# My Workspace → 2 Rounds AI Call

**Portal:** Manage · **Route:** `manage.my-work.ai-rounds.*` (`/manage/my-work/ai-rounds`) · **Nav:** My Assistant → My Workspace →
My Task → third tab **2 Rounds AI Call** · **Parent:** [My Workspace](/docs/modules_handbook/manage/my-work/readMe.md)

## What it does

The boss's flow for signal-grade **B** leads ("some signal, not yet clear"; see
[signal-grade.md](/docs/modules_handbook/manage/leads/signal-grade.md)), owner decisions 2026-10-15 — and, since
2026-10-16, for **new leads who just registered for a class**, whose round 1 is the funnel's registration call (see
*Funnel registrants* below). Renamed from "Grade B - AI 2 Rounds" to **2 Rounds AI Call** on 2026-10-22 for that
reason; the route, code names and `ai-rounds` keys are unchanged.

1. **AI call 1** asks the lead what is holding them back on property (the Stage 1 Needs questions), and the AI
   scores their interest **1–5 with a one-line reason**.
2. **The AI writes two things** from call 1 (owner, 2026-10-21): a **brief** for the account manager (the situation,
   what to pitch, what to verify, what call 2 should ask) and the short **script** call 2 will speak. Both come from
   **Claude Sonnet**, the same model for the old rows and every new lead. The account manager opens the row's
   **slide-out**, reads the brief, edits the script (the AI's original stays underneath), ticks the **projects** that
   fit from the **property guideline** (still has 90% quota? → loan-eligibility band → project), and may have the AI
   rewrite the script around them. The catalogue search of 2026-10-20 is **off** until the catalogue is ready
   (`ai_rounds.catalogue_search`).
3. **Approve**, then **AI call 2** explains that script and asks **when the account manager may call them**. **Every
   call 2 needs an approved answer** — single or bulk (owner, 2026-10-22).
4. The agreed time becomes a **callback task** (Action Item) and a **calendar entry** for the account manager, and the
   lead moves to **Done**.

**Every call is a button a person presses**, behind a confirm box — one lead's button, or **Bulk AI call** for many
(see below). Nothing rings on a schedule of its own. Two exceptions: the lead's own tap — a round's first
unanswered call sends a WhatsApp with a **"Call me back now"** button (once per round), and a tap rings them straight
back on the same round, instantly and without a limit — and a **funnel registration** whose AI-call rule picks the
round-1 profile, which rings round 1 itself (see *Funnel registrants* below).

**Who sees whom:** the list is the viewer's leads as **account manager** (`leads.assigned_admin_id`) that are grade B
now, or whose rounds have already started (a lead keeps its row when its grade moves, so no work is lost — and a
funnel registrant joins the list the moment the funnel rings their round 1). A lead who said no or was Excluded moves
to the **Stopped** view instead, whatever their stage ([what that stops](../leads/signal-grade.md)). A leader with view-leads-all (or a super
admin) can open a colleague's list with the person picker, as on Sales steps.

**Language:** a Hong Kong number (+852) is called in **Cantonese** and gets the Cantonese WhatsApp (traditional
characters); everyone else hears **Malaysian Mandarin with English and local slang**. The agent always says it is an
AI assistant, as Retell's terms and the house rule require ("Never deny being an AI").

**The plan this replaced:** [b-grade-ai-nurture-plan.md](/docs/modules_handbook/manage/leads/b-grade-ai-nurture-plan.md)
(an automatic flow inside Sales steps). The owner chose this manual table instead on 2026-10-15.

## How it works

**Its own page.** `Manage\MyWork\AiRoundsController` renders `Pages/Manage/MyWork/AiRounds.vue`; it builds none of
Sales steps' or Action Item's data and shares no code with them. On My Work's page the third tab is a link
(`Index.vue` `workspaceTabs` + one line in the tab watcher); on this page the first two tabs link back to
`/manage/my-work?tab=workflow|tasks`.

**One row per lead — `lead_ai_rounds` ([`LeadAiRound`](/src/Lead/LeadAiRound.php)).** Opened by the first call. Keeps
the stage (`STAGES`: round 1 called → answered → round 2 called → answered → completed), the language, what each
answered call learnt (`r1_data` + `r1_score` + `r1_score_reason` + `r1_summary`; `r2_data` + `r2_summary`), the answer
(`answer`, the AI's `answer_draft`, who edited it and when), each round's one WhatsApp stamp (`r1/r2_missed_sent_at`),
and the booking (`callback_at`, `callback_note`, `callback_task_id`, `callback_activity_id`, `completed_at`). Writes:
[`LeadAiRoundRepository`](/src/Lead/Repositories/LeadAiRoundRepository.php).

**The calls stay in `ai_voice_calls`**, tagged `meta.ai_round_id` + `meta.ai_round` (+ `ai_round_trigger` staff /
call_me_back, `ai_round_by`). The page reads every attempt, its time and status from there; the call cell lists them
and opens the summary, recording and transcript (`GET …/ai-rounds/calls/{uuid}`).

**Placing a call — [`PlaceAiRoundCall`](/app/Actions/PlaceAiRoundCall.php).** Refuses with a reason when: no phone ·
the number is outside `AI_ROUNDS_ALLOWED_NUMBERS` (the dev box's safety net) · round 2 has no answer · a call to the
lead is still ringing · the round's profile is missing or unsynced. Then it calls `PlaceAiVoiceCall` — every existing
guard still applies (do-not-call list, contact suppression, the daily budget). Dynamic variables: `lead_name`,
`account_manager`, `known_context` (the grade's signal), `today` (so "tomorrow 3pm" becomes a date the agent confirms
aloud), `is_repeat_call` + `previous_summary`; round 2 adds `approved_answer` (≤ `answer_max_chars`),
`round1_summary`, `round1_problem`. Round 2 may repeat a figure that is in the approved answer (as an example) and
nothing beyond it.

**After the call — [`AiRoundsAftermath`](/src/Lead/AiRounds/AiRoundsAftermath.php)**, reached from
`AiCallAftermath` for any call with `isAiRoundCall()` instead of the funnel's thanks / missed WhatsApps and the funnel
journey hooks:
- not reached (no answer, line busy, voicemail, declined, or picked up and silent) →
  `SendAiRoundMissedMessage` once per round. **A call that never rang gets nothing** —
  `AiVoiceCall::NEVER_RANG_REASONS` (a country not enabled on Twilio, an unreachable / invalid number, a routing
  failure, flagged as spam). On 2026-10-15, 27 Singapore / Hong Kong leads got "we just called you" for calls that never
  left Twilio, until Voice → Settings → Geo permissions had SG and HK switched on. (The funnel's own missed follow-up
  still sends for these on purpose — see the voice-agent handbook.)
- round 1 picked up but only **"busy / call me later"** → also *not reached*: the lead stays in "Round 1 · no answer yet"
  (no score, no answer draft, the round's one WhatsApp), so the next bulk round 1 tries again. The AI tells us through
  the round-1 extraction field **`call_outcome`** (`talked` · `call_later` · `not_interested` · `do_not_call` ·
  `wrong_person`, `LeadAiRound::CALL_OUTCOMES`); for an agent synced before that field, a call where not one Needs
  question got an answer counts the same (`LeadAiRound::roundOneNotReached()`). **Since 2026-10-20 a `talked` call
  where not one Needs question got a real answer counts the same too** (owner: no answer is drafted from nothing — the
  lead goes back to round 1); only `not_interested` / `do_not_call` / `wrong_person` stand without one. Rows recorded
  before the fix (busy, 1/5, nothing learnt) are put back with `php artisan ai-rounds:reopen-busy` (lists them;
  `--apply` reopens — only rows nobody edited or approved, with no round 2);
- **`do_not_call`** → round 1 is kept, and both rounds are locked ("Asked not to be called again": grey buttons, bulk
  skips it, `PlaceAiRoundCall` refuses);
- round 1 answered → keep the fields + score (+ `call_outcome`); queue `DraftAiRoundAnswer` unless a person has already
  edited the answer;
- round 2 answered → keep the fields; a usable `callback_time` (ISO, Malaysia time, not in the past) creates the
  CALL task (high priority, due that day, assigned to the account manager) + a 30-minute calendar Activity, then
  completes the row — in one transaction.

**The answer — [`DraftAiRoundAnswer`](/app/Jobs/AiRounds/DraftAiRoundAnswer.php)** (an `AiJob` on the `ai` lane)
through `AiClient` under the prompt key **`ai_rounds_answer`** ([resources/prompts/ai_rounds_answer.md](/resources/prompts/ai_rounds_answer.md)).
**The knowledge base is that prompt's `## Knowledge` section** — written 2026-10-15 from Wai Kit's webinar of
2026-10-14: the five-question buying formula (net price below market value · whole-unit rental covers the installment,
ROI ≥ 7% and "rental × 10 ≥ a year's installment" · several rental strategies · own-stay buyers would want it ·
developer track record · the "big pie" last), earning three times, holding power, his own win and loss stories, the
90% loan quota, new project / subsale / auction / bulk purchase, Airbnb, room rental, renovation, landed, foreign
buyers, and what the team does. The draft may only use what it says, quotes figures only as Wai Kit's rule of thumb or
someone's example, and never mentions the membership, its price, refunds, grants or listing plans. **Edit it on
Manage → AI Prompts** (an edit there wins over the file). Gemini 3 thinks before it writes and the thinking shares the
output budget, so the job sends `thinking: low` + `max_tokens: 4000` — at 900 the draft stopped after ~40 characters.
A later draft never replaces a person's edit unless they pressed Rewrite.

