Phone line and hand-off architecture

A directory entry

Three customers, one table

Shapes per Fede's ruling, Oct 28. Maintenance and emergency: unchanged at all three.

1. Clara answers with the open greeting. The classifier decides the topic or the person asked for. (Sean, Oct 18 and Oct 27; Fede, Oct 27)
2. Topic of a module that is on: that module's agent. Maintenance and emergency: their flows, unchanged.
3. Topic of a module that is off, or a person ask, after hours: Clara helps or takes a message. No transfer (Fede, Oct 28). Maintenance and emergency keep their own paths.
4. In hours: the resolver reads the directory, the named entry or that topic's entries in order.
5. One dial path rings them; no answer means the next entry, then a message.

• Nothing set means a message (Fede, Oct 7); a number the customer forwards to Clara is never a destination (Sep 22).
• A transfer to the customer's team shows Clara's line; only an outside service, like the call center, sees the caller's number. (Fede, Oct 21)
• No "press 1" menus for callers. (Fede, Oct 20)
• Proposed: each entry says "ring me" or "take a message for me". Fede, Oct 15, recalling Western Slope: "if they asked to speak to someone explicitly ... we were allowed to forward." Western Slope, Oct 21: "if they're asking for me ... she should take a message."

Fede, Oct 29: "right now how it works is you have to be added to PropFlow as a user with a role, right, and then that would make you discoverable in the system… we'll need to check if we require a phone number when people sign up. We should."

• Source of truth: PropFlow people with a role. Today a list row is a name and a number; the link to a person's record is optional. Western Slope's hand-off rows have it, and the call dials that person's current phone; Camellia's ring-list rows have no link.
• Phone at sign-up: required when a person is added through the company's Team settings or a ring list; not asked on the invite-accept page, optional on the admin user API. Fix: required on every path that creates a staff person.
• The phone-system export (Situs's user list) is a one-time import helper at onboarding, never a second source. Group lines and outside services are entries without a PropFlow login.

Fede, Oct 29: "once we interact with someone we should remember their language preference… we shouldn't ask again on the next call." Read on 2026-10-29 from the code, the change history and 30 days of real calls at the three customers.

Every channel, 30 days

Only 4 of 346 automated texts and emails to prospects and tenants went out in Spanish, all to one person. Clara's own written replies follow the caller's language; the sends made on her behalf read the saved language, which is usually missing.

What to change (fixes, no switch): one offer policy at every customer instead of per-settings luck; a Spanish "yes" ("sí", "sí gracias", "sí por favor") counts as a switch; a saved language is used on the next call, text and email with no re-ask in either direction, and one sentence to switch back; the audit rows above marked broken get the same treatment.

Unverified: RingCentral ring strategies and overflow, outside numbers inside a Crexendo queue, dial by name; the Comcast guide predates a rename. Not borrowed: menus, dial by extension, hold queues, agent presence. Clara is the attendant.

Goes away: the leasing ring list and the older leasing hand-off setting; the two "everything else" settings; the "two modes" of Oct 7; the call platform's built-in transfer for person asks. Maintenance and emergency keep their settings, lists and dial paths.

1. Resident services: one switch instead of two (voice already reads the lease and billing one).
2. Directory table and card on the company page: today's ring-list people plus the three new fields; phone required on every staff-creation path.
3. Resolver: one function picks when the caller asks, proven by replaying real transfer history unchanged.
4. One dial path for person and topic asks.
5. Ask by name, honoring each entry's choice.
6. Delete the old settings after a supervised week.

Nothing turns on soon (Fede, Oct 28). Steps 1 to 4 must replay today's calls unchanged; step 5 and anything callers notice wait for Fede.

• Oct 15 incident: a law firm's call filed as a leasing lead.
• Oct 7 evening: inspection, six root problems.
• Oct 7 inspection: per-client state, two modes, route by name.
• Earlier design: two modes, one core, superseded.
• Sep 22 vision; Sep 21 emergency line proposal.

Where things stand

Incident Oct 15: a law firm's call was filed as a leasing lead Fix in review

A law firm called Western Slope and asked for Jay by name. Clara sent the call to the leasing ring list, and after the call she filed the caller as a prospect and armed the leasing follow-up texts. The caller's number was already on file as a vendor. The card was archived and its follow-ups stopped the same day. Nothing was written to AppFolio, because the front-door line has no AppFolio property behind it.

Root cause. #19632 (our own session, 10/14) widened the signal and relied on the AI reason alone, with no deterministic gate and no logging. The rule it widened had already been wrong 5 out of 5 times: every card it made was not a leasing call. The cards were the lawyer, Jay's own ask, a photographer's sales call, a job applicant and a hardware-store vendor.

Fede's decision order (Oct 15)

What the 90-day pull shows

Next: pick the outcome, not just "make a card?" Proposed — pending review

Fede: "Do I create a guest card? Do I create a tour? Do I send an application?" The after-call step should choose one outcome per call: guest card, tour, application link, team note, or nothing. Identity runs first, as above. The verdict and its reason are logged as ids and labels, never the call's words. The 90-day pull is the test set: 256 Western Slope calls and 134 Camellia transfers, each already labeled with the right outcome. The bar is 98% of real leads caught and 95% of cards correct.

Harness: evals/fast-tier/postcall-prospect-capture/replay/ (13 scrubbed real calls) and src/__tests__/postcall-capture-replay.test.ts; subscription runner scripts/replay-postcall-capture-subscription.ts. Before the fix, 7 of the calls made a wrong card. After, all pass, and all 7 planted breaks go red. Not yet proven on a live Willows call. Incident: #20208.

Decisions for you

Decided (Fede, Oct 7) "Take a message" is the default when nothing is set, and onboarding confirms each call type was looked at.

1. Leasing's hidden fallback to the office phone. (a) keep, (b) replace with explicit settings copied from each client's number, (c) delete. Recommend (b).
2. The ordered fix list (table below). Recommend yes to the "no switch" rows once the overnight proof passes.
3. Caller ID on transfers. (a) keep the caller's number everywhere (today), (b) a choice per destination on the Calls page: the caller's number or the line the call came in on, default = today, (c) always the client's own Clara line. Recommend (b). Nobody in the field decides by destination type; everyone makes it a setting. The only fix for a silenced personal phone is a known number saved once.
4. Emergency work-order chain (Western Slope, Oct 7 call): build it as the leasing ring list with a different trigger, an emergency-category work order. Recommend yes once the inventory below confirms what exists.
5. The three "your call" items: a property default list for on-site clients; retire the live note tool; move Saturday from office hours to tour hours at Camellia and Yale. Recommend deciding each after the morning results.

