Execute control-flow tool actions, and take the cold start off the first turn

[My-CMDhub]

The dispatcher answers hang_up_call and transfer_to_staff with an
action field, but the cascaded tool loop only serialized the result
back to the LLM and never read it. The model therefore narrated both
actions while nothing performed them: it said goodbye and left the line
open, and promised a transfer that never happened.

Verified live — hang_up_call fired 3 times on one call (11:13:00,
11:15:33, 11:15:41 AEST) and the leg only ended when the caller hung up
at 566s.

• read action after dispatcher.execute
• terminate via Twilio REST Status=completed, idempotent (the model
  retries a tool that produces no observable effect)
• transfer via a TwiML update with a redirect back to the AI on
  no-answer
• both fire after the turn's audio has drained, so the farewell and the
  handoff line are heard first; a barge-in during the farewell aborts
  the hangup

---

Second commit — perf(voice): remove first-call cold start from the voice hot path:

First call after a dyno boot reached "Call context ready" 4,316ms after the
Twilio stream started; the second call on the same dyno took 2ms. Whoever
calls first pays the difference, and voice_settings (voice, TTS model, Flux
turn thresholds) are not applied until it clears, so the opening turn also
ran on Deepgram's defaults rather than the tenant's configuration.

Two causes, both removed:

• Tenant rows are keyed by an Appwrite document id, not by their slug, so
  the primary GET .../documents/coalcreek always 404'd before the slug
  query found the row — ~920ms of pure waste per cache miss, logged as a
  success. Slug query is now primary; document-id lookup stays as fallback.
• openai and the dispatcher chain were imported inside
  _build_call_context(), so the first call paid a blocking cold import on
  the event loop. Startup now warms those imports and pre-fetches the tenant
  config into its 5-minute cache.

Warm-up failures are logged and swallowed — a cold first call is slow, not
broken.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ovela-ai	Ready	Preview	Aug 19, 2026 10:50am