Reply Agent - Muted Leads

[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_280":281,"_282":283},"routes/_app.x.$handle.$name",{"_94":93,"_32":45,"_102":84,"_27":96,"_13":14,"_103":-5,"_11":104,"_105":106,"_107":108,"_109":110,"_111":112,"_20":-5,"_113":-5,"_114":-5,"_115":25,"_116":117,"_118":119,"_120":-5,"_121":-5,"_122":123},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_43":44,"_32":45,"_38":46,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":-5,"_53":-5,"_18":54,"_55":-5,"_56":25,"_57":-5,"_58":-5,"_59":60,"_61":62,"_63":-5,"_64":-5,"_65":-5,"_66":-5,"_67":68,"_69":68,"_70":71,"_72":71,"_73":71,"_74":68,"_75":68,"_76":77,"_78":79,"_80":81,"_82":25},"version",93,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":25,"_27":28,"_29":28},"file",{"_30":31,"_32":28,"_33":34,"_35":37,"_38":39,"_40":14,"_41":42,"_24":25},"previous",{"_30":31,"_32":28,"_33":34,"_35":36},"deleted",false,"deletedHere","path","playbook2.ts","latestPath","fileId","f4e1e332-8315-11f1-9891-1607ee4eb77e","name","fileType","script","content","// AUTO-GENERATED by scripts/deploy.py from skill/*.md - do not edit here.\nexport const PLAYBOOK_B = \"\\n# Swipe File - Real Replies with Sam's Sample Responses (2026-07-16 rewrite)\\n\\nSource of truth together with templates.md - these override every older\\ninstruction. CONSULT THIS FILE BEFORE EVERY RESPONSE: find the example(s) closest\\nto the prospect's reply and mirror how Sam adapted, blended, and worded the\\ntemplates. Nearly every positive reply is a template used almost verbatim - the\\nchanges are the greeting, the connective tissue, and making it fit the thread and\\nthe first cold email's framing. But cold email isn't deterministic: when a\\ntemplate truly doesn't fit what they said, don't box yourself in - adapt\\n(Example 7 shows a lost meeting from following a template too closely).\\n\\nPlaceholders ({LENGTH_OF_CAMPAIGN}, {SENDER_FIRST_NAME}, {LOWER_BOUND},\\n{TREATMENT_TYPE}, {CURRENCY}, times) and all rules (signature, P.S., verified\\ntimes, offer variants) are defined in templates.md.\\n\\n---\\n\\n## Example 1 - bare \\\"Yes\\\" to the cold email's question\\n\\nThe cold email is question-led (\\\"...would you be interested / like to hear\\nmore?\\\") - so a bare \\\"Yes\\\" / \\\"Sure\\\" / \\\"I'd be interested\\\" / \\\"Ok send info\\\"\\nmeans TELL ME MORE, not \\\"book me\\\".\\n\\nCold email (for context): \\\"Monica - if we could bring your practice 8 implant\\ncases in the next 2 weeks, would you be interested? For this first campaign,\\nwe waive our agency fee entirely. If we don't hit those numbers, we reimburse\\nyou for your ad spend. ... P.S. I just ask for a quick video testimonial if we\\ndeliver :)\\\"\\n\\nProspect: \\\"Yes\\\"\\n\\nSam's response: the Tell Me More template, verbatim (no website line - that's\\nCompany Details only):\\n\\n> Hey, of course.\\n>\\n> We're Lumina and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first 2 weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely. If we don't bring in at least 8 implant cases, we reimburse your ad spend.\\n>\\n> Let's hop on a quick call to share more details about the pilot, answer any questions, and go over the results you can expect.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: This is the most common reply shape. They asked to hear more, so tell\\nthem more - short - then pivot to the call. The guarantee sentence MIRRORS the\\nfirst email (\\\"8 implant cases in 2 weeks\\\" -> \\\"at least 8 implant cases\\\",\\nreimburse wording because that's what the email promised). Just the short\\nsentence - no reimbursement rate, no ad-spend numbers.\\n\\n---\\n\\n## Example 2 - string of questions (how, where, cost)\\n\\nProspect: \\\"Thanks Sam\\nhow do you do this, where are you located & how much does it cost?\\\"\\n\\nSam's response (Tell Me More + the Price pricing line, one seam added for\\nlocation - no website line since they didn't ask who we are):\\n\\n> Hey - happy to answer your questions.\\n>\\n> We're Lumina, based in California, and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first {LENGTH_OF_CAMPAIGN}, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely. If we don't bring in at least {LOWER_BOUND} {TREATMENT_TYPE} cases, we reimburse your ad spend.\\n>\\n> If you'd like to continue with us after the pilot campaign, our flagship plan starts at $497/month or $2,197 for six months paid upfront - that's 25% off, about 6 weeks free.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Three questions, three short answers, one paragraph each - then the CTA.\\nThe fee-waiver block answers \\\"how much\\\" for the free campaign; the flagship\\nline covers what comes after. No padding.\\n\\n---\\n\\n## Example 3 - pure price question\\n\\nProspect: \\\"Hello Sam, What would the cost be for the below?\\\"\\n\\nSam's response: the \\\"Need More Info: Price\\\" template, copied directly (greeting\\n\\\"Hey, totally fair question.\\\" - fee-waiver block with the mirrored guarantee\\nsentence - flagship pricing line - CTA).\\n\\nNotes: Directly copy and paste the Price template. As always, double check that\\nthe greeting, CTA, and structure make sense given their reply and the first\\ncold email. No ad-spend or reimbursement-rate numbers unless they asked.\\n\\n---\\n\\n## Example 4 - referral WITH new contact provided\\n\\nProspect: \\\"Hello Sam,\\nThank you for reaching out to us. We appreciate your offer.\\nFor inquiries regarding marketing opportunities, please directly contact our marketing coordinator, Courtney, at courtney@petitedental.com.\\nBest regards, Monica\\nPetite Dental & Orthodontics\\\"\\n\\nSam's response (to the NEW contact - INTERIM: agent drafts, Sam sends):\\n\\n> Hey Courtney, got in touch with Monica and they told me that you're the right person to talk to about this.\\n>\\n> TLDR: We help dental clinics build a predictable pipeline of high-value treatments with Meta ads, and it all starts with a free {LENGTH_OF_CAMPAIGN} campaign so you can see exactly what working with us looks like before paying us a cent.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Since the initial prospect gave the contact of the new prospect, directly\\nuse the applicable part of the \\\"Wrong Person\\\" template. The greeting uses the new\\nprospect's name.\\n\\n---\\n\\n## Example 5 - \\\"schedule the call\\\" + cost question (template blend)\\n\\nProspect: \\\"OK , Schedule the call . I need to know : * Estimated cost ? *\\\"\\n\\nSam's response: the \\\"Need More Info: Price\\\" template, but the CTA line becomes:\\n\\n> As for the call, how about {X_DAY} at {TIME_1} or {TIME_2} or {Y_DAY} at {TIME_3} or {TIME_4}?\\n>\\n> Or if it's easier, here's our Calendly: [calendar]\\n\\nNotes: Mix \\\"Price\\\" and \\\"Ask for Call\\\". Replace the CTA section of \\\"Price\\\" with\\n\\\"Ask for Call\\\" and use language like \\\"As for the call\\\" to remind them that they\\nasked for the call, leveraging consistency bias.\\n\\n---\\n\\n## Example 6 - \\\"Sure we can book\\\" (explicit booking interest)\\n\\nSam's response:\\n\\n> Hey, great to hear!\\n>\\n> During the call, I'll give you a rundown of the free campaign, what results you can expect, and get the info we need to get your campaign live.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Classic \\\"Interested\\\" reply - they explicitly want to book, so no Tell Me\\nMore needed. Since they said \\\"book\\\", the CTA opens \\\"During the call\\\"\\ninstead of \\\"Let's do a quick call\\\". No P.S. (the old \\\"not a sales\\ncall\\\" P.S. is retired); the first-time-offering-times P.S. may be appended per\\nthe P.S. rules.\\n\\n---\\n\\n## Example 7 - PMoC mixed with a meeting request (the Norton case - REAL, post-mortem)\\n\\nCold email (for context): \\\"Hey Dr. Norton - if I could bring your dental office\\n12-16 patients in 2 weeks on our dime through Meta ads, would that be worth a\\nquick convo? ... p.s. my only ask is an honest video shoutout if I deliver\\\"\\n\\nProspect: \\\"Yes, it would be. Probably have sometime Fri. afternoon but not\\nbefore. 770-401-0102\\\"\\n\\nWHAT THE AGENT SENT (WRONG - this lost the meeting):\\n\\n> Hey Dr. Norton, sounds good! Expect a call from me later today.\\n\\nIt stuck to the old PMoC template verbatim: \\\"later today\\\" directly contradicted\\nthe \\\"Friday afternoon, not before\\\" he had just stated - clearly demonstrating we\\ndidn't listen.\\n\\nSam's sample response (CORRECT - use as the exact reference):\\n\\n> Hey Dr. Norton, sounds good! Expect a few texts from our founder Sam (his number is 925-730-9373) to set up a time to talk Friday afternoon.\\n>\\n> Thanks,\\n> Jennifer\\n> SDR, Lumina\\n\\nTAKEAWAYS:\\n- ALL PMoC acknowledgments include Sam's name and phone number like the above.\\n- MIRROR their stated timing. \\\"Later today\\\" is only for when they gave none.\\n- A PMoC + meeting-request mix is still PMoC: the meeting does NOT get set up\\n over email - acknowledge, and Sam sets it up over text/call (he gets the\\n alert with their number and timing).\\n- You can't stick to templates verbatim in every case. When a reply fits, go\\n 100% template (that's most cases) - but don't box yourself in.\\n\\n---\\n\\n## Example 8 - ad budget question\\n\\nProspect: \\\"Hi Sam,\\nI am open to hearing more! How much should I expect to pay for the ad budget?\\\"\\n\\nSam's response: the \\\"Need More Info: Ad Budget\\\" template:\\n\\n> Hey, good question.\\n>\\n> For your first {LENGTH_OF_CAMPAIGN}, we guarantee 1 patient shows up to their {TREATMENT_TYPE} consultation (capped at {LOWER_BOUND} patients) for every {CURRENCY}150 you spend on ads, or we reimburse your ad spend.\\n>\\n> We'd recommend budgeting {CURRENCY}750 towards Meta ads. At that level of spend, you can expect 5 {TREATMENT_TYPE} patients showing up to their consultations, give or take.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: They explicitly asked about ad budget, so the dedicated template applies -\\nthis is the only reply shape where the {CURRENCY}750 recommendation and the\\nper-{CURRENCY}150 guarantee rate are volunteered together. Short answer, no\\nwhat-we-do padding. (On a legacy thread with no reimburse promise in the first\\nemail: keep just the budgeting paragraph.)\\n\\n---\\n\\n## Example 9 - process + \\\"how do you profit?\\\" (the Macedo case - template blend)\\n\\nProspect: \\\"Hello Abigail, Can you tell me more about your process and how would\\nyou be profiting from this deal?\\nThank you\\\"\\n\\nSam's sample response (blend of \\\"Process\\\" + \\\"Too Good to Be True\\\"):\\n\\n> Hey, totally fair questions.\\n>\\n> During your first {LENGTH_OF_CAMPAIGN} with us, here's what we do free of charge.\\n>\\n> - Produce 25-30 high-performing {TREATMENT_TYPE} ads\\n> - Build a conversion-optimized landing page that turns clicks into treatment-ready buyers\\n> - Install a simple lead management system so your front desk can connect with new leads within 60 seconds\\n>\\n> I'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n>\\n> If you'd like to continue with us after the pilot campaign, our flagship plan starts at $497/month or $2,197 for six months paid upfront - that's 25% off, about 6 weeks free.\\n>\\n> For us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n>\\n> Let's do a quick call so we can give you a rundown of the free campaign, what results you can expect, and get the info we need to create your ads.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Two questions, two templates: the Process bullets answer \\\"your process\\\",\\nthe Too Good to Be True paragraphs answer \\\"how do you profit\\\". The AI's earlier\\nattempt at this reply was rejected as WAY too long - it explained the whole\\nfunnel in dense paragraphs and buried the profit answer. Bullets + short\\nparagraphs win.\\n\\n---\\n\\n## Example 10 - \\\"How can we talk?\\\"\\n\\nSam's response:\\n\\n> Hey, thanks for getting back to me.\\n>\\n> We usually do a quick video call so we can give you a rundown of the free campaign, what results you can expect, and get the info we need to get your campaign live.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: They asked HOW to connect, so answer that directly (\\\"a quick video\\ncall\\\") and go straight to times. No P.S.\\n\\n---\\n\\n## Example 11 - \\\"sam from which company please ?\\\" / website + company asks\\n\\nSam's response: the \\\"Company Details\\\" template verbatim, greeting adapted\\n(\\\"Hey, thanks for getting back to me.\\\"). The \\\"We're Lumina. You can find us at\\nluminamg.com.\\\" line answers who-are-you and website questions in one stroke; if\\nthey also asked about pricing, add the flagship pricing line from the Price\\ntemplate (see Example 2).\\n\\nNotes: The website line lives ONLY here (split from Tell Me More 2026-07-16):\\ncompany-info asks and website asks get Company Details; a plain \\\"yes tell me\\nmore\\\" gets Tell Me More with no website line.\\n\\n---\\n\\n## Example 12 - reschedule request with a stated window (real, Sam-approved)\\n\\nProspect (they had a booked call; office manager replying):\\n\\n\\\"Hi Kim!\\nI am sorry, I had miscommunication with Dan.\\nWe need to reschedule our appointment for some time next week.\\nWill you please cancel our appointment for tomorrow and send me the link to book for a different time next week. Thank you!\\nWarm Regards,\\nAngela Yeager\\nMarketing Coordinator & Business Team Member | Wynkoop Dentistry\\\"\\n\\nAI draft Sam loved:\\n\\n> Hey Angela,\\n>\\n> No worries at all - happy to reschedule. Here's our calendar so you can grab a time that works next week: https://calendly.com/luminamg/pilot\\n>\\n> Thanks,\\n> Kimberly\\n> SDR, Lumina\\n\\nNotes: Short, warm, zero friction - and it MIRRORS THEIR CONSTRAINT (\\\"a time\\nthat works next week\\\"). \\\"Cancel and send the link to book a different time\\\" is\\na RESCHEDULE, not a cancellation. When proposing concrete times for a reply\\nlike this, every offered time must fall inside their stated window (next week\\nmeans nothing this week) - the verified times list is already filtered for\\nthat, and the wording should still echo their constraint like this draft does.\\n\\n---\\n\\n## Example 13 - \\\"how does the reimbursement work?\\\" (direct mechanics question)\\n\\nProspect: \\\"Sounds interesting. How exactly does the ad spend reimbursement\\nwork?\\\"\\n\\nSam's response: the \\\"Reimbursement: direct question\\\" template:\\n\\n> Hey, totally fair question.\\n>\\n> For your first {LENGTH_OF_CAMPAIGN}, you get the same fully-managed service our ongoing clients receive. The only difference is that we waive our service fee entirely, and for every {CURRENCY}150 you spend on ads, we guarantee 1 patient shows up to their {TREATMENT_TYPE} consultation (capped at {LOWER_BOUND} patients), or we reimburse your ad spend.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: The {CURRENCY}150-per-patient rate appears ONLY when they explicitly ask\\nhow the reimbursement/guarantee works - everyone else gets just the short\\nguarantee sentence. No recommended-budget number here (that's the Ad Budget\\ntemplate, for ad-spend questions). If their message is broader skepticism\\n(\\\"what's the angle?\\\"), use Too Good to Be True and fold the guarantee sentence\\nin instead.\\n\\n---\\n\\n## Example 14 - LEGACY: bare \\\"Yes\\\" on an UNTIL WE DO thread (Sam-approved 2026-07-15)\\n\\nOlder threads only - the first email promised to keep running until the\\nnumbers hit (no reimbursement wording existed then).\\n\\nCold email (for context): \\\"...if we could bring your practice 5-10 implant\\ncases in 2 weeks, and you only cover Meta's ad budget, would you be\\ninterested? ... p.s. If we don't hit those numbers, we'll keep it running on\\nour dime until we do\\\"\\n\\nProspect: \\\"Yes\\\"\\n\\nSam's approved response - Tell Me More plus ONE extra line for the clause:\\n\\n> Hey, of course.\\n>\\n> We're Lumina and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first 2 weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely, so you only cover the ad costs paid to Meta.\\n>\\n> And yep - if we don't bring in at least 5 implant cases, we keep running your campaign free of charge until we do. Ad spend included.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: \\\"at least 5\\\" is the LOWER end of the promised 5-10. \\\"free of charge\\\"\\ninstead of the email's \\\"on our dime\\\" - clarity beats verbatim echo (replies\\naren't bound by cold-email spam-word constraints). ONE line, same cadence -\\nnever longer. The fee-waiver copy keeps the old \\\"you only cover the ad costs\\\"\\nclause here because that's what THIS thread's first email promised - never the\\nreimburse sentence. On a BASE thread (no guarantee at all) neither line exists.\\n\\n---\\n\\n## Example 15 - LEGACY: they ask about the Until We Do clause directly (Sam-approved 2026-07-15)\\n\\nProspect: \\\"Wait - you'd really keep running it free until we get those\\nnumbers? How does that work?\\\"\\n\\nSam's approved response:\\n\\n> Hey, totally fair question.\\n>\\n> Yep - exactly what it sounds like. If the first 2 weeks don't get you those 5-10 implant cases, we keep the campaign running on our dime until you have them. That covers the ad spend too, so your total cost stays capped at the first 2 weeks of ad budget.\\n>\\n> I'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n>\\n> For us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: the capped-cost sentence is the strongest line - it turns the promise\\ninto a concrete risk ceiling. No flagship pricing unless they also asked\\nabout price. UNTIL WE DO threads only - on current (reimburse) threads the\\ndedicated Reimbursement template applies instead (Example 13).\\n\\n---\\n\\n## Corrections log (Sam's feedback on AI drafts - newest first)\\n\\n- 2026-07-16: REIMBURSE AD SPEND campaigns launched (new copy overhaul, this\\n file + templates.md + facts.md). (1) The fee-waiver block everywhere is now\\n \\\"...we waive our service fee entirely. If we don't bring in at least\\n {LOWER_BOUND} {TREATMENT_TYPE} cases, we reimburse your ad spend.\\\" - the\\n guarantee sentence mirrors the first email's variant (reimburse / legacy\\n until-we-do line / nothing on legacy base threads). (2) Tell Me More and\\n Company Details SPLIT: website line only in Company Details and when they\\n ask for it. (3) One-liner is now \\\"a predictable pipeline of high-value\\n treatments\\\". (4) Ad budget is {CURRENCY}750 (was 500), 5 patients give or\\n take, still ask-only; reimbursement rate ({CURRENCY}150/patient) only for\\n direct mechanics questions. (5) New Process bullets ({TREATMENT_TYPE} ads,\\n conversion-optimized landing page, 60-second front-desk connect). (6) The\\n phone-ask line (\\\"And what's the best number to reach you at?\\\") is RETIRED -\\n dentists don't use phone numbers much; never include it even if an earlier\\n message in the thread did.\\n- 2026-07-15 (night): UNTIL WE DO variant launched (see Examples 14-15 and\\n templates.md rules). Detect from the first email's P.S.; when present, ONE\\n casual clause line after the fee-waiver copy; direct questions get the\\n dedicated template with the capped-cost sentence. NEVER on base threads.\\n Replies may clarify wording vs the spintax (no spam-word constraints) but\\n mirror numbers/treatment/duration exactly.\\n- 2026-07-15: FRICTION PASS. (1) \\\"15-minute\\\" is DEAD everywhere - it's just \\\"a\\n quick call\\\" / \\\"During the call, I'll...\\\" / \\\"a quick video call\\\". (2) Ad spend\\n is ASK-ONLY: never volunteered in fee-waiver copy or price answers; the\\n Ad Budget template (explicit ad-spend questions only) keeps it.\\n- 2026-07-14 (late): WEBSITE REMOVED FROM SIGNATURES (tentative - Sam is\\n reworking the site). Signature is now just persona name + \\\"SDR, Lumina\\\".\\n Website mentions in template BODIES stay - now Company Details only (2026-07-16).\\n- 2026-07-14: SHORTENED TEMPLATES rewrite. Replies were too long - the\\n \\\"educate them\\\" doctrine is dead; short, simple sentences win. Bare-\\\"Yes\\\"\\n replies get Tell Me More. The \\\"not a sales call\\\" P.S. is RETIRED - only the\\n booked-invite and first-times P.S. lines survive. PMoC acks name founder\\n Sam + 925-730-9373 and mirror stated timing (see Example 7). New templates:\\n Process, Too Good to Be True.\\n- Still in force from earlier corrections: pricing is $497/month or $2,197 for\\n six months upfront (25% off - about 6 weeks free) - any older number ($749,\\n $1,797, \\\"three months\\\") is DEAD. Times are real (verified against the live\\n calendar) - never invent one. Signature block on every reply. \\\"thanks for\\n getting back to me\\\" is a fine greeting. Sign-off \\\"Thanks,\\\" (\\\"Take care,\\\" on\\n reschedule/cancel). Back-and-forth replies go short and loose instead of\\n re-running a full template.\\n\\n---\\n\\n## 2026-07-18 corrections (from the week's real sends - Sam-graded)\\n\\n### Example 16 - \\\"Ok let set up a meeting\\\" is ASK FOR CALL, not a pitch moment\\n\\nThey replied (from info@bocadentalgroup.com): \\\"Ok let set up a meeting Sent from my iPhone\\\"\\n\\nWRONG (what shipped): opened with the full CTA pitch line (\\\"Let's do a quick\\ncall so we can give you a rundown...\\\") before the times. They already asked\\nfor the meeting - re-pitching it is noise.\\n\\nRIGHT:\\n\\n> Hey, great to hear!\\n>\\n> How about Monday at 10am or 1:30pm ET, or Tuesday at 12pm or 4pm ET?\\n>\\n> Or if it's easier, here's our Calendly: [CALENDAR]\\n>\\n> Thanks,\\n> Bethany\\n> SDR, Lumina\\n>\\n> P.S. If those times don't work, feel free to send over other openings and I'll book you in directly\\n\\nSame grading applies to \\\"can we schedule a call to further discuss\\\" (Example\\nfrom the Mariana thread) - short greeting (\\\"Hey Mariana - of course!\\\"),\\ntimes, Calendly line, new P.S. Nothing else.