Where the phone line is going, and the gaps Proposed — pending review

At altitude: two facts carry the rest. First, anything set inside the voice vendor's agent configuration is shared by every client, so it can never differ per client. At dozens of customers with different phone setups, every hand-off has to run through one dial path we own, where who to ring, which number shows and whether a person answered are data, not code. Second, that data is the settings the customer sees on the Calls & messages page, read through Gera's shared lookup, so the page and the engine can never disagree. Everything else is implementation inside those two.

Pillars, each a ruling already made

Caller ID lineage: 9/20 open trade-off (caller's number vs Clara's line) → 9/23 caller's number everywhere → 10/8 a choice per destination. Older text in the history below that says "press 1 stays the fallback" or "staff still see the original caller's number" is superseded.

Gaps between today and that picture

Nothing in this section is built or switched on. Each gap closes as a fix with no switch, proven by replay and real calls on our own lines, in the order on the fix list above.

The caller ID setting, as it will look

One new line under each person on the Call routing card: Shows: the caller's number (today's behavior, the default) or the line the call came in on. Only PropFlow staff see the select; the customer sees the list as before. Change Joanna's select below and the phone shows what she would see.

Proved on real calls, Oct 9

Two real calls into the Fairhaven test line, each asking for a person. The call was passed to a PropFlow test phone. We read what that phone showed from Twilio.

Turning the setting on changes what the person’s phone shows, and turning it off puts it back. The setting was cleared afterward.

Turned on at Camellia, Oct 9, 2:04 PM Mountain. Fede’s call: Joanna’s row at JP&Co is set to “the line”. Fede then called the Camellia line from his cell at 2:16 PM and asked for a person; Clara rang Joanna and her phone showed the Camellia line, (844) 510-1007. Twilio shows the leg to Joanna leaving from the Camellia line.

Known gap: the app does not show this choice for Camellia today. The control renders only beside people on the company-level ring lists, and JP&Co stores Joanna on the building. Follow-up issue 18329; the redesign (one routing card at both the company and the building level, each row labeled with where it is stored) is paused pending Fede.

Every route still reaches a person after the shared lookup went live, Oct 9

The routing change that passes the call’s group to every lookup (the third routing PR) is live in production. We re-proved the Willows test company’s routes with real calls: the proof script resolved each route from the stored settings, then dialed the answer. Every dial reached our test phone. Two placeholder emergency numbers on the test buildings were refused by the script’s own safety list, which is the guard working.

These calls were placed at 2:01 PM Mountain from the code on main that includes the routing change. An earlier set at 1:10 PM ran from a stale checkout without it and was replaced; those five calls (CA903bd4…, CA7f7cad…, CAf1eb74…, CA113d28…, CA1dd74a…) reached the same test phone and stand as the “before”.

Before: the same five answers, resolved without the group step. After: the same five answers through the shared lookup, each one proven by a real call. No customer’s routing changed; the before/after on real customer settings was byte-identical in the PR. One note: the proof script does not look at the clock, so running it after closing time would dial the same five numbers and prove nothing new. The after-hours message desk is covered by its own eval scenarios, not by this script. The Fairhaven test company could not be used for this run because one of its buildings is missing the test flag. That building (“Company line bench building (live)”) is not created by any code in the repo, so we left its record alone for whoever owns it; the Willows test company covers the same routes.

What customers asked for on the phone line (client calls Sep 1 – Oct 7)

Read from the raw per-speaker transcripts of nine client calls: Western Slope (9/4, 9/9, 9/16, 9/22, 9/30, 10/7), MMG Denver (9/18), ConAm/Yale (9/29, 10/1). No JP&Co or Situs call in the window. Asks we already meet are left out (forward the main line, ring a list then take a message, written summary with transcript and recording, staff texting Clara, no duplicate guest cards).

Two signals. Nobody asked about caller ID, but Jay saved the after-hours number "so I know it's an emergency", which is the per-destination setting above. Nobody wants a key press to accept a call, but staff do expect to acknowledge an emergency, so the emergency chain needs a defined "got it" step that is not press-1.

The design

Superseded 2026-10-28 by the design at the top of the page: the "two modes, one core" shape below is replaced by the one-engine design (one shape for every customer). Kept for history.

Never infer, always observe (Fede, Oct 8)

Every setting: where it lives, what happens when it's empty

"All three" = property, group, company; nearest set wins. The group level is never read on a call today.

Per client: what each call reaches

System design (Oct 8): "one way to hand off a call" is a decision on the Clara platform vision page, beside "one picture / one door". This page stays the owner of the call-type routing design, the fix list and the proofs.

Call routing: end-to-end inspection and holistic design (Oct 7 evening) Proposed — pending review

Read-only, Oct 7 evening: no calls, texts or emails. Times Mountain. "Inferred" = not proven live.

TL;DR

Today: one call, end to end

Six root problems

The section above's "8 of 15 reached nobody" is Camellia only; 19 here is Camellia plus Western Slope.

Why it ended up this way

Moved to the top of the page (2026-10-07 restructure): the target design diagram, the "every setting" table and the per-client table are under The design.

Decisions for you

1. "Take a message": explicit setting or default?
   (a) An explicit "message" setting per call type. (b) The default when nothing is set. (c) Default, and onboarding confirms each call type was looked at.
   Recommend (c): no new switch per client, matches Jay's "other → message", catches a forgotten emergency. Decided (Fede, 10/07): (c)
2. Leasing's hidden fallback to the property office phone.
   (a) Keep. (b) Replace with explicit settings copied from each client's office phone. (c) Delete, no copy.
   Recommend (b): fits the no-silent-fallback ruling and keeps Camellia's and Yale's reach until you change it.
3. The ordered fix list (the inspection's 9 steps; supersedes "Fix order" above).
   Recommend: yes to the "no switch" rows once proven; the three "your call" rows wait on the morning results.

Moved to the top of the page (2026-10-07 restructure): the 9-step fix list is under Decisions for you; the overnight box is under Running overnight.

Inspection 2026-10-07: current state, what's broken, what's missing Proposed — pending review

Inspected 2026-10-07. All times Mountain. Facts are verified unless marked "inferred".

TL;DR

Two operating modes Decided (Fede, 10/07)

Current state per client

No client is fully on the new model. Only the triage and unknown-caller agents use the new connect step; five other agents still carry the old transfer button with blank legs. JP&Co's office hours (Mon–Fri 9–5, weekends closed) stand, matching Camellia's website: Fede decided 10/07 that calls respect the configured office hours and nobody is rung on Saturdays.

What's broken, worst first

No cross-client leak: an Oct 1 "missed call" email to Western Slope was a Western Slope call. Recipients are always read from the call's own property or organization.