**Project knowledge + "Suggest" (owner, 2026-10-23: "let admin upload details of the projects so the AI knows everything
about our recommended projects … AI recommended projects, admin chooses").** Until then the only thing the AI knew about
a guideline line's project was its one-line card. Each sales project now has an **AI knowledge** tab
(`Pages/Manage/SalesProjects/Show.vue` → [`Partials/KnowledgeTab.vue`](/resources/js/Pages/Manage/SalesProjects/Partials/KnowledgeTab.vue);
[`ProjectKnowledgeController`](/app/Http/Controllers/Manage/Engagements/ProjectKnowledgeController.php), `manage-projects`):
typed text (title + body), an uploaded `.txt` / `.md` / PDF whose text is pulled out
([`ProjectKnowledgeExtractor`](/src/Property/Services/ProjectKnowledgeExtractor.php), `pdftotext` through the brochure
extractor), or a picture kept for a later vision step (not read yet). Rows in `project_knowledge_items`
([`ProjectKnowledgeItem`](/src/Property/ProjectKnowledgeItem.php), `KIND_TEXT / FILE / PICTURE`, `media_id` → a `media`
row in collection `knowledge` owned by the project; [`ProjectKnowledgeItemRepository`](/src/Property/Repositories/ProjectKnowledgeItemRepository.php)
deletes the file with the item). **`Project::aiKnowledge($maxChars)`** = the row's own facts line (name, developer,
location, price range, avg psf, rental yield, highlights) + every readable item, oldest first, capped — and it now
reaches the AI in three places: the **combine** brief (each ticked card + "What we know of this project", 1,200 chars;
`ai_rounds_combine.md` may describe with it, never price from it), **call 2** as `{{projects_context}}` (the ticked
projects' knowledge, 700 chars each, only while *Mention in the answer* is on; the round-2 prompts gained a section —
answer questions about a named project from it, say the account manager will explain anything else), and the new
**Suggest** button on the slide-out's Projects picker: `POST …/{id}/property-options/suggest` stamps
`property_options.suggesting_at` and queues [`SuggestAiRoundProjects`](/app/Jobs/AiRounds/SuggestAiRoundProjects.php)
(prompt `ai_rounds_suggest_projects`, Sonnet), which reads call 1's summary, the brief, the financing on record, the
branch / band picked so far and every active guideline line with its project's knowledge (900 chars), and writes up to
three `{option_id, reason}` to `property_options.suggested` (+ `suggested_at`). **It proposes only** (the boss's rule
of 2026-10-21 — a person picks the projects): the drawer shows them in a violet "AI suggests" box with the reason, the
manager ticks what suits the lead, and *Mention in the answer* stays the switch for whether call 2 names them. The
Guideline modal shows each project's "AI knowledge · N" link (amber when nothing is written) so a line with an empty
project is visible. Migration [2026_10_23_160000_create_project_knowledge_items_table.php](/database/migrations/2026_10_23_160000_create_project_knowledge_items_table.php).
Tests: [`ProjectKnowledgeTest`](/tests/Feature/Engagement/ProjectKnowledgeTest.php) and the suggest test in `AiRoundsTest`.

**Move the lead to another sales admin (owner, 2026-11-04).** The slide-out's title carries a **Move** button beside
`AM · name` (prop `canAssign` = `manage-leads` and not a group user, as on the Lead page; `assignableAdmins` = the
same staff list). It opens the shared `Components/AssignLeadManagerModal.vue` in its `modal` host mode, posting to
`POST leads/{id}/assign` — which now moves the WHOLE lead (seat + the previous manager's open action items + roles, via
`LeadHandover::move()`). The row then belongs to the new manager's list, so the drawer closes when the page comes back.

**Test call 2 in the browser (owner, 2026-10-26: "sales admin need to test call 2 — use the Test Audio function, no
call to them").** The Answer and Call 2 steps offer **Test call 2**: `POST …/{id}/call/2/web-test` (throttled)
resolves the round-2 profile for the lead's language (`PlaceAiRoundCall::profileFor()`), builds exactly the variables a
real call 2 would carry (`PlaceAiRoundCall::variablesFor()` — the script as it stands, approved or not, the questions,
`{{projects_context}}`, the loan context) and starts a Retell web call through the new shared
[`Src\VoiceAgent\Services\WebTestCallStarter`](/src/VoiceAgent/Services/WebTestCallStarter.php) — the same starter
the AI Profiles Test lab's *Test Audio* now uses. The run is a `SOURCE_WEB_TEST` row on the round-2 profile (so it shows
in that profile's Kept runs with recording + transcript) with `meta.lead_uuid` / `ai_round_uuid` for reference, and
never carries `META_ROUND_ID`, so it is not one of the lead's calls and moves nothing on the round. The browser side is
the shared [`Components/VoiceAgent/WebTestCallModal.vue`](/resources/js/Components/VoiceAgent/WebTestCallModal.vue)
(mic check, WebRTC session, live transcript; `startUrl` + words as props), which the Test lab tab also mounts now.
Retell bills the minutes; no Twilio leg; the lead hears nothing.

**AI review of every script, scored, never a block (owner, 2026-11-03; changed 2026-11-10).** Wai Kit sent 11 of the
first 14 answers back for six reasons, and the resubmissions repeated three of them (digits/RM in the script, "（同意后）"
stage directions, empty questions). So every script is read by the AI against **the approver's rules**. The rules are
free text on the page's ⚙ Settings modal (*AI review*), `Setting::AI_ROUNDS_APPROVER_RULES`, edited by the final
approver or a leader (`PUT …/approver-rules`, [`UpdateApproverRulesRequest`](/app/Http/Requests/Manage/MyWork/AiRounds/UpdateApproverRulesRequest.php));
empty = the shipped default [`resources/prompts/ai_rounds_check_rules_default.md`](/resources/prompts/ai_rounds_check_rules_default.md)
(his six reasons). The review is [`CheckAiRoundAnswer`](/app/Jobs/AiRounds/CheckAiRoundAnswer.php) on the AI lane
(prompt `ai_rounds_check`, `resources/prompts/ai_rounds_check.md`): it reads the script and the questions with the lead's
phone country and nationality, call 1, the financing on record, the brief, the WHOLE property guideline (each line with
its branch / band and the project's AI knowledge or "EMPTY"), the ticked lines and the rules, and answers seven fixed
checks — `who_buys`, `facts`, `framework_first`, `projects`, `read_aloud`, `closing_ask`, `questions` — each
pass / warn / fail with a one-line note and fix. The result lands in `lead_ai_rounds.answer_check` (json: verdict,
checks, **`text_hash`** of the script + questions it judged) with `answer_check_requested_at` / `answer_checked_at` /
`answer_check_error` (migration 2026_11_03_100000); `LeadAiRound::answerCheck()` returns it only while the hash still
matches, so an edit makes it stale rather than wrong.

**The score.** `LeadAiRound::checkScore()` turns the seven lines into a score out of 100: a pass is full marks, a ⚠️
half, a ❌ none (5 ✅ + 1 ⚠️ + 1 ❌ = 79). It is computed from the stored lines, so every review ever written has one, with
no migration; null when the review could not run. The card's title reads *AI review · 79/100 · 1 to improve*; the list's
Answer cell carries an **AI 79** badge (green ≥ 80, amber ≥ 50, red below) for a review of the CURRENT text only.
Everyone who opens the page sees both (owner, 2026-11-10: "all can view" — Shawn, Waikit, Zen and Kexin are all
super-admin on production, so no role could have drawn a narrower line).

**Why it no longer blocks (owner, 2026-11-10).** From 2026-11-03 to 11-10 a ❌ held Submit until the text changed, and a
pass submitted by itself. The team found it too strict: on production 65 of 69 reviews said *fix*, only 4 passed, and
53 of the 67 unsubmitted answers sat behind a ❌. Now **Submit** (`PUT …/answer`) saves the text and hands the row to
[`AiRoundSubmission`](/src/Lead/AiRounds/AiRoundSubmission.php), which submits it at once
(`LeadAiRoundRepository::submitted()`, the approvers emailed) whatever the review says. When the text has no review of
its own and none is running (the manager edited the AI's draft), `checkRequested()` + the job are queued, so the
approver reads the score of the version that was actually submitted (the page polls every 10 s while
`answer.check.checking`). The job only records the verdict; it no longer submits anything. The review also runs on its
own after every AI draft and every project combine, and on demand (*Run AI review* / *Review again*,
`POST …/answer/check`, which saves the text only when it CHANGED — an unchanged text is reviewed as it stands, so an
answer waiting for the approver stays submitted; a changed one is an edit and withdraws the submission as always).
**Answers that were submitted before the review existed** have none until one runs: `php artisan
ai-rounds:check-answers` lists every open script without a current review (`--submitted` = only those waiting for the
approver), `--apply` queues them — nothing is submitted or withdrawn by it
([`CheckAiRoundAnswers`](/app/Console/Commands/CheckAiRoundAnswers.php)). The repository's `update()` clears the
manager's approval when no `answer_approved_by` is given — an edit withdraws a submission until it is submitted again.

**Own stay or invest (owner, 2026-11-04).** Whoever says on call 1 that they are looking, want to buy or want to find a
property is always asked "是买来自己住，还是投资？". The answer lands in the existing `buy_intent` field (own_stay /
investment / both / undecided), which the slide-out shows as *Buying for*. Before this no prompt asked it, and 5 of the 12
leads who said they were looking had none on record. The prompts reach the agents only through
`php artisan ai-rounds:install-profiles --update --sync`, **run on production**: on this dev box the profile rows carry
production's agent ids.

**Singapore numbers (owner, 2026-11-03: "for 65 phone number need ask the lead is Malaysian or Singapore … and all
these leads need assign to Zen in the round robin").** `PhoneNumber::callingCode()` / `countryName()` / `isSingapore()`
read the lead's number; every call gets `phone_country`, `is_singapore_number` and `nationality` as dynamic variables.
Round 1's prompts ask a +65 lead "你是马来西亚人、新加坡人，还是马来西亚人拿新加坡 PR？" (and a Singaporean "有没有马来西亚的家人可以一起买？"; a Malaysian with Singapore PR buys as a Malaysian, so is not asked)
before question 1; round 2 asks the same when `nationality` is still unknown. The answers are the new extraction
fields `nationality` (`LeadAiRound::NATIONALITY_FIELD`, malaysian / malaysian_sg_pr / singaporean / other / unknown) and
`co_buyer_malaysian` on both rounds — **run `php artisan ai-rounds:install-profiles --update --sync` on production
after this deploy**. The brief's dossier names the phone country so 要核实 covers nationality and the foreigner
minimum; the AI check's `who_buys` fails a foreign number whose nationality is neither known nor asked. Routing:
Setting → Lead Distribution → Round Robin → *Singapore (+65) numbers go to* (`Setting::LEAD_ROTATION_SG_OWNER`) —
`LeadRotationRepository::assignManager()` gives such a new lead to that person with
`LeadAssignment::REASON_COUNTRY_ROUTE` (12) without spending a turn (group rule kept); empty = turns as usual.