\\n\\n### Example 17 - blend TMM + Too Good to Be True; no price on skepticism\\n\\nThey replied: \\\"Let's do it! I'd love to learn more about your offer. I hope\\nto hear the catch too - Thanks.\\\"\\n\\nWRONG (what shipped): Company-Details-style body plus \\\"the only thing we ask\\nfor is a short testimonial video\\\" as the catch answer, wrong greeting.\\n\\nRIGHT (relevant parts of Tell Me More, then relevant parts of Too Good to Be\\nTrue, NO price):\\n\\n> Hey, of course!\\n>\\n> We're Lumina and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first 2 weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely, so you only cover the ad costs paid to Facebook.\\n>\\n> There's no catch :)\\n>\\n> I'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n>\\n> For us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n>\\n> [CTA BLOCK]\\n\\nNote: \\\"Of course!\\\" (not \\\"great to hear!\\\") is the natural response to \\\"I'd\\nlove to learn more about your offer.\\\"\\n\\n### Example 18 - \\\"I'm out next week, book me the week after\\\" gets REAL times\\n\\nThey replied (deep in thread, from Angela): \\\"I am out of the office all next\\nweek and will be unable to meet. Please let me know when your calendar is\\nextended into the following week.\\\"\\n\\nWRONG (what shipped): \\\"Since I don't have any open slots to offer just yet\\nfor the week after, go ahead and grab whatever works best for you here once\\nit's open: [link]\\\" - punted to a link instead of booking her.\\n\\nRIGHT (the system now fetches that week's availability - propose it; note\\nthe deep-thread phrasing, looser than the standard shape):\\n\\n> Hey Angela,\\n>\\n> No worries at all - enjoy the time off. I can book you in manually for the following week. Would {DAY 1} at {TIME 1} or {TIME 2} work or {DAY 2} at {TIME 3} or {TIME 4}?\\n>\\n> Feel free to suggest other times that work for you.\\n>\\n> Take care,\\n> Kimberly\\n> SDR, Lumina\\n\\n### Doctrine updates that ride with these examples\\n\\n- New universal first-times P.S.: \\\"If those times don't work, feel free to\\n send over other openings and I'll book you in directly.\\\"\\n- Booking confirmations are the two-variant template in templates.md (office\\n email check, Meet link, two Add-to-Calendar links, gated reminders-phone\\n ask). The old one-liner + \\\"better email\\\" P.S. is dead.\\n- CTA phone line (\\\"And just for convenience, what's a good number to reach\\n you directly?\\\") - first proposing reply only, system-gated, never once a\\n number is known. Max two asks per lead ever (CTA + booking confirmation).\\n- Tell Me More now carries the \\\"Three quick bullets\\\" block - the only place\\n price is volunteered outside the Price template.\\n\\n\\n\\n# Mechanism Deep Context (from Sam's founder videos)\\n\\nDistilled from the three YouTube walkthroughs. Use ONLY when a prospect asks a\\nspecific how-does-it-work question that facts.md doesn't cover, and translate to\\nplain, first-read-understandable language - never technical, never jargon. Items\\nmarked \\\"(general context)\\\" are instructional content, not necessarily what we do for\\nevery client. If anything here conflicts with facts.md, facts.md wins (e.g. the\\nvideos mention a \\\"free advisory call\\\" and typical ad spend of $1,500-3,000/mo; the\\ncurrent offer is the free first campaign, then $497/mo or $2,197 for six months\\nupfront (25% off), with $750 of ad spend recommended during the free\\ncampaign).\\nNEVER dump this wholesale into a reply - pull the one or two ideas that\\nanswer the question, simply. When in doubt, share the right video link in the\\nbody instead of explaining.\\n\\n---\\n\\n## Video 1: How the lead-generation system works (youtu.be/WlAQcaskL28)\\n\\n### The core problem we solve\\nDentists' number-one fear with paid ads is being flooded with junk leads that waste the front desk's and treatment coordinator's time. Our system is built around deliberately making it slightly harder to book - the right amount of \\\"hoops\\\" filters out tire-kickers and price-shoppers while keeping real patients. Too few hoops brings in junk; too many scares off good patients. We aim for the middle.\\n\\n### The mechanism, step by step\\n1. **Ad on Facebook/Instagram** promotes a specific treatment offer (e.g., implants, Invisalign).\\n2. **A dedicated offer page** (not the practice's homepage) built to get one action: request a consult. It includes a short video explaining the treatment, **the price shown up front** (for a $6,000 Invisalign case, price is the top reason deals die - showing it early makes non-buyers remove themselves), financing options (to bring back good patients who just need payment plans), and a clear treatment timeline so people know exactly what they're signing up for.\\n3. **A qualifying questionnaire before anyone can see the calendar.** The booking calendar is treated as sacred - only people who pass budget questions (\\\"Invisalign runs $4,000-$8,000 - does that work?\\\") and health-readiness questions (gum disease, loose teeth, last checkup, how soon they want to start) ever get to book. People who don't qualify are routed to a follow-up sequence - often offered a checkup or a lower-cost treatment first, then brought back to the bigger treatment later.\\n4. **A small refundable deposit to lock the appointment.** Serious patients don't mind (they get it back); non-serious ones won't put a card down. Result: far fewer no-shows and higher treatment acceptance.\\n5. **Teaching Facebook who your best patients are.** Tracking code on the page reports back only *qualified* actions - a qualified lead, a booked consult, a started treatment. Facebook (which holds ~52,000 data points per person) then goes and finds more people like your paying patients, not like your junk leads. When your coordinator moves a patient to \\\"showed up qualified\\\" or \\\"started treatment\\\" in the pipeline, that gets sent back to Facebook automatically, so targeting keeps improving.\\n6. **Fast follow-up:** automated texting contacts new leads within about 60 seconds, while they're still warm.\\n\\n### What we specifically do for clients\\n- We never change a funnel until it has run **at least 3 days and $300 of spend** - enough data for an informed decision, not guesswork.\\n- For new clients (or polluted past data), we start by having Facebook find leads, then step down to optimizing for booked consults, then for started treatments, as data accumulates. If ad costs rise later, we step back up to widen the pool.\\n- We diagnose *where* to add or remove hoops using benchmarks: **1.5-3%** of ad viewers clicking, **5-7%** of visitors booking, **70-80%** showing up, **~30%** starting treatment. Whichever number lags tells us what to fix.\\n\\n### Objections this answers\\n- **\\\"Leads will be junk.\\\"** The whole system exists to pre-filter them, medically and financially.\\n- **\\\"Referral leads are better.\\\"** (general context) True - ads reach strangers, so ad leads will always feel colder; judge them against benchmarks, not word-of-mouth.\\n- **\\\"Low booking or close rates mean bad leads.\\\"** Often it's slow follow-up or an untrained coordinator - we address both with automation, scripts, and training.\\n\\n---\\n\\n## Video 2: How leads become booked patients (youtu.be/PvFJuJSNZiI)\\n\\n### The core idea: speed to lead\\nThe single biggest factor in turning a lead into a booked patient is how fast someone contacts them. A Harvard Business Review study found that responding within 5 minutes makes you **21x more likely** to convert a lead into a patient versus responding within 30 minutes. Yet the average dental practice takes **47 hours** to respond to new inquiries.\\n\\n(general context) Most practices lose patients not because the leads are bad, but because leads aren't followed up fast enough or persistently enough. Each lead costs real money (roughly a $100 bill) and can become a patient worth $10,000+ over their lifetime.\\n\\n### Why practices think this doesn't apply to them (common objection)\\nDentists often say \\\"my front desk is always on the phone.\\\" Two problems: (1) calls cluster at peak times - start of day, lunch, after hours - so more than one comes in at once and some get missed; (2) most practices have no hard data on missed calls or response time, so they go by feel.\\n\\n### The follow-up system\\nThree ways a lead can book, all offered right after they submit their info:\\n1. **We call them** - an outbound call a few minutes after they opt in, from a clear script: answer basic questions about candidacy, treatment, and pricing, then steer toward booking.\\n2. **They call the office** - handled the same way, plus capturing contact info for follow-up.\\n3. **They book themselves online** - self-scheduled patients have a **41% lower no-show rate** because they chose a time that truly works for them.\\n\\nIf a lead doesn't answer: same day, call twice, leave a voicemail, and text several times spaced hours apart. Days 1-2: two calls and two texts daily. Days 3-6: two calls and one text daily. After a week: a series of helpful emails (about the team, promotions, procedures) that builds enough trust to book later.\\n\\n**Personal touches that lift booking rates:** tailor messages to what the lead said on the intake questions (e.g., financing needs get monthly payment plan info), and send a short video of the doctor welcoming them - putting a face to the practice builds trust fast.\\n\\n### Scheduling practices that raise show rates\\n(general context) Keep plenty of open slots (availability directly drives bookings; aim for weekday 9am-3pm coverage); keep consults to ~45 minutes or less; 15-minute booking intervals; book no more than ~3 business days out and pull appointments earlier when possible - the further out, the more no-shows.\\n\\n### What Lumina actually does for clients\\nFor every practice we work with, we implement all these automations and hand over all the scripts and standard operating procedures - the whole follow-up and booking system is done for the practice, alongside running their ads.\\n\\n---\\n\\n## Video 3: The ads themselves (youtu.be/X_Cc8Iy3Hgg)\\n\\n### The short answer\\nFor each practice we build and run **25-30 genuinely different ads at all times** - not one ad copied 30 times with small tweaks. Each ad combines three ingredients: an **offer** (what brings a new patient in, like a new-patient special or a free consultation for a big-ticket treatment), an **angle** (the specific reason that patient cares - comfort, speed, price, trust), and a **format** (patient testimonial, before-and-after, the dentist talking to camera, a photo, or footage of the office and chairside care).\\n\\n### Why so many ads?\\n(general context) In 2025, Facebook/Instagram got dramatically better at matching each person with an ad they'll respond to (\\\"Meta Andromeda\\\" update) - and dental ad costs jumped **97% in one year**, the steepest rise of any industry on the platform. Practices still running two or three ads now pay far more per patient.\\n\\n(general context) Even for the same treatment, patients want different things: one Invisalign prospect is weighing clear aligners against metal braces (needs to hear discreet, faster, comfortable); another is price-shopping (wants the discount and freebies). One generic ad can't speak to both. Without variety you get \\\"ad fatigue\\\" - paying more per patient and getting fewer of them.\\n\\n### How we build them\\n- **Start broad:** launch a mix of offers, angles, and formats.\\n- **Find winners:** when one element clearly works, build ~30 fresh variations around it - new angles (credentials, friendly team, treatment timelines, what the first visit is like) and new formats.\\n- **Keep winners running, keep testing:** a proven winner gets its own budget and often keeps producing for months or years.\\n- **Give every ad a fair shot:** at least 3 days and ~$300 of spend before judging; we test big, meaningful differences, not tiny word tweaks.\\n\\n### Targeting by treatment\\nAds are built per treatment (Invisalign, veneers, implants, whitening, etc.) and, within each treatment, per patient motivation - the price-shopper, the anxious patient, the appearance-focused patient - so the ad each person sees feels written for them. That personal fit is how a smaller practice competes against big spenders in crowded markets.\\n\\n### Common questions this answers\\n- *\\\"Do you just run one ad for my practice?\\\"* No - 25-30 unique ones, refreshed as winners emerge.\\n- *\\\"Won't people get sick of seeing my ads?\\\"* That's exactly what the volume and variety prevent.\\n\\n\";\n","// AUTO-GENERATED by scripts/deploy.py from skill/*.md - do not edit here.\nexport const PLAYBOOK_B = \"\\n# Swipe File - Real Replies with Sam's Sample Responses (2026-07-16 rewrite)\\n\\nSource of truth together with templates.md - these override every older\\ninstruction. CONSULT THIS FILE BEFORE EVERY RESPONSE: find the example(s) closest\\nto the prospect's reply and mirror how Sam adapted, blended, and worded the\\ntemplates. Nearly every positive reply is a template used almost verbatim - the\\nchanges are the greeting, the connective tissue, and making it fit the thread and\\nthe first cold email's framing. But cold email isn't deterministic: when a\\ntemplate truly doesn't fit what they said, don't box yourself in - adapt\\n(Example 7 shows a lost meeting from following a template too closely).\\n\\nPlaceholders ({LENGTH_OF_CAMPAIGN}, {SENDER_FIRST_NAME}, {LOWER_BOUND},\\n{TREATMENT_TYPE}, {CURRENCY}, times) and all rules (signature, P.S., verified\\ntimes, offer variants) are defined in templates.md.\\n\\n---\\n\\n## Example 1 - bare \\\"Yes\\\" to the cold email's question\\n\\nThe cold email is question-led (\\\"...would you be interested / like to hear\\nmore?\\\") - so a bare \\\"Yes\\\" / \\\"Sure\\\" / \\\"I'd be interested\\\" / \\\"Ok send info\\\"\\nmeans TELL ME MORE, not \\\"book me\\\".\\n\\nCold email (for context): \\\"Monica - if we could bring your practice 8 implant\\ncases in the next 2 weeks, would you be interested? For this first campaign,\\nwe waive our agency fee entirely. If we don't hit those numbers, we reimburse\\nyou for your ad spend. ... P.S. I just ask for a quick video testimonial if we\\ndeliver :)\\\"\\n\\nProspect: \\\"Yes\\\"\\n\\nSam's response: the Tell Me More template, verbatim (no website line - that's\\nCompany Details only):\\n\\n> Hey, of course.\\n>\\n> We're Lumina and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first 2 weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely. If we don't bring in at least 8 implant cases, we reimburse your ad spend.\\n>\\n> Let's hop on a quick call to share more details about the pilot, answer any questions, and go over the results you can expect.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: This is the most common reply shape. They asked to hear more, so tell\\nthem more - short - then pivot to the call. The guarantee sentence MIRRORS the\\nfirst email (\\\"8 implant cases in 2 weeks\\\" -> \\\"at least 8 implant cases\\\",\\nreimburse wording because that's what the email promised). Just the short\\nsentence - no reimbursement rate, no ad-spend numbers.\\n\\n---\\n\\n## Example 2 - string of questions (how, where, cost)\\n\\nProspect: \\\"Thanks Sam\\nhow do you do this, where are you located & how much does it cost?\\\"\\n\\nSam's response (Tell Me More + the Price pricing line, one seam added for\\nlocation - no website line since they didn't ask who we are):\\n\\n> Hey - happy to answer your questions.\\n>\\n> We're Lumina, based in California, and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first {LENGTH_OF_CAMPAIGN}, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely. If we don't bring in at least {LOWER_BOUND} {TREATMENT_TYPE} cases, we reimburse your ad spend.\\n>\\n> If you'd like to continue with us after the pilot campaign, our flagship plan starts at $497/month or $2,197 for six months paid upfront - that's 25% off, about 6 weeks free.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Three questions, three short answers, one paragraph each - then the CTA.\\nThe fee-waiver block answers \\\"how much\\\" for the free campaign; the flagship\\nline covers what comes after. No padding.\\n\\n---\\n\\n## Example 3 - pure price question\\n\\nProspect: \\\"Hello Sam, What would the cost be for the below?\\\"\\n\\nSam's response: the \\\"Need More Info: Price\\\" template, copied directly (greeting\\n\\\"Hey, totally fair question.\\\" - fee-waiver block with the mirrored guarantee\\nsentence - flagship pricing line - CTA).\\n\\nNotes: Directly copy and paste the Price template. As always, double check that\\nthe greeting, CTA, and structure make sense given their reply and the first\\ncold email. No ad-spend or reimbursement-rate numbers unless they asked.\\n\\n---\\n\\n## Example 4 - referral WITH new contact provided\\n\\nProspect: \\\"Hello Sam,\\nThank you for reaching out to us. We appreciate your offer.\\nFor inquiries regarding marketing opportunities, please directly contact our marketing coordinator, Courtney, at courtney@petitedental.com.\\nBest regards, Monica\\nPetite Dental & Orthodontics\\\"\\n\\nSam's response (to the NEW contact - INTERIM: agent drafts, Sam sends):\\n\\n> Hey Courtney, got in touch with Monica and they told me that you're the right person to talk to about this.\\n>\\n> TLDR: We help dental clinics build a predictable pipeline of high-value treatments with Meta ads, and it all starts with a free {LENGTH_OF_CAMPAIGN} campaign so you can see exactly what working with us looks like before paying us a cent.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Since the initial prospect gave the contact of the new prospect, directly\\nuse the applicable part of the \\\"Wrong Person\\\" template. The greeting uses the new\\nprospect's name.\\n\\n---\\n\\n## Example 5 - \\\"schedule the call\\\" + cost question (template blend)\\n\\nProspect: \\\"OK , Schedule the call . I need to know : * Estimated cost ? *\\\"\\n\\nSam's response: the \\\"Need More Info: Price\\\" template, but the CTA line becomes:\\n\\n> As for the call, how about {X_DAY} at {TIME_1} or {TIME_2} or {Y_DAY} at {TIME_3} or {TIME_4}?\\n>\\n> Or if it's easier, here's our Calendly: [calendar]\\n\\nNotes: Mix \\\"Price\\\" and \\\"Ask for Call\\\". Replace the CTA section of \\\"Price\\\" with\\n\\\"Ask for Call\\\" and use language like \\\"As for the call\\\" to remind them that they\\nasked for the call, leveraging consistency bias.\\n\\n---\\n\\n## Example 6 - \\\"Sure we can book\\\" (explicit booking interest)\\n\\nSam's response:\\n\\n> Hey, great to hear!\\n>\\n> During the call, I'll give you a rundown of the free campaign, what results you can expect, and get the info we need to get your campaign live.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Classic \\\"Interested\\\" reply - they explicitly want to book, so no Tell Me\\nMore needed. Since they said \\\"book\\\", the CTA opens \\\"During the call\\\"\\ninstead of \\\"Let's do a quick call\\\". No P.S. (the old \\\"not a sales\\ncall\\\" P.S. is retired); the first-time-offering-times P.S. may be appended per\\nthe P.S. rules.\\n\\n---\\n\\n## Example 7 - PMoC mixed with a meeting request (the Norton case - REAL, post-mortem)\\n\\nCold email (for context): \\\"Hey Dr. Norton - if I could bring your dental office\\n12-16 patients in 2 weeks on our dime through Meta ads, would that be worth a\\nquick convo? ... p.s. my only ask is an honest video shoutout if I deliver\\\"\\n\\nProspect: \\\"Yes, it would be. Probably have sometime Fri. afternoon but not\\nbefore. 770-401-0102\\\"\\n\\nWHAT THE AGENT SENT (WRONG - this lost the meeting):\\n\\n> Hey Dr. Norton, sounds good! Expect a call from me later today.\\n\\nIt stuck to the old PMoC template verbatim: \\\"later today\\\" directly contradicted\\nthe \\\"Friday afternoon, not before\\\" he had just stated - clearly demonstrating we\\ndidn't listen.\\n\\nSam's sample response (CORRECT - use as the exact reference):\\n\\n> Hey Dr. Norton, sounds good! Expect a few texts from our founder Sam (his number is 925-730-9373) to set up a time to talk Friday afternoon.\\n>\\n> Thanks,\\n> Jennifer\\n> SDR, Lumina\\n\\nTAKEAWAYS:\\n- ALL PMoC acknowledgments include Sam's name and phone number like the above.\\n- MIRROR their stated timing. \\\"Later today\\\" is only for when they gave none.\\n- A PMoC + meeting-request mix is still PMoC: the meeting does NOT get set up\\n over email - acknowledge, and Sam sets it up over text/call (he gets the\\n alert with their number and timing).\\n- You can't stick to templates verbatim in every case. When a reply fits, go\\n 100% template (that's most cases) - but don't box yourself in.\\n\\n---\\n\\n## Example 8 - ad budget question\\n\\nProspect: \\\"Hi Sam,\\nI am open to hearing more! How much should I expect to pay for the ad budget?\\\"\\n\\nSam's response: the \\\"Need More Info: Ad Budget\\\" template:\\n\\n> Hey, good question.\\n>\\n> For your first {LENGTH_OF_CAMPAIGN}, we guarantee 1 patient shows up to their {TREATMENT_TYPE} consultation (capped at {LOWER_BOUND} patients) for every {CURRENCY}150 you spend on ads, or we reimburse your ad spend.\\n>\\n> We'd recommend budgeting {CURRENCY}750 towards Meta ads. At that level of spend, you can expect 5 {TREATMENT_TYPE} patients showing up to their consultations, give or take.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: They explicitly asked about ad budget, so the dedicated template applies -\\nthis is the only reply shape where the {CURRENCY}750 recommendation and the\\nper-{CURRENCY}150 guarantee rate are volunteered together. Short answer, no\\nwhat-we-do padding. (On a legacy thread with no reimburse promise in the first\\nemail: keep just the budgeting paragraph.)\\n\\n---\\n\\n## Example 9 - process + \\\"how do you profit?\\\" (the Macedo case - template blend)\\n\\nProspect: \\\"Hello Abigail, Can you tell me more about your process and how would\\nyou be profiting from this deal?