Why nobody knew

What's missing to put every client on one model

Fede: "we want every client and customer in the same new architecture." The decided design (one routing core, people lists, no keypress, caller's own number, message with full context, one notice builder, proven destinations at onboarding) still lacks five things:

New: route by name Design only, not building yet

Open questions, with a recommendation:

1. Who is routable by name? (a) only people on a ring list; (b) every property manager, leasing agent and maintenance person on the staff record; (c) anyone on staff. Recommend (b), yes/hold.
2. Unknown name, in hours: (a) message for "the team"; (b) fall through to the property default list; (c) ask once, then (b). Recommend (c), yes/hold.
3. Hours for a named person: (a) the property's office hours; (b) per-person hours on the staff record, else (a). Recommend (b), yes/hold.

Test harness first, before any fix

The 10-06 page already proposes one harness: a grid of company × conversation kind × channel × situation, run on the real decision code with nothing live, red first with planted-wrong controls. This folds into that grid; no second plan.

Fix order

Fede has approved none of these yet. Decided already: the two operating modes, calls respect configured office hours, and the website tool limited to concessions.

1. Engineering alerts: nobody to connect, transfer outcome unknown, unkept "I'll leave a note". Deduped per client per hour, live clients only. Live 10/7 ("nobody to connect" and "outcome unknown"; the unkept-promise alert is still to do).
2. Hang-up backstop becomes the always-on never-strand path. Dark 28 days, no proof, already listed as "delete now?" on the feature-flag page. It is a fix-shaped switch: delete the switch, keep the path. Recommend: yes awaiting Fede
3. "No message, no email" reversed. Known caller and reason always produce a notice. Recommend: yes awaiting Fede
4. "Everything else" falls back to a property default list, then the company list, then a message. Fixes Yale before go-live; Western Slope keeps no catch-all person by its own choice. Recommend: yes awaiting Fede
5. Screening and voicemail detection: "record your name" is not a pickup. Recommend: yes awaiting Fede
6. Promise backup, team side only: an unkept promise emails the team again, never the caller. Recommend: yes awaiting Fede
7. Website tool: concessions only. Delete the dead code that could overwrite office hours, phone or amenities from the website; fix the "website" source label on manually entered concessions and the "Refresh from website" button that saves nothing. Office hours come from the company setting (property override only on the client's word); tour availability stays separate. Recommend: yes awaiting Fede
8. Camellia: move Saturday from office hours to tour hours; same check for Yale 25 Station (Saturday 9–4). Recommend: yes awaiting Fede
9. Finish moving office hours to one company home (the Sep 22 cleanup, never done). Recommend: yes awaiting Fede
10. Old transfer button retired on the five remaining agents. Recommend: hold awaiting Fede until the baseline is green on 1–5.
11. Route by name. After Fede's three answers. Recommend: hold awaiting Fede

Doc drift: this page's older caller-ID and pickup-mode entries predate the 9/23 rulings (original caller number shown everywhere; no keypress ever). Read those two rulings as current.

Decisions

• Two operating modes, one routing core: centralized clients keep destinations at the company level, on-site clients at the property level, with one default list per property. Inspection 10/07 2026-10-07
• Calls respect the configured office hours; nobody is rung on Saturdays at JP&Co. Tour availability is not office hours. 2026-10-07
• The website tool is for concessions only and never writes a setting Clara acts on; office hours have one home (company setting, property override only on the client's word). Dead write path deleted, no switch. 2026-10-07
• Ring-time text is off by default; the core is a caller reaching the primary, then the secondary. Automatic pickup stays off; press-1 accept is the mode. No more test calls to Fede until a stand-in run passes clean twice (done 9/20 night). 2026-09-20, superseded 2026-09-22 by #10835 — automatic pickup (AMD) is now the default, press-1 the fallback
• Record every leg, always 2026-09-19
• Staff use their own personal cell, not a company line 2026-09-19
• Automatic pickup first; press-a-key is the fallback, not the default 2026-09-20
• Whisper says who and what only — no other content 2026-09-20
• Handoffs always go out from Clara's line; raw-number passthrough removed 2026-09-20, supersedes the 2026-09-19 "option A" decision
• Clara is the permanent middleman on both legs — no raw cell numbers exchanged either direction 2026-09-20
• One ring list per role plus office hours; a maintenance emergency always rings; leasing calls after hours take a message 2026-09-20
• Office hours are not the same setting as tour availability 2026-09-20
• Ring-time text stays on; setting exists as a per-customer off switch. Shipped as a per-customer switch, on by default (PR #10250, merged 9/20). 2026-09-20
• Western Slope turns on today, 9/20, after a clean Willows run 3 (Fede's call, same-day rule waived). Camellia stays on today's single-number forwarding to Joanna until a Willows rehearsal of that shape and Fede's word; a read-only guard checks Camellia's routing before and after every test run. 2026-09-20

Caller ID on every handoff: what the vendors support (2026-09-23) Decided (9/23)

Chosen combination: blind transfer for single outside lines. For the staff ring list, our hunt step pushes TwiML <Dial callerId={From} sequential> onto the live call — Twilio's own published ElevenLabs handoff pattern (Twilio blog).

Must be proven by a real call:

• What a chained second webhook (blind transfer into our own number) actually receives as "From" and whether it carries a fresh CallToken.
• What caller ID a conference transfer presents — undocumented by ElevenLabs.
• How the staff-leg caller ID and attestation land on real cell carriers (Fede's and Gera's cells, checking spam labels).
• Whether a whisper url and machineDetection coexist on the same <Number>, and in what order.
• Whether a <Dial> forward carries the inbound SHAKEN PASSporT onward.

Risks: STIR/SHAKEN attestation C on passed-through numbers (spam-label exposure on staff cells); error 21210 if the binding is lost (retry once from the business number); CallToken TTL undocumented; the first endpoint that connects swallows a sequential hunt.

Status: Proof in progress via the voice fast lane; PR #11075 + hunt-route PR; merge = deploy to every production agent, on Fede's yes.

At a glance: today and where it's going

These two pictures are the whole page in miniature — what a call does today, and where the design is headed. Everything below is the backup for these two pictures.

What changes between the two

Showing the caller's own number when Clara forwards a call Decided (9/22)

Decision for Fede

Proof plan

Real calls between our own lines only, before anything touches a customer or a personal cell: Clara test line +1 720 807 1724, robot +1 720 594 2061, bench +1 970 715 6952. Show the destination's actual caller-ID display for each option, screenshotted or logged. Then a Willows rehearsal. Then Fede's yes, per customer, before it goes live anywhere real — same ladder as every other change on this page.