**Brief-only drafts (owner, 2026-10-26: "Wai Kit also needs to see the AI advice, even after approval").** An answer
written before the brief existed (the 10-22 P0 — and the bulk `ai-rounds:redraft` skips not-now / not-a-buyer /
not-interested leads, so theirs never got one) shows *No brief yet* with a **Write the brief (keeps the script)**
button: `POST …/answer/rewrite` with `brief_only=1` queues `DraftAiRoundAnswer($id, replace: false, briefOnly: true)`,
whose `LeadAiRoundRepository::briefDrafted()` sets `answer_brief` (and `answer_questions` only when none were written),
stamps `answer_drafted_at`, and leaves the script, the AI's original and every approval exactly as they are — so the
approver sees the advice without un-approving anything. The brief shows in every state (pending, approved, rejected).
Backfill: `php artisan ai-rounds:redraft --brief-only` lists every open round with a script and no brief, any stance or
approval; `--apply` queues them (production had 30 of 49 on 2026-10-26).

**Questions call 2 asks (owner, 2026-10-23: "will the AI ask what I write, like 'what job do his sons do'?").** Until
then the only thing call 2 received was the script, so a question written into it was asked only if the model took
the hint, and the lead's reply landed nowhere but the transcript. Now the drafting AI returns a third part,
`questions` (`ai_rounds_answer.md`: the brief's 第二通要问, phrased as the AI will say them aloud, two or three, in the
script's language; an older prompt without the list falls back to `LeadAiRound::questionsFromBrief()` — the lines under
that heading). They live in `lead_ai_rounds.answer_questions` (json, migration
[2026_10_23_150000_add_answer_questions_to_lead_ai_rounds_table.php](/database/migrations/2026_10_23_150000_add_answer_questions_to_lead_ai_rounds_table.php)),
follow the draft's replace rule in `answerDrafted()`, and are **edited and approved with the script**: the Answer tab has a
second box, *Questions call 2 will ask — one per line* (≤ 10, ≤ 200 chars each, `questionsFromText()` splits and
caps; empty = ask nothing), and `PUT …/answer` carries `questions`; a change to either box clears the approvals like
any edit. Call 2 gets them as `{{questions}}`, numbered one per line, and the round-2 prompts gained a section *讲完建议后要问他的问题* — ask them in order AFTER the advice and BEFORE asking for a time, one at a time, waiting
for each answer, no comment; skip the section when empty. The round-2 profile gained the field **`question_answers`**
(`InstallAiRoundsProfiles::roundTwoFields()`: "问题 → 回答", one per line; `ROUND_TWO_FIELDS` carries it into `r2_data`),
shown as *Answers to our questions* on the Call 2 tab and labelled in `FIELD_LABELS`. **Prod needs
`php artisan ai-rounds:install-profiles --update --sync` after this deploy** (new prompt section + new field on the
round-2 agents); until then the variable is simply empty in the published prompt.

**v2 — the owner's review of 2026-10-21 (built 2026-10-22).** Everything below the table is a slide-out
([RoundDrawer.vue](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/RoundDrawer.vue), `Components/Drawer`, 4xl).
It was first laid out as one long page in the order of the account manager's job, then — the same evening, after the
owner asked that "when users open the drawer he is very clear on everything … use tab to let people click that step" —
rebuilt as **four step TABS** in the drawer's `toolbar` slot (the slot never scrolls), drawn as a **chevron rail**
(2026-10-23, owner: "above need clearly show the step and current step, as arrow"; the same geometry as
`Components/StageTabs`): **Call 1 · Learn their problem → Answer · Check & approve → Call 2 · Explain + book a time →
Round 3 · The human.** The arrow shape carries the order; the fill carries the state — dark navy with a tick = done,
brand blue with a **Now** tag = the step that needs doing, amber **Missed** = tried and not reached, grey = later,
faded + struck through **Skipped** = they said not now / not a buyer / not interested — and a thin bar under a segment
marks the tab being read; the segment's tooltip holds the one-line sub (the last call's outcome and length, who
approved and whether final approval is pending, the booked time, the Zoom). **The drawer opens on the step that needs
doing.** **No button sits in the toolbar** (owner, 10-23: "make CTA button all below") — **every action is in the
footer**, chosen by the step that needs doing: AI call 1 / Call 1 again · Approve (Save & approve after an edit) or
**Final approve — unlock call 2** for a named approver, else "Waiting for final approval", then the **AI call 2
group** — one labelled control with two ways to place the call, **Call now** and **Book a time** (the datetime form
opens above the footer buttons), or **Booked · time** with Cancel once booked — so the booking reads as a booking of
this call · Start Zoom / Join link / Book Zoom, with a quiet "Call 2 again"; a skipped lead gets **Write the answer
anyway**. There is no Close in the footer; the header's × closes. **A guidance line at the top of every tab says
what to do now**, toned blue (to do), green (done, move on), amber (waiting on a call, the AI or the boss) or grey (not
yet possible) — e.g. *"Read the brief, fix the script if anything is off, pick projects if you want them named, then
press Approve"*, *"Call 2 is booked for … Nothing to do — Cancel if plans change"*. The tabs: **Call 1** — the AI's
summary of call 1, the facts as chips, the biggest worry and the help wanted in an amber box, financing and stance side
by side (each editable in place — Enter / Correct), every raw answer folded, then **every try of call 1**, each with
**Recording & transcript** opening [`CallDetailPanel`](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/CallDetailPanel.vue)
INLINE (audio player + the turns; the table's `CallDetailModal` frames the same panel — a modal opened from the slide-out
used to sit behind it, both at z-50; `Modal` / `ConfirmModal` now sit at z-60, above any drawer). **Answer** — the
**brief** as its four boxes (客户情况 / 建议方向 / 要核实 / 第二通要问, parsed from `answer_brief`; for the account manager,
never read to the lead), the **projects** line with the guideline picker folded and a **Guideline** button that opens
the owner's property guideline (`PropertyOptionsModal`, above the drawer), the **script** to edit, the **questions
call 2 will ask** (below; one per line) and one status line (`answer`; the
AI's `answer_draft` folded underneath) and one status line (approved by whom, final approval given or waiting); a lead
who said not now / not a buyer / not interested sees "No call 2" instead — with **Write the answer anyway** (the Rewrite
endpoint) and a pointer to **Correct** on the Call 1 tab, because the bulk redraft skips them but a person may want call
2 regardless; an answer that exists is always shown (with an amber note). The stance that decides is the newest on record
(`row.said`, a colleague's Correct wins), else call 1's reading. **Everyone hears by email** (owner, 2026-10-26):
a submission emails every final approver but the submitter; a final approval and a rejection (reason included, no
length limit — `answer_reject_reason` is `text` since migration
[2026_10_26_100000](/database/migrations/2026_10_26_100000_widen_answer_reject_reason_on_lead_ai_rounds_table.php), the
request caps at 20,000) email the submitter and the lead's account manager but the actor —
[`AiRoundApprovalMailer`](/src/Lead/AiRounds/AiRoundApprovalMailer.php) queues one
[`AiRoundAnswerMail`](/app/Mail/AiRoundAnswerMail.php) (markdown `emails/ai-round-answer`: lead, reason, the script's
first 800 chars, the questions, a button to the right Status view) per recipient, staff only, fail-soft. **Call 2** — every try (recording + transcript inline),
the booked time or the bulk-queue note, what they said on call 2, the "who will call you and when" WhatsApp. **Round 3**
— the round-3 card ([`Round3Cell`](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/Round3Cell.vue), full) and the
**human conversations on record**. The previous step sits at the far left of the footer and the next step beside Close, so the footer walks the
steps too (10-23: "all clickable buttons at the sticky part below"). The
table row is a one-line summary per cell (stance chip + score + the problem, answer status + projects, the latest
call's chip with the count and date, the agreed time) and **the whole row opens the slide-out** (`DataTable`'s
`row-action`; the lead's name still opens the lead). The table borrows the Leads index's shape (2026-10-22):
`sticky-header` + `top-scrollbar` + `compact`, so the header stays put, a wide table scrolls sideways from its top
edge, and more leads fit on one screen. `RoundCallCell` has a `compact` prop for the table; the slide-out mounts it
full, with every call of the round. A **Stance** filter joins Score and Stage, and a **Who** row sits above them
(2026-10-22): **All** · **Grade B** (signal grade B now) · **From funnel** (round 1 was the funnel's registration
call — `LeadAiRound::funnelTriggeredSql()`, the call's `ai_round_trigger = funnel`), each with its count; the funnel
chip also names the nearest class on the list (`counts.funnel_next`). A lead can be both. The default order puts
**registrants with a class ahead first, nearest class first** (`LeadAiRound::nextClassSql()`), then everyone by the
AI's score as before — a registrant's call 2 must land before their class.