\\nThank you\\\"\\n\\nSam's sample response (blend of \\\"Process\\\" + \\\"Too Good to Be True\\\"):\\n\\n> Hey, totally fair questions.\\n>\\n> During your first {LENGTH_OF_CAMPAIGN} with us, here's what we do free of charge.\\n>\\n> - Produce 25-30 high-performing {TREATMENT_TYPE} ads\\n> - Build a conversion-optimized landing page that turns clicks into treatment-ready buyers\\n> - Install a simple lead management system so your front desk can connect with new leads within 60 seconds\\n>\\n> I'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n>\\n> If you'd like to continue with us after the pilot campaign, our flagship plan starts at $497/month or $2,197 for six months paid upfront - that's 25% off, about 6 weeks free.\\n>\\n> For us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n>\\n> Let's do a quick call so we can give you a rundown of the free campaign, what results you can expect, and get the info we need to create your ads.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: Two questions, two templates: the Process bullets answer \\\"your process\\\",\\nthe Too Good to Be True paragraphs answer \\\"how do you profit\\\". The AI's earlier\\nattempt at this reply was rejected as WAY too long - it explained the whole\\nfunnel in dense paragraphs and buried the profit answer. Bullets + short\\nparagraphs win.\\n\\n---\\n\\n## Example 10 - \\\"How can we talk?\\\"\\n\\nSam's response:\\n\\n> Hey, thanks for getting back to me.\\n>\\n> We usually do a quick video call so we can give you a rundown of the free campaign, what results you can expect, and get the info we need to get your campaign live.\\n>\\n> [times + calendar CTA]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: They asked HOW to connect, so answer that directly (\\\"a quick video\\ncall\\\") and go straight to times. No P.S.\\n\\n---\\n\\n## Example 11 - \\\"sam from which company please ?\\\" / website + company asks\\n\\nSam's response: the \\\"Company Details\\\" template verbatim, greeting adapted\\n(\\\"Hey, thanks for getting back to me.\\\"). The \\\"We're Lumina. You can find us at\\nluminamg.com.\\\" line answers who-are-you and website questions in one stroke; if\\nthey also asked about pricing, add the flagship pricing line from the Price\\ntemplate (see Example 2).\\n\\nNotes: The website line lives ONLY here (split from Tell Me More 2026-07-16):\\ncompany-info asks and website asks get Company Details; a plain \\\"yes tell me\\nmore\\\" gets Tell Me More with no website line.\\n\\n---\\n\\n## Example 12 - reschedule request with a stated window (real, Sam-approved)\\n\\nProspect (they had a booked call; office manager replying):\\n\\n\\\"Hi Kim!\\nI am sorry, I had miscommunication with Dan.\\nWe need to reschedule our appointment for some time next week.\\nWill you please cancel our appointment for tomorrow and send me the link to book for a different time next week. Thank you!\\nWarm Regards,\\nAngela Yeager\\nMarketing Coordinator & Business Team Member | Wynkoop Dentistry\\\"\\n\\nAI draft Sam loved:\\n\\n> Hey Angela,\\n>\\n> No worries at all - happy to reschedule. Here's our calendar so you can grab a time that works next week: https://calendly.com/luminamg/pilot\\n>\\n> Thanks,\\n> Kimberly\\n> SDR, Lumina\\n\\nNotes: Short, warm, zero friction - and it MIRRORS THEIR CONSTRAINT (\\\"a time\\nthat works next week\\\"). \\\"Cancel and send the link to book a different time\\\" is\\na RESCHEDULE, not a cancellation. When proposing concrete times for a reply\\nlike this, every offered time must fall inside their stated window (next week\\nmeans nothing this week) - the verified times list is already filtered for\\nthat, and the wording should still echo their constraint like this draft does.\\n\\n---\\n\\n## Example 13 - \\\"how does the reimbursement work?\\\" (direct mechanics question)\\n\\nProspect: \\\"Sounds interesting. How exactly does the ad spend reimbursement\\nwork?\\\"\\n\\nSam's response: the \\\"Reimbursement: direct question\\\" template:\\n\\n> Hey, totally fair question.\\n>\\n> For your first {LENGTH_OF_CAMPAIGN}, you get the same fully-managed service our ongoing clients receive. The only difference is that we waive our service fee entirely, and for every {CURRENCY}150 you spend on ads, we guarantee 1 patient shows up to their {TREATMENT_TYPE} consultation (capped at {LOWER_BOUND} patients), or we reimburse your ad spend.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: The {CURRENCY}150-per-patient rate appears ONLY when they explicitly ask\\nhow the reimbursement/guarantee works - everyone else gets just the short\\nguarantee sentence. No recommended-budget number here (that's the Ad Budget\\ntemplate, for ad-spend questions). If their message is broader skepticism\\n(\\\"what's the angle?\\\"), use Too Good to Be True and fold the guarantee sentence\\nin instead.\\n\\n---\\n\\n## Example 14 - LEGACY: bare \\\"Yes\\\" on an UNTIL WE DO thread (Sam-approved 2026-07-15)\\n\\nOlder threads only - the first email promised to keep running until the\\nnumbers hit (no reimbursement wording existed then).\\n\\nCold email (for context): \\\"...if we could bring your practice 5-10 implant\\ncases in 2 weeks, and you only cover Meta's ad budget, would you be\\ninterested? ... p.s. If we don't hit those numbers, we'll keep it running on\\nour dime until we do\\\"\\n\\nProspect: \\\"Yes\\\"\\n\\nSam's approved response - Tell Me More plus ONE extra line for the clause:\\n\\n> Hey, of course.\\n>\\n> We're Lumina and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first 2 weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely, so you only cover the ad costs paid to Meta.\\n>\\n> And yep - if we don't bring in at least 5 implant cases, we keep running your campaign free of charge until we do. Ad spend included.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: \\\"at least 5\\\" is the LOWER end of the promised 5-10. \\\"free of charge\\\"\\ninstead of the email's \\\"on our dime\\\" - clarity beats verbatim echo (replies\\naren't bound by cold-email spam-word constraints). ONE line, same cadence -\\nnever longer. The fee-waiver copy keeps the old \\\"you only cover the ad costs\\\"\\nclause here because that's what THIS thread's first email promised - never the\\nreimburse sentence. On a BASE thread (no guarantee at all) neither line exists.\\n\\n---\\n\\n## Example 15 - LEGACY: they ask about the Until We Do clause directly (Sam-approved 2026-07-15)\\n\\nProspect: \\\"Wait - you'd really keep running it free until we get those\\nnumbers? How does that work?\\\"\\n\\nSam's approved response:\\n\\n> Hey, totally fair question.\\n>\\n> Yep - exactly what it sounds like. If the first 2 weeks don't get you those 5-10 implant cases, we keep the campaign running on our dime until you have them. That covers the ad spend too, so your total cost stays capped at the first 2 weeks of ad budget.\\n>\\n> I'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n>\\n> For us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: the capped-cost sentence is the strongest line - it turns the promise\\ninto a concrete risk ceiling. No flagship pricing unless they also asked\\nabout price. UNTIL WE DO threads only - on current (reimburse) threads the\\ndedicated Reimbursement template applies instead (Example 13).\\n\\n---\\n\\n## Corrections log (Sam's feedback on AI drafts - newest first)\\n\\n- 2026-08-05: NEW APPOINTMENT/DEPOSIT COPIES launched (5 copies; this file +\\n templates.md + facts.md + voice.md). (1) Guarantee sentence is now\\n \\\"...If we don't get at least {LOWER_BOUND} {TREATMENT_TYPE} patients\\n showing up to their appointments, we reimburse your ad spend.\\\" on ALL\\n current-era threads (2026-07-16-era threads keep their \\\"cases\\\" sentence;\\n legacy variants unchanged). (2) {PRACTICE_TYPE} one-liner: \\\"we help\\n [family dental practices / solo practices / ...] build a predictable\\n pipeline of high-value treatments\\\" - pluralize the last word; fallback\\n \\\"dental clinics\\\". (3) TREATMENT NARROWING: when they name their treatment,\\n every mention narrows to it. (4) TMM bullet 1 + Process bullet 3 are now\\n the call-and-transfer truth (\\\"We call new leads within 60 seconds and\\n transfer them to your front desk to get booked in\\\") - the old\\n \\\"call and follow up on your behalf\\\" / \\\"lead management system\\\" claims are\\n DEAD. Never say \\\"AI\\\". (5) Ads count is 10-15 (was 25-30). (6) Ad budget is\\n {CURRENCY}500 for the two weeks with a FLAT 3-shows guarantee in ad-spend\\n answers (never {LOWER_BOUND} there; $750/5-patients is DEAD). (7) Booking\\n deposits: ask-only facts in facts.md (under $50, practice keeps it,\\n flexible - never argued over email). (8) PMoC acks RETIRED - the agent\\n sends nothing on contact-preference replies; Sam is alerted and reaches\\n out himself. (9) Current copies promise NO duration - say \\\"two weeks\\\".\\n- 2026-07-16: REIMBURSE AD SPEND campaigns launched (new copy overhaul, this\\n file + templates.md + facts.md). (1) The fee-waiver block everywhere is now\\n \\\"...we waive our service fee entirely. If we don't bring in at least\\n {LOWER_BOUND} {TREATMENT_TYPE} cases, we reimburse your ad spend.\\\" - the\\n guarantee sentence mirrors the first email's variant (reimburse / legacy\\n until-we-do line / nothing on legacy base threads). (2) Tell Me More and\\n Company Details SPLIT: website line only in Company Details and when they\\n ask for it. (3) One-liner is now \\\"a predictable pipeline of high-value\\n treatments\\\". (4) Ad budget is {CURRENCY}750 (was 500), 5 patients give or\\n take, still ask-only; reimbursement rate ({CURRENCY}150/patient) only for\\n direct mechanics questions. (5) New Process bullets ({TREATMENT_TYPE} ads,\\n conversion-optimized landing page, 60-second front-desk connect). (6) The\\n phone-ask line (\\\"And what's the best number to reach you at?\\\") is RETIRED -\\n dentists don't use phone numbers much; never include it even if an earlier\\n message in the thread did.\\n- 2026-07-15 (night): UNTIL WE DO variant launched (see Examples 14-15 and\\n templates.md rules). Detect from the first email's P.S.; when present, ONE\\n casual clause line after the fee-waiver copy; direct questions get the\\n dedicated template with the capped-cost sentence. NEVER on base threads.\\n Replies may clarify wording vs the spintax (no spam-word constraints) but\\n mirror numbers/treatment/duration exactly.\\n- 2026-07-15: FRICTION PASS. (1) \\\"15-minute\\\" is DEAD everywhere - it's just \\\"a\\n quick call\\\" / \\\"During the call, I'll...\\\" / \\\"a quick video call\\\". (2) Ad spend\\n is ASK-ONLY: never volunteered in fee-waiver copy or price answers; the\\n Ad Budget template (explicit ad-spend questions only) keeps it.\\n- 2026-07-14 (late): WEBSITE REMOVED FROM SIGNATURES (tentative - Sam is\\n reworking the site). Signature is now just persona name + \\\"SDR, Lumina\\\".\\n Website mentions in template BODIES stay - now Company Details only (2026-07-16).\\n- 2026-07-14: SHORTENED TEMPLATES rewrite. Replies were too long - the\\n \\\"educate them\\\" doctrine is dead; short, simple sentences win. Bare-\\\"Yes\\\"\\n replies get Tell Me More. The \\\"not a sales call\\\" P.S. is RETIRED - only the\\n booked-invite and first-times P.S. lines survive. PMoC acks name founder\\n Sam + 925-730-9373 and mirror stated timing (see Example 7). New templates:\\n Process, Too Good to Be True.\\n- Still in force from earlier corrections: pricing is $497/month or $2,197 for\\n six months upfront (25% off - about 6 weeks free) - any older number ($749,\\n $1,797, \\\"three months\\\") is DEAD. Times are real (verified against the live\\n calendar) - never invent one. Signature block on every reply. \\\"thanks for\\n getting back to me\\\" is a fine greeting. Sign-off \\\"Thanks,\\\" (\\\"Take care,\\\" on\\n reschedule/cancel). Back-and-forth replies go short and loose instead of\\n re-running a full template.\\n\\n---\\n\\n## 2026-07-18 corrections (from the week's real sends - Sam-graded)\\n\\n### Example 16 - \\\"Ok let set up a meeting\\\" is ASK FOR CALL, not a pitch moment\\n\\nThey replied (from info@bocadentalgroup.com): \\\"Ok let set up a meeting Sent from my iPhone\\\"\\n\\nWRONG (what shipped): opened with the full CTA pitch line (\\\"Let's do a quick\\ncall so we can give you a rundown...\\\") before the times. They already asked\\nfor the meeting - re-pitching it is noise.\\n\\nRIGHT:\\n\\n> Hey, great to hear!\\n>\\n> How about Monday at 10am or 1:30pm ET, or Tuesday at 12pm or 4pm ET?\\n>\\n> Or if it's easier, here's our Calendly: [CALENDAR]\\n>\\n> Thanks,\\n> Bethany\\n> SDR, Lumina\\n>\\n> P.S. If those times don't work, feel free to send over other openings and I'll book you in directly\\n\\nSame grading applies to \\\"can we schedule a call to further discuss\\\" (Example\\nfrom the Mariana thread) - short greeting (\\\"Hey Mariana - of course!\\\"),\\ntimes, Calendly line, new P.S. Nothing else.\\n\\n### Example 17 - blend TMM + Too Good to Be True; no price on skepticism\\n\\nThey replied: \\\"Let's do it! I'd love to learn more about your offer. I hope\\nto hear the catch too - Thanks.\\\"\\n\\nWRONG (what shipped): Company-Details-style body plus \\\"the only thing we ask\\nfor is a short testimonial video\\\" as the catch answer, wrong greeting.\\n\\nRIGHT (relevant parts of Tell Me More, then relevant parts of Too Good to Be\\nTrue, NO price):\\n\\n> Hey, of course!\\n>\\n> We're Lumina and we help dental clinics build a predictable pipeline of high-value treatments.\\n>\\n> For your first 2 weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely, so you only cover the ad costs paid to Facebook.\\n>\\n> There's no catch :)\\n>\\n> I'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n>\\n> For us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n>\\n> [CTA BLOCK]\\n\\nNote: \\\"Of course!\\\" (not \\\"great to hear!\\\") is the natural response to \\\"I'd\\nlove to learn more about your offer.\\\"\\n\\n### Example 18 - \\\"I'm out next week, book me the week after\\\" gets REAL times\\n\\nThey replied (deep in thread, from Angela): \\\"I am out of the office all next\\nweek and will be unable to meet. Please let me know when your calendar is\\nextended into the following week.\\\"\\n\\nWRONG (what shipped): \\\"Since I don't have any open slots to offer just yet\\nfor the week after, go ahead and grab whatever works best for you here once\\nit's open: [link]\\\" - punted to a link instead of booking her.\\n\\nRIGHT (the system now fetches that week's availability - propose it; note\\nthe deep-thread phrasing, looser than the standard shape):\\n\\n> Hey Angela,\\n>\\n> No worries at all - enjoy the time off. I can book you in manually for the following week. Would {DAY 1} at {TIME 1} or {TIME 2} work or {DAY 2} at {TIME 3} or {TIME 4}?\\n>\\n> Feel free to suggest other times that work for you.\\n>\\n> Take care,\\n> Kimberly\\n> SDR, Lumina\\n\\n### Example 19 - bare \\\"Yes\\\" on a NEW-ERA thread (2026-08-05 appointment/deposit copies)\\n\\nCold email (for context): \\\"Bomi - if we could book your practice 4-8 implant,\\ninvisalign or bonding appointments with patients that put down a booking\\ndeposit, would you be interested? We're opening 5 spots for a program where\\nyou don't pay us for your first campaign.\\\" (lead's practice type: family\\ndental practice)\\n\\nProspect: \\\"Yes\\\"\\n\\nResponse - Tell Me More with the new-era pieces:\\n\\n> Hey, of course.\\n>\\n> We're Lumina and we help family dental practices build a predictable pipeline of high-value treatments.\\n>\\n> Three quick bullets\\n>\\n> - We call new leads within 60 seconds and transfer them to your front desk to get booked in\\n> - Your campaign goes live within 72 hours of onboarding\\n> - After the pilot, our service fee is $497 a month\\n>\\n> For your first two weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely. If we don't get at least 4 implant, invisalign or bonding patients showing up to their appointments, we reimburse your ad spend.\\n>\\n> [CTA + times + calendar]\\n>\\n> Thanks,\\n> {SENDER_FIRST_NAME}\\n> SDR, Lumina\\n\\nNotes: the shows-guarantee sentence goes on EVERY current-era thread - even\\nthe copies whose body carried no explicit reimburse P.S. (Sam's 2026-08-05\\nruling). \\\"at least 4\\\" = the LOWER end of the promised 4-8. Treatments are\\nmirrored exactly as the first email joined them; the moment they name one\\n(\\\"we mainly want Invisalign\\\") every treatment mention narrows to it. \\\"two\\nweeks\\\" is stated even though the copy promised no duration - it's the real\\npilot length and shorter than what they'd assume. The one-liner uses the\\nlead's practice type, pluralized.\\n\\n### Doctrine updates that ride with these examples\\n\\n- New universal first-times P.S.: \\\"If those times don't work, feel free to\\n send over other openings and I'll book you in directly.\\\"\\n- Booking confirmations are the two-variant template in templates.md (office\\n email check, Meet link, two Add-to-Calendar links, gated reminders-phone\\n ask). The old one-liner + \\\"better email\\\" P.S. is dead.\\n- CTA phone line (\\\"And just for convenience, what's a good number to reach\\n you directly?\\\") - first proposing reply only, system-gated, never once a\\n number is known. Max two asks per lead ever (CTA + booking confirmation).\\n- Tell Me More now carries the \\\"Three quick bullets\\\" block - the only place\\n price is volunteered outside the Price template.\\n\\n\\n\\n# Mechanism Deep Context (from Sam's founder videos)\\n\\nDistilled from the three YouTube walkthroughs. Use ONLY when a prospect asks a\\nspecific how-does-it-work question that facts.md doesn't cover, and translate to\\nplain, first-read-understandable language - never technical, never jargon. Items\\nmarked \\\"(general context)\\\" are instructional content, not necessarily what we do for\\nevery client. If anything here conflicts with facts.md, facts.md wins (e.g. the\\nvideos mention a \\\"free advisory call\\\" and typical ad spend of $1,500-3,000/mo; the\\ncurrent offer is the free first campaign, then $497/mo or $2,197 for six months\\nupfront (25% off), with $750 of ad spend recommended during the free\\ncampaign).\\nNEVER dump this wholesale into a reply - pull the one or two ideas that\\nanswer the question, simply. When in doubt, share the right video link in the\\nbody instead of explaining.\\n(2026-08-05: facts.md now says 10-15 ads and the we-call-then-transfer lead\\nflow - the video summaries below predate both; facts.md wins.)\\n\\n---\\n\\n## Video 1: How the lead-generation system works (youtu.be/WlAQcaskL28)\\n\\n### The core problem we solve\\nDentists' number-one fear with paid ads is being flooded with junk leads that waste the front desk's and treatment coordinator's time. Our system is built around deliberately making it slightly harder to book - the right amount of \\\"hoops\\\" filters out tire-kickers and price-shoppers while keeping real patients. Too few hoops brings in junk; too many scares off good patients. We aim for the middle.\\n\\n### The mechanism, step by step\\n1. **Ad on Facebook/Instagram** promotes a specific treatment offer (e.g., implants, Invisalign).\\n2. **A dedicated offer page** (not the practice's homepage) built to get one action: request a consult. It includes a short video explaining the treatment, **the price shown up front** (for a $6,000 Invisalign case, price is the top reason deals die - showing it early makes non-buyers remove themselves), financing options (to bring back good patients who just need payment plans), and a clear treatment timeline so people know exactly what they're signing up for.\\n3. **A qualifying questionnaire before anyone can see the calendar.** The booking calendar is treated as sacred - only people who pass budget questions (\\\"Invisalign runs $4,000-$8,000 - does that work?\\\") and health-readiness questions (gum disease, loose teeth, last checkup, how soon they want to start) ever get to book. People who don't qualify are routed to a follow-up sequence - often offered a checkup or a lower-cost treatment first, then brought back to the bigger treatment later.\\n4. **A small refundable deposit to lock the appointment.** Serious patients don't mind (they get it back); non-serious ones won't put a card down. Result: far fewer no-shows and higher treatment acceptance.\\n5. **Teaching Facebook who your best patients are.** Tracking code on the page reports back only *qualified* actions - a qualified lead, a booked consult, a started treatment. Facebook (which holds ~52,000 data points per person) then goes and finds more people like your paying patients, not like your junk leads. When your coordinator moves a patient to \\\"showed up qualified\\\" or \\\"started treatment\\\" in the pipeline, that gets sent back to Facebook automatically, so targeting keeps improving.\\n6. **Fast follow-up:** automated texting contacts new leads within about 60 seconds, while they're still warm.\\n\\n### What we specifically do for clients\\n- We never change a funnel until it has run **at least 3 days and $300 of spend** - enough data for an informed decision, not guesswork.\\n- For new clients (or polluted past data), we start by having Facebook find leads, then step down to optimizing for booked consults, then for started treatments, as data accumulates. If ad costs rise later, we step back up to widen the pool.\\n- We diagnose *where* to add or remove hoops using benchmarks: **1.5-3%** of ad viewers clicking, **5-7%** of visitors booking, **70-80%** showing up, **~30%** starting treatment. Whichever number lags tells us what to fix.\\n\\n### Objections this answers\\n- **\\\"Leads will be junk.\\\"** The whole system exists to pre-filter them, medically and financially.\\n- **\\\"Referral leads are better.