing

• Full suite on the pre-commit hook: 380 tests / 62 files pass, typecheck
  and lint clean. 8 of those are new, covering the gateway being off, an empty
  line, a non-absolute base URL, a refusing gateway, a throwing fetch, a 200
  that is not audio, token path-walking (../../etc/passwd), tokens being tied
  to the secret, and the TTL sweep.
• The updated alert-flow test now asserts the call speaks
  Telnyx.KokoroTTS.am_michael rather than a generic voice.
• Synthesis itself verified against the live gateway on dev2 before writing
  this: 15.56 s of 24 kHz audio in am_michael, auctioned to a Kokoro node.

🤖 Generated with Claude Code

[github-actions]

ThreatCrush Security Scan

14 finding(s)

HIGH/CRITICAL: 2 | MEDIUM: 4 | LOW: 8

Severity	Rule	Location
HIGH	sh-unquoted-expansion-destructive	public/install.sh:67
HIGH	sh-remote-script-execution	public/install.sh:89
MEDIUM	sh-remote-script-execution	public/install.sh:133
MEDIUM	js-open-redirect	src/app/numbers/page.tsx:180
MEDIUM	js-open-redirect	src/components/inbox-client.tsx:35
MEDIUM	js-unescaped-html-sink	src/components/pit-footer.tsx:17
LOW	secret-generic-credential	src/lib/providers/__tests__/twilio.test.ts:24
LOW	secret-generic-credential	src/lib/providers/__tests__/twilio.test.ts:46
LOW	secret-generic-credential	src/lib/providers/__tests__/twilio.test.ts:60
LOW	secret-generic-credential	src/lib/providers/__tests__/twilio.test.ts:78
LOW	secret-generic-credential	src/lib/vault/__tests__/vault.test.ts:117
LOW	secret-generic-credential	src/lib/webhooks/__tests__/supabase-store.test.ts:11
LOW	secret-generic-credential	supabase/config.toml:287
LOW	secret-generic-credential	supabase/config.toml:319

Snippets are redacted; ThreatCrush never prints matched credential material.