- **Model.** `ai_rounds_answer`, `ai_rounds_combine` and `ai_rounds_criteria` carry `provider` + `model` in
  `config/ai_prompts.php` (Claude Sonnet, `claude-sonnet-5-5`); `AiPrompt::registryPin()` reads that in
  `AiClient::prepare()` after an admin pin and before the global default, so an admin can still re-pin a key on
  Manage → AI Prompts. The Anthropic transport drops `temperature` and `thinking` for Sonnet 5+ — expected. The owner
  compared the models on 2026-10-21: Gemini Flash missed "3 months vs 1 year, 2 showrooms" and one model mis-computed a
  DSR, so **the prompt forbids computing any loan, instalment, DSR or deposit** — those numbers come from a person.
- **Brief + script** are one JSON reply (`{"brief","script"}`, `DraftAiRoundAnswer::parts()`); a prose reply (an older
  prompt edit on AI Prompts) is taken whole as the script. The script names no project; the brief names none either
  (the account manager picks them). `LeadAiRoundRepository::answerDrafted()` stores the brief with the draft.
- **Property guideline** — [`AiRoundPropertyOption`](/src/Lead/AiRounds/AiRoundPropertyOption.php), table
  `ai_round_property_options`, seeded from the owner's two slides by
  [`AiRoundPropertyOptionsSeeder`](/database/seeds/AiRoundPropertyOptionsSeeder.php)
  (`php artisan db:seed --class='\AiRoundPropertyOptionsSeeder'` — note the leading backslash, the seeders are global
  classes). A line = quota branch (`QUOTA_BRANCHES`: still has 90% quota / none left — two home loans per person at 90%,
  70% from the third) × LE band (`BANDS`: < 550k · 550–750k · 750k–1m · > 1m, the owner's bands; the public Property
  Match quiz and the Leads list use other cut-offs) × title, price label, structure (HDA / non-HDA), note, and an
  optional link to the **sales project** (`projects`, looked up by name when seeding — ids differ between boxes; the
  slide-out opens its page, a later booking lands on it). **Everyone on the page may edit it** (owner): the "Property
  guideline" button → [PropertyOptionsModal.vue](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/PropertyOptionsModal.vue),
  routes `property-options` (store / update / destroy). Figures are TEXT copied verbatim into the answer — the AI never
  computes from them.
- **Picks per lead** — `lead_ai_rounds.property_options` JSON (`OPTIONS_QUOTA`, `OPTIONS_BAND`, `OPTIONS_IDS`,
  `OPTIONS_INCLUDE`, `OPTIONS_COMBINED_AT`), saved by `PUT {id}/property-options`
  (`LeadAiRoundRepository::propertyOptionsSaved()`). With "Mention these projects in the answer" ticked and
  `combine`, [`CombineAiRoundAnswer`](/app/Jobs/AiRounds/CombineAiRoundAnswer.php) (prompt
  [ai_rounds_combine.md](/resources/prompts/ai_rounds_combine.md)) rewrites the script around the ticked cards —
  `answerCombined()` replaces `answer`, **clears the approval** (a person reads it again) and leaves `answer_draft`
  alone. Unticked, the picks are internal reference only.
- **Approval gates every call 2**: `LeadAiRound::canCallRoundTwo()` now requires `answerApproved()`; the single
  button says "Approve the answer first", bulk already skipped unapproved rows.
- **Round 1 asks two more questions** (prompt files `round1_zh.md` / `round1_yue.md`, Q7 owns property / 90% quota,
  Q8 loan eligibility). **Round 1 is read by OUR OWN AI, not Retell** (owner, 2026-10-22): the round-1 profiles are on
  `analysis_mode = own` (`ai-rounds:install-profiles` sets it), so when a call ends
  [`ReadAiVoiceCall`](/app/Jobs/Calls/ReadAiVoiceCall.php) (Claude Sonnet, prompt `ai_call_reading`) reads the
  transcript against the profile's 16 fields — the 7 Needs answers, `will_attend`, `quota_said`, `call_outcome`, the
  score + reason, and the four past Retell's cap: **`stance`** (`LeadAiRound::STANCES`), **`persona`**
  (`PERSONAS`), `owns_property`, `le_said` — writes them where Retell's analysis used to land, and the aftermath
  runs unchanged. `r1_data` keeps `ROUND_ONE_KEPT`; a stance of busy / nothing said is "not reached"
  (`roundOneNotReached()`), like `call_later`. Retell's own summary stays in the call's `meta.retell_original`.
  Round 2 still uses Retell's analysis (its 5 fields fit) — switch it on the profile form when wanted. The mechanism
  is in the [voice-agent handbook](/docs/modules_handbook/shared/voice-agent/readMe.md) (*Who reads the call*).
  **Takes effect only after `ai-rounds:install-profiles --update --sync`**, which republishes the live agents (with
  an empty extraction schema) — run it before the next bulk round 1, not in the middle of one.
- **The lead's financing — [`LeadLoanProfile`](/src/Lead/LeadLoanProfile.php)** (`lead_loan_profiles`, one row per
  lead; built 2026-10-22): 90% quota left (0 / 1 / 2), loan eligibility (amount + the guideline's band), properties
  owned — each with its SOURCE. [`LoanProfileResolver`](/src/Lead/Support/LoanProfileResolver.php) reads every source
  and the strongest wins per fact: **WhoPay** (CCRIS `facility_house_loan` through `ResidentialLoanQuota`,
  `elig_new_home_loan`) > **staff** (the slide-out's ② "Enter", kept in `sources.staff`) > **said** (round 1's
  `quota_said` / `le_said` / `owns_property`; a spoken amount is parsed — 六百千, 60万, 1.2 million, RM650,000) >
  **quiz** (Property Match `quota` / `loan`, whose 550 / 650 / 850 cut-offs are mapped to the nearest band) > **poll**
  (`webinar_property_count`). Re-resolved by `LeadLoanProfileRepository::sync()` after a round-1 answer and whenever
  a WhoPay report is parsed; `leads:sync-loan-profiles` backfills every lead that has any source (run once on
  deploy). The slide-out's ④ picker starts from the profile's branch and band when nothing was picked, and **call 2
  gets `{{loan_context}}`** (`LeadLoanProfile::spokenContext()`, in the call's language) — the round-2 prompts say it
  aloud only when it is not empty, in the words given, never adding a number.
- **What they said — the stance** (`lead_stances`, see the
  [signal-grade handbook](/docs/modules_handbook/manage/leads/signal-grade.md) *What they SAID*): every AI call our
  reading gave a `stance` to records one (`StanceRecorder::fromAiCall`, with the `stance_quote` and `return_when`
  fields) and regrades the lead at once — not now / not a buyer / not interested leave the S/A/B/C ladder while they
  hold, a WhatsApp STOP excludes. The slide-out shows the newest one with its channel and quote; **Correct** records a
  staff stance that wins from then on. The table's grade chip (B → Not now …) follows.
- **Every channel feeds the stance (2026-10-22).** A **phone call, showroom visit or Zoom meeting** with a transcript
  gets one small Sonnet reading of its own — [`ReadConversationStance`](/app/Jobs/Conversation/ReadConversationStance.php),
  prompt `lead_stance_reading`, queued by `ConversationAutomationService::queueFor()` on every transcript / lead-link
  event whatever `CONVERSATION_AUTOMATION_ENABLED` says (`conversation_automation.stance_reading`), idempotent per
  recording (an existing stance row, or a 60-day cache marker when the reading was "unclear"), head + tail of a long
  transcript (6k + 18k chars). It is **separate from the big conversation analysis on purpose**: that prompt is
  overridden per installation on AI Prompts (production runs its own body on Gemini), so a field added to it would never
  reach production. The webinar poll **"最近有没有考虑买房？" answered No** is a "not now" stance from the day they
  answered ([`WebinarBuyingPlanSync`](/src/Lead/Services/WebinarBuyingPlanSync.php), run beside the property-count
  sync after every poll import and by the hourly `leads:sync-webinar-property-counts`); a Yes is already the
  `poll_buy_yes` signal, and being the lead's own act it outranks an earlier No by itself.
- **The WhatsApp after call 2 (owner, 2026-10-21).** When call 2 agrees a time, `bookCallback()` queues
  [`SendAiRoundSummaryMessage`](/app/Jobs/AiRounds/SendAiRoundSummaryMessage.php): the approved UTILITY template in the
  lead's language (`config('ai_rounds.summary_templates')` — `ai_round2_summary_zh` / `ai_round2_summary_yue`) with
  {{1}} the lead's name, {{2}} the account manager, {{3}} the agreed time spoken ("10月23日（星期四）下午 3:00",
  `spokenTime()`). Once per row (`summary_sent_at`, shown in ⑤); skipped while unapproved, for an opted-out or blocked
  contact, and outside the allowlist. The wording lives in
  [`SubmitAiRoundSummaryTemplates`](/app/Console/Commands/SubmitAiRoundSummaryTemplates.php)
  (`php artisan ai-rounds:submit-summary-templates`), which authored and submitted both to Meta on 2026-10-22 (PENDING);
  the WABA is shared, so any other box sees them after `php artisan whatsapp:sync-templates`.
- **The round-1 opener says why we call (owner, 2026-10-22).** `PlaceAiRoundCall::variables()` hands the agent
  `opening_context` + `opening_kind`: a funnel registrant hears **"谢谢你报名我们的{class}！"** (`registered_ahead`,
  and the prompt then asks whether they will attend); a Grade B lead who registered for a class that already ran hears
  **"你之前报名了我们的{class}，所以打给你，"** (`registered_before`, the newest past registration —
  `pastClass()`); anyone else **"你之前有留过资料给我们，所以打给你，"** (`contact`). Cantonese twins in the same
  places; the class name is the event title without its date (`className()`). The begin message in
  `InstallAiRoundsProfiles::OPENERS` splices the sentence in; the round-1 prompts branch on `opening_kind` — so a
  prompt change here needs `ai-rounds:install-profiles --update --sync` on production again.
- **The boss approves too (owner, 2026-10-22).** Leaders name the **final approvers** on the page's panel
  (`Setting::AI_ROUNDS_FINAL_APPROVERS`, user ids, `PUT …/final-approvers`; empty = off). While anyone is named, the
  account manager's approval is not enough: `LeadAiRound::canCallRoundTwo()` also needs `answer_final_approved_at`
  (`POST …/{id}/answer/final-approve`, approvers only, `LeadAiRoundRepository::finalApproved()`), and bulk round 2
  skips with `SKIP_NOT_FINAL_APPROVED`. An edit, a new draft or a rewrite with projects clears the final approval
  like the manager's. **A final approver's list IS the queue** (`approverList` prop): pick them on *View someone's
  list* — or they sign in — and its STATUS row has only **Pending approval** (every manager's submitted-but-not-final
  answer, across all leads the viewer may see) · **Rejected** (what they sent back) · **Approved**; leads of their own
  are not listed there. Each row's Lead cell and the slide-out's title then name the lead's **account manager**
  (`row.manager`; shown whenever the list is not the viewed person's own) and, when a different colleague engages the
  lead most, the **top REP** (`row.top_rep`, from `LeadTopRep::query()` — the same measure the Leads list's Rep column
  reads; the seat follows it only when clearly ahead and on the round robin, see the Leads handbook) (an old `?view=todo` / `done` link maps to pending / approved). **Approving is SUBMITTING
  while an approver is named** (owner, 2026-10-23): the manager's footer button reads *Submit to approve* (*Save &
  submit to approve* after an edit), the row then says *Pending approval*, and the manager may **Take back & edit**
  (`DELETE …/{id}/answer/approval`, `LeadAiRoundRepository::submissionWithdrawn()` — both approvals cleared, nothing
  else). The approver's footer shows **Final approve** and **Reject** — a reason form opens above the buttons
  (`POST …/{id}/answer/reject`, `RejectAnswerRequest` `reason` 3–500 chars, approvers only, the answer must be
  submitted; `LeadAiRoundRepository::rejected()` stamps `answer_rejected_by / _at / answer_reject_reason` and clears
  both approvals, so call 2 stays locked). The manager then sees **Rejected** on the row (red chip, the reason in its
  tooltip), a red *Rejected* segment on the slide-out's rail, the reason in a red box at the top of the Answer card
  and in the guidance line, and the same *Submit to approve* button; **submitting again clears the rejection**
  (`update()` nulls the three columns whenever `answer_approved_by` is set). The AI call 2 group (Call now | Book a
  time) is **hidden until the answer is unlocked**, so nothing in the footer suggests a call that cannot ring. The
  table's answer chip reads *Rejected* / *Pending approval* / *Approved ✓✓*; the names of the approver, the final
  approver and the rejecter ride on the payload (`answer.rejected = {at, by, reason}`). Migration
  [2026_10_23_100000_add_answer_rejection_to_lead_ai_rounds_table.php](/database/migrations/2026_10_23_100000_add_answer_rejection_to_lead_ai_rounds_table.php).
- **Call 2 booked for a time (owner, 2026-10-22).** **Book a time** beside AI call 2 (`POST …/{id}/call/2/schedule`,
  `ring_at` in Malaysia time; `DELETE` unbooks) writes one bulk-ledger row with `ring_at`
  (`lead_ai_round_bulk_calls.ring_at`); `PlaceBulkAiRoundCall` waits for that minute, then the usual window, budget
  and pacing checks apply, and the call carries `ai_round_trigger = scheduled`. Needs call 2 unlocked, a time still
  ahead and inside the bulk hours; a new booking replaces the old. The row's Call 2 cell and the stepper say
  "Booked · when".
- **Round 3 is a Zoom, booked by call 2 (owner, 2026-10-22).** When call 2 agrees a time, `bookCallback()` first asks
  [`AiRoundZoomBooker`](/src/Lead/AiRounds/AiRoundZoomBooker.php) to book the account manager's **Zoom sales meeting**
  with the lead at that time — Zoom API first, then `ZoomMeetingRepository::create()`, exactly like a meeting scheduled
  from the lead page, hosted by the manager (so they must be a Zoom account user, `ZoomServerService::isAccountUser()`),
  30 minutes, topic "2 Rounds AI Call · {name}", the lead e-mailed the invitation when they have an address. The round
  keeps the link (`callback_zoom_meeting_id`); the Zoom IS the calendar entry, so the round writes no Activity of its
  own in that case. When it cannot be booked (no Zoom user, Zoom down) nothing changes from before: the CALL task +
  calendar entry stand in, and the reason is logged. The table's new **ROUND 3 · HUMAN** band
  ([`Round3Cell.vue`](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/Round3Cell.vue), also in the slide-out's ⑤)
  shows the meeting's state (Upcoming / Live / Ended · minutes / Past / Cancelled, "by AI" when call 2 booked it),
  **Start** for the host while it is ahead, and once it ran the stance our reading took from it and an Open link; a
  meeting booked by hand since the round began shows the same way. Call 2 answered **without** a time shows a
  **Book Zoom** button (the lead page's `ZoomMeetingModal`, for a viewer who is a Zoom account user) — a person books
  round 3. `AiRoundsController::roundThree()` builds the payload; `canBookZoom` / `meetingTypes` are page props.
- **⑤ shows the human conversations on record**: the newest three phone calls / showroom visits / Zoom meetings with a
  transcript or analysis (`AiRoundsController::conversations()`, three light queries, never the transcript), each with
  the AI's summary, the stance read from it and an **Open** link to its own page — so the account manager sees what a
  colleague already heard before calling.