\\\"** (general context) True - ads reach strangers, so ad leads will always feel colder; judge them against benchmarks, not word-of-mouth.\\n- **\\\"Low booking or close rates mean bad leads.\\\"** Often it's slow follow-up or an untrained coordinator - we address both with automation, scripts, and training.\\n\\n---\\n\\n## Video 2: How leads become booked patients (youtu.be/PvFJuJSNZiI)\\n\\n### The core idea: speed to lead\\nThe single biggest factor in turning a lead into a booked patient is how fast someone contacts them. A Harvard Business Review study found that responding within 5 minutes makes you **21x more likely** to convert a lead into a patient versus responding within 30 minutes. Yet the average dental practice takes **47 hours** to respond to new inquiries.\\n\\n(general context) Most practices lose patients not because the leads are bad, but because leads aren't followed up fast enough or persistently enough. Each lead costs real money (roughly a $100 bill) and can become a patient worth $10,000+ over their lifetime.\\n\\n### Why practices think this doesn't apply to them (common objection)\\nDentists often say \\\"my front desk is always on the phone.\\\" Two problems: (1) calls cluster at peak times - start of day, lunch, after hours - so more than one comes in at once and some get missed; (2) most practices have no hard data on missed calls or response time, so they go by feel.\\n\\n### The follow-up system\\nThree ways a lead can book, all offered right after they submit their info:\\n1. **We call them** - an outbound call a few minutes after they opt in, from a clear script: answer basic questions about candidacy, treatment, and pricing, then steer toward booking.\\n2. **They call the office** - handled the same way, plus capturing contact info for follow-up.\\n3. **They book themselves online** - self-scheduled patients have a **41% lower no-show rate** because they chose a time that truly works for them.\\n\\nIf a lead doesn't answer: same day, call twice, leave a voicemail, and text several times spaced hours apart. Days 1-2: two calls and two texts daily. Days 3-6: two calls and one text daily. After a week: a series of helpful emails (about the team, promotions, procedures) that builds enough trust to book later.\\n\\n**Personal touches that lift booking rates:** tailor messages to what the lead said on the intake questions (e.g., financing needs get monthly payment plan info), and send a short video of the doctor welcoming them - putting a face to the practice builds trust fast.\\n\\n### Scheduling practices that raise show rates\\n(general context) Keep plenty of open slots (availability directly drives bookings; aim for weekday 9am-3pm coverage); keep consults to ~45 minutes or less; 15-minute booking intervals; book no more than ~3 business days out and pull appointments earlier when possible - the further out, the more no-shows.\\n\\n### What Lumina actually does for clients\\nFor every practice we work with, we implement all these automations and hand over all the scripts and standard operating procedures - the whole follow-up and booking system is done for the practice, alongside running their ads.\\n\\n---\\n\\n## Video 3: The ads themselves (youtu.be/X_Cc8Iy3Hgg)\\n\\n### The short answer\\nFor each practice we build and run **25-30 genuinely different ads at all times** - not one ad copied 30 times with small tweaks. Each ad combines three ingredients: an **offer** (what brings a new patient in, like a new-patient special or a free consultation for a big-ticket treatment), an **angle** (the specific reason that patient cares - comfort, speed, price, trust), and a **format** (patient testimonial, before-and-after, the dentist talking to camera, a photo, or footage of the office and chairside care).\\n\\n### Why so many ads?\\n(general context) In 2025, Facebook/Instagram got dramatically better at matching each person with an ad they'll respond to (\\\"Meta Andromeda\\\" update) - and dental ad costs jumped **97% in one year**, the steepest rise of any industry on the platform. Practices still running two or three ads now pay far more per patient.\\n\\n(general context) Even for the same treatment, patients want different things: one Invisalign prospect is weighing clear aligners against metal braces (needs to hear discreet, faster, comfortable); another is price-shopping (wants the discount and freebies). One generic ad can't speak to both. Without variety you get \\\"ad fatigue\\\" - paying more per patient and getting fewer of them.\\n\\n### How we build them\\n- **Start broad:** launch a mix of offers, angles, and formats.\\n- **Find winners:** when one element clearly works, build ~30 fresh variations around it - new angles (credentials, friendly team, treatment timelines, what the first visit is like) and new formats.\\n- **Keep winners running, keep testing:** a proven winner gets its own budget and often keeps producing for months or years.\\n- **Give every ad a fair shot:** at least 3 days and ~$300 of spend before judging; we test big, meaningful differences, not tiny word tweaks.\\n\\n### Targeting by treatment\\nAds are built per treatment (Invisalign, veneers, implants, whitening, etc.) and, within each treatment, per patient motivation - the price-shopper, the anxious patient, the appearance-focused patient - so the ad each person sees feels written for them. That personal fit is how a smaller practice competes against big spenders in crowded markets.\\n\\n### Common questions this answers\\n- *\\\"Do you just run one ad for my practice?\\\"* No - 25-30 unique ones, refreshed as winners emerge.\\n- *\\\"Won't people get sick of seeing my ads?\\\"* That's exactly what the volume and variety prevent.\\n\\n\";\n","createdAt",["D",1786127124429],"branchVersion","branchId","54481d2d-a3f3-4dfa-ab51-1abb2a0d7de9","id","e3dafc71-738c-40fe-ab70-9c5ae76a5d02","bison-reply-agent",["D",1783490909384],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","imageUrl","352b7b81-2d40-4ef8-babc-d1dd56767822","pinnedFileId","pinPreview","pinMigrated","screenshotVersion","likeCount",0,"orgId","019f404e-e73e-709b-8be4-84c9207a4140","tursoName","tursoHostname","tursoMeasuredAt","pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_94":93,"_101":96},"org",{"_94":93,"_95":96,"_97":96,"_18":54,"_98":99,"_100":25},"branch",{"_13":83,"_32":84,"_43":42,"_85":44,"_86":-5,"_87":-5,"_88":-5,"_89":-5,"_90":-5,"_91":45,"_92":93},"isOwned",97,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","samy","handle","displayName","","avatar","tier","free","canRestrictVals","profileImageUrl","branchName","adoptOrgId",{"_43":44,"_32":45,"_38":276,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":-5,"_53":-5,"_18":54,"_55":-5,"_56":25,"_57":-5,"_58":-5,"_59":60,"_61":62,"_63":-5,"_64":-5,"_65":-5,"_66":-5,"_67":68,"_69":68,"_70":71,"_72":71,"_73":71,"_74":68,"_75":68,"_76":277,"_78":278,"_80":279,"_82":25},"tree",[231,232,233,234,235,236,237,238,239,240,241],"files",[231,232,233,234,235,236,237,238,239,240,241],"counts",{"_220":60,"_221":173,"_222":60,"_223":224,"_225":226},"contents",[124,125,126,127,128,129,130,131,132,133,134],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles",true,"grants","orgs","customDomains",{},{"_43":213,"_32":214,"_30":215,"_138":-5,"_41":42,"_40":216,"_140":25,"_38":217,"_13":173,"_33":34,"_35":218,"_24":25,"_145":-5,"_27":219},{"_43":206,"_32":207,"_30":208,"_138":-5,"_41":42,"_40":209,"_140":119,"_38":210,"_13":173,"_33":34,"_35":211,"_24":25,"_145":-5,"_27":212},{"_43":199,"_32":200,"_30":201,"_138":-5,"_41":42,"_40":202,"_140":119,"_38":203,"_13":173,"_33":34,"_35":204,"_24":25,"_145":-5,"_27":205},{"_43":191,"_32":192,"_30":193,"_138":-5,"_41":42,"_40":194,"_140":25,"_38":195,"_13":196,"_33":34,"_35":197,"_24":25,"_145":-5,"_27":198},{"_43":183,"_32":184,"_30":185,"_138":-5,"_41":42,"_40":186,"_140":119,"_38":187,"_13":188,"_33":34,"_35":189,"_24":25,"_145":-5,"_27":190},{"_43":175,"_32":176,"_30":177,"_138":-5,"_41":42,"_40":178,"_140":119,"_38":179,"_13":180,"_33":34,"_35":181,"_24":25,"_145":-5,"_27":182},{"_43":171,"_32":28,"_30":31,"_138":-5,"_41":42,"_40":14,"_140":119,"_38":172,"_13":173,"_33":34,"_35":37,"_24":25,"_145":-5,"_27":174},{"_43":163,"_32":164,"_30":165,"_138":-5,"_41":42,"_40":166,"_140":119,"_38":167,"_13":60,"_33":168,"_35":169,"_24":25,"_145":-5,"_27":170},{"_43":155,"_32":156,"_30":157,"_138":-5,"_41":42,"_40":158,"_140":119,"_38":159,"_13":160,"_33":34,"_35":161,"_24":25,"_145":-5,"_27":162},{"_43":147,"_32":148,"_30":149,"_138":-5,"_41":42,"_40":150,"_140":119,"_38":151,"_13":152,"_33":143,"_35":153,"_24":25,"_145":-5,"_27":154},{"_43":135,"_32":136,"_30":137,"_138":-5,"_41":42,"_40":139,"_140":25,"_38":141,"_13":142,"_33":143,"_35":144,"_24":25,"_145":-5,"_27":146},"6f5a76dd-e0ae-4ab7-80ae-74bb38e33137","webhook.ts","a767a496-7a93-11f1-ad89-1607ee4eb77e","parentId",86,"latest",["D",1784845729511],32,"http","// Email Bison \"reply received\" -> analyze with Claude -> schedule/book via GHL\n// -> draft + send reply / alert Sam.\n// Two-phase model flow: call 1 classifies and extracts scheduling facts (no\n// drafting); code then resolves the timezone, checks real calendar availability,\n// books when appropriate, and computes a verified time menu; call 2 drafts the\n// reply and may only use times from that menu.\n// The playbook (voice, templates, facts, swipefile, mechanism) is compiled from\n// skill/*.md by scripts/deploy.py into playbook.ts.\n// DRY_RUN=true: classify + compute + draft + log + [DRY RUN] alerts, but no\n// Bison sends and no GHL bookings. Set DRY_RUN=false to go live.\nimport { sqlite } from \"https://esm.town/v/std/sqlite\";\nimport { email } from \"https://esm.town/v/std/email\";\nimport { PLAYBOOK_A } from \"./playbook.ts\";\nimport { PLAYBOOK_B } from \"./playbook2.ts\";\nexport const PLAYBOOK = PLAYBOOK_A + \"\\n\\n\" + PLAYBOOK_B;\nimport {\n bookWithRecovery,\n buildCalendarLinks,\n calendarConfigured,\n cancelAppointment,\n fetchSlotsForDate,\n getMeetLink,\n getScheduledEvent,\n getUpcomingAppointment,\n rescheduleAppointment,\n} from \"./calendly.ts\";\nimport { legacyGhlAppointment } from \"./ghl.ts\";\nimport { buildCallPrep } from \"./callprep.ts\";\nimport {\n checkDueEmails,\n type ContactEmails,\n contactEmailsFor,\n contactName,\n ensureContact,\n leadVar,\n lookupContact,\n markBooked,\n mutedSince,\n normalizePhone,\n prospectPhoneKnown,\n reconcile,\n stripHtml,\n syncPositiveReply,\n websiteIntel,\n} from \"./crm.ts\";\nimport {\n counterProposal,\n findExactSlot,\n loadSlots,\n localNow,\n menuForPrompt,\n minutesToLabel,\n type ProposalMenu,\n slotMinutes,\n standardProposal,\n tzAbbreviation,\n weekdayOf,\n windowProposal,\n} from \"./scheduling.ts\";\n\nconst MAX_AI_REPLIES_PER_LEAD = 7;\nconst MODEL = \"claude-sonnet-5\";\nconst BOOKING_LINK = \"https://calendly.com/luminamg/pilot\";\nconst FALLBACK_TZ = \"America/Los_Angeles\";\n\n// ---------- Claude plumbing ----------\n\nasync function claudeTool(\n system: string,\n messages: { role: string; content: string }[],\n tool: Record,\n maxTokens = 1500,\n): Promise {\n const resp = await fetch(\"https://api.anthropic.com/v1/messages\", {\n method: \"POST\",\n headers: {\n \"x-api-key\": Deno.env.get(\"ANTHROPIC_API_KEY\") ?? \"\",\n \"anthropic-version\": \"2023-06-01\",\n \"content-type\": \"application/json\",\n },\n body: JSON.stringify({\n model: MODEL,\n max_tokens: maxTokens,\n system,\n tools: [tool],\n tool_choice: { type: \"tool\", name: (tool as any).name },\n messages,\n }),\n signal: AbortSignal.timeout(45_000),\n });\n if (!resp.ok) throw new Error(`anthropic ${resp.status}: ${await resp.text()}`);\n const data = await resp.json();\n const toolUse = (data.content ?? []).find((b: { type: string }) => b.type === \"tool_use\");\n if (!toolUse) throw new Error(\"no tool_use block in Claude response\");\n return toolUse.input;\n}\n\n// ---------- phase 1: classify + extract ----------\n\nimport { ANALYSIS_PROMPT, ANALYSIS_TOOL, type Analysis } from \"./analysis.ts\";\n\n// ---------- phase 2: draft ----------\n\nexport const DRAFT_TOOL = {\n name: \"submit_draft\",\n description: \"Submit the drafted reply.\",\n input_schema: {\n type: \"object\",\n properties: {\n reply_text: { type: \"string\", description: \"Plain text with real line breaks.\" },\n },\n required: [\"reply_text\"],\n },\n};\n\nexport const DRAFT_PROMPT_HEAD = `\nYou draft replies to cold-email responses for Lumina (Facebook ads for dental\nclinics). You write AS the sender persona named in the input (the pen-name the\nprospect knows) - never any other name.\n\nRULES:\n- Follow the playbook below: voice.md governs tone, templates.md gives the exact\n shapes (blend when the reply spans several - swipefile.md shows how), facts.md\n is the only source of claims, MIRROR RULE for the free-campaign framing (the\n first email we sent in the thread is the source of truth for duration and any\n case-count promise - never invent one). Guarantee language mirrors the OFFER\n VARIANT named in the input: the reimburse sentence only on REIMBURSE threads,\n the \"until we do\" line only on UNTIL WE DO threads, and neither - in any\n form - on BASE threads.\n- READ THE ENTIRE THREAD FIRST. The templates fit ~95% of replies; when a thread\n has gone somewhere no template covers, compose naturally from voice.md and\n facts.md instead of forcing a template.\n- NEVER produce a reply that mirrors the structure of one of our previous\n messages in this thread. If an earlier message already used a template's\n skeleton (same greeting shape, same paragraph order, same CTA phrasing), vary\n this one - different opener, reordered or merged paragraphs, tighter length -\n while keeping the voice natural and non-verbose. Identical-looking emails\n read as a bot.\n- Greeting: \"Hey,\" plus their name when known (\"Hey Monica,\" / \"Hey Dr. Song,\"),\n then a short remark that fits the thread.\n- TIMES: you may ONLY mention meeting times that appear in the VERIFIED TIMES\n section of the input - never invent, shift, or add times. Write them with the\n weekday name and timezone label exactly as given (e.g. \"Friday at 10am ET\").\n Never say \"today\"/\"tomorrow\" - use the weekday name.\n- reply_text is PLAIN TEXT with real line breaks: blank line between paragraphs,\n the sign-off (\"Thanks,\" normally; \"Take care,\" on the reschedule/cancel\n templates) on its own line, persona first name on the next line, then exactly\n \"SDR, Lumina\" on the next. NO website in the signature (body mentions like\n \"You can find us at luminamg.com.\" are fine where a template has them). The\n signature goes on EVERY reply (short acks and back-and-forth follow-ups\n included), before any P.S. line. No HTML, no markdown, no em-dashes. Links\n are bare URLs.\n- The input specifies DRAFT MODE (which template shape applies), the P.S. line\n to use (if any), and whether this is a back-and-forth follow-up. In\n back-and-forth mode: drop the full template structure - short, looser,\n one-liner-ish per voice.md, but keep the calendar link when offering times.\n\nReturn via submit_draft.\n\n======================= PLAYBOOK =======================\n` + PLAYBOOK;\n\n// ---------- draft sanitation ----------\n\nconst BANNED_SNIPPETS = [\n \"i hope this finds you well\", \"furthermore,\", \"moreover,\", \"click here\",\n \"as an ai\", \"i'm an ai\", \"language model\",\n];\n\n// Old/alternate calendar URLs that may leak into drafts from earlier thread\n// messages (old cold emails carry the GHL links forever) - normalized to the\n// canonical Calendly link.\nconst STALE_LINKS = [\n \"https://links.luminamg.com/widget/bookings/lumina-intro-call\",\n \"https://links.luminamg.com/widget/booking/1nbt58J9c0CBZFyqCBQk\",\n];\n\nexport function mechanicalFixes(text: string): string {\n let out = text\n .replace(/[—–]/g, \" - \")\n .replace(/<[^>]+>/g, \"\")\n .replace(/\\r\\n/g, \"\\n\")\n .trim();\n for (const stale of STALE_LINKS) out = out.replaceAll(stale, BOOKING_LINK);\n return out;\n}\n\n// Every reply ends with this directly under the persona's first name (before\n// any P.S. line). Website removed from the signature 2026-07-14 (tentative -\n// Sam reworking the site); isSigLine still strips stray \"luminamg.com\" lines\n// the model may copy from older messages in the thread.\nconst SIGNATURE = \"SDR, Lumina\";\n\nexport function ensureSignature(text: string, senderName: string, withCompany = true): string {\n const out = text.replace(/\\s+$/, \"\");\n if (!out) return out;\n const isSigLine = (l: string) =>\n /^\\s*SDR,?\\s*(?:at\\s+|@\\s*)?Lumina\\s*$/i.test(l) ||\n /^\\s*(?:https?:\\/\\/)?(?:www\\.)?luminamg\\.com\\/?\\s*$/i.test(l);\n // Drop whatever signature lines the model wrote (partial or misplaced),\n // then re-insert the canonical block right under the sign-off name line.\n const lines = out.split(\"\\n\").filter((l) => !isSigLine(l));\n const name = senderName.trim().toLowerCase();\n let nameIdx = -1;\n for (let i = lines.length - 1; i >= 0; i--) {\n if (lines[i].trim().toLowerCase().replace(/[,.]\\s*$/, \"\") === name) {\n nameIdx = i;\n break;\n }\n }\n if (nameIdx === -1) {\n // No sign-off at all (e.g. a short ack) - add one with the signature.\n return lines.join(\"\\n\").replace(/\\s+$/, \"\") +\n (withCompany ? `\\n\\nThanks,\\n${senderName}\\n${SIGNATURE}` : `\\n\\nThanks,\\n${senderName}`);\n }\n if (withCompany) lines.splice(nameIdx + 1, 0, SIGNATURE);\n // Removing misplaced sig lines can leave stray blanks - tidy without touching content.\n return lines.join(\"\\n\").replace(/\\n{3,}/g, \"\\n\\n\").replace(/\\s+$/, \"\");\n}\n\nfunction normTimeToken(tok: string): string {\n return tok.toLowerCase().replace(/\\s+/g, \"\").replace(\":00\", \"\");\n}\n\nexport function violations(\n text: string,\n requireBooking: boolean,\n allowedTimes: Set | null,\n guarantee: \"none\" | \"until_we_do\" | \"reimburse\",\n phoneAsks?: { cta: boolean; confirm: boolean },\n): string[] {\n const problems: string[] = [];\n const lower = text.toLowerCase();\n for (const p of BANNED_SNIPPETS) {\n if (lower.includes(p)) problems.push(`banned phrase: \"${p}\"`);\n }\n if (requireBooking && !lower.includes(\"calendly.com/luminamg/pilot\")) problems.push(\"booking link (bare URL) missing\");\n // Old phone-ask line RETIRED forever (2026-07-16); the two 2026-07-18 lines\n // are allowed ONLY on the reply their gate opened (max 2 asks per lead,\n // zero once we hold a number).\n if (lower.includes(\"best number to reach you\")) {\n problems.push(\"the old phone-ask line is retired - remove it\");\n }\n if (!phoneAsks?.cta && lower.includes(\"number to reach you directly\")) {\n problems.push(\"the CTA phone ask is not allowed on this reply - remove it\");\n }\n if (!phoneAsks?.confirm && lower.includes(\"number for reminders\")) {\n problems.push(\"the reminders phone ask is not allowed on this reply - remove it\");\n }\n // Guarantee language promises REAL MONEY - each family may only appear on\n // threads whose first email carried that exact promise (model-detected).\n // Prompt-level gating alone proved insufficient (2026-07-15 test), hence\n // these hard rails.\n if (guarantee !== \"until_we_do\" && /until (we|you|they) (do|get|have|hit)|free of charge until|on our dime|ad spend included|keep (it|the campaign|your campaign) running|keep running (it|the campaign|your campaign)/i.test(text)) {\n problems.push(\"this thread's first email made NO 'until we do' promise - remove every trace of a keep-running/free-extension promise\");\n }\n // \"we cover/pay back your ad spend\" is a reimbursement promise; \"you (only)\n // cover the ad costs\" is the prospect paying - legitimate legacy copy.\n if (guarantee !== \"reimburse\" && /reimburs|refund|money.back|we('ll| will|'d)? (cover|pay (you )?back for) (your|the) ad (spend|budget|costs)/i.test(text)) {\n problems.push(\"this thread's first email made NO reimbursement promise - remove every trace of reimburse/refund/money-back language\");\n }\n if (allowedTimes) {\n for (const tok of text.match(/\\b\\d{1,2}(?::\\d{2})?\\s*(?:am|pm)\\b/gi) ?? []) {\n if (!allowedTimes.has(normTimeToken(tok))) {\n problems.push(`time \"${tok}\" is not in the verified times list`);\n }\n }\n }\n return problems;\n}\n\n// ---------- Bison API ----------\n\nfunction bisonHeaders() {\n return {\n \"Authorization\": `Bearer ${Deno.env.get(\"BISON_API_KEY\")}`,\n \"Accept\": \"application/json\",\n \"Content-Type\": \"application/json\",\n };\n}\n\n// The first campaign email is the source of truth for what was promised (spintax).\nasync function fetchFirstSentEmail(leadId: number): Promise {\n try {\n const resp = await fetch(\n `${Deno.env.get(\"BISON_BASE_URL\")}/api/leads/${leadId}/sent-emails`,\n { headers: bisonHeaders(), signal: AbortSignal.timeout(15_000) },\n );\n if (!resp.ok) return \"\";\n const data = await resp.json();\n const items = data?.data ?? [];\n if (!items.length) return \"\";\n items.sort((a: any, b: any) =>\n String(a.sent_at ?? a.created_at ?? \"\").localeCompare(String(b.sent_at ?? b.created_at ?? \"\"))\n );\n const first = items[0];\n const body = stripHtml(String(first.email_body ?? first.body ?? \"\")).slice(0, 1500);\n return `Subject: ${first.email_subject ?? first.subject ?? \"\"}\\n${body}`;\n } catch (e) {\n console.error(\"fetchFirstSentEmail failed:\", e);\n return \"\";\n }\n}\n\n// Bison's own interested-detection misses many positives, so we push our\n// classification back: marking the reply interested flips the lead's\n// per-campaign \"interested\" flag that Sam filters on in Bison's lead data.\nasync function markReplyInterested(replyId: number) {\n const resp = await fetch(\n `${Deno.env.get(\"BISON_BASE_URL\")}/api/replies/${replyId}/mark-as-interested`,\n { method: \"PATCH\", headers: bisonHeaders(), signal: AbortSignal.timeout(10_000) },\n );\n if (!resp.ok) throw new Error(`bison mark-as-interested ${resp.status}: ${(await resp.text()).slice(0, 300)}`);\n}\n\n// Bison wraps the message in HTML, so raw newlines render as nothing, and Gmail\n// strips

margins (both verified in Gmail).
