What's new in Orbit

**What changed** — When a visitor signs up on your branded (white-label) domain, Devotel Orbit adds them to your account as a team member instead of creating a separate workspace for them. They land in your workspace at the lowest permission seat, ready for you to promote, and the signup hands straight off to your branded dashboard.

Previously these signups were instead routed onto the standard workspace form, which asked for a workspace name — a step the visitor had typically already answered once during signup. Today we also correct an earlier changelog entry that announced a different arrangement for branded-domain signups (a new intermediate workspace under your account): that arrangement could break on branded domains owned by subaccounts, so it has been replaced by the team-join behaviour above. The entry has been rewritten; you can treat this one as the current reference.

The 24-hour customer-service window is the most-asked-about WhatsApp failure mode, and until now the error code that marks a refusal to send outside it had no dedicated fix page. A new runbook — **Troubleshooting: WhatsApp 24-hour session expired** — walks the failure end to end in Devotel Orbit's docs:

- What the refusal looks like on a send request, a Delivery Log row, a campaign recipient failure, and a webhook event.
- The cause table: window lapsed with no pre-flight, free-form sent from a template-only state, and first contact before any window ever opened.
- The fixes: gate your composer on the window-status check (GET /api/v1/messages/whatsapp/window-status), fall back to an approved template, and let the recipient's reply re-open free-form messaging.
- What not to do: the refusal is deterministic, so tight retries change nothing — re-route to a template or another channel instead — and when to escalate to support.

The FAQ under Operator Messaging Questions gains a matching entry that hands off to the runbook, and the page is wired into the docs navigation under Reference → Troubleshooting → Channels and integrations.

When Meta returns its generic "Something went wrong" rejection on a WhatsApp
send (Meta error 131000 or the similarly-worded 141000), Devotel Orbit used
to treat that response as a genuine provider failure. A handful of these
could lock the channel's internal circuit breaker, which then made
subsequent sends fail with a generic "Channel temporarily unavailable"
page — even though the WhatsApp account itself was fine. The catalog now
recognizes this fault as the recoverable blip Meta documents it as: the
send retries normally, the channel stays available, and your message only
surfaces as a clean retry rather than a hard failure. No configuration
change is needed on your account.

The Voice Agent Templates page in the dashboard puts a small icon on every pre-built voice agent card. Three template pairs shared the same icon, and every template category colored its icon differently, so eleven cards read as though several templates were duplicates of one another and the page lacked a consistent visual register.

Each template now declares a single-purpose icon in the catalog — the retail order-status template moved off the shared shopping cart onto a package-with-search glyph, the clinic-scheduling template moved off the shared calendar onto a calendar-with-heart, and the banking card-freeze template moved off the shared credit card onto a shield — and the card component renders every icon in the same accent color. The category badge keeps its per-category tint so grouping is still scannable, and a regression test on the catalog now fails if two templates ever collapse onto the same icon again.

When your account had fewer than 20 finalised calls, the Voice hub's Success Rate tile rendered the sentence "Awaiting more calls (N/20)" as its primary value. That was the only KPI on the row that displayed a full sentence at the large metric font — every sibling tile ("Active Calls", "Total Minutes", "SIP Trunks", and the mature Success Rate percentage) shows a short numeric value. The result looked visually inconsistent and was flagged as a poorly styled font.

The tile now always shows a short numeric value: the percentage once 20 or more calls have finalised (unchanged), or the progress "N/20" count while sampling (e.g. "5/20"). The explanatory sentence has been moved into the muted hint line below the value, so the tile still communicates that the percentage appears once more calls have finalised — it just no longer renders that explanation at the big metric font.

Voice real-time text (RTT) is the RFC 4103/T.140 equivalent of TTY/TDD relay:
a deaf or hard-of-hearing caller can type text on the same voice leg, and the
agent reads and responds in the softphone's RTT panel. Until now the only
reference to it was in the API route summary — no guide ever walked through
how to enable or use it.

The new **Real-time text (RTT/TTY-TDD) accessibility on live voice calls**
guide closes that gap:

- **Arm the channel on a live call.** Flip the softphone's **Real-time text
(TTY)** toggle, or post to the internal RTT config route from an IVR step
before the call reaches a queue. The agent's send endpoint also accepts
the first outgoing line and auto-flips the config, so a dashboard agent
typing is itself the per-call opt-in.
- **Exchange frames on the live call.** Caller text streams over the SSE
/rtt/stream connection; agent replies post to /rtt/send and push back
over the caller's endpoint. Frames carry an is_final marker and a
call-relative timestamp.
- **Replay the transcript.** Persisted rows are tagged source = "rtt_relay"
and land on the same transcript model as the post-call speech-to-text
rows — tenant-isolated, soft-deletable, and exportable through the
existing caption path. A 2000-character per-frame ceiling guards the
relay endpoint.