**The catalogue search (owner, 2026-10-20).** The answers were too general, so before writing, the job makes two AI
requests instead of one:

1. **Criteria** — prompt key **`ai_rounds_criteria`** ([resources/prompts/ai_rounds_criteria.md](/resources/prompts/ai_rounds_criteria.md),
   JSON mode, temperature 0) turns call 1 into `states` / `areas` / `types` (landed · high-rise) / `budget_min` /
   `budget_max` / `purpose` / `searchable` (100万 → 1,000,000; an instalment alone → × 200).
   [`PropertyMatcher::clean()`](/src/Lead/AiRounds/PropertyMatcher.php) keeps only known states, short plain area words
   and a believable budget.
2. **Search** — `PropertyMatcher::match()` reads the **catalogue connection only** (the shared master on production —
   never the site DB, never a cross-connection join) and returns up to 8 projects from three pools, best evidence first:
   - **New launch · analysed** (`layout_analyses` joined to `catalog_projects`, same connection) — the best layout per
     project, below-market first, then cash flow; a layout our engine called **OVERPRICED is never suggested**. Only
     KL / Selangor have analyses today (66 MY projects on 2026-10-20).
   - **Completed / subsale** — `price_median` in budget, **≥ 5 recorded sales**, "completed" read off the completion
     year (the master's subsale rows do not carry the `subsale` segment tag). Ranked by a believable yield, then most
     traded — **never by highest yield**, which only finds data errors.
   - **New launch** — any other `new_project` whose launch range overlaps the budget, published first.

   Rules the numbers obey: a yield outside **2–8 %** (`SANE_YIELD`) is a data error and is dropped, and on an analysed
   layout the rent and cash flow derived from it go too; a **landed or high-rise search drops a record that lists the
   other kind as well** (a Taman filed "Flat,Terrace House" has one median blended across both — a landed search found
   flats at a flat's price before this rule), and shops / offices / land / blank types never match; a township filed
   twice (same name, or same median in the same area) is shown once; prices below `CatalogProject::MIN_CREDIBLE_UNIT_PRICE` are
   ignored; the budget window is ±10 % / +5 % (a one-sided budget gets a sensible other side); fewer than 2 matches in
   the areas named → widen to their states; no place named → KL + Selangor. East Malaysia has almost no data, so those
   leads honestly get few or none.
3. **The answer brief** gets only `PropertyMatcher::summary()` — count, the kinds found ("Terrace House ×5"), price /
   rent / yield ranges (a launch range is clamped to the budget; prices rounded to RM1,000, rents to RM10, as they are
   said), the best below-market %, how many already cover their instalment — **no names**, so the model cannot say one.
   `ai_rounds_answer` now also says: be specific (their area, budget, worry, in their words; no filler that fits any
   lead), give the database figures as "records, for reference" in one or two sentences, say what kind they are (never
   call terraces a semi-D), never name or hint at a project, never call an investment safe or guaranteed, stay within
   450 characters (one Knowledge point when the catalogue sentence is there), and say nothing about matches when none
   were found. The criteria prompt maps "central KL" / 市中心 to KLCC · Bukit Bintang · City Centre.

Stored on the row in the same write as the draft (`LeadAiRoundRepository::answerDrafted()`'s `$input`):
`property_criteria` (+ the one-line `searched`), `property_matches`, `properties_matched_at` — kept even when the answer
itself failed; cleared by `reopenRoundOne()`. A criteria request that fails transiently retries the job; anything else
(bad JSON, the catalogue unreachable) writes the answer without the search. **The page shows the matches** in the
*Matching projects · staff only* column ([PropertyMatchesCell.vue](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/PropertyMatchesCell.vue)):
pool, name → `/my/projects/{slug}`, area · type · layout, price, rent / yield / below-market / cash flow / sales on
record. A *not public* project's link opens for staff only (`CatalogueDetailService` preview) — never send it to the
customer. Existing untouched answers are re-written with `php artisan ai-rounds:redraft` (lists; `--apply` queues —
only rows waiting for call 2 that nobody edited or approved, and not a lead whose `call_outcome` is a final verdict —
`LeadAiRound::FINAL_OUTCOMES`; run `ai-rounds:reopen-busy --apply` first).

**Round 1 re-read by hand (2026-10-22).** Retell's post-call analysis (gpt-4.1-mini) misread some calls — a "not needed"
scored 4/5 (#8), a seller's loan worry was recorded as a buyer's (#227) — and 34 "answered" calls were only a hello.
Every round-1 transcript (87 answered + 39 unanswered with a voice) was read again and saved as
`storage/app/ai-rounds-reanalysis/round1-2026-10-22.json` (git-ignored: it holds customer remarks; copy it to the box).
[`ai-rounds:apply-reanalysis {file}`](/app/Console/Commands/ApplyAiRoundReanalysis.php) (dry run; `--apply` writes) puts it
on the rows: new fields / score / summary and a **`stance`** in `r1_data` (`LeadAiRound::STANCES` — wants to buy ·
exploring · not now · not a buyer · not interested · busy · nothing said); busy and nothing-said rows go back to round 1;
the call shows the new summary with Retell's original kept in `meta.retell_original`; a row is skipped when its round-1
call is not the one in the file, an answer was edited / approved, or round 2 started. **Run order after deploy:**
`apply-reanalysis --apply` → `redraft --apply` (which now also skips the not now / not a buyer / not interested
stances — `STANCES_NO_ANSWER`). Result on production's 87: 5 wants to buy, 11 exploring, 16 not now, 5 not a buyer,
10 not interested, 40 back to round 1 (34 nothing said + 6 busy) — so only 16 get a new answer.