per newline is the email-native\n// way: a blank line in the draft becomes

(visible empty line), a single\n// newline (Thanks, / name) becomes one
. URLs become clickable links.\nexport function textToHtml(text: string, anchors?: Record): string {\n const escaped = text.replace(/&/g, \"&\").replace(//g, \">\")\n // Hyphen bullets render indented (Sam's style: \"-\" lists, not disc\n // bullets) - nbsp indent survives every mail client.\n .split(\"\\n\").map((l) => l.replace(/^- /, \" - \")).join(\"\\n\");\n let linked = escaped.replace(/(https?:\\/\\/[^\\s]+)/g, '$1');\n // Bare website/booking-page mentions become real links; full URLs above and\n // the preceding-char guard keep these from double-wrapping.\n linked = linked.replace(\n /(^|[\\s(])luminamg\\.com(?=$|[\\s).,!?:;])/gm,\n '$1luminamg.com',\n );\n linked = linked.replace(\n /(^|[\\s(])calendly\\.com\\/luminamg\\/pilot(?=$|[\\s).,!?:;])/gm,\n '$1calendly.com/luminamg/pilot',\n );\n // Named anchor lines (\"Add to Google Calendar\" etc.) get their URLs here -\n // the model never writes them.\n for (const [label, url] of Object.entries(anchors ?? {})) {\n linked = linked.replaceAll(label, `${label}`);\n }\n return linked.replace(/\\n/g, \"
\");\n}\n\nasync function sendBisonReply(\n replyId: number,\n message: string,\n senderEmailId?: number,\n leadEmail?: string,\n anchors?: Record,\n) {\n // reply_all picks the right sender + recipients; we also pass them explicitly\n // (duplicates are ignored by Bison) so a strict validator can never reject us.\n const payload: Record = {\n reply_all: true,\n message: textToHtml(message, anchors),\n content_type: \"html\",\n inject_previous_email_body: true,\n };\n if (senderEmailId) payload.sender_email_id = senderEmailId;\n if (leadEmail) payload.to_emails = [{ name: null, email_address: leadEmail }];\n const resp = await fetch(\n `${Deno.env.get(\"BISON_BASE_URL\")}/api/replies/${replyId}/reply`,\n {\n method: \"POST\",\n headers: bisonHeaders(),\n body: JSON.stringify(payload),\n signal: AbortSignal.timeout(30_000),\n },\n );\n const text = await resp.text();\n if (!resp.ok) throw new Error(`bison reply ${resp.status}: ${text.slice(0, 500)}`);\n return text;\n}\n\nfunction cleanSubject(s: string): string {\n return String(s ?? \"\").replace(/^\\s*((re|fwd?|fw)\\s*:\\s*)+/i, \"\").trim();\n}\n\n// ---------- storage ----------\n\nasync function ensureTables() {\n await sqlite.execute(\n `CREATE TABLE IF NOT EXISTS webhook_events (\n id INTEGER PRIMARY KEY AUTOINCREMENT,\n received_at TEXT DEFAULT (datetime('now')),\n method TEXT, body TEXT)`,\n );\n await sqlite.execute(\n `CREATE TABLE IF NOT EXISTS processed_replies (\n reply_uuid TEXT PRIMARY KEY,\n status TEXT DEFAULT 'processing',\n processed_at TEXT DEFAULT (datetime('now')))`,\n );\n await sqlite.execute(\n `CREATE TABLE IF NOT EXISTS ai_actions (\n id INTEGER PRIMARY KEY AUTOINCREMENT,\n ts TEXT DEFAULT (datetime('now')),\n lead_id INTEGER, lead_email TEXT,\n reply_id INTEGER, reply_uuid TEXT,\n category TEXT, action TEXT,\n reply_body TEXT, detail TEXT)`,\n );\n}\n\nasync function logAction(row: {\n lead_id?: number; lead_email?: string; reply_id?: number; reply_uuid?: string;\n category?: string; action: string; reply_body?: string; detail?: string;\n}) {\n await sqlite.execute({\n sql: `INSERT INTO ai_actions (lead_id, lead_email, reply_id, reply_uuid, category, action, reply_body, detail)\n VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,\n args: [\n row.lead_id ?? null, row.lead_email ?? null, row.reply_id ?? null,\n row.reply_uuid ?? null, row.category ?? null, row.action,\n row.reply_body ?? null, row.detail ?? null,\n ],\n });\n}\n\n// Alerts are best-effort: a slow/down email service must never kill reply\n// processing (it caused 60s gateway-timeout deaths in testing).\nasync function safeEmail(msg: { subject: string; text: string }): Promise {\n try {\n await Promise.race([\n email(msg),\n new Promise((_, rej) => setTimeout(() => rej(new Error(\"email timeout after 8s\")), 8000)),\n ]);\n return true;\n } catch (e) {\n console.error(\"alert email failed:\", e);\n await logAction({ action: \"alert_failed\", detail: `${msg.subject} :: ${String(e).slice(0, 300)}` }).catch(() => {});\n return false;\n }\n}\n\n// Admin endpoint (test-ghl.ts) - the mute page lives there. Links carry the\n// shared token so Sam can act straight from an alert email; alerts only go to\n// his own inbox, so this exposes nothing new.\nconst ADMIN_URL = \"https://samy--54d1c1fc7bf711f1bb541607ee4eb77e.web.val.run\";\n\nfunction adminLink(action: string, emailAddr?: string): string {\n const p = new URLSearchParams({ action, token: Deno.env.get(\"WEBHOOK_SECRET\") ?? \"\" });\n if (emailAddr) p.set(\"email\", emailAddr);\n return `${ADMIN_URL}?${p}`;\n}\n\n// Appended to lead alerts so \"I'm taking this thread over\" is one click away.\nfunction muteFooter(leadEmail?: string): string {\n if (!leadEmail) return \"\";\n return `\\n\\n---\\nMute this lead (stop auto-replies): ${adminLink(\"mute\", leadEmail)}\\nAll mutes: ${adminLink(\"mutes\")}`;\n}\n\nasync function priorAiMessages(leadId: number): Promise {\n const res = await sqlite.execute({\n sql: `SELECT reply_body FROM ai_actions\n WHERE lead_id = ? AND action IN ('sent', 'drafted', 'booked_and_sent', 'rescheduled_and_sent', 'cancelled_and_sent') AND reply_body IS NOT NULL\n ORDER BY id ASC`,\n args: [leadId],\n });\n return res.rows.map((r: unknown[]) => String(r[0]));\n}\n\n// ---------- scheduling orchestration ----------\n\nconst WINDOWS: Record = {\n morning: [8 * 60, 11 * 60 + 30],\n afternoon: [12 * 60, 17 * 60],\n evening: [17 * 60, 20 * 60 + 30],\n};\n\ntype SchedulingOutcome = {\n mode:\n | \"booked\"\n | \"counter\"\n | \"propose\"\n | \"already_booked\"\n | \"rescheduled\"\n | \"reschedule_propose\"\n | \"cancelled\"\n | \"no_reply\"\n | \"no_calendar\"\n | \"booking_stuck\" // open slot, calendar unresponsive after retries -> alert Sam, NO reply\n | \"legacy_appt\" // pre-migration GHL appointment -> alert Sam, NO reply\n | \"booked_info\"; // they have a booking; reply is a question, not a schedule change\n menu?: ProposalMenu;\n bookedLabel?: string; // \"Tuesday at 2pm ET\" (booked or rescheduled-to)\n proposedLabel?: string; // their unavailable ask, e.g. \"10am\"\n apptId?: string; // Calendly event uuid (booked/rescheduled) - feeds calendar links\n joinUrl?: string; // Google Meet link for the booked event\n startUtc?: string; // booked start in UTC ISO - feeds calendar links\n tz: string;\n tzAbbr: string;\n tzEvidence: string;\n detail: string; // for logs/alerts\n};\n\nasync function runScheduling(\n a: Analysis,\n lead: any,\n emails: ContactEmails,\n dryRun: boolean,\n): Promise {\n // Resolve timezone: model -> website intel fallback -> PT.\n let tz = a.timezone ?? \"\";\n let evidence = a.timezone_evidence ?? \"\";\n if (!tz) {\n const intel = await websiteIntel(lead);\n if (intel?.timezone) {\n tz = intel.timezone;\n evidence = `website: ${intel.evidence ?? \"site content\"}`;\n }\n }\n if (!tz) {\n tz = FALLBACK_TZ;\n evidence = \"no signal anywhere - defaulted to PT (labeled in copy)\";\n }\n // Guard against a hallucinated zone name blowing up Intl downstream.\n try {\n new Intl.DateTimeFormat(\"en-US\", { timeZone: tz });\n } catch {\n evidence = `model returned invalid zone \"${tz}\" - defaulted to PT`;\n tz = FALLBACK_TZ;\n }\n const tzAbbr = tzAbbreviation(tz);\n\n if (!calendarConfigured()) {\n return { mode: \"no_calendar\", tz, tzAbbr, tzEvidence: evidence, detail: \"Calendly not configured - calendar-link-only reply\" };\n }\n\n // They already moved it themselves - the new confirmation flow covers them,\n // no reply needed.\n if (a.self_rescheduled) {\n return { mode: \"no_reply\", tz, tzAbbr, tzEvidence: evidence, detail: \"self-rescheduled - no reply per Sam's rule\" };\n }\n\n // Look up any existing upcoming appointment whenever the reply might be about\n // one (booked claim, time affirmation, reschedule/cancel ask, or a concrete\n // proposed time that may actually be a move). Checked by email (reply-From +\n // lead) against Calendly.\n const mightHaveAppt = a.claims_booked || a.says_time_works || a.wants_reschedule ||\n a.wants_cancel || (a.proposed_date && a.proposed_time) || a.category === \"info_request\";\n const emailList = [emails.primary, emails.additional];\n let existing: { id: string; startTime: string } | null = null;\n if (mightHaveAppt) {\n existing = await getUpcomingAppointment(emailList).catch(() => null);\n // Pre-migration appointments live on the retired GHL calendar. Acting on\n // them (or ignoring them) confuses everyone - a reschedule/cancel/booked\n // claim OR a fresh proposed time from someone with a legacy appointment\n // (that's a move) alerts Sam instead; the prospect gets nothing. Dead code\n // once LEGACY_GHL_APPTS is deleted.\n if (!existing && Deno.env.get(\"LEGACY_GHL_APPTS\") === \"true\") {\n const legacy = await legacyGhlAppointment(emails).catch(() => null);\n if (legacy) {\n return {\n mode: \"legacy_appt\", tz, tzAbbr, tzEvidence: evidence,\n detail: `pre-migration GHL appointment ${legacy.id} at ${legacy.startTime} (location time) - manual handling`,\n };\n }\n }\n }\n\n // Cancel with no intent to rebook: cancel it, then try to save the meeting\n // anyway (Sam's Cancel template proposes new times).\n if (a.wants_cancel) {\n let detail = \"cancel request; no upcoming appointment found\";\n if (existing) {\n if (!dryRun) {\n await cancelAppointment(existing.id);\n detail = `cancelled appointment ${existing.id} (was ${existing.startTime} location time)`;\n } else detail = `[dry run] would cancel ${existing.id} (${existing.startTime})`;\n }\n let menu: ProposalMenu | undefined;\n try {\n menu = standardProposal(await loadSlots(tz), tz, a.earliest_date ?? \"\");\n } catch { /* save attempt without times is still fine */ }\n return { mode: \"cancelled\", tz, tzAbbr, tzEvidence: evidence, menu, detail };\n }\n\n // Pure \"I booked\" / \"that works\" with an appointment on the books and no move\n // requested -> courteous close.\n if ((a.claims_booked || a.says_time_works) && existing && !a.wants_reschedule &&\n !(a.proposed_date && a.proposed_time)) {\n return {\n mode: \"already_booked\", tz, tzAbbr, tzEvidence: evidence,\n detail: `existing appointment ${existing.id} at ${existing.startTime} (location time)`,\n };\n }\n // Question from someone who ALREADY has a booking (and isn't moving it):\n // answer + re-surface the meeting details; never pitch times at them\n // (2026-07-18 \"where's my invite\" case).\n if (existing && a.category === \"info_request\" && !a.wants_reschedule && !a.wants_cancel &&\n !(a.proposed_date && a.proposed_time)) {\n const ev = await getScheduledEvent(existing.id).catch(() => null);\n return {\n mode: \"booked_info\", tz, tzAbbr, tzEvidence: evidence,\n apptId: existing.id, joinUrl: ev?.location?.join_url ?? \"\", startUtc: ev?.start_time ?? \"\",\n bookedLabel: existing.startTime,\n detail: `existing appointment ${existing.id} (${existing.startTime} location time) - info follow-up, no scheduling`,\n };\n }\n if (a.claims_booked && !existing) {\n // They say they booked but nothing is on the calendar (lag, other email) -\n // a courteous close beats pitching them times again.\n return { mode: \"already_booked\", tz, tzAbbr, tzEvidence: evidence, detail: \"claims booked; no appointment found - courteous close\" };\n }\n\n let slots;\n try {\n slots = await loadSlots(tz);\n } catch (e) {\n console.error(\"loadSlots failed:\", e);\n return { mode: \"no_calendar\", tz, tzAbbr, tzEvidence: evidence, detail: `calendar slots unavailable: ${String(e).slice(0, 200)}` };\n }\n\n // Far-out asks WITHOUT a concrete time (\"out next week - book me the week\n // after\") propose real times from THAT week via the twin's 60-day window\n // (2026-07-18 swipefile case: never punt to a bare link when we can offer\n // times). The explicit-time branch below does its own targeted fetch.\n const lastNear = Object.keys(slots).sort().pop() ?? \"\";\n const farAnchor = [a.earliest_date, !a.proposed_time ? a.proposed_date : \"\"]\n .filter((d): d is string => Boolean(d && d > lastNear))\n .sort()[0];\n if (farAnchor && Math.round((Date.parse(`${farAnchor}T12:00:00Z`) - Date.now()) / 86_400_000) <= 60) {\n try {\n const far = await fetchSlotsForDate(farAnchor, tz);\n for (const [day, list] of Object.entries(far)) slots[day] = list;\n } catch (e) {\n console.error(\"far-week fetch failed:\", e);\n }\n }\n\n // Reschedule ask without a concrete new time -> propose times (Sam's\n // Reschedule template). The existing appointment stays until they pick.\n if (a.wants_reschedule && !(a.proposed_date && a.proposed_time)) {\n const floor = a.earliest_date ?? \"\";\n const menu = a.proposed_date && a.proposed_window && WINDOWS[a.proposed_window]\n ? windowProposal(slots, tz, a.proposed_date, WINDOWS[a.proposed_window])\n : standardProposal(slots, tz, floor);\n return {\n mode: \"reschedule_propose\", tz, tzAbbr, tzEvidence: evidence,\n menu: menu.days.length ? menu : standardProposal(slots, tz, floor),\n detail: `reschedule request without a time; proposing${floor ? ` (floor ${floor})` : \"\"} (existing appt ${existing?.id ?? \"none\"} kept)`,\n };\n }\n\n // Explicit proposed time -> move an existing appointment, or book fresh, or counter.\n if (a.proposed_date && a.proposed_time) {\n const mins = parseInt(a.proposed_time.slice(0, 2), 10) * 60 + parseInt(a.proposed_time.slice(3, 5), 10);\n const label = minutesToLabel(mins);\n\n // Far-out proposals are visible through the secret twin's 60-day window:\n // one targeted fetch for that week, merged into the near-term slots, then\n // the exact-slot path below is the single validator. (GHL's blind\n // out-of-window booking has no Calendly equivalent - verified 2026-07-18,\n // the API rejects window-external times with \"already_filled\".)\n const lastSlotDate = Object.keys(slots).sort().pop() ?? \"\";\n const daysOut = Math.round((Date.parse(`${a.proposed_date}T12:00:00Z`) - Date.now()) / 86_400_000);\n if (a.proposed_date > lastSlotDate && daysOut <= 60) {\n try {\n const farSlots = await fetchSlotsForDate(a.proposed_date, tz);\n for (const [day, list] of Object.entries(farSlots)) slots[day] = list;\n } catch (e) {\n // Counter from the near-term slots we DO have.\n console.error(\"far-out slot fetch failed:\", e);\n }\n }\n\n const exact = findExactSlot(slots, a.proposed_date, mins);\n if (exact) {\n const weekday = weekdayOf(a.proposed_date, tz);\n const bookedLabel = `${weekday} at ${label} ${tzAbbr}`;\n // Calendly books by invitee identity: the invite goes to the reply-From\n // address (must reach the actual human, not the office inbox), with the\n // twin's required Mobile Phone / Clinic Name answers from the best data\n // we hold.\n const nm = contactName(a, lead);\n const bookInfo = {\n email: emails.primary,\n name: [nm.firstName, nm.lastName].filter(Boolean).join(\" \") || a.clinic_name ||\n emails.primary.split(\"@\")[0],\n startTime: exact,\n timezone: tz,\n phone: a.phone_number ?? (leadVar(lead, \"phone\") || undefined),\n clinic: a.clinic_name ?? (lead.company || undefined),\n };\n // An open proposed time from someone who already has an appointment is a\n // MOVE, whether or not they used the word \"reschedule\".\n if (existing) {\n if (!dryRun) {\n // Book-new-first, then cancel-old (calendly.ts): a failed booking\n // keeps their current appointment.\n const moved = await rescheduleAppointment({ oldEventUuid: existing.id, ...bookInfo });\n if (moved.outcome === \"moved\") {\n return {\n mode: \"rescheduled\", tz, tzAbbr, tzEvidence: evidence, bookedLabel,\n apptId: moved.apptId, startUtc: new Date(exact).toISOString(),\n joinUrl: await getMeetLink(moved.apptId).catch(() => \"\"),\n detail: `moved appointment ${existing.id} to ${exact} (${bookedLabel})${moved.note}`,\n };\n }\n if (moved.outcome === \"taken\") {\n const menu = counterProposal(slots, tz, a.proposed_date, mins);\n return {\n mode: \"counter\", tz, tzAbbr, tzEvidence: evidence, menu, proposedLabel: label,\n detail: `move to ${exact} rejected (${String(moved.error).slice(0, 120)}); countered (existing appt ${existing.id} KEPT)`,\n };\n }\n return {\n mode: \"booking_stuck\", tz, tzAbbr, tzEvidence: evidence, bookedLabel,\n detail: `move to ${exact} stuck - Calendly unresponsive through retries (existing appt ${existing.id} kept; ${String(moved.error).slice(0, 160)})`,\n };\n }\n return {\n mode: \"rescheduled\", tz, tzAbbr, tzEvidence: evidence, bookedLabel,\n detail: `[dry run] would move ${existing.id} to ${exact} (${bookedLabel})`,\n };\n }\n if (!dryRun) {\n // GHL contact enrichment continues (GHL stays the CRM) but no longer\n // feeds the booking; a CRM hiccup must not block the calendar.\n try {\n const nmc = contactName(a, lead);\n await ensureContact(emails, {\n firstName: nmc.firstName,\n lastName: nmc.lastName,\n phone: a.phone_number ?? (leadVar(lead, \"phone\") || undefined),\n clinicName: a.clinic_name ?? undefined,\n });\n } catch (e) {\n console.error(\"ensureContact failed (booking proceeds):\", e);\n }\n // Double-book guard: booked under either email between our lookup and now.\n const raced = await getUpcomingAppointment(emailList).catch(() => null);\n if (raced) {\n return {\n mode: \"already_booked\", tz, tzAbbr, tzEvidence: evidence,\n detail: `double-book guard: appointment ${raced.id} already at ${raced.startTime}`,\n };\n }\n const booked = await bookWithRecovery(bookInfo);\n if (booked.outcome === \"booked\") {\n return {\n mode: \"booked\", tz, tzAbbr, tzEvidence: evidence, bookedLabel,\n apptId: booked.apptId, startUtc: new Date(exact).toISOString(),\n joinUrl: await getMeetLink(booked.apptId).catch(() => \"\"),\n detail: `booked ${booked.apptId} at ${exact} (${bookedLabel})${booked.note}`,\n };\n }\n if (booked.outcome === \"taken\") {\n const menu = counterProposal(slots, tz, a.proposed_date, mins);\n return {\n mode: \"counter\", tz, tzAbbr, tzEvidence: evidence, menu, proposedLabel: label,\n detail: `Calendly rejected ${exact} (${String(booked.error).slice(0, 120)}); countered`,\n };\n }\n return {\n mode: \"booking_stuck\", tz, tzAbbr, tzEvidence: evidence, bookedLabel,\n detail: `booking ${exact} stuck - Calendly unresponsive through retries (${String(booked.error).slice(0, 160)})`,\n };\n }\n return {\n mode: \"booked\", tz, tzAbbr, tzEvidence: evidence, bookedLabel,\n detail: `[dry run] would book ${exact} (${bookedLabel})`,\n };\n }\n // Unavailable (taken, past, or under the 2h notice - availability excludes all).