Current state (as of 9/22) Proposed — pending Fede's review

How a call to a Western Slope‑style client works today, end to end.

Known gaps (named in the 9/22 Fede–Sean discussion or open on this page already)

Long-term vision (from Fede + Sean's discussion, 9/22) Proposed — pending Fede's review

Source: Slack #transcripts, thread starting 1790089098.984989 (Sep 22, ~1–2 PM MT), live call notes with Fede, Sean and Gera. Quotes are verbatim from the transcript; anything not a direct quote is marked inference.

What they discussed

Greeting is a setting; routing is one core Proposed — pending Fede's review

Worth building now (low-hanging fruit) Proposed — pending Fede's review, nothing built

Each row is one piece of the vision above that's buildable with what exists today, roughly ranked by effort. Recommendation only — nothing here ships without Fede's go, per the standard rollout ladder.

A topic model that grows: what people really call about, and who handles it Proposed — for discussion, nothing built

What our own callers actually call about

Read-only pull from production, 2026-09-22. Western Slope = 40 voice calls in 30 days across 70 properties; Camellia/JP&Co = 658 voice calls over 5 months (May 1 – Sep 22). Willows (1,394 calls) is our test property and is excluded. Dead air is the single biggest bucket — 317 of Camellia's 658 calls and 9 of Western Slope's 40 had no caller speech at all. The table below counts only calls where someone actually said something: 341 at Camellia, 31 at Western Slope. Where the stored topic was the catch-all "none", 51 Camellia calls and all 13 Western Slope calls were read and labelled by hand.

What the numbers say. (1) Leasing is our biggest real bucket and it is growing at Camellia — 10% of volume in May, 22% in September. (2) Vendors are the surprise: a fifth of Camellia's real calls, and the current menu has no bucket for them. (3) Maintenance is only 6% at Camellia because their calls go to a call centre — at Western Slope, where Clara takes them, it is 19%. (4) Renewals are lumpy, not seasonal: they cluster on lease-expiry dates. (5) Half of everything that reaches us is silence; any topic model must have a clean "nothing to act on" exit.

What the outside data says

Where we differ from the industry picture. Every vendor taxonomy is built around maintenance and leasing. Ours is too — but vendors calling in (21% at Camellia) and verification desks (10% of Western Slope's real calls) appear in no published taxonomy we could find, and both are the exact calls that misroute today. Nobody publishes a real call-mix benchmark, so our own numbers are the only honest baseline we have. RealPage Knock and Zuma both hand off on the same rule — if the answer is not in the knowledge base, roll to a person rather than guess (Zuma, RealPage); heating emergencies spike ~300% in December–January and leasing peaks May–August at month-end (seasonality, leasing season). The one thing everyone agrees on and we do not do yet: identify who is calling before routing on what they said.

The routing table

Identity beats the stated topic. A tenant who says "leasing" is asking a renewal or lease question, never a prospect — they never get a guest card. A caller on a vendor's number who says "maintenance" is a vendor, not a work-order intake. Identity is resolved first, then topic, then the module switch.

There are exactly four outcomes. Clara handles it end to end. Transfer to a person — a role's ring list. Forward — the client's call centre, emergency line, or office line. Take a message — full context, emailed to the team.

The promotion ladder

Every topic sits at a level, per client. A topic moves up one level at a time, never two, and only on Fede's word.

The gate to promote a topic is all four of: the module is on for that organisation; a replay set of at least 25 real recorded calls on that topic scores 100% on the right outcome; zero "I'm not sure" or "I can't look that up" endings anywhere in the eval; and a one-switch drop back to level 1 that a session can throw without a deploy.

The "never embarrassing" guardrail

Risks, and what not to change now

Four questions for Fede

Routing V2 proof of concept — what we found (Sep 22) Built dark — nothing turned on at any client

Built tonight, switch off

Five pieces shipped tonight behind a company switch that is off everywhere, so nothing changed for any client. We also built a proof runner: it lists every number a company's routing could dial, then dials each one.

Number fence: the runner dials only numbers we own and hand it — the bench line +1 970-715-6952 and two of our test phones. Anything else is refused before dialing.

The policy, as the transcripts state it

Design flaws, biggest first

1. We don't know what the call is about when we pick the destination. The topic never steers the route. Fix: carry the topic forward into the transfer decision.
2. "Timely" is not defined anywhere. A non-emergency police call or a vendor at a door falls into "something else". Fix: Fede's call below.
3. An emergency rings one person and stops. Nobody else is tried, nobody is told. Fix: go down the list, then alert.
4. Clara can't ring the on-call list mid-call. Only the after-hours maintenance path reaches it. Fix: one transfer tool usable anywhere, with a test.
5. A company with no office hours on file has no row. It reads as closed all day and never rings anyone. Fix: add the row, pick the default.
6. Two of the three lookups can't tell "nobody set this up" from "the read timed out." A blip reads as an empty list, so Clara takes a message when she should ring. Fix: all three report the same three answers.
7. The older person-transfer path skips the middle rung. A group-level list is silently ignored. Fix: walk the same three rungs everywhere.
8. Nothing writes the emergency route. It can only be configured by hand. Fix: a setter, like the ring lists have.
9. The voice log prints staff names to the console. Names land in raw logs nobody scrubs. Fix: log an id.

Decisions for Fede

• a) Where the switch lives. Company level or per property? Recommendation: company level — every routing setting already reads company → group → property, so a property still wins where it is set. Needs a note against the constitution's per-property default.
• b) A middle "urgent, not an emergency" tier? Add one, or leave those calls in "something else"? Recommendation: leave it — "something else" already rings the property manager in hours. Log the cases for two weeks, then decide.
• c) A machine picking up counts as unsure. Already decided on the priorities page; we just follow it.

Where we can prove it

Fairhaven TEST is ready: one destination, the bench line, and the runner plans a real dial to it. The Willows cannot be proven yet — its person-transfer row still points at two personal cells, so the fence refuses every leg. Point that row at the bench line first; until then there is no Willows proof.

Facts that corrected the brief

• The maintenance and catch-all lists were already wired and read. The gap was meaning, not wiring.
• No emergency destination has ever been proven for any client.
• The middle "group" rung has no rows anywhere, for any customer.

Where the code is

Emergency line Proposed — pending review (2026-09-21)

Every call gets sorted into one of three tiers:

When it's ambiguous, Clara asks one question, once: "So I get you the right person — do you need someone right now, tonight?" Authority callers skip the question. If anyone is in danger, Clara says to call 911 first, before anything else.