*Dry run on production's 87 answered rows (2026-10-20):* 32 go back to round 1 (`reopen-busy` — "talked", nothing
learnt), 4 are final verdicts (skipped), **51 are re-drafted**, of which ~15 gave a budget or a place and get catalogue
matches; the rest get the stricter answer with "no project found". Each re-draft took 5–12 s (criteria ~1.5 s, search
≤ 1.3 s warm, answer 3–10 s); the answer prompt averaged US$0.01 per call on production. Rows recorded before
`call_outcome` existed where the AI wrote "in a meeting, call at 7 PM" into a Needs field (#61, #69, #78, #93) still
count as answered — the account manager presses round 1 again for those.

**The WhatsApp — [`SendAiRoundMissedMessage`](/app/Jobs/AiRounds/SendAiRoundMissedMessage.php).** From the Cloud API
number `AI_ROUNDS_WHATSAPP_PHONE` (+60103368411), as the round's approved template in the lead's language
(`config/ai_rounds.php` → `templates`), `{{1}}` = name, one quick-reply button whose payload is
`AIR:{row uuid}:{round}`. Skipped (and logged) while the template is not approved, for a lead who opted out of
WhatsApp marketing or is blocked, and outside the allowlist. The tap is handled by
[`AiRoundsCallback`](/src/Lead/AiRounds/AiRoundsCallback.php) from `ProcessInboundWhatsAppWebhook::considerAiCallButton`
— only from the lead's own number.

**The four AI profiles** — `ai-rounds-{1,2}-{zh,yue}` (`config/ai_rounds.php` → `profiles`), made by
`php artisan ai-rounds:install-profiles --from=call` from
[resources/prompts/ai_rounds/](/resources/prompts/ai_rounds/). Round 1 extracts the Stage 1 Needs fields plus
`lead_score` (1–5) and `score_reason`; round 2 extracts `callback_time` (ISO), `callback_time_said`, `keen`,
`main_objection`, `next_step`. The Mandarin pair copies the `--from` profile's voice; the Cantonese pair uses
`yue-CN` + `minimax-May` on `speech-2.8-turbo`. Without `--sync` nothing reaches Retell (review on Calls → AI Profiles,
Save publishes); `--update` rewrites existing ones.

**The opener and the questions (owner, 2026-10-15).** Each profile's `begin_message` introduces straight away —
"Hi {{lead_name}}, I am the AI assistant of PropertyLab Wai Kit's team…" — and never asks "is this X?". Round 1 asks,
in this order: looking now? · how can Wai Kit's team help? · the biggest worry (with "why?") · when to buy ·
must-haves and area · budget (the price range, not cash). Openers live in `InstallAiRoundsProfiles::OPENERS`; re-run
the command with `--update --sync` after changing them or the prompt files.

**Money is said, not written (owner, 2026-10-15).** Malaysians never say the currency, so every prompt — the four call
scripts and the answer draft — writes amounts in Chinese characters with no RM / ringgit / 令吉 and no digits: Mandarin
counts in 千 below a million (850,000 → 八百五十千, 3,500 → 三千五), a million and up as 一点零八 million, percent as
巴仙 (七巴仙); the Cantonese pair counts in 萬 (八十五萬) and says 馬幣 once at the first price so a Hong Kong listener
does not hear Hong Kong dollars. Retell's own `normalize_for_speech` is ignored by Retell (see `RetellAgentSync`), so the
prompt is the only place this can live.

**The Settings modal** (the gear button beside *Bulk AI call*, `settingsOpen`; until 2026-10-23 a collapsible
"AI call scripts & WhatsApp messages" panel above the table — the owner wanted the head to be the list, so the setup
moved behind a gear that shows an amber dot while something is not set up; `kit` prop): a link per profile to its Test
tab (`/manage/calls/ai-profiles/{uuid}?tab=test`, new browser tab), each template's wording, button and Meta status,
and the **AI answer brain** — the `###` topics of the knowledge base as it resolves now, with a Read / edit link to
`/manage/integrations/ai/prompts?key=ai_rounds_answer` for whoever may view integrations. So a manager hears and reads
exactly what the lead will. The modal also holds the bulk hours and the final-approval pickers, and lists what is
still missing when `setup.ready` is false. The same day the one-line intro above the table went, and the sky bulk
row now shows only while calls still WAIT (`bulk.waiting > 0`: "N leads are waiting for AI call 2 · about X ring a
minute / they ring from HH:MM", with Cancel waiting calls); a finished batch's placed / skipped tally stays in the
bulk modal. The search box is the Campaigns page's rounded one with a clear button. **The filters are one row per
question, every row the same segmented control** (grey rail, white chosen segment, a quiet count; owner, 2026-10-23:
"one design, one row each"), the label a fixed column: **Status** · Who · AI score · Stance · Stage. **Status is the
list itself** (`?view=`, `AiRoundsController::VIEWS`): **To do** = everything still open · **Check & approve** = an
answer is written, nobody has approved it, not sent back (the row's amber chip) · **Pending approval** = submitted,
waiting for the final approver (shown while an approver is named; empty otherwise) · **Rejected** = sent back with a
reason (shown while an approver is named, or while any exist) · **Approved** = unlocked for call 2 (the account
manager's approval, plus the final one while required) · **Done** = round 3 booked · **Stopped** = Excluded or said no,
shown while anyone is in it: no bulk call rings them, the row names why, and a leader can lift a stop-all with *Correct*.
The four middle ones are slices of To do, not a partition of it; the stage / score / stance / who counts describe the view on screen
(`counts()` takes the view; `counts.total` is its size, the Who row's *All*), and a row with nothing to pick hides
label and all. Stage names are shortened on the control (`STAGE_SHORT`, the full name in the tooltip). Ticks and
*Bulk AI call* show on every open view. The *Property guideline* button left the head the same day: the guideline
opens from the slide-out's Projects line.

**Live rows — no refresh needed.** The page reloads `leads` + `counts` + `bulk` every **10 s** while anything is under
way: a call ringing, a call that ended with a conversation but whose Retell analysis has not landed yet (`settling` —
the cell says "Call ended — reading the results…"; this gap between call-ended and call-analyzed used to stop the
refresh before the score arrived), an answer being written, or a bulk call waiting. Otherwise every **60 s**, so a
lead's own "Call me back now" tap or a colleague's change appears on its own. It pauses while the browser tab is hidden
and refreshes at once when it comes back. The refresh is a **background (`async`) partial reload** — no progress bar,
and DataTable skips its loading overlay for async visits — so only the cells whose data changed re-render; scroll,
ticked rows, an open box and a half-typed answer stay. The **default order is the most valuable customer first** (owner,
2026-10-15): the AI's round-1 score, highest first, unscored leads after, ties by the freshest grade signal — so a row
moves up when its score lands (the only movement left; round activity alone no longer reorders anything). Score /
Updated stay as header sorts. (A dev box sees its own calls settle only with the two settings below.)

**Bulk AI call (owner, 2026-10-15).** The toolbar button opens `Partials/AiRounds/BulkCallModal.vue`, two ways in:
- **Choose by call** — pick a round and *which try of it*: round 1 · 1st call (never called), 2nd call (missed once),
  3rd, 4th or later; the same for round 2 — each with how many leads it would ring (lazy prop `bulkOptions`, built by
  [`AiRoundsBulk`](/src/Lead/AiRounds/AiRoundsBulk.php)) — then **how many** (10 / 50 / 100 / all, or a number). Order:
  round 2 by the AI score, highest first; then the freshest signal.
- **The ticked rows** — tick leads on the page (header box = the whole page), choose round 1 or 2; each gets its own
  next try, and leads that do not fit are skipped and listed in the flash.

Who fits: bulk round 1 = round 1 not answered yet; bulk round 2 = round 1 answered, round 2 not answered yet, **and the
answer approved** (`LeadAiRound::bulkBlock()`). **Approving** = the answer cell's "Approve" / "Save & approve"
(`answer_approved_by/at`, set by the same Save endpoint); a new AI draft that replaces the answer clears it, and existing
human edits counted as approved when the column arrived. A single call 2 is not held to it — its confirm box shows the
text.

A press writes one row per lead to **`lead_ai_round_bulk_calls`** ([`LeadAiRoundBulkCall`](/src/Lead/LeadAiRoundBulkCall.php):
batch, round, try, status Waiting → Called / Skipped (+ reason) / Cancelled, the call) and queues
[`PlaceBulkAiRoundCall`](/app/Jobs/AiRounds/PlaceBulkAiRoundCall.php) on the paced `redis-broadcast` lane. That job shares
the funnel AI calls' pacer (`RateLimited('ai-calls')`, `RETELL_CALLS_PER_MINUTE`, default 10) and calls the same
`PlaceAiRoundCall` a button press does (trigger `bulk`), so every guard still applies. It **waits instead of failing**:
outside the bulk window (**09:00–21:00 Malaysia** by default) it re-queues for the next opening; when the
daily budget is full it re-checks every 10 minutes (no refused call is written per retry); a Retell concurrency refusal
gets one more try 10 minutes later. A lead whose state moved on since the press is Skipped with the reason.

**The bulk hours are set on the page**, in the Settings modal behind the gear (the gear's and the Bulk AI call
button's tooltips show them): `PUT …/ai-rounds/bulk-window` (`UpdateBulkWindowRequest`: `HH:MM`, end after start, same day),
**leaders only** (super admin / view-all-leads — the hours apply to everyone's bulk calls; others see them read-only).
Stored in `settings` (`Setting::AI_ROUNDS_BULK_START` / `_END`), read by `LeadAiRoundBulkCall::window()`; with no row
it falls back to `config('ai_rounds.bulk_window')` (`AI_ROUNDS_BULK_START` / `AI_ROUNDS_BULK_END` in `.env`). A call
already waiting for the old opening wakes then and re-checks against the new hours. Call length is unchanged by owner
decision 2026-10-15: both rounds wrap up from minute 4 (`target_call_minutes`) and are cut at 8 (`max_call_minutes`);
45 s of silence ends the call, voicemail hangs up, ringing stops at 25 s.

On the page: a blue strip shows the latest batch (Waiting / Called / Skipped / Cancelled) and **Cancel waiting calls**
(ConfirmModal → `DELETE …/bulk-calls`, every Waiting row on this list); a row's call cell says "In bulk queue · 2nd
call", or why the bulk call skipped it. The page refreshes every 10 s while anything waits.