\n const menu = counterProposal(slots, tz, a.proposed_date, mins);\n return {\n mode: \"counter\", tz, tzAbbr, tzEvidence: evidence, menu, proposedLabel: label,\n detail: `their ${a.proposed_date} ${label} not bookable; countered${existing ? ` (existing appt ${existing.id} kept)` : \"\"}`,\n };\n }\n\n // Day + rough window, no clock time.\n if (a.proposed_date && a.proposed_window && WINDOWS[a.proposed_window]) {\n const menu = windowProposal(slots, tz, a.proposed_date, WINDOWS[a.proposed_window]);\n if (menu.days.length) {\n return { mode: \"propose\", tz, tzAbbr, tzEvidence: evidence, menu, detail: `window proposal for ${a.proposed_date} ${a.proposed_window}` };\n }\n const fallback = counterProposal(slots, tz, a.proposed_date, (WINDOWS[a.proposed_window][0] + WINDOWS[a.proposed_window][1]) / 2);\n return { mode: \"counter\", tz, tzAbbr, tzEvidence: evidence, menu: fallback, proposedLabel: a.proposed_window, detail: `their ${a.proposed_date} ${a.proposed_window} had no slots; countered` };\n }\n\n // No timing from them -> standard proposal (floored to any stated window).\n const menu = standardProposal(slots, tz, a.earliest_date ?? \"\");\n const now = localNow(tz);\n return {\n mode: \"propose\", tz, tzAbbr, tzEvidence: evidence, menu,\n detail: `standard proposal (their local time ${Math.floor(now.minutes / 60)}:${String(now.minutes % 60).padStart(2, \"0\")}${a.earliest_date ? `, floor ${a.earliest_date}` : \"\"})`,\n };\n}\n\n// ---------- handler ----------\n\nexport default async function (req: Request): Promise {\n const url = new URL(req.url);\n if (url.searchParams.get(\"token\") !== Deno.env.get(\"WEBHOOK_SECRET\")) {\n return new Response(\"unauthorized\", { status: 401 });\n }\n const dryRun = (Deno.env.get(\"DRY_RUN\") ?? \"true\") !== \"false\";\n const tag = (s: string) => `${dryRun ? \"[DRY RUN] \" : \"\"}[Reply Agent] ${s}`;\n\n let body: any;\n try {\n body = await req.json();\n } catch {\n return Response.json({ ok: false, error: \"non-json body\" }, { status: 400 });\n }\n\n let processedUuid: string | null = null;\n await ensureTables();\n await sqlite.execute({\n sql: `INSERT INTO webhook_events (method, body) VALUES (?, ?)`,\n args: [req.method, JSON.stringify(body)],\n });\n\n try {\n if (body?.event?.type !== \"LEAD_REPLIED\") {\n return Response.json({ ok: true, skipped: \"not a LEAD_REPLIED event\" });\n }\n const reply = body?.data?.reply ?? {};\n const lead = body?.data?.lead ?? {};\n const campaign = body?.data?.campaign ?? {};\n const sender = body?.data?.sender_email ?? {};\n const senderName = (sender.name || \"Sam\").split(\" \")[0];\n const leadLabel = `${lead.first_name ?? \"\"} ${lead.last_name ?? \"\"} <${lead.email ?? \"?\"}>`.trim();\n const contactEmails = contactEmailsFor(reply, lead);\n\n if (reply.automated_reply === true) {\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id,\n reply_uuid: reply.uuid, action: \"skipped\", detail: \"automated_reply flag\",\n });\n return Response.json({ ok: true, skipped: \"automated reply\" });\n }\n\n // Crash-safe dedup: 'processing' rows older than 10 min are treated as dead\n // (the run was killed mid-flight) and the event is picked up again on Bison's\n // automatic retry. 'done' rows are never reprocessed.\n const replyUuid = reply.uuid ?? crypto.randomUUID();\n const dup = await sqlite.execute({\n sql: `SELECT status, processed_at >= datetime('now', '-10 minutes') AS fresh\n FROM processed_replies WHERE reply_uuid = ?`,\n args: [replyUuid],\n });\n if (dup.rows.length > 0) {\n const [status, fresh] = dup.rows[0] as [string, number];\n if (status === \"done\" || fresh) {\n return Response.json({ ok: true, skipped: \"duplicate delivery\" });\n }\n await sqlite.execute({\n sql: `UPDATE processed_replies SET processed_at = datetime('now'), status = 'processing' WHERE reply_uuid = ?`,\n args: [replyUuid],\n });\n } else {\n await sqlite.execute({\n sql: `INSERT INTO processed_replies (reply_uuid) VALUES (?)`,\n args: [replyUuid],\n });\n }\n processedUuid = replyUuid;\n const markDone = () =>\n sqlite.execute({\n sql: `UPDATE processed_replies SET status = 'done' WHERE reply_uuid = ?`,\n args: [replyUuid],\n });\n\n // Scheduled one-off emails (reminders): due-check is one cheap sqlite\n // query, so it runs on every event - whichever trigger fires first sends.\n try {\n const dueNote = await checkDueEmails(dryRun);\n if (dueNote !== \"no due emails\" && dueNote !== \"no GHL key\") {\n console.log(\"due emails:\", dueNote);\n await logAction({ action: \"scheduled_email\", detail: dueNote.slice(0, 900) });\n }\n } catch (e) {\n console.error(\"checkDueEmails failed:\", e);\n }\n\n // Self-booking reconciler piggybacks on webhook traffic (throttled to\n // >=15 min inside; ~2-5s when it actually runs). Keeps the Cold Email\n // pipeline in sync with the calendar even for widget self-bookings that\n // never touch this webhook.\n try {\n const recNote = await reconcile(dryRun);\n if (recNote !== \"throttled\") {\n console.log(\"reconcile:\", recNote);\n if (/->|created|matched/.test(recNote)) {\n await logAction({ action: \"crm_reconciled\", detail: recNote.slice(0, 900) });\n }\n }\n } catch (e) {\n console.error(\"reconcile failed:\", e);\n await logAction({ action: \"crm_failed\", detail: `reconcile: ${String(e).slice(0, 300)}` }).catch(() => {});\n }\n\n // Muted lead: Sam is handling this thread manually (admin page on the\n // test endpoint) - no auto-reply, just a heads-up so nothing gets missed.\n // Matches the campaign lead email AND the address they replied from.\n const muted = await mutedSince([lead.email, reply.from_email_address]);\n if (muted) {\n await safeEmail({\n subject: tag(`Muted lead replied - nothing sent: ${leadLabel}`),\n text:\n `This lead is MUTED (since ${muted.slice(0, 16).replace(\"T\", \" \")} UTC) - the agent stayed silent.\\n\\n` +\n `Campaign: ${campaign.name}\\nReply from: ${reply.from_email_address ?? \"?\"}\\n\\n` +\n `Their message:\\n${(reply.text_body ?? stripHtml(reply.html_body ?? \"\")).slice(0, 3000)}\\n\\n` +\n `---\\nUnmute (auto-replies back on): ${adminLink(\"unmute\", lead.email ?? reply.from_email_address)}\\n` +\n `All mutes: ${adminLink(\"mutes\")}`,\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id,\n reply_uuid: reply.uuid, action: \"muted_skip\", detail: `muted since ${muted}`,\n });\n await markDone();\n return Response.json({ ok: true, skipped: \"muted\" });\n }\n\n const prior = await priorAiMessages(lead.id ?? -1);\n if (prior.length >= MAX_AI_REPLIES_PER_LEAD) {\n await safeEmail({\n subject: tag(`Reply cap hit - take over: ${leadLabel}`),\n text: `The AI already wrote ${prior.length} replies to this lead and is stopping.\\n\\nCampaign: ${campaign.name}\\nTheir latest message:\\n${reply.text_body ?? \"\"}` + muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id,\n reply_uuid: reply.uuid, action: \"alerted\", detail: \"reply cap reached\",\n });\n await markDone();\n return Response.json({ ok: true, skipped: \"reply cap\" });\n }\n\n const firstEmail = await fetchFirstSentEmail(lead.id);\n\n // Current time in the common US zones so the model can resolve \"tomorrow\"\n // etc. in the prospect's local zone.\n const nowRef = [\"America/New_York\", \"America/Chicago\", \"America/Denver\", \"America/Phoenix\", \"America/Los_Angeles\"]\n .map((z) => `${z}: ${new Date().toLocaleString(\"en-US\", { timeZone: z, weekday: \"long\", year: \"numeric\", month: \"2-digit\", day: \"2-digit\", hour: \"2-digit\", minute: \"2-digit\", hour12: false })}`)\n .join(\"\\n\");\n\n const leadVars = (lead.custom_variables ?? [])\n .map((v: any) => `${v.name}: ${v.value}`).join(\"\\n\");\n\n const context = [\n `Sender persona (reply as this name): ${senderName}`,\n `Campaign: ${campaign.name ?? \"?\"}`,\n `Prospect: ${leadLabel}${lead.company ? ` - ${lead.company}` : \"\"}`,\n leadVars ? `Lead data (from Email Bison):\\n${leadVars}` : \"\",\n `CURRENT TIME REFERENCE:\\n${nowRef}`,\n firstEmail\n ? `FIRST EMAIL WE SENT IN THIS THREAD (source of truth for the free-campaign framing and any patient-count promise):\\n${firstEmail}`\n : `(first sent email unavailable - use \"the free campaign\" with no duration/numbers)`,\n prior.length\n ? `Our previous AI-written messages in this thread:\\n${prior.map((b, i) => `--- message ${i + 1} ---\\n${b}`).join(\"\\n\")}`\n : \"\",\n `Subject: ${reply.email_subject ?? \"\"}`,\n reply.from_name || reply.from_email_address\n ? `Reply From header: ${reply.from_name ?? \"\"} <${reply.from_email_address ?? \"?\"}>`.trim()\n : \"\",\n `PROSPECT'S LATEST REPLY (may include quoted history below it):\\n${(reply.text_body ?? stripHtml(reply.html_body ?? \"\")).slice(0, 6000)}`,\n ].filter(Boolean).join(\"\\n\\n\");\n\n // ---- phase 1: classify + extract ----\n const analysis: Analysis = await claudeTool(ANALYSIS_PROMPT, [{ role: \"user\", content: context }], ANALYSIS_TOOL, 1000);\n console.log(\"analysis:\", JSON.stringify(analysis));\n\n // Offer variant (2026-07-15 boolean, 2026-07-16 three-way): the MODEL\n // judges how the first email backed its promise - any wording, any\n // position (Sam moves copy around and spintax varies; string-matching his\n // templates is too brittle). \"reimburse\" = current campaigns (refund ad\n // spend on a miss), \"until_we_do\" = legacy keep-running promise, \"none\" =\n // legacy base. The draft prompt is told the verdict and violations()\n // rails the output - a guarantee the first email never made costs real\n // money. Missing/unknown verdicts fall back to \"none\" (the safe rail).\n const guarantee: \"none\" | \"until_we_do\" | \"reimburse\" =\n analysis.offer_guarantee === \"reimburse\" || analysis.offer_guarantee === \"until_we_do\"\n ? analysis.offer_guarantee\n : \"none\";\n\n if (analysis.category === \"not_positive\") {\n // Even a hard \"cancel, not interested\" must still clear the calendar.\n let cancelNote = \"\";\n if (analysis.wants_cancel && !dryRun && calendarConfigured()) {\n try {\n const appt = await getUpcomingAppointment([contactEmails.primary, contactEmails.additional]);\n if (appt) {\n await cancelAppointment(appt.id, \"Cancelled per your reply - no problem at all\");\n cancelNote = `; cancelled appointment ${appt.id} (${appt.startTime})`;\n } else if (Deno.env.get(\"LEGACY_GHL_APPTS\") === \"true\") {\n const legacy = await legacyGhlAppointment(contactEmails).catch(() => null);\n if (legacy) {\n await safeEmail({\n subject: tag(`Cancel request on pre-migration GHL appointment - handle manually: ${lead.email}`),\n text: `A not-interested reply asked to cancel, and the appointment lives on the RETIRED GHL calendar (${legacy.id} at ${legacy.startTime} location time). Cancel it in GHL yourself - the agent no longer touches that calendar.${muteFooter(lead.email)}`,\n });\n cancelNote = `; LEGACY GHL appointment ${legacy.id} - alerted Sam for manual cancel`;\n }\n }\n } catch (e) {\n console.error(\"cancel on not_positive failed:\", e);\n cancelNote = `; CANCEL FAILED: ${String(e).slice(0, 150)}`;\n }\n }\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"skipped\", detail: analysis.reasoning + cancelNote,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: `none${cancelNote ? \" (appointment cancelled)\" : \"\"}` });\n }\n\n // Positive categories are pushed into Bison right away (even if drafting\n // later fails, the interest is real). Referral is excluded: Sam's templates\n // file groups \"Wrong Person\" under Uninterested - the person who replied\n // isn't the interested party. Best-effort: a Bison hiccup here must never\n // block the reply itself.\n if (!dryRun && reply.id && analysis.category !== \"referral\") {\n try {\n await markReplyInterested(reply.id);\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"marked_interested\",\n });\n } catch (e) {\n console.error(\"markReplyInterested failed:\", e);\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"mark_interested_failed\", detail: String(e).slice(0, 300),\n });\n }\n }\n\n // GHL is Sam's CRM (2026-07-15): every positive reply creates/enriches a\n // contact + a Positive-stage opportunity in the Cold Email pipeline.\n // Best-effort - a CRM hiccup never blocks the reply.\n let crmContactId: string | undefined;\n try {\n const crm = await syncPositiveReply({ analysis, lead, reply, dryRun });\n crmContactId = crm.contactId;\n console.log(\"crm sync:\", crm.note);\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"crm_synced\", detail: crm.note.slice(0, 900),\n });\n } catch (e) {\n console.error(\"crm sync failed:\", e);\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"crm_failed\", detail: String(e).slice(0, 300),\n }).catch(() => {});\n }\n\n // Sam calls EVERY interested reply now (2026-07-23), so every positive\n // category gets a call-prep alert: best number (personal beats office),\n // decisionmaker (extracted, or researched off the practice website when\n // the replier reads as staff), reply time in PT, offer, their message.\n // Best-effort and self-degrading - the notification must never be lost.\n {\n let prep: { subject: string; text: string; detail: string };\n try {\n prep = await buildCallPrep({ analysis, lead, reply, leadLabel, senderEmail: sender.email, senderName });\n } catch (e) {\n console.error(\"buildCallPrep failed:\", e);\n prep = {\n subject: `Interested - call prep: ${leadLabel}`,\n text: `Call prep couldn't be fully built (${String(e).slice(0, 200)}) - raw facts:\\n\\n` +\n `Phone: ${analysis.phone_number ?? \"?\"}\\n` +\n `Name: ${analysis.prospect_full_name ?? analysis.prospect_first_name ?? \"?\"}\\n` +\n `Subject line: ${cleanSubject(reply.email_subject) || \"?\"}\\n` +\n `Sender email: ${reply.from_email_address ?? lead.email ?? \"?\"}\\n\\n` +\n `THEIR MESSAGE\\n-------------\\n${reply.text_body ?? stripHtml(reply.html_body ?? \"\")}`,\n detail: `build failed: ${String(e).slice(0, 200)}`,\n };\n }\n await safeEmail({ subject: tag(prep.subject), text: prep.text + muteFooter(lead.email) });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"call_prep_alerted\", detail: prep.detail.slice(0, 900),\n });\n // prospectPhoneKnown() gates the phone asks on 'phone_alerted' rows -\n // keep recording every extracted number under that action name.\n if (analysis.phone_number) {\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"phone_alerted\",\n detail: `${analysis.phone_number} (${analysis.phone_number_type ?? \"unknown\"})`,\n });\n }\n }\n\n if (analysis.category === \"info_request\" && analysis.answerable === false) {\n await safeEmail({\n subject: tag(`Question outside playbook - take over: ${leadLabel}`),\n text:\n `A prospect asked something the playbook can't answer. The AI is NOT replying.\\n\\n` +\n `Prospect: ${leadLabel}\\nCampaign: ${campaign.name}\\nPersona they know: ${senderName}\\n` +\n `Why escalated: ${analysis.reasoning}\\n\\nTheir message:\\n${reply.text_body ?? \"\"}\\n\\n` +\n `Tip: answer them in Bison, then tell Claude Code the Q&A so it goes into facts.md for next time.` +\n muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"alerted\", detail: `unanswerable: ${analysis.reasoning}`,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"alerted\" });\n }\n\n // ---- scheduling (meeting_request / info_request only) ----\n const needsScheduling = analysis.category === \"meeting_request\" || analysis.category === \"info_request\";\n let sched: SchedulingOutcome | null = null;\n if (needsScheduling) {\n // Scheduling must NEVER kill the reply: a GHL hiccup here once aborted\n // the whole run AFTER the booking had landed server-side (Cornerstone,\n // 2026-07-17) - orphaned appointment, prospect never answered. Degrade\n // to a calendar-link-only reply and tell Sam to check the calendar.\n try {\n sched = await runScheduling(analysis, lead, contactEmails, dryRun);\n } catch (e) {\n console.error(\"runScheduling crashed:\", e);\n let fbTz = analysis.timezone ?? FALLBACK_TZ;\n try {\n new Intl.DateTimeFormat(\"en-US\", { timeZone: fbTz });\n } catch {\n fbTz = FALLBACK_TZ;\n }\n sched = {\n mode: \"no_calendar\", tz: fbTz, tzAbbr: tzAbbreviation(fbTz),\n tzEvidence: \"scheduling crashed - calendar-link-only reply\",\n detail: `scheduling crashed: ${String(e).slice(0, 200)}`,\n };\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"sched_failed\", detail: String(e).slice(0, 500),\n }).catch(() => {});\n await safeEmail({\n subject: tag(`Scheduling failed - ${lead.email} got a calendar-link-only reply`),\n text: `runScheduling crashed for ${lead.email}; the reply still went out with just the calendar link.\\n\\nError: ${String(e)}\\n\\nCHECK THE CALENDAR: if a booking or reschedule was in flight when this failed, it may have landed anyway (orphaned appointment the reply doesn't mention).${muteFooter(lead.email)}`,\n });\n }\n console.log(\"scheduling:\", JSON.stringify({ ...sched, menu: sched.menu ? menuForPrompt(sched.menu) : undefined }));\n }\n\n // Pipeline follows the calendar: agent-booked, agent-moved, and\n // detected self-bookings all land the opportunity in Booked.\n if (sched && (sched.mode === \"booked\" || sched.mode === \"rescheduled\" || sched.mode === \"already_booked\")) {\n try {\n const note = await markBooked(\n crmContactId ?? await lookupContact(contactEmails) ?? undefined,\n analysis.clinic_name ?? lead.company ?? lead.first_name ?? lead.email,\n dryRun,\n );\n console.log(\"crm booked:\", note);\n } catch (e) {\n console.error(\"markBooked failed:\", e);\n await logAction({ action: \"crm_failed\", detail: `markBooked: ${String(e).slice(0, 300)}` }).catch(() => {});\n }\n }\n\n // Open slot but Calendly wouldn't confirm the booking through retries\n // (~3 min): anything we could write might contradict what actually landed,\n // so the prospect gets NOTHING and Sam takes over (his rule, 2026-07-17).\n if (sched?.mode === \"booking_stuck\") {\n await safeEmail({\n subject: tag(`BOOKING STUCK - take over: ${leadLabel} wants ${sched.bookedLabel ?? \"a specific time\"}`),\n text:\n `Calendly wouldn't confirm a booking for an OPEN slot after 2 attempts + calendar checks. ` +\n `The AI sent the prospect NOTHING - reply and book manually.\\n\\n` +\n `Prospect: ${leadLabel}\\nCampaign: ${campaign.name}\\nPersona they know: ${senderName}\\n` +\n `Their ask: ${sched.bookedLabel ?? \"(see reply)\"}\\nError: ${sched.detail}\\n\\n` +\n `CHECK THE CALENDLY CALENDAR FIRST: an aborted booking call can still land minutes later. ` +\n `If the appointment is there, just send the confirmation.\\n\\n` +\n `Their message:\\n${(reply.text_body ?? stripHtml(reply.html_body ?? \"\")).slice(0, 3000)}` +\n muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"alerted\", detail: `booking stuck: ${sched.detail}`,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"alerted (booking stuck)\" });\n }\n\n // Pre-migration appointment on the retired GHL calendar: the agent no\n // longer reads or writes that calendar, so Sam handles the thread himself.