One rule for every agent (Fede, 2026-09-21): the same ask gets the same outcome no matter which Clara agent takes the call — triage, maintenance, an unknown caller, leasing, or the after-hours desk. One shared tier rule, not one per agent.

Rollout: built dark. Proven overnight on the Fairhaven test line, bench number only — no calls to real people. Turning it on for a client, and who is on their list, is the client's and Fede's call, made after that proof.

Questions for the client, tomorrow

• Who's on the emergency on-call list, in order? Recommend: the same two people already on the leasing list, to start — can change later.
• What does Clara do at 3am if nobody answers? Recommend: ring the list once more, then take the full message, email/text it marked EMERGENCY, and give the caller the 911 advice.
• Which non-maintenance situations count as an emergency for them? Recommend: police/fire/court access, break-ins, welfare checks. Deliveries and vendors go to the office line, not the on-call list.

PR chain

Frozen contract and the Smith/Gera revision behind this section: session 001-59 scratchpad, 2026-09-21. No new registry keys — this reuses Gera's existing emergency.route.* keys with the people-list schema.

Recording and record-keeping fixes (2026-09-22)

Live for all customers, bug fixes, no gate:

• #10706 — Call recordings are never deleted.
• #10713 — Every recording is transcribed after the call.
• #10707 — A forward that never settles still writes the conversation.
• #10709 — The office line works from every phone agent.

What's proven

September 21: first day live at Western Slope, what broke and the fix

Western Slope's staff phones run on RingCentral. Their network answers a call within 1 to 2 seconds. Our "press 1 to accept" prompt waited 5 seconds, heard nothing, and dropped the staff member. Callers who asked for a person never reached Kat or Beverly.

Decision (Fede, Sept 21): plain forward is the default. Ring the first person 20 seconds. Whoever or whatever answers is connected, with no prompt and no voicemail detection. No answer or busy moves to the next person. If nobody answers, Clara takes the full message. If the client's phone system sends a call to voicemail, that is their system doing its job. Press 1 and automatic detection remain as per-company options, off by default.

After: live in production at 4:14 PM MT Sept 21 (#10567). Proven at 4:16 to 4:27 PM MT on our own test numbers only.

Also shipped the same day: missed-call emails carry the whole message with nothing cut (#10591); the conversation page shows who was rung and what happened (#10575, #10587); every staff leg is recorded from the first second (#10563); playback starts at the real pickup (#10585).

Still open: "something else" callers go to the office main line instead of staff cells (#10586, in review); a phone tree on the forwarded line gets labeled as a voicemail (issue #10626); confirmed on the real Western Slope line Sept 21 at 4:34 PM MT: Kat answered in about 15 seconds, no prompt, both sides recorded.

Today vs. new design

Details and history

Everything below is the research, run logs and vendor mechanics behind the summary above. Tables, ledgers and call/session IDs are kept in full; surrounding prose is cut.

The problem in one paragraph

Unless the team member opens Clara's follow-up email, the call is simply lost — the most-reported failure in AI-receptionist setups like ours, including two RingCentral community threads describing our exact configuration.

Deep dive for 2026-09-08: how the market does it, and the vendor tools we are not using

1. The four target behaviors are standard features across the market

The mature human-receptionist services (Smith.ai, Ruby, Abby, Nexa, PATLive) don't publish what caller ID staff see on transfer. PM-native AI vendors (AppFolio Realm-X, Knock, Entrata, Colleen, ResMan, Yardi) document almost none of the four; only EliseAI and Funnel are specific. Keypad-during-AI-conversation is a 2024–25 feature most AI-receptionist startups dropped.

2. What best-in-class sounds like when the human does not answer

1. Clara waits for live speech before speaking, so her whisper never plays into voicemail (Synthflow's pattern).
2. No live pickup in the window → the attempt is marked failed with a reason (timeout, voicemail, busy, error).
3. Clara returns to the caller: "No one was available. I can take a message, or help you directly?" (EliseAI's line.)
4. The message becomes a structured task with the caller's number, not a raw voicemail.
5. A repeat caller isn't made to re-explain (Funnel).
6. Final fallback: the office main line, never a dead end.
7. One continuous recording throughout.

3. Extensions: what modern desks do instead

• Extensions serve returning callers/vendors who want to skip the front door, and staff who put one on outbound material. Western Slope's tree offers extensions but publishes none — 100 and 999 both probed invalid, so only insiders can use them today.
• Vendors treat frequent "0 for operator" as a menu-friction signal, not a feature to preserve (Nextiva).
• What replaces the unpublished list everywhere: recognize the caller by number and route straight to their person (Smith.ai VIP list); say the name (Dialpad, RingCentral, Nextiva, GoTo); keep extension dialing in parallel for those who know one; accept keypad digits mid-call; keep "0" as a measured escape.
• How legacy phone systems stay untouched: the AI sits on the main number only (Ooma, Nextiva). RingCentral supports routing to an "AI extension" on an existing PBX; some platforms register the AI as a plain PBX extension (3CX, Yeastar, Telnyx SIP Attach) — the deepest coexistence pattern, needing a PBX that allows it. RingCentral doesn't offer that to small accounts.
• No vendor publishes an extension list to fix this; the direction is recognition plus say-the-name, keypad as fallback.

4. Vendor tools we are not using

Confirmed gaps on ElevenLabs: no documented caller ID on conference transfers, no ring timeout, no resume-after-failed-transfer, no transfer outcome in the post-call webhook, and an open bug where the transfer cuts Clara off mid-sentence.

5. Three ways to get there, sized

6. Proposed staging (for Fede to accept or change)

Sources for this section

• Retell transfer · Retell keypad · Retell voicemail
• Vapi transfer · Vapi keypad · Vapi forwarding
• Synthflow transfers · Synthflow transfer to humans
• Bland warm transfer
• EliseAI call center guide · EliseAI transfer fallback
• Funnel voice
• My AI Front Desk transfer · recording
• Smith.ai VIP list · Smith.ai transfer by SMS
• PATLive transfers
• Dialpad directories · Dialpad transfer
• RingCentral dial-by-name · RingCentral AI receptionist with PBX
• Nextiva extensions · Nextiva auto attendant · Nextiva direct dial · Nextiva on "0"
• GoTo AI receptionist · Ooma direct dial
• AI as PBX extension · case · Telnyx SIP Attach
• Amazon Connect recording caveat
• ElevenLabs keypad input changelog · voicemail detection · transfer · post-call webhooks · workflows · status history
• Twilio Dial · AMD · AMD FAQ · failover · screening · SHAKEN/STIR · Lookup · ConversationRelay
• Western Slope phone tree map (internal): ~/westernslope-data/phone-tree.md, probed 2026-09-05.