**Limits on a bulk run** (none of them is Retell alone): Retell's concurrency (hence the pacer — 700 leads ≈ 70 min);
our daily budget fuse, which reserves `max_call_minutes` × `combined_per_minute_usd` (15 × 0.25 = US$3.75) for every
call still ringing; Meta's daily unique-recipient tier and quality rating on +60103368411 (each miss sends one
template); carriers flagging one caller number that dials many people.

**Money.** The shared daily AI-call fuse `RETELL_DAILY_BUDGET_USD` defaults to **US$100** since 2026-10-15 (owner),
was 15. A box whose `.env` sets it keeps its own value — and **a line set to nothing (`RETELL_DAILY_BUDGET_USD=`) reads
as 0, which switches the fuse OFF** (production had exactly that on 2026-10-15); delete the line or give it a number.

## Funnel registrants — round 1 as the funnel's AI call (owner, 2026-10-16)

The Stage 1 funnel (1. Trust Nurturing — AI买房训练营 and HK 數據講座) wanted its registration call to be these two
rounds. Nothing new is switched on in code: on the funnel's Automation tab, the slot's **AI call** rule picks the
profile **`ai-rounds-1-zh`** instead of its welcome-call profile. Then:

- **The registration rings round 1** ([`SendFunnelAiCall`](/app/Jobs/Automation/SendFunnelAiCall.php) →
  `PlaceAiRoundCall`, trigger `funnel`), ~20 s after sign-up as before, Cantonese for +852. The call carries the
  registration's Session (`LeadAiRound::FUNNEL_META_KEYS`), the opener congratulates them on the class
  (`{{opening_context}}` — empty on every other call), `known_context` tells the agent the class and when it starts,
  and round 1 asks **will they attend** (`will_attend`) before the usual questions.
- **The row opens under the account manager** like any other — a sky **Registered** chip under the name says which
  class. From here it is the ordinary page: the score, the answer draft, Approve, round 2, bulk.
- **The funnel keeps its own WhatsApps** while the class is still ahead: thanks after an answered round 1, the funnel's
  missed-call message (M3) after a miss — its button rings round 1 again — and the M2 offer when the box was unticked.
  The rounds' own "Call me back now" message is NOT sent for these calls: one message, not two.
- **The sales flow hears of every round-1 try** before the class (a press or a bulk retry carries the registration
  too — `LeadAiRound::funnelCall()`): missed → *AI call missed*; answered → the answers are filed, but **the
  colleague's call waits for round 2** (owner's choice): round 2 answered before the class starts hands off at once,
  with the agreed time as the task's appointment line, the task moved to that day and the calendar entry made — the
  row then completes on that task. If round 2 has not come within the usual wait (`handoff_hours`, 6 h, pulled to
  3 h before the class) the journey hands off without it, as for a missed call; a round 2 after that updates the same
  task. After the class has started, the rounds are plain AI 2 Rounds again.
- **Approval is still the gate**: round 2 needs an approved answer, exactly as for grade B leads — an answer nobody
  approves within the wait means the colleague calls without round 2.
- **The Stage 1 flowchart shows it** (My Workspace → Sales flow, `jstage=1`): *AI Call · Round 1* → *Approve the
  answer* (opens this page) → *AI Call · Round 2* → *Human Call / Zoom* at the agreed time — see the
  [sales-flow chart handbook](/docs/modules_handbook/shared/sales-flow-chart/readMe.md). While a lead waits for round 2
  their My Workspace row stays grey ("the system is handling it"): the work is the Approve here, not a call.

## Setup checklist

1. `php artisan migrate` (tables `lead_ai_rounds` + its `summary_sent_at` / `callback_zoom_meeting_id`, `lead_ai_round_bulk_calls`,
   `ai_round_property_options`, `lead_loan_profiles`, `lead_stances`, the profiles' `analysis_mode`), then
   `php artisan db:seed --class='\AiRoundPropertyOptionsSeeder'` for the owner's property guideline and
   `php artisan leads:sync-loan-profiles` once for the financing facts the CRM already holds.
2. `php artisan ai-rounds:install-profiles --from=call`, review the four profiles on Calls → AI Profiles, press Save on
   each (that publishes to Retell).
3. Submit the four templates (`ai_round1_missed_zh`, `ai_round1_missed_yue`, `ai_round2_missed_zh`,
   `ai_round2_missed_yue`) on the Cloud API number and wait for Meta's approval. Until then no WhatsApp is sent.
   The two after-call-2 templates (`ai_round2_summary_zh` / `_yue`) are authored by
   `php artisan ai-rounds:submit-summary-templates` (idempotent); a box that did not submit them runs
   `php artisan whatsapp:sync-templates` to pull them (and their approval) from the shared WABA.
4. The knowledge base ships in the `ai_rounds_answer` prompt; add to it on Manage → AI Prompts as new material comes.
5. The panel's header chip says "Not fully set up yet" and, opened, lists whatever is still missing.
6. **Round 3 Zoom:** every account manager who should get Zooms booked for them must be a user on the connected Zoom
   account (Integrations → Zoom); otherwise their rounds fall back to the call task.
6b. **Final approval:** a leader opens the page's Settings modal (the gear beside *Bulk AI call*) → *Final approval* and
   ticks the boss (Wai Kit). From then on every call-2 answer waits for his approval after the manager's; his own
   list on the page (or picking him on *View someone's list*) is the queue of submitted answers.
6c. **Questions call 2 asks + project knowledge (2026-10-23):** run `php artisan ai-rounds:install-profiles --update --sync`
   on production once more after this deploy — the round-2 prompts gained the *讲完建议后要问他的问题* and *关于答案里提到的项目*
   sections and the round-2 agents the `question_answers` field. Until it runs, `{{questions}}` and `{{projects_context}}`
   are empty in the published prompt: nothing is asked and the agent knows only the card.
6c-bis. **Briefs for old answers (2026-10-26):** `php artisan ai-rounds:redraft --brief-only --apply` once on production —
   the 30 open answers written before the brief existed get the AI advice without their scripts or approvals moving.
6d. **Write the projects' AI knowledge:** Sales Projects → a project → *AI knowledge* tab, for each project on the guideline
   (the Guideline modal's "AI knowledge · N" link shows which are still empty).
6e. **AI review + Singapore numbers (2026-11-03):** `php artisan migrate` (the `answer_check` columns), then
   `php artisan ai-rounds:install-profiles --update --sync` once more (the nationality question and the two new fields).
   Wai Kit opens ⚙ Settings → *AI review* and edits the rules in his own words (the default is his six
   reasons of 2026-10-26). A leader opens Setting → Lead Distribution → Round Robin and picks Zen under *Singapore
   (+65) numbers go to*.
6f. **AI feedback on the answers submitted before the check existed (2026-11-04):** `php artisan ai-rounds:check-answers`
   (list) then `--apply` once on production — every open script without a check gets one; submissions stay as they are.
6g. **Agents, refunds and "stop contacting me" (2026-11-04):** `php artisan leads:apply-contact-review` (list) then
   `--actor=<e-mail> --apply` once on production. The rules are in [signal-grade.md](../leads/signal-grade.md).
7. **Funnel registrants (2026-10-16):** run `php artisan ai-rounds:install-profiles --update --sync` on production
   once (the round-1 opener's `{{opening_context}}` and the `will_attend` question), then pick `ai-rounds-1-zh` on the
   slot's AI-call rule. (Checked 2026-10-23: petav3-dev's four `ai-rounds-*` profiles point at their OWN Retell agents, distinct from production's, so a `--sync` on that box republishes only those; a dev box whose profiles carry production's agent ids must not sync.)

**Each call's chip says why** (2026-10-15): a failed call shows the reason in plain words
(`AiVoiceCall::reasonLabel()` — "Country not enabled on Twilio", "Phone unreachable", "Invalid number", "Declined",
"Line busy", "Voicemail"…), and a "busy, call me later" pick-up shows an amber **"Busy — call later"**.

**Round 1's button after an answer (owner, 2026-10-15).** Once round 1 is answered, its button follows the AI score
(`LeadAiRound::roundOneOutcome()`, threshold `ROUND_ONE_DONE_SCORE` = 2): **score 2 or more → grey and not pressable**
("Round 1 done · 4/5 — no need to call again"); **below 2 → orange**, still pressable, with "Low score — a call from you
may work better than the AI" — the AI learnt little, the account manager may do better calling in person. Unanswered
or unscored rows keep the blue button. Round 2's button is untouched.

**Test mode** — a dev box sets `AI_ROUNDS_ALLOWED_NUMBERS` (comma-separated, with or without `+`, e.g.
`60108685352,60122794886`): it binds the real Retell account, so nothing may ring a customer. While it is set
(`PlaceAiRoundCall::testMode()`), **every row's call button can be pressed**: the confirm box lists the test phones
(`testPhones` prop; the browser remembers the last pick, and a lead whose own number is a test phone picks itself) and
the call **speaks as that lead** (name, round, answer, language) but **rings the picked test phone**
(`PlaceAiRoundCall::target()`; a pick outside the list is refused). The call stays tied to the lead's row and carries
`meta.ai_round_test`; its history line shows "Test → +60…". The missed-call WhatsApp goes to **the number that rang**
(`ai_voice_calls.to_number`) and a "Call me back now" tap from a test phone rings that phone back — so a test runs the
whole flow on the tester's phone.

**Making test mode work end to end on a testing site (2026-10-15).** Two pieces, nothing changes on production's
behaviour:
1. **Call results** — every Retell agent, the four `ai-rounds-*` included, keeps sending its events to production
   (owner: "change back to production"). Production relays a copy to each testing site listed on **Messages → Settings →
   Delivery APIs → Forward Retell call events to testing sites** (`https://<site>/webhooks/retell/voice`; see the
   webhook-forwarding handbook), and the testing site settles its own calls from the copy. The in-call `check_time` tool
   cannot be relayed (Retell waits for its answer), so on a testing site the agent still ends by the profile's hard
   limit. Needs this code on production.
2. **The three jobs** (answer draft, missed-call WhatsApp, bulk call) — `AI_ROUNDS_QUEUE=ai-rounds` puts them on their own
   lane (`App\Jobs\AiRounds\Concerns\RunsOnAiRoundsLane`, on `redis-ai`). Shawn's box runs that lane as
   `peta-zs-queue@airounds` (`worker.sh` lane `airounds`) with a systemd drop-in that gives it network — the only zs
   worker that has any; each of the three jobs refuses numbers outside the allowlist. Production leaves the setting
   unset.