\n if (sched?.mode === \"legacy_appt\") {\n await safeEmail({\n subject: tag(`Pre-migration GHL appointment - handle manually: ${leadLabel}`),\n text:\n `This prospect's reply is about an appointment booked on the RETIRED GHL calendar ` +\n `(before the Calendly migration). The AI sent them NOTHING - reply and manage the ` +\n `appointment in GHL yourself.\\n\\n` +\n `Prospect: ${leadLabel}\\nCampaign: ${campaign.name}\\nPersona they know: ${senderName}\\n` +\n `Appointment: ${sched.detail}\\n\\n` +\n `Their message:\\n${(reply.text_body ?? stripHtml(reply.html_body ?? \"\")).slice(0, 3000)}` +\n muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"alerted\", detail: sched.detail,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"alerted (legacy GHL appt)\" });\n }\n\n // They already rebooked themselves - their new confirmation flow covers it.\n if (sched?.mode === \"no_reply\") {\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"skipped\", detail: sched.detail,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"none (self-rescheduled)\" });\n }\n\n // ---- phase 2: draft ----\n const isFollowUp = prior.length > 0;\n // Phone-ask doctrine (Sam, 2026-07-18): at most TWO asks ever - the CTA\n // ask on the first time-proposing reply, and the reminders ask in the\n // booking confirmation. The moment we hold a number they gave us\n // (this reply or any earlier one), both are dead forever.\n const phoneKnown = await prospectPhoneKnown(lead.id, analysis.phone_number);\n const askPhoneCta = !phoneKnown && !isFollowUp &&\n (sched?.mode === \"counter\" || sched?.mode === \"propose\" || sched?.mode === \"no_calendar\");\n const askPhoneConfirm = !phoneKnown && sched?.mode === \"booked\" &&\n !prior.some((p) => p.includes(\"reminders before the call\"));\n let draftMode: string;\n let psLine = \"\";\n if (analysis.category === \"contact_preference\") {\n draftMode = \"PMOC_ACK: the short PMoC acknowledgment from templates.md - tell them to expect the {text/call/...} they asked for from our founder Sam (his number is 925-730-9373). MIRROR any timing they stated in their reply (e.g. \\\"to set up a time to talk Friday afternoon\\\") - say \\\"later today\\\" ONLY if they gave no timing. Nothing else, no links.\";\n } else if (analysis.category === \"referral\" && analysis.new_contact_provided === true) {\n draftMode = \"REFERRAL_INTRO: the Wrong Person template, second part - the intro email TO THE NEW PERSON, greeting them by name. Sam sends it manually, so make it final, no placeholders. Do not mention meeting times (no verified times available for this draft) - end the CTA with the calendar link.\";\n } else if (analysis.category === \"referral\") {\n draftMode = \"REFERRAL_ASK: the Wrong Person template, first part - ask the person who replied for the new contact's best email. No booking link, no times.\";\n } else if (sched?.mode === \"booked_info\") {\n draftMode = `BOOKED_INFO: they already have a booked call and this reply is a question, not a scheduling change. Answer briefly per the playbook. If they can't find the invite/details, use the Post-Booking Details template: light apology, point them to the Calendly email (notifications@calendly.com) + spam folder, then \"In the meantime, here are the meeting details:\" followed by EXACTLY these three lines, each on its own line (the system swaps them for real links - NEVER write actual URLs): \"Google Meet: [MEET_LINK]\" then \"Add to Google Calendar\" then \"Add to Apple / Outlook Calendar\", then a closing line like \"Everything you need is in this thread now, but I'm happy to resend the Calendly invite if it still hasn't come through.\" NO times, NO calendar link, no pitch, no phone ask, no P.S.`;\n } else if (sched?.mode === \"already_booked\") {\n draftMode = \"ALREADY_BOOKED: they booked (or say they did). One short, warm line confirming you saw it and you're looking forward to the call - e.g. \\\"saw the booking come through - talk then!\\\" adapted to the thread. No times, no calendar link, no pitch, no P.S.\";\n } else if (sched?.mode === \"booked\") {\n draftMode = `BOOKED_CONFIRMATION: you booked their proposed time. Use the Booking Confirmation template, ${\n analysis.invite_email_generic\n ? `GENERIC-EMAIL variant: open with \"Hey - ${sched.bookedLabel} it is! Just sent over a confirmation email with meeting details to ${contactEmails.primary}. It looks like this may be an office email, so if there's a better address, just let me know and I'll send it there instead.\"`\n : `RIGHT-EMAIL variant: open with \"Hey - ${sched.bookedLabel} it is! Just sent over a confirmation email with meeting details.\"`\n } Then EXACTLY these three lines, each on its own line (the system swaps them for real links - NEVER write actual URLs for them): \"Google Meet link for your convenience: [MEET_LINK]\" then \"Add to Google Calendar\" then \"Add to Apple / Outlook Calendar\".${\n askPhoneConfirm\n ? prior.some((p) => p.includes(\"number to reach you directly\"))\n ? ` Then ask exactly: \"And what's the best number for reminders before the call? Promise you won't get spammed :)\"`\n : ` Then ask exactly: \"And what's the best number for reminders before the call?\"`\n : \" Do NOT ask for a phone number.\"\n } If their reply also asked questions, answer them briefly first per the playbook. No other times, no calendar link, no P.S.`;\n } else if (sched?.mode === \"rescheduled\") {\n draftMode = `RESCHEDULED_CONFIRMATION: their appointment is already MOVED to ${sched.bookedLabel} - done deal. Quick courteous confirmation (\"No worries at all - just moved us to ${sched.bookedLabel}.\" adapted to the thread), then EXACTLY these three lines, each on its own line (the system swaps them for real links - NEVER write actual URLs for them): \"Google Meet link for your convenience: [MEET_LINK]\" then \"Add to Google Calendar\" then \"Add to Apple / Outlook Calendar\". No other times, no calendar link, no pitch, no phone ask.`;\n } else if (sched?.mode === \"reschedule_propose\") {\n draftMode = `RESCHEDULE_PROPOSE: they want to move their existing appointment but didn't name a new time. Use the \"Written Reschedule Request\" template: \"Hey, No worries at all - happy to reschedule. Does [VERIFIED TIMES] work for you instead?\" then the calendar line (\"Alternatively, have your pick at any available time on our calendar:\"), then \"Feel free to reach out if those times don't work and we can work something out.\" Sign off \"Take care,\". Their current appointment stays until they pick.`;\n } else if (sched?.mode === \"cancelled\") {\n draftMode = `CANCELLED_SAVE: their appointment is cancelled as they asked. Use the \"Written Cancellation Request\" template: \"Hey, No worries at all.\" then - IF their message leaves any door open - try to save it: \"Would [VERIFIED TIMES] work for you instead?\" plus the calendar line, closing \"Let me know if you have any questions or concerns.\" Sign off \"Take care,\". If their message is clearly final (firmly out, unhappy), skip the times and calendar entirely - just a warm, brief, door-open close.`;\n } else if (sched?.mode === \"counter\") {\n draftMode = `COUNTER: what they asked for (${sched.proposedLabel ?? \"their ask\"}) isn't open. Use the \"IF SLOT ISN'T OPEN\" template from templates.md, adapting the opener naturally to what they asked for, then offer the VERIFIED TIMES by day, then the Calendly line (\"Or if it's easier, here's our Calendly: ...\").${askPhoneCta ? ` Before the Calendly line ask exactly: \"And just for convenience, what's a good number to reach you directly?\"` : \"\"}`;\n } else if (sched?.mode === \"no_calendar\") {\n draftMode = \"PROPOSE_LINK_ONLY: the calendar couldn't be checked, so do not mention any specific times - answer per the playbook and end the CTA with the Calendly link only.\";\n } else if (analysis.asked_for_call && analysis.category === \"meeting_request\") {\n draftMode = `ASK_FOR_CALL: they explicitly asked to set up the call - do NOT re-pitch it (no \"Let's hop on a quick call...\" line; they're already sold on talking). Use the Ask for Call template: a short natural greeting fitted to their words (\"Hey, great to hear!\" / \"Hey - of course!\"), then straight to \"How about [VERIFIED TIMES by day]?\", then \"Or if it's easier, here's our Calendly: ${BOOKING_LINK}\".${askPhoneCta ? ` Then ask exactly: \"And just for convenience, what's a good number to reach you directly?\"` : \"\"}`;\n } else {\n draftMode = `PROPOSE: answer their reply per the playbook templates (when the Tell Me More template applies - bare interest / hear-more replies - it MUST include the full \"Three quick bullets\" block, never skipped or trimmed), then close with the CTA BLOCK from templates.md (\"Let's hop on a quick call to share more details about the pilot, answer any questions, and go over the results you can expect.\" then the VERIFIED TIMES by day${askPhoneCta ? ` then ask exactly: \"And just for convenience, what's a good number to reach you directly?\"` : \"\"}). No inline calendar-link line - the system appends the self-booking line under the signature.`;\n }\n // Only fully-interested (meeting_request) first replies get a P.S.; the old\n // \"not a sales call\" P.S. for info requests is retired (2026-07-14).\n if (!psLine && !isFollowUp && analysis.category === \"meeting_request\" && sched && (sched.mode === \"counter\" || sched.mode === \"propose\" || sched.mode === \"no_calendar\")) {\n psLine = \"P.S. If those times don't work, feel free to send over other openings and I'll book you in directly\";\n }\n\n const offerVariantLine =\n guarantee === \"reimburse\"\n ? `OFFER VARIANT: REIMBURSE - the first email promised to reimburse/cover their ad spend if the numbers don't hit. Use the fee-waiver block's short reimburse guarantee sentence per templates.md (mirror their numbers/treatment exactly; rate and budget numbers only in the dedicated Reimbursement / Ad Budget templates when they explicitly asked). Never use \"until we do\" keep-running language.`\n : guarantee === \"until_we_do\"\n ? `OFFER VARIANT: UNTIL WE DO (legacy) - the first email's P.S. promised to keep the campaign running until the numbers hit (ad spend included). Apply the Until We Do rules in templates.md (info-style replies get the one casual clause line; direct questions get the dedicated legacy template). Never use reimbursement language.`\n : `OFFER VARIANT: BASE (legacy) - the first email made NO guarantee. Never write or imply any guarantee, reimbursement, free extension, or \"until we do\" language.`;\n\n const draftInput = [\n context,\n `=== DRAFTING INSTRUCTIONS ===`,\n `DRAFT MODE: ${draftMode}`,\n offerVariantLine,\n sched?.menu ? `VERIFIED TIMES (the ONLY times you may mention, prospect-local ${sched.tzAbbr}):\\n${menuForPrompt(sched.menu)}` : \"\",\n sched ? `Prospect timezone: ${sched.tz} (${sched.tzAbbr}) - evidence: ${sched.tzEvidence}` : \"\",\n analysis.prospect_first_name ? `Greet them as: ${analysis.prospect_first_name}` : \"Their name is unknown - greet with just \\\"Hey,\\\"\",\n askPhoneCta || askPhoneConfirm\n ? `Phone ask: include ONLY the exact phone question your draft mode specifies - no other phone request anywhere (the old \"best number to reach you\" line stays dead).`\n : `Do NOT ask for their phone number in ANY form (we have it, or already asked) - and the old \"best number to reach you\" line is retired forever.`,\n psLine ? `End with exactly this P.S. line:\\n${psLine}` : \"No P.S. line on this reply.\",\n isFollowUp ? \"BACK-AND-FORTH MODE: this is a follow-up exchange - short and loose, no full template structure, no repeated P.S. boilerplate.\" : \"\",\n `Calendar link (bare URL when the mode calls for it): ${BOOKING_LINK}`,\n ].filter(Boolean).join(\"\\n\\n\");\n\n const messages: { role: string; content: string }[] = [{ role: \"user\", content: draftInput }];\n let draft = mechanicalFixes(String((await claudeTool(DRAFT_PROMPT_HEAD, messages, DRAFT_TOOL)).reply_text ?? \"\"));\n\n // Sanitation + one repair round. Counter/propose/no_calendar drafts no\n // longer need an inline link (the self-book line is appended under the\n // signature by code) - only templates with a written-in Calendly line\n // still hard-require it.\n const requireBooking = sched?.mode === \"reschedule_propose\" ||\n (analysis.category === \"referral\" && analysis.new_contact_provided === true);\n const allowedTimes: Set | null = sched?.menu || sched?.bookedLabel || sched?.proposedLabel\n ? new Set([\n ...(sched?.menu?.days.flatMap((d) => d.times) ?? []),\n ...(sched?.bookedLabel ? [sched.bookedLabel.split(\" at \")[1]?.split(\" \")[0] ?? \"\"] : []),\n ...(sched?.proposedLabel ? [sched.proposedLabel] : []),\n ].filter(Boolean).map(normTimeToken))\n : null;\n const strictDraft = analysis.category === \"meeting_request\" || analysis.category === \"info_request\" ||\n (analysis.category === \"referral\" && analysis.new_contact_provided !== true);\n\n let problems = violations(draft, requireBooking, allowedTimes, guarantee, { cta: askPhoneCta, confirm: askPhoneConfirm });\n if (draft === \"\") problems.push(\"empty draft\");\n if (problems.length) {\n messages.push(\n { role: \"assistant\", content: `Draft: ${draft}` },\n {\n role: \"user\",\n content:\n `Your draft failed these checks:\\n- ${problems.join(\"\\n- \")}\\n` +\n `Rewrite it to fix every item, same draft mode, and resubmit via the tool.`,\n },\n );\n draft = mechanicalFixes(String((await claudeTool(DRAFT_PROMPT_HEAD, messages, DRAFT_TOOL)).reply_text ?? \"\"));\n problems = violations(draft, requireBooking, allowedTimes, guarantee, { cta: askPhoneCta, confirm: askPhoneConfirm });\n if (draft === \"\") problems.push(\"empty draft\");\n if (problems.length) {\n const stateChanged = sched?.mode === \"booked\" || sched?.mode === \"rescheduled\" || sched?.mode === \"cancelled\";\n if (strictDraft && !stateChanged) {\n await safeEmail({\n subject: tag(`Draft failed checks - take over: ${leadLabel}`),\n text: `The AI's draft failed checks twice, so nothing was sent.\\n\\nProblems:\\n- ${problems.join(\"\\n- \")}\\n\\nDraft:\\n${draft}\\n\\nTheir message:\\n${reply.text_body ?? \"\"}` + muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"alerted\",\n reply_body: draft, detail: `failed checks: ${problems.join(\"; \")}`,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"alerted (draft failed checks)\" });\n }\n // Calendar state already changed - we MUST tell them something, so fall\n // back to the bare template. Soft drafts (PMoC/intro) just degrade.\n if (sched?.mode === \"booked\") {\n draft = `Hey - ${sched.bookedLabel} it is! Just sent over a confirmation email with meeting details.\\n\\nGoogle Meet link for your convenience: [MEET_LINK]\\n\\nAdd to Google Calendar\\nAdd to Apple / Outlook Calendar\\n\\nThanks,\\n${senderName}`;\n } else if (sched?.mode === \"rescheduled\") {\n draft = `Hey - no worries at all, just moved us to ${sched.bookedLabel}.\\n\\nGoogle Meet link for your convenience: [MEET_LINK]\\n\\nAdd to Google Calendar\\nAdd to Apple / Outlook Calendar\\n\\nThanks,\\n${senderName}`;\n } else if (sched?.mode === \"cancelled\") {\n draft = `Hey - no worries at all, just cancelled it. If you ever want to grab another time, here's our calendar: ${BOOKING_LINK}\\n\\nTake care,\\n${senderName}`;\n } else {\n draft = \"\";\n }\n }\n }\n if (draft) draft = ensureSignature(draft, senderName, !(sched?.mode === \"already_booked\" || sched?.mode === \"booked_info\"));\n // Real URLs are injected by code, never typed by the model. [MEET_LINK]\n // becomes the event's Meet URL (line dropped if Google hasn't minted one);\n // the self-book line lands directly under \"SDR, Lumina\" on CTA replies\n // that didn't already include an inline Calendly link.\n if (draft) {\n draft = sched?.joinUrl\n ? draft.replaceAll(\"[MEET_LINK]\", sched.joinUrl)\n : draft.split(\"\\n\").filter((l) => !l.includes(\"[MEET_LINK]\")).join(\"\\n\");\n const ctaMode = sched?.mode === \"counter\" || sched?.mode === \"propose\" || sched?.mode === \"no_calendar\";\n if (ctaMode && !draft.toLowerCase().includes(\"calendly.com/luminamg/pilot\")) {\n const lines = draft.split(\"\\n\");\n const sigIdx = lines.findIndex((l) => /^\\s*SDR,\\s*Lumina\\s*$/i.test(l));\n const selfBook = \"Book a time yourself at calendly.com/luminamg/pilot\";\n if (sigIdx >= 0) lines.splice(sigIdx + 1, 0, selfBook);\n else lines.push(selfBook);\n draft = lines.join(\"\\n\");\n }\n }\n const calLinks = sched?.apptId && sched?.startUtc\n ? buildCalendarLinks(sched.apptId, sched.joinUrl ?? \"\", sched.startUtc, ADMIN_URL)\n : null;\n const htmlAnchors = calLinks\n ? { \"Add to Google Calendar\": calLinks.gcal, \"Add to Apple / Outlook Calendar\": calLinks.ics }\n : undefined;\n console.log(\"draft:\", JSON.stringify({ mode: sched?.mode ?? analysis.category, draft }));\n\n // ---- routing ----\n\n // PMoC: send the one-line acknowledgment, then hand the actual outreach to Sam.\n if (analysis.category === \"contact_preference\") {\n let ackNote = \"no acknowledgment sent (draft failed) - consider sending one yourself\";\n if (draft) {\n if (dryRun) ackNote = \"[dry run] acknowledgment drafted but NOT sent\";\n else {\n try {\n await sendBisonReply(reply.id, draft, sender.id, lead.email);\n ackNote = \"the AI already sent them this acknowledgment\";\n } catch (e) {\n console.error(\"PMoC ack send failed:\", e);\n ackNote = `acknowledgment send FAILED - send one yourself (${String(e).slice(0, 200)})`;\n }\n }\n }\n await safeEmail({\n subject: tag(`Contact preference - reach out personally: ${leadLabel}`),\n text:\n `They gave a contact preference - the actual texting/calling is on you.\\n\\n` +\n `Prospect: ${leadLabel}\\nCampaign: ${campaign.name}\\nPersona they know: ${senderName}\\n` +\n `Details: ${analysis.contact_details ?? \"(see message)\"}\\n\\n` +\n `Acknowledgment (${ackNote}):\\n${draft || \"(none)\"}\\n\\nTheir message:\\n${reply.text_body ?? \"\"}` +\n muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: draft ? (dryRun ? \"drafted\" : \"sent\") : \"alerted\",\n reply_body: draft || undefined, detail: analysis.contact_details ?? analysis.reasoning,\n });\n await markDone();\n return Response.json({\n ok: true, category: analysis.category,\n action: draft && !dryRun ? \"ack sent + alerted\" : \"alerted\",\n });\n }\n\n // Referral with the new contact's email: the intro draft goes to Sam to send.