What we run today (Camellia, Yale 25 Station, Western Slope test line)

• Number hosting: the customer keeps their number and forwards to a Twilio number we own, imported into ElevenLabs, so ElevenLabs owns the call end to end. Our only hooks are the call-start and post-call webhooks.
• Hand-off: ElevenLabs "conference" transfer, one destination per property. Clara whispers the caller's name, then drops off.
• Caller ID on the team phone: not controllable, not documented by ElevenLabs. Our own Western Slope test-line bridge explicitly shows a PropFlow number.
• No-answer detection: none from ElevenLabs — reconstructed after the fact from the two-channel recording. That's why the missed-call email arrives late, and why eleven Camellia hand-offs died silently in August.
• After a miss: one email with the caller's number. No text (removed 2026-08-31 at Fede's request). Nothing to the caller.
• Recording: two-channel from answer, both sides, because the call never leaves Twilio during a conference transfer. We keep only the human leg.
• Failover: no fallback URL on any production number; an ElevenLabs outage means dead air (a 16-hour inbound-routing incident hit 2026-07-31).
• Recording disclosure: none on inbound agents. Colorado is one-party consent, so this is legal but flagged in our portfolio architecture notes.

What the vendors actually support (documented)

ElevenLabs transfer to number

• Three modes: conference (default, ours), blind, SIP REFER.
• Blind is the only mode documented to preserve the caller's number; it can't whisper.
• No parameter for caller ID, ring timeout, or no-answer behavior; the post-call webhook doesn't report transfer success.
• Known bug: conference transfer can cut Clara off before her "please hold" line finishes. Field fix: fire the transfer from our own backend with a 2–5s delay.

Twilio (we own the number and account)

• A bridged call can present the real caller's number — standard forwarding, legal under the Truth in Caller ID Act.
• Trade-off: forwarded calls get the lowest STIR/SHAKEN attestation (C, diverted), so carriers may label it "Spam Likely." Mitigation: Voice Integrity registration, low complaint rates, per-property choice of caller ID.
• Full control on the human leg: ring timeout, simultaneous ring, a whisper before bridging, answering-machine detection, and an instant callback on busy/no-answer/failed.
• Two-channel recording of the whole call including the human leg, plus a fallback URL if the primary handler fails.

RingCentral

• Forwarding the main number to an outside number passes the original caller's number through — no setting changes it. This hop works today.
• Forwarded minutes draw from the plan's outbound minutes, no separate fee.
• RingCentral's own AI receptionist has the same issue: transfers show the AI's number, not the caller's.

How peers do it

• Forwarding the existing number is the market-standard entry path; porting is a later upgrade (Smith.ai, My AI Front Desk).
• Warm transfer with a spoken summary is the norm; blind transfer is the fallback.
• Showing the original caller to staff is an explicit option on Retell and Synthflow, and only works on a Twilio-purchased AI number — ours are.
• Missed transfers everywhere become voicemail plus a staff notification. Synthflow tracks six outcomes and can retry or fall back; nobody automatically re-rings.
• PM-native agents (AppFolio Realm-X, Knock, Entrata) sidestep caller ID entirely by opening the guest card for the receiving agent instead.

Most-reported pitfalls in the field

Options

Option C was called "own the dial" in the June design note, deferred then as too much rewiring.

Recommendation

Answers to Fede's direct questions

Shipped 2026-09-19 — before → after

Proof

• #9921 ring list · #9972 ring line webhook · #9933 forward to client's line · #9956 vendors to a person · #10001 ring-line ticket fix
• Ring line webhook set 2026-09-19 21:22 UTC · Willows list written same time
• Live test calls, Fede's cell to the Willows line: 2026-09-19 23:43 UTC (24s) and 2026-09-20 00:02 UTC (20s). The second reached the ring line (Twilio CA16cded30f4, 5s) but nobody's phone rang: the ticket was keyed on the caller's number while the leg arriving at the ring line carries the property line's number, so the lookup missed.
• Fix #10001 merged 2026-09-20 01:24 UTC, live in production, two new regression tests.
• Retest pending (Fede): call the Willows line, ask for a person, let it ring out, confirm the message and email.

How the stack actually works (vendor docs, 2026-09-19)

Clara as the phone system (target architecture, 2026-09-19)

Listen: 4-minute episode

Harness: a real-call ring-chain harness (#10030, #10033 and two siblings) is the acceptance test for everything below; nothing turns on at Western Slope until two clean sweeps.

How the whole call gets recorded

Every dial or transfer leg carries <Dial record="record-from-answer-dual"> — the caller-facing leg, the ring-line leg, and the leg to whichever cell finally picks up. Each leg produces its own recording, listed separately on the conversation record — Twilio does the recording, not code we wrote.

Root cause on the transfer leg: the hand-off opens a new outbound call from the property line to a PropFlow bridge number, with no ParentCallSid link back to the inbound call, and the bridge TwiML never had a record attribute. ElevenLabs never sees this leg either. Example: 2026-09-18 23:58 UTC, 351s with Clara then 323s with the call center — both real human legs, but the Clara-side recording was misclassified and deleted, and the transfer leg was never recorded; 2026-09-18 19:36 UTC, same double gap, caller hung up after 37s. Both examples predate #9565/#9716. The Sentry "not eligible for recording" issue is harness noise from a fake CallSid, not a real call. Both fixes ship dark, default off, Fede turns on per customer.

Live proof, 2026-09-20 (Willows)

Real calls from PropFlow tester numbers into the Willows ring line (+1 970-615-5171) with a seeded handoff ticket, staff list Fede then Gera — Twilio's own call records are the evidence.

• Automatic pickup — Twilio AMD, mode Enable, no keypress — becomes the pickup method; press 1 stays the fallback.
• Both ship as one platform-wide setting.
• Whisper is "who, about what" only (example: "Maria Lopez, tenant, about a leak in 204.").
• Detection runs on the direct dial, so the staff phone still shows the original caller's number. Superseded 10/8: caller ID is a choice per destination.

Open trade-off — "Silence Unknown Callers" drops the ring unless the number is a saved contact. Showing Clara's line works once staff save it once; showing the real caller (the platform passthrough setting) gets silenced. Fede to decide: (A) real caller's number, onboarding tells staff to allow unknowns; (B) Clara's line, staff save one contact.

Automatic pickup, live run 2 (2026-09-20 afternoon)

Second round over a tunnel into the Willows hunt chain, testing automatic pickup end to end: a plain human pickup, caller-number passthrough, a bench "voicemail," and a decline. Fede answered B and C normally and declined E; test D (the product path through fastlane) wasn't run — Willows was still in message mode. All times MT.