The guide also describes the tenant-owned rollout posture — a queue or DID
lurking on an accessibility lane opens the channel before an agent picks up,
while the default per-call opt-in keeps RTT off until something arms it.

Until now, the codes QUEUE_DEPTH_CAP_REACHED and OVERFLOW_QUEUE_NOT_FOUND existed on the error-codes reference with only "Platform-defined" placeholder text, and no runbook told you which gate actually refused the request. The new [queue capacity gates](/troubleshooting/queue-capacity-gates) page covers the full gate set: the tenant-set depth cap that yields a 503 on enqueue, the overflow-chain validation that yields a 422 on queue config writes, and the 409 you get when you try to re-prioritise an entry that already left the queue. Each cause has its fix — raise the per-org acd_queue_max_depth cap, point overflowQueueId at a resolved non-cycling target, or divert into the press-1 virtual-callback gate — plus a decision checklist and a "what not to do" list.

The Billing › Voice Destinations page completes the rate-inventory hierarchy for voice: enter a destination prefix, map it to an MCC/MNC, operator, and country, and the inventory's voice-rate console can then key on that prefix instead of silently falling back to a country-level rate. This is the last mile of the report that asked for both SMS Destination Management (MCC/MNC — already live at /dashboard/mccmnc) and Voice Destination Management (prefix — built here).

**What changed** — On the Voice Clones page, creation modes (clone from a recording, or design a voice from a text description) previously appeared unconditionally. Where the platform's voice-synthesis provider credential was not provisioned, clicking **Design voice** resulted in a 503 Service Unavailable error with no way to resolve it from the dashboard. The page now checks a capability notice and hides those modes when the feature cannot succeed, replacing the dead-end button with a clear "Voice clone creation is unavailable" message (retry included if the capability check itself was unreachable). Available modes behave exactly as before.

Voice › Call Logs had a regression: successful calls — including completed calls with durations over 40 seconds — rendered $0.00 in the Cost column, even though the wallet debit for each call had been posted correctly. The accounting finalize step failed to stamp the final charge onto the call record itself, so the dashboard had nothing real to show and the Cost total for the period was understated.

With this fix the finalize step writes the actual charge, the per-minute rate, and the destination onto the call record, so the Cost column shows the per-call amount that was debited, and the Total cost summary matches it.

A new post joins the blog's comparison series: **Vapi vs Retell vs Bland vs Synthflow** — the four-way question a voice-AI buyer weighing finalists actually asks, which the existing one-vendor head-to-heads each answer only partially.

The post names the specialist category taxonomy first (pipeline constructors, wrapper constructors, and the speech-model vendor), then holds all four specialists, ElevenLabs, and Orbit to the same three evaluation dimensions — a published per-stage voice latency budget, model-cost governance, and handback plus outbound tooling — and closes with the single table a finalist comparison needs.

Product behavior is unchanged. If you're already on Orbit, nothing in the dashboard, the API, or your account moves — the comparison coverage this post adds lives on the blog, next to the sibling /compare registry pages it links back to.

A new playbook joins the A2P SMS series on the blog: sending SMS into both the US and Canada, the two-country sender-id map.

Teams shipping into both markets often inherit US-only documentation and misconfigure the Canadian sender type at provisioning time. The playbook lays out the carrier-by-carrier split (AT&T/T-Mobile/Verizon against Bell/Rogers/Telus), registration windows, sender-id posture — including where registered alphanumeric sender IDs fit, and why they are one-way — and opt-out handling under TCPA in the US and CASL in Canada.

No product behavior changed. If you are already on Orbit, nothing in the dashboard, the API, or your account moves — the gap this post closes is in the blog's regional-playbook coverage.

The announcements hub previously covered regulatory and fraud explainers, but had no reactive explainer for competitor event cycles — the Twilio Signal 2026 announcements in particular. That gap is now closed with a new industry-news explainer that recaps the dated Signal items (the conversation layer, Conversation Orchestrator, Conversation Memory, Conversation Intelligence, Agent Connect, the new console, and Conversation Relay updates), states which /compare/twilio rows each announcement confirms or shifts, and names the tenant-owned controls that close the gaps the incumbent's cycle left open.