**Bulk calls are not redirected**: in test mode they only ring leads whose own number
is a test phone (otherwise a batch would ring one phone hundreds of times). Production has no allowlist, so none of this
applies there.

## Related files

- [app/Http/Controllers/Manage/MyWork/AiRoundsController.php](/app/Http/Controllers/Manage/MyWork/AiRoundsController.php) · [QueryRequest](/app/Http/Requests/Manage/MyWork/AiRounds/QueryRequest.php) · [UpdateAnswerRequest](/app/Http/Requests/Manage/MyWork/AiRounds/UpdateAnswerRequest.php)
- [app/Actions/PlaceAiRoundCall.php](/app/Actions/PlaceAiRoundCall.php) · bulk: [src/Lead/AiRounds/AiRoundsBulk.php](/src/Lead/AiRounds/AiRoundsBulk.php) · [app/Jobs/AiRounds/PlaceBulkAiRoundCall.php](/app/Jobs/AiRounds/PlaceBulkAiRoundCall.php) · [src/Lead/LeadAiRoundBulkCall.php](/src/Lead/LeadAiRoundBulkCall.php) · [LeadAiRoundBulkCallRepository](/src/Lead/Repositories/LeadAiRoundBulkCallRepository.php) · [BulkCallRequest](/app/Http/Requests/Manage/MyWork/AiRounds/BulkCallRequest.php)
- [src/Lead/LeadAiRound.php](/src/Lead/LeadAiRound.php) · [src/Lead/Repositories/LeadAiRoundRepository.php](/src/Lead/Repositories/LeadAiRoundRepository.php)
- Final approval + booked call 2: `AiRoundsController::finalApproveAnswer / updateFinalApprovers / scheduleCall / cancelScheduledCall`, requests `UpdateFinalApproversRequest` / `ScheduleCallRequest`, `Setting::AI_ROUNDS_FINAL_APPROVERS`, [2026_10_22_150000_add_final_approval_and_scheduled_calls_to_ai_rounds.php](/database/migrations/2026_10_22_150000_add_final_approval_and_scheduled_calls_to_ai_rounds.php)
- Reject with a reason + take a submission back + the Status views (2026-10-23): `AiRoundsController::rejectAnswer / withdrawAnswer`, `VIEWS` + `scope()` / `counts()`, `RejectAnswerRequest`, `LeadAiRoundRepository::rejected / submissionWithdrawn`, `LeadAiRound::pendingFinalApproval / answerRejected`, [2026_10_23_100000_add_answer_rejection_to_lead_ai_rounds_table.php](/database/migrations/2026_10_23_100000_add_answer_rejection_to_lead_ai_rounds_table.php)
- [src/Lead/AiRounds/AiRoundsAftermath.php](/src/Lead/AiRounds/AiRoundsAftermath.php) · [src/Lead/AiRounds/AiRoundsCallback.php](/src/Lead/AiRounds/AiRoundsCallback.php) · round 3: [src/Lead/AiRounds/AiRoundZoomBooker.php](/src/Lead/AiRounds/AiRoundZoomBooker.php) · [Round3Cell.vue](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/Round3Cell.vue) · [2026_10_22_140000_add_callback_zoom_meeting_id_to_lead_ai_rounds_table.php](/database/migrations/2026_10_22_140000_add_callback_zoom_meeting_id_to_lead_ai_rounds_table.php)
- [app/Jobs/AiRounds/DraftAiRoundAnswer.php](/app/Jobs/AiRounds/DraftAiRoundAnswer.php) · [app/Jobs/AiRounds/SendAiRoundMissedMessage.php](/app/Jobs/AiRounds/SendAiRoundMissedMessage.php) · [app/Jobs/AiRounds/SendAiRoundSummaryMessage.php](/app/Jobs/AiRounds/SendAiRoundSummaryMessage.php) · [app/Console/Commands/SubmitAiRoundSummaryTemplates.php](/app/Console/Commands/SubmitAiRoundSummaryTemplates.php) · [2026_10_22_130000_add_summary_sent_at_to_lead_ai_rounds_table.php](/database/migrations/2026_10_22_130000_add_summary_sent_at_to_lead_ai_rounds_table.php)
- Stance from every channel: [app/Jobs/Conversation/ReadConversationStance.php](/app/Jobs/Conversation/ReadConversationStance.php) · [resources/prompts/lead_stance_reading.md](/resources/prompts/lead_stance_reading.md) · [src/Lead/Services/WebinarBuyingPlanSync.php](/src/Lead/Services/WebinarBuyingPlanSync.php) · hook `ConversationAutomationService::queueFor()` · [tests/Feature/Lead/ConversationStanceTest.php](/tests/Feature/Lead/ConversationStanceTest.php)
- Financing + stance: [src/Lead/LeadLoanProfile.php](/src/Lead/LeadLoanProfile.php) · [src/Lead/Support/LoanProfileResolver.php](/src/Lead/Support/LoanProfileResolver.php) · [LeadLoanProfileRepository](/src/Lead/Repositories/LeadLoanProfileRepository.php) · [app/Console/Commands/SyncLeadLoanProfiles.php](/app/Console/Commands/SyncLeadLoanProfiles.php) · [src/Lead/LeadStance.php](/src/Lead/LeadStance.php) · [src/Lead/Support/StanceRecorder.php](/src/Lead/Support/StanceRecorder.php) · requests `UpdateLoanProfileRequest` / `UpdateStanceRequest` · [tests/Feature/Lead/LoanProfileResolverTest.php](/tests/Feature/Lead/LoanProfileResolverTest.php)
- Re-reading: [app/Console/Commands/ApplyAiRoundReanalysis.php](/app/Console/Commands/ApplyAiRoundReanalysis.php)
- v2 (2026-10-22): [RoundDrawer.vue](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/RoundDrawer.vue) · [PropertyOptionsModal.vue](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/PropertyOptionsModal.vue) · [src/Lead/AiRounds/AiRoundPropertyOption.php](/src/Lead/AiRounds/AiRoundPropertyOption.php) · [AiRoundPropertyOptionRepository](/src/Lead/AiRounds/Repositories/AiRoundPropertyOptionRepository.php) · [AiRoundPropertyOptionsSeeder](/database/seeds/AiRoundPropertyOptionsSeeder.php) · [app/Jobs/AiRounds/CombineAiRoundAnswer.php](/app/Jobs/AiRounds/CombineAiRoundAnswer.php) · [resources/prompts/ai_rounds_combine.md](/resources/prompts/ai_rounds_combine.md) · [2026_10_22_100000_add_brief_and_property_options_to_lead_ai_rounds_table.php](/database/migrations/2026_10_22_100000_add_brief_and_property_options_to_lead_ai_rounds_table.php) · requests `UpdatePropertyOptionsRequest` / `StorePropertyOptionRequest` / `UpdatePropertyOptionRequest`
- Catalogue search (off by default since 2026-10-21, `ai_rounds.catalogue_search`): [src/Lead/AiRounds/PropertyMatcher.php](/src/Lead/AiRounds/PropertyMatcher.php) · [resources/prompts/ai_rounds_criteria.md](/resources/prompts/ai_rounds_criteria.md) · [app/Console/Commands/RedraftAiRoundAnswers.php](/app/Console/Commands/RedraftAiRoundAnswers.php) · [2026_10_20_120000_add_property_matches_to_lead_ai_rounds_table.php](/database/migrations/2026_10_20_120000_add_property_matches_to_lead_ai_rounds_table.php)
- [app/Console/Commands/InstallAiRoundsProfiles.php](/app/Console/Commands/InstallAiRoundsProfiles.php) · [resources/prompts/ai_rounds/](/resources/prompts/ai_rounds/) · [resources/prompts/ai_rounds_answer.md](/resources/prompts/ai_rounds_answer.md)
- [config/ai_rounds.php](/config/ai_rounds.php) · [database/migrations/2026_10_15_100000_create_lead_ai_rounds_table.php](/database/migrations/2026_10_15_100000_create_lead_ai_rounds_table.php) · [2026_10_15_120000_create_lead_ai_round_bulk_calls_table.php](/database/migrations/2026_10_15_120000_create_lead_ai_round_bulk_calls_table.php)
- Shared hooks touched: `PlaceAiVoiceCall::dailyBudgetExhausted()` (now public, for the bulk job's wait), `AiVoiceCall::isAiRoundCall()`, `AiCallAftermath` (settled / analyzed branches), `ProcessInboundWhatsAppWebhook::considerAiCallButton` (the `AIR:` payload)
- Funnel registrants: [app/Jobs/Automation/SendFunnelAiCall.php](/app/Jobs/Automation/SendFunnelAiCall.php) · [app/Actions/PlaceFunnelAiCallback.php](/app/Actions/PlaceFunnelAiCallback.php) · `FunnelJourneyHooks::aiCallAnswered()` / `aiRoundTwoAnswered()` · [src/Lead/Journey/Rules/HandOffAfterAiCall.php](/src/Lead/Journey/Rules/HandOffAfterAiCall.php) · `FunnelAutomationSend::SKIP_AI_ROUND_REFUSED`
- Frontend: [Pages/Manage/MyWork/AiRounds.vue](/resources/js/Pages/Manage/MyWork/AiRounds.vue) · [Partials/AiRounds/](/resources/js/Pages/Manage/MyWork/Partials/AiRounds/)
- Tests: [tests/Feature/Manage/AiRoundsTest.php](/tests/Feature/Manage/AiRoundsTest.php) · [tests/Feature/Lead/AiRoundPropertyMatchTest.php](/tests/Feature/Lead/AiRoundPropertyMatchTest.php) · the funnel hand-off: `FunnelJourneyTest::test_an_ai_2_rounds_round_one_leaves_the_colleague_call_to_round_two_on_one_task`