Headline: automatic human pickup PASS (test A3, 3.6s). Passthrough PASS at the carrier level (test C). Decline FAIL — reaching carrier voicemail scores unknown, and unknown is today treated as accept, so the chain stops instead of advancing (test E). Two mechanical defects: the AMD status-callback 403s in prod (test A2), and the accept verdict can land after <Dial>'s action already moved on (test C's late-verdict race). Separately, the writer script arming AMD mode wrote to a local settings file instead of prod DynamoDB on its first run (test A) — the prod-table assert passed while the actual write missed the table.

Design change needed, at the time: only a confirmed human verdict should auto-accept; unknown should fall back to whisper-plus-press-any-key, not silent accept; machine should advance the chain, not stop it. Update, 2026-09-22 — shipped in PR #10835: automatic pickup detection on staff ring-list calls is now LIVE. Default is Twilio AMD (human_handoff.acceptMode absent = amd) — no keypress needed on a confirmed human. Press-any-key is the fallback, used only when AMD returns unknown. Still open: machine_start currently hangs up the call; Fede is deciding whether to treat it as unsure instead. Per-client routing (which mode a given customer runs) is tracked on the Clara onboarding playbook, the source of truth for that.

Write ledger (all restored, verified): ring line webhook pointed at the tunnel 10:44, restored to prod 12:44. Bench line webhook pointed at the tunnel 12:29, restored 12:44. huntAcceptMode flipped press1→amd in prod Dynamo 12:23, restored to press1 12:45; callerIdPassthrough untouched (false) — the signature and unknown-handling defects are still open. No Camellia or customer data touched; dev server and tunnel stopped, worktree clean.

Run 3 and the stand-in rehearsal (2026-09-20 night)

Fede tested from his own phone first. What happened, in order:

• Call A (answer and say hello): the answering-machine detector scored his voice as unknown, the "press any key" window was 4 s and he had been told to press nothing, so his leg dropped after 14 s and the chain moved on. Gera's phone rang 22 s later (gap filed as #10337).
• Call B (decline): Fede declined, Gera rang 0 s later, nobody answered, the caller was offered a message. Pass.
• Call A3: the no-pickup email went out twice for one call. Fixed by a once-per-call lock (#10350).
• Ring-time text: arrived, but the reply and call-back path did not hold up. Fede: turn it off, focus on the core. Default off (#10336).

Fede's ruling: prove it on lines we own before calling him again. The rest of the night ran on the stand-in line (+1 970 715 6952, a test number that answers and presses 1, plays a voicemail greeting, or reports busy on request). Blocker found and fixed on the way: the site's login wall was refusing Twilio's calls to that test line (#10346).

Every test setting (stand-in lists at Willows and the test company, Sunday hours, stand-in line behaviours) was restored afterward; Camellia's routing checked identical before and after. Harness grader gaps filed as #10362.

Maintenance forward and its recording (03:24Z)

Setup: the Willows company's maintenance route pointed at the stand-in line for the test (written 03:24:06Z, ended 03:35:28Z, back to unset). Recording of transferred legs was already on for Willows.

• A robot tenant called the Willows line and asked to be connected to maintenance about a leaking sink.
• Clara: "I'll connect you with the maintenance line now — one moment." Then she transferred to the stand-in number with the reason "Tenant in unit 102 reporting a leaking kitchen sink".
• Twilio: the transfer leg from Clara's line to the stand-in connected and lasted 17 s. The caller's leg lasted 39 s in total.
• Recording: one recording of 39 s on the caller's leg, started when the call was answered, so it holds Clara's part and the maintenance part. The conversation record keeps it (transfer stamped at 03:25:01Z), so it is not deleted by the no-transfer cleanup.
• Known shape: a direct forward with no phone-menu step creates no second recording on the forwarded leg. The audio is still captured on the caller's leg. PASS.

Western Slope switched on (2026-09-20, 21:35 MT)

Fede's words: "once you're highly confident this works, then you can go ahead and do Kat first, Bev second." Written at the company level, so every Western Slope property gets it:

• Ring list: Kat Barker first, Beverly Villar second, 20 s each, numbers from their staff profiles. Was: unset (Clara took a message for everyone).
• Office hours: Monday to Friday 9:00 to 17:00 Mountain, Saturday and Sunday closed. Was: none on file, which Clara treats as closed, so the ring list would never have fired.
• Unchanged: maintenance forwards to the client's own maintenance line; ring-time text stays off; press-1 accept.
• Proof, read-only: the routing resolver for two Western Slope properties answers transfer, Kat then Bev, 20 s; Monday 10:00 MT reads open, Tuesday 18:30 MT reads after hours; maintenance route unchanged. No test call was placed on the Western Slope line, so as not to ring real staff or leave a note in their records.

What others do, and what Clara should do (2026-09-20)

Decided (Fede, 2026-09-20)

Fede's framing: vendor behavior is a reference, not a template — Clara is AI-first, not a phone tree.

Next build: text-on-ring (who, why, caller's number, tap-to-call), dark at Willows, plus a contact-card onboarding step. Filed for later: staff call Clara back for a summary of a handed-off call.

AI receptionists and leasing assistants converge on a warm transfer with a briefing, no keypress:

• iPhone Silence Unknown Callers lets through Contacts, Recents and Siri Suggestions; everything else goes to voicemail.
• Pixel Call Screen and call forwarding can't coexist.
• iOS 26 Call Screening answers unknown calls with a reason prompt (PIRG) — with the real caller's number shown, automatic pickup would hear that prompt, score it human, and bridge the caller into the screening bot.
• 75% of US consumers never answer unknown calls (TNS 2022); ~40–50% of business users enable unknown-caller silencing (GetVoIP).
• Our own proof: Fede's silenced phone dropped four legs in 2–3s until he saved the ring line as a contact.

Recommendation (AI-first):

1. The staff phone shows Clara's line; staff save it once at onboarding, and Clara texts the contact card herself.
2. When the phone rings, Clara also texts who, why, the caller's number, and one tap to call back.
3. Automatic pickup runs on the direct dial; press 1 is the fallback.
4. Later: staff reply to the text ("call back in 10") and Clara relays it.

Consequence: the platform passthrough setting stays built and OFF, not the default. Option A (real number as default) requires staff to disable unknown-caller silencing and breaks under iOS 26 screening.

One routing engine for every call — Proposed, pending Fede (2026-09-20)

Proposed — pending Fede

Problem: three separate settings answer one question today — message-vs-transfer mode, the ring list, and the company office-line catch-all. Clients differ only in headcount: Camellia one (Joanna), Western Slope two, others three or four.