The post carries a buyer-facing FAQ (published as FAQPage JSON-LD), attributes every Signal item to the vendor's own public writing, and deliberately omits the restructuring storyline when a public source for it cannot be cited. It follows the same quick-answer → recap → row-shift map → tenant-owned controls → FAQ → takeaway shape the other industry news explainers on the announcements hub use.

Devotel Orbit's error reference lists every code the API can return, but
until now the generic system codes carried placeholder descriptions and no
runbook. A new troubleshooting page,
[Troubleshoot system and 5xx error
codes](https://docs.orbit.devotel.io/reference/troubleshooting-system-5xx),
fills that gap: it shows how to read the retryable flag and Retry-After
window a response envelope may carry, classifies each system code as
transient (safe to retry) or deterministic (fix the payload, don't retry),
and lays out the decision checklist — check the status page, retry once
where safe, escalate with the request_id. Worked payload examples show
what each class looks like on the wire.

When a comparison run in Voice → STT Playground fails for one provider (for
example, a missing or invalid Deepgram key), the failing provider card now
shows a "Manage the STT API key" link next to the error. The link takes you
straight to Settings → Voice, which is where you register, activate, and
rotate the speech-to-text credential the comparison uses.

Previously the error text appeared with no guidance, so you had to leave the
playground and find the right settings page yourself. The fix only touches
that error row on the comparison results — successful transcripts and the
"not yet available" provider rows are unchanged.

Once you've registered your own Deepgram API key on Settings → Voice →
Speech-to-text and activated it, every comparison you run in Voice →
STT Playground uses that key — the same routing the rest of call
transcription already follows.

Previously the playground kept using the platform-owned credential even
after your key showed active, so a comparison run could report a credential
error even though your key was valid. The playground now resolves your
registered key on every run, and falls back to the platform credential only
when no active key is registered.

The Channels › SMS compose phone preview mocked an iOS/Android frame with status-bar elements that were oversized for its narrow device chassis — at some message lengths the trailing clock and the Dynamic Island overlapped the leading 5G label, and the iOS island itself swallowed the row into which the time should sit. The preview now renders both the clock and the signal label at the smallest sanctioned text size with the iOS island shrunk and the status row laid out to guarantee no overlap at any fallback label width; the live delivery of the composed text into the frame is unchanged.

**What's new** — [Single-tenant vs multi-tenant CPaaS: which isolation model fits](/compare/single-tenant-vs-multi-tenant) is a buyer's guide for the question that follows how many vendors you should have: how many customer workloads should share one communications platform. It defines both postures in one answer-first block, scores a dedicated instance against a shared platform on the five criteria that actually decide the choice — tenant isolation, per-account SLA, data residency, upgrade cadence, and audit scope — and closes with a scenario guide and an FAQ buyers and answer engines can cite directly.

**Why it matters** — "Would a dedicated instance isolate us better?" usually costs dedicated-instance money to harden a boundary a well-built shared platform already enforces. The guide shows the checklist that settles it, links into the [white-label reseller](/white-label) and [PaaS infrastructure](/paas-infrastructure) pages for the isolation options Orbit ships, and points readers who started from the provider-count question to the [one provider vs many vendors](/vendors-comparison) comparison for that decision.

Session cookies are scoped to the primary Devotel Orbit domain, and a reseller's branded domain cannot read them. The visible result: a customer signed in on the primary who clicked through to the branded site was treated as signed out, pushed to its sign-in page, and often landed right back on the primary.

Branded domains now perform the session handshake with the primary on first load. An existing sign-in is picked up automatically, and the visitor lands signed in on the branded domain. Visitors who are genuinely signed out are unaffected — they are asked to sign in as before — and nothing changes on the primary domain itself.

The [Embedded Web SDK API reference](/docs/api-reference/sdk) closed its coverage gap for the endpoints your embeds actually call. Alongside identify, track, and personalize, the page now documents:

- **In-app message feed** — GET /sdk/in-app/feed pulls the visitor's eligible content cards and in-app messages, with per-visitor segment targeting honored in the response.
- **In-app engagement receipts** — POST /sdk/in-app/events reports impressions, clicks, and dismissals (including which button a click hit) into the same event stream as your tracked events.
- **Unload-flush event intake** — POST /sdk/track/anonymous accepts the pending batch on page unload over a key-free path, so older browsers that lack fetch({ keepalive: true }) still deliver final events without a key ever touching a sendBeacon URL.

If you render the in-app channel with your own components instead of the SDK's built-in renderer, the request and response shapes for both directions are now specified end to end.