Proposal: one ring list per role (leasing, maintenance) at the company level, property-level only when it truly differs. Business hours decide, not a mode: during hours a caller who wants a person is always rung; nights/weekends Clara takes a message by default. Decided (Fede, 2026-09-20): a maintenance emergency after hours always rings, weekends included. Leasing is never an emergency — after hours it takes a message. Client inputs are the two lists plus office hours, a setting of its own, separate from tour availability.

Removes: the message-vs-transfer switch, the separate catch-all setting, per-property duplicates. Keeps: the ring chain as built, recording, automatic pickup.

Fede's module layer (2026-09-20): leasing, maintenance, and renewals sell separately. A client can have leasing on and maintenance off. The ring list above only runs once a call has already been routed to the right function — the missing layer decides, per client, whether that function is Clara's to handle at all, or should forward straight to the client's own existing line for it.

Today these five pieces are separate and only two of them (maintenance, human handoff) actually decide anything on a live call. The module switch exists at the company level but nothing in the voice path reads it yet.

The engine: one decision per call, run in order — topic, then is that module on for this client, then is the office open, then the result.

Smith's registry review 2026-09-20: resolver reuses the existing forward-to shape (maintenance.intake_route generalized per function), emergency resolves via emergency.route.*, renewals stays a module; two questions with Gera (renewals row vs topic; ring_list as successor of transfer.<role>).

What staff receive on a handoff

Two people can be pinged about one missed call (the rung person by text, the owner by email); the text stays off until Fede turns it on.

What changes in the build chain: one routing resolver that reads the module switch, a per-module forward-to number (new company-level setting, one per module), office hours, and the ring lists. The maintenance handle-or-forward setting and the office-line catch-all fold into it — they stop being separate settings and become the module-off destination for their own module. Camellia's typed number is one row in that resolver, not a special case.

Module off = never Clara; a client with a few buildings on Clara models as module on + company forward-to + those buildings set to Clara.
If module status can't be read during a call, Clara does today's behavior, never forwards to nothing.

Current state

The routing model

• Roles, not a single list: leasing, maintenance, emergency, everything else, each with its own ordered ring list of plain cell numbers.
• On-call rotation, per role: two people per role today, static (Western Slope: leasing agent, then an admin backup). Weekly rotation optional later; whoever's on call rings first, the rest follow.
• Hot lead flow: a caller asks for a person → the role's ring chain runs → nobody accepts → Clara records a message → the team gets an email with a short callback window for a live prospect.
• Everything else — vendors, after hours, anything not leasing/maintenance/emergency — forwards to the customer's own main office line, a company-level setting.

Phases Sep 19 plan, superseded

Kept for history. Phase 1 shipped (our own dial on the ring chain). Phase 2 (on-call rotation) is now an onboarding answer on the ring list, not a build. Phase 3 (Twilio TaskRouter) is dropped: call-center queues are legacy VoIP, and the target is one dial path on Gera's lookup, described under Where the phone line is going.

1. Phase 1 (now): replace the hop-by-hop hunt webhook with Twilio's own <Dial sequential="true"> across the ordered ring list, each <Number> carrying AMD, record="record-from-answer-dual", and a 20s timeout. The "press 1" whisper is now the fallback, not the default — Twilio's own docs confirm sequential is a native ring-chain, not something we're building.
2. Phase 2: on-call rotation settings, per role — drag order, an "on call this week" pill, optional weekly auto-rotate, rendered into the ring list at call time.
3. Phase 3: Twilio TaskRouter, once a team outgrows a short list. Workers with role attributes, queues per role, per-target timeouts and escalation. ~$0.06/task after 100 free/month (unconfirmed on Twilio's pricing page). No native scheduling — our own setting still decides who's on call.

Later — with the renewals lane (Fede, 2026-09-19)

Not a phase of this rollout — picked up later. Facts gathered now so the option is scoped when that lane resumes.

RingCentral / call center comparison

Decided (Fede, 2026-09-20, supersedes 2026-09-19): handoff calls always come from Clara's line.

Decisions (Fede)

• D1. Where the role model lives. Proposed (a) Recommended: extend "talk to a person" into roles, same storage, same page. (b) A new "team routing" page. (c) Wait for a customer to ask.
• D2. On-call rotation shape. Proposed (a) Recommended: weekly primary per role, set on the page, no calendar sync. (b) Read from the customer's calendar. (c) Both, calendar wins.
• D3. Hot leads. Proposed (a) Recommended: the missed-call email says "prospect asked for a person," sent at once. (b) Also text the leasing role's primary. (c) Email only, no marking.
• D4. Porting the customer's number. Proposed (a) Recommended: not before a month of clean forwarded-line operation. (b) Offer at onboarding.

Reply with a letter per decision, or say nothing and the recommendations ship, dark, on the ladder above.

Context block for agents

Related pages

• Voice agent architecture decision — the agent-stack decision (which voice model, which orchestration); this page is about phone hosting and the hand-off path instead.
• RCA: Camellia voicemails, 2026-08-31
• Missed-call email design, 2026-08-31
• Client intake — Western Slope (draft)
• Onboarding plan — Situs / Western Slope

Sources

• ElevenLabs transfer docs
• ElevenLabs post-call webhooks
• ElevenLabs cut-off bug and fix
• ElevenLabs SIP incident, 2026-07-31
• Twilio Dial verb (callerId, action, machineDetection, record)
• Twilio STIR/SHAKEN attestation
• Twilio spam labeling remediation
• Twilio failover best practices
• Twilio call-forwarding charges
• RingCentral thread: caller ID replaced after AI transfer
• RingCentral thread: warm transfer failing from AI platform
• RingCentral AI receptionist transfer behavior
• RingCentral call forwarding
• Retell transfer options
• Retell caller ID 403 thread
• Synthflow call transfers
• EliseAI voice configuration
• Recording consent by state
• RingCentral + Twilio + ElevenLabs caller ID write-up
• Retell audio after cold transfer

Internal: docs/planning/unknown-caller-transfer-no-answer-detection.md (June 2026 design, options 2a–2d); agents/clara/lib/voice/post-transfer.ts (recording at answer, outcome classification); config/phone-registry.json; agents/clara/lib/voice-agents/prototypes/western-slope-portfolio-triage.ts (bridge callerId).

Old subtitle and change stamps

How we host a property's phone number and hand a caller off to a person — what's proven live, what's decided, and the research behind it.

Proposed 2026-09-07 · deep dive 2026-09-08 · live-tested and restructured 2026-09-20 · current state + long-term vision added 2026-09-22 · topic model proposed 2026-09-22 · inspection and two-mode design 2026-10-07 · call routing end-to-end inspection and holistic design 2026-10-07 evening