\n if (analysis.category === \"referral\" && analysis.new_contact_provided === true) {\n await safeEmail({\n subject: tag(`Referral - intro drafted, send it to the new contact: ${leadLabel}`),\n text:\n `They referred us to someone else and included their contact. Sending to a new address isn't automated yet - copy the draft below into a fresh email.\\n\\n` +\n `New contact: ${analysis.contact_details ?? \"(see message)\"}\\nCampaign: ${campaign.name}\\nPersona they know: ${senderName}\\n\\n` +\n `Ready-to-send draft:\\n${draft || \"(draft failed - use the Wrong Person template)\"}\\n\\n` +\n `Their message:\\n${reply.text_body ?? \"\"}` + muteFooter(lead.email),\n });\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"alerted\",\n reply_body: draft || undefined, detail: analysis.contact_details ?? analysis.reasoning,\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"alerted (intro drafted for Sam)\" });\n }\n\n // Everything else (referral-ask, already_booked, booked, counter, propose): send.\n if (!draft) {\n // soft-draft failure with nothing to send - nothing sensible to do beyond log\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"skipped\", detail: \"no draft after repair (soft)\",\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: \"skipped (no draft)\" });\n }\n\n const actionName = sched?.mode === \"booked\"\n ? \"booked_and_sent\"\n : sched?.mode === \"rescheduled\"\n ? \"rescheduled_and_sent\"\n : sched?.mode === \"cancelled\"\n ? \"cancelled_and_sent\"\n : \"sent\";\n if (dryRun) {\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: \"drafted\", reply_body: draft,\n detail: `${sched?.detail ?? \"\"} | tz: ${sched?.tz ?? \"n/a\"} (${sched?.tzEvidence ?? \"\"})`.trim(),\n });\n await markDone();\n return Response.json({\n ok: true, category: analysis.category, action: \"drafted (dry run)\",\n mode: sched?.mode, draft,\n });\n }\n\n await sendBisonReply(reply.id, draft, sender.id, lead.email, htmlAnchors);\n await logAction({\n lead_id: lead.id, lead_email: lead.email, reply_id: reply.id, reply_uuid: reply.uuid,\n category: analysis.category, action: actionName, reply_body: draft,\n detail: `${sched?.detail ?? analysis.reasoning} | tz: ${sched?.tz ?? \"n/a\"} (${sched?.tzEvidence ?? \"\"})`.trim(),\n });\n await markDone();\n return Response.json({ ok: true, category: analysis.category, action: actionName, mode: sched?.mode });\n } catch (err) {\n console.error(\"handler error:\", err);\n if (processedUuid) {\n await sqlite.execute({\n sql: `DELETE FROM processed_replies WHERE reply_uuid = ? AND status = 'processing'`,\n args: [processedUuid],\n }).catch(() => {});\n }\n await logAction({ action: \"error\", detail: String(err).slice(0, 1000) });\n await safeEmail({\n subject: tag(\"Error - agent did nothing for this event\"),\n text: `The reply agent hit an error and took no action.\\n\\n${String(err)}\\n\\nPayload snippet:\\n${JSON.stringify(body).slice(0, 1500)}`,\n });\n return Response.json({ ok: false, error: String(err).slice(0, 300) }, { status: 500 });\n }\n}\n","label",[136],"6eafbb06-64e4-4a3a-bb3f-6d7155a8df5a","test-ghl.ts","54d1c1fc-7bf7-11f1-bb54-1607ee4eb77e",82,["D",1784420098092],8,"// Token-gated test + admin endpoint. Never touches Bison. ?action=menus\n// (default) computes proposal menus from the live calendar backend;\n// ?action=calmenus does the same explicitly against Calendly. booktest (GHL)\n// and caltest (Calendly twin) run full book+cancel cycles on sam@luminamg.com\n// - safe, but momentary real appointments + emails. ?action=mutes is Sam's\n// bookmarkable mute admin page. ?action=calsetup discovers Calendly URIs and\n// (&sub=1) creates the webhook subscription whose receiver is\n// ?action=calwebhook (alerts Sam on prospect self-cancels).\nimport {\n counterProposal,\n isoForLocal,\n loadSlots,\n localNow,\n menuForPrompt,\n standardProposal,\n windowProposal,\n} from \"./scheduling.ts\";\nimport {\n bookAppointment,\n cancelAppointment,\n getUpcomingAppointment,\n upsertContact,\n} from \"./ghl.ts\";\nimport { checkDueEmails, coldEmailPipeline, listMutes, muteLead, reconcile, unmuteLead, websiteIntel } from \"./crm.ts\";\nimport {\n bookWithRecovery as calBook,\n calendarConfigured,\n cancelAppointment as calCancel,\n createWebhookSubscription,\n fetchFreeSlots as calFreeSlots,\n getScheduledEvent,\n getUpcomingAppointment as calUpcoming,\n listWebhookSubscriptions,\n orgUri,\n storeSigningKey,\n verifyWebhookSignature,\n} from \"./calendly.ts\";\nimport { email as valEmail } from \"https://esm.town/v/std/email\";\n\n// Per-booking .ics for the \"Add to Apple / Outlook Calendar\" link in booking\n// confirmations. Tokenless by design (prospects click it): the event uuid is\n// the secret, and the file holds only what their confirmation email already\n// contains. Calendly hosts no linkable .ics (verified 2026-07-18).\nasync function icsResponse(eventUuid: string): Promise {\n if (!/^[0-9a-f-]{36}$/.test(eventUuid)) return new Response(\"bad id\", { status: 400 });\n try {\n const ev = await getScheduledEvent(eventUuid);\n if (!ev) return new Response(\"not found\", { status: 404 });\n const fmt = (iso: string) => new Date(iso).toISOString().replace(/[-:]/g, \"\").replace(/\\.\\d+/, \"\");\n const meet = ev.location?.join_url ?? \"\";\n const cancelled = ev.status !== \"active\";\n const ics = [\n \"BEGIN:VCALENDAR\",\n \"VERSION:2.0\",\n \"PRODID:-//Lumina//Pilot Campaign Call//EN\",\n cancelled ? \"METHOD:CANCEL\" : \"METHOD:PUBLISH\",\n \"BEGIN:VEVENT\",\n `UID:lumina-${eventUuid}@luminamg.com`,\n `DTSTAMP:${fmt(new Date().toISOString())}`,\n `DTSTART:${fmt(ev.start_time)}`,\n `DTEND:${fmt(ev.end_time)}`,\n `SUMMARY:Pilot Campaign Call with Lumina${cancelled ? \" (CANCELLED)\" : \"\"}`,\n meet ? `DESCRIPTION:Google Meet: ${meet}` : \"DESCRIPTION:Details in your confirmation email.\",\n meet ? `LOCATION:${meet}` : \"LOCATION:Google Meet\",\n cancelled ? \"STATUS:CANCELLED\" : \"STATUS:CONFIRMED\",\n \"END:VEVENT\",\n \"END:VCALENDAR\",\n ].join(\"\\r\\n\");\n return new Response(ics, {\n headers: {\n \"Content-Type\": \"text/calendar; charset=utf-8\",\n \"Content-Disposition\": 'attachment; filename=\"lumina-call.ics\"',\n },\n });\n } catch (e) {\n console.error(\"ics failed:\", e);\n return new Response(\"unavailable\", { status: 500 });\n }\n}\n\nexport default async function (req: Request): Promise {\n const url = new URL(req.url);\n if (url.searchParams.get(\"action\") === \"ics\") {\n return await icsResponse(url.searchParams.get(\"e\") ?? \"\");\n }\n if (url.searchParams.get(\"token\") !== Deno.env.get(\"WEBHOOK_SECRET\")) {\n return new Response(\"unauthorized\", { status: 401 });\n }\n const action = url.searchParams.get(\"action\") ?? \"menus\";\n const out: Record = {\n ghlKeySet: !!Deno.env.get(\"GHL_API_KEY\"),\n calendlyConfigured: calendarConfigured(),\n };\n try {\n if (action === \"menus\") {\n for (const tz of [\"America/New_York\", \"America/Chicago\", \"America/Los_Angeles\"]) {\n const slots = await loadSlots(tz);\n const dates = Object.keys(slots).sort();\n const date = url.searchParams.get(\"date\") ?? dates[1];\n const mins = parseInt(url.searchParams.get(\"mins\") ?? \"600\", 10); // default 10:00\n const minDate = url.searchParams.get(\"minDate\") ?? \"\";\n out[tz] = {\n localNow: localNow(tz),\n days: dates.map((d) => `${d}: ${slots[d].length} slots`),\n standard: menuForPrompt(standardProposal(slots, tz, minDate)),\n counterFor: `${date} @ ${mins} mins`,\n counter: menuForPrompt(counterProposal(slots, tz, date, mins)),\n window_afternoon: menuForPrompt(windowProposal(slots, tz, date, [12 * 60, 17 * 60])),\n };\n }\n } else if (action === \"booktest\") {\n const tz = \"America/New_York\";\n const slots = await loadSlots(tz);\n const dates = Object.keys(slots).sort();\n const lastDate = dates[dates.length - 1];\n const slot = slots[lastDate][slots[lastDate].length - 1];\n const contactId = await upsertContact({\n email: \"sam@luminamg.com\",\n firstName: \"ReplyAgent\",\n lastName: \"IntegrationTest\",\n clinicName: \"Test Dental Clinic\",\n });\n const existing = await getUpcomingAppointment(contactId);\n let booked: string | null = null;\n let during: unknown = null;\n if (!existing) {\n const appt = await bookAppointment({\n contactId,\n startTime: slot,\n title: \"[TEST] scheduling engine e2e - auto-cancelled\",\n });\n booked = appt.id;\n during = await getUpcomingAppointment(contactId);\n await cancelAppointment(appt.id);\n }\n out.booktest = {\n slot,\n contactId,\n skippedBecauseExisting: !!existing,\n booked,\n upcomingDuringBooking: during,\n upcomingAfterCancel: await getUpcomingAppointment(contactId),\n };\n } else if (action === \"crmcheck\") {\n // Scope + pipeline discovery report. oppScope flips to true once Sam\n // adds the opportunities scopes to the Private Integration.\n try {\n const pipe = await coldEmailPipeline();\n out.oppScope = true;\n out.pipeline = pipe ?? \"scope OK but 'Cold Email' pipeline / Positive / Booked stages not found by name\";\n } catch (e) {\n out.oppScope = false;\n out.error_detail = String(e).slice(0, 300);\n }\n } else if (action === \"reconcile\") {\n // Manual reconciler run (bypasses the 15-min throttle). dry=1 previews.\n // Also the target of the GHL booking/cancellation workflows - piggyback\n // the due-email check so workflow pings deliver reminders too.\n out.dueEmails = await checkDueEmails(url.searchParams.get(\"dry\") === \"1\").catch((e) => String(e));\n out.reconcile = await reconcile(url.searchParams.get(\"dry\") === \"1\", true);\n } else if (action === \"intel\") {\n // Website-intel probe: ?site=https://example.com\n const site = url.searchParams.get(\"site\") ?? \"\";\n out.intel = await websiteIntel({ custom_variables: [{ name: \"website\", value: site }] });\n } else if (action === \"mute\" || action === \"unmute\" || action === \"mutes\") {\n // Per-lead mute admin (Sam takes a thread over manually - the webhook\n // stays silent for muted leads). mute/unmute then redirect back to the\n // bookmarkable ?action=mutes page.\n const token = url.searchParams.get(\"token\") ?? \"\";\n const mutesUrl = `${url.origin}${url.pathname}?action=mutes&token=${encodeURIComponent(token)}`;\n if (action !== \"mutes\") {\n const email = (url.searchParams.get(\"email\") ?? \"\").trim().toLowerCase();\n if (!email.includes(\"@\")) {\n return Response.json({ error: \"valid ?email= required\" }, { status: 400 });\n }\n if (action === \"mute\") await muteLead(email);\n else await unmuteLead(email);\n return Response.redirect(mutesUrl, 302);\n }\n const esc = (s: string) => s.replace(/&/g, \"&\").replace(/\n `

${esc(m.email)} muted ${esc(m.muted_at.slice(0, 16).replace(\"T\", \" \"))} UTC ` +\n `Unmute

`\n ).join(\"\");\n const html = `\nReply Agent - Muted Leads\n\n

Muted leads

Muted leads get no auto-replies - you handle those threads yourself (you still get an\nalert email when they reply). Matching covers the campaign lead email and the address they\nreply from; muting either works.

\n\n

(none muted right now)

`;\n return new Response(html, { headers: { \"Content-Type\": \"text/html; charset=utf-8\" } });\n } else if (action === \"calsetup\") {\n // Calendly discovery + one-time webhook subscription. &sub=1 creates the\n // invitee.created/canceled subscription pointing back at ?action=calwebhook;\n // the generated signing key is shown ONCE - store it in the val env as\n // CALENDLY_WEBHOOK_SIGNING_KEY immediately (lost key = delete+recreate).\n out.org = await orgUri();\n out.eventType = Deno.env.get(\"CALENDLY_EVENT_TYPE\") ?? \"(unset)\";\n const subs = await listWebhookSubscriptions();\n out.subscriptions = subs.map((s: any) => ({ url: s.callback_url, state: s.state, events: s.events }));\n if (url.searchParams.get(\"sub\") === \"1\") {\n const cb = `${url.origin}${url.pathname}?action=calwebhook&token=${\n encodeURIComponent(Deno.env.get(\"WEBHOOK_SECRET\") ?? \"\")\n }`;\n const existing = subs.find((s: any) => s.callback_url === cb && s.state === \"active\");\n if (existing) {\n out.subscription = \"already active - signing key persisted in crm_state at creation\";\n } else {\n const keyBytes = new Uint8Array(32);\n crypto.getRandomValues(keyBytes);\n const signingKey = Array.from(keyBytes).map((b) => b.toString(16).padStart(2, \"0\")).join(\"\");\n await storeSigningKey(signingKey); // BEFORE subscribing: first webhook must verify\n const created = await createWebhookSubscription(cb, signingKey);\n out.subscription = created?.uri ?? created;\n out.signingKey = \"generated + stored in crm_state automatically\";\n }\n }\n } else if (action === \"calwebhook\") {\n // Calendly webhook receiver (invitee.created / invitee.canceled). The\n // registered URL carries ?token= (gate above) + HMAC signature (here).\n // Both events force a reconcile -> near-real-time pipeline sync,\n // replacing the old GHL workflow pings.\n const raw = await req.text();\n const ok = await verifyWebhookSignature(raw, req.headers.get(\"Calendly-Webhook-Signature\"));\n if (!ok) return Response.json({ ok: false, error: \"bad signature\" }, { status: 401 });\n let evt: any = {};\n try {\n evt = JSON.parse(raw);\n } catch { /* signature passed; treat as ping */ }\n const kind = evt?.event ?? \"unknown\";\n const p = evt?.payload ?? {};\n const dryRun = (Deno.env.get(\"DRY_RUN\") ?? \"true\") !== \"false\";\n const due = await checkDueEmails(dryRun).catch((e) => String(e));\n const rec = await reconcile(dryRun, true).catch((e) => String(e));\n if (kind === \"invitee.canceled\" && p?.cancellation?.canceler_type !== \"host\") {\n // Quiet self-cancel through Calendly's links - no email ever reaches\n // Bison, so Sam gets told here. Agent-initiated (host) cancels stay silent.\n await valEmail({\n subject: `[Reply Agent] Prospect canceled via Calendly - ${p?.email ?? \"unknown\"}`,\n text: `${p?.name ?? \"Unknown\"} <${p?.email ?? \"?\"}> canceled their booking themselves.\\n\\n` +\n `Event: ${p?.scheduled_event?.name ?? \"?\"} at ${p?.scheduled_event?.start_time ?? \"?\"} UTC\\n` +\n `Reason given: ${p?.cancellation?.reason || \"(none)\"}\\n\\n` +\n `Pipeline sync has already run. The thread in Bison is untouched - decide there whether to save it.`,\n }).catch((e) => console.error(\"cancel alert failed:\", e));\n }\n return Response.json({ ok: true, handled: kind, dueEmails: due, reconcile: rec });\n } else if (action === \"calmenus\") {\n // Calendly-backed proposal menus (works pre-cutover; diff against\n // ?action=menus which follows loadSlots' current backend).\n for (const tz of [\"America/New_York\", \"America/Chicago\", \"America/Los_Angeles\"]) {\n const slots = await calFreeSlots(6, tz);\n const dates = Object.keys(slots).sort();\n out[tz] = {\n localNow: localNow(tz),\n days: dates.map((d) => `${d}: ${slots[d].length} slots`),\n standard: menuForPrompt(standardProposal(slots, tz, \"\")),\n counter: menuForPrompt(counterProposal(slots, tz, dates[1] ?? \"\", 600)),\n };\n }\n } else if (action === \"caltest\") {\n // Full Calendly book+cancel cycle on the SECRET TWIN for sam@luminamg.com.\n // Creates a momentary real booking: Sam gets a confirmation + cancellation\n // email - expected. Validates: availability, bookWithRecovery, required\n // question answers, upcoming lookup by email, cancellation.\n const tz = \"America/New_York\";\n const slots = await calFreeSlots(6, tz);\n const dates = Object.keys(slots).sort();\n const lastDate = dates[dates.length - 1];\n const slot = slots[lastDate][slots[lastDate].length - 1];\n const existing = await calUpcoming([\"sam@luminamg.com\"]);\n if (existing) {\n out.caltest = { skippedBecauseExisting: existing };\n } else {\n const booked = await calBook({\n email: \"sam@luminamg.com\",\n name: \"ReplyAgent IntegrationTest\",\n startTime: slot,\n timezone: tz,\n phone: \"+19257309373\",\n clinic: \"Test Dental Clinic\",\n });\n let during: unknown = null;\n if (booked.outcome === \"booked\") {\n during = await calUpcoming([\"sam@luminamg.com\"]);\n await calCancel(booked.apptId, \"caltest auto-cancel - disregard\");\n }\n out.caltest = {\n slot,\n booked,\n upcomingDuringBooking: during,\n upcomingAfterCancel: await calUpcoming([\"sam@luminamg.com\"]),\n };\n }\n } else if (action === \"iso\") {\n const date = url.searchParams.get(\"date\") ?? \"2026-07-20\";\n const mins = parseInt(url.searchParams.get(\"mins\") ?? \"600\", 10);\n out.iso = {};\n for (const tz of [\"America/New_York\", \"America/Chicago\", \"America/Denver\", \"America/Phoenix\", \"America/Los_Angeles\"]) {\n (out.iso as Record)[tz] = isoForLocal(date, mins, tz);\n }\n } else {\n return Response.json({ error: `unknown action ${action}` }, { status: 400 });\n }\n return Response.json(out);\n } catch (e) {\n return Response.json({ ...out, error: String(e) }, { status: 500 });\n }\n}\n",[148],"7815d70a-2ec3-46d1-9d0d-87aabc8b72cf","scheduling.ts","5484ccb2-7bf7-11f1-bbeb-1607ee4eb77e",75,["D",1784412835045],3,"// Deterministic proposal engine for the reply agent. All decisions about WHICH\n// times to offer live here (code, testable), not in the model - the model only\n// phrases them. Rules from Sam, 2026-07-09:\n// - Times are always the prospect's local timezone; slots fetched from GHL in it.\n// - Ideal windows: 10:00, 12:00, 13:00, 16:00. Same-day special: 15:00 (only\n// when their reply lands before 10:00 their time -> 3-day shape).\n// - A taken time slides to the nearest 30-min interval, later preferred.\n// - Proposals use business days only (skip Sat/Sun); weekend times are still\n// BOOKED if the prospect asks for one - that rule is enforced by the caller\n// simply checking availability, not here.\n// - Counter-proposals (\"IF SLOT ISN'T OPEN\" template): proposed day at nearest\n// time, then two more days at their-time-if-free (else nearest) + an\n// alternate ideal, giving spread for flexibility.\n\nimport { fetchFreeSlots } from \"./calendly.ts\";\n\nexport type DayOptions = {\n date: string; // \"YYYY-MM-DD\" (prospect-local)\n weekday: string; // \"Friday\"\n times: string[]; // human labels: \"10am\", \"1:30pm\" (prospect-local)\n slotIso: Record; // label -> exact ISO slot for booking\n};\n\nexport type ProposalMenu = {\n timezone: string;\n tzAbbr: string; // \"PT\", \"ET\", ...\n days: DayOptions[];\n};\n\nconst IDEAL_MINUTES = [10 * 60, 12 * 60, 13 * 60, 16 * 60]; // 10a, 12p, 1p, 4p\nconst SAME_DAY_MINUTES = 15 * 60; // 3pm\n\n// ---------- small time helpers (no Date-math across zones; slots arrive\n// already rendered in the requested timezone, e.g. \"2026-07-10T13:15:00-07:00\") ----------\n\nexport function slotDate(slot: string): string {\n return slot.slice(0, 10);\n}\n\nexport function slotMinutes(slot: string): number {\n return parseInt(slot.slice(11, 13), 10) * 60 + parseInt(slot.slice(14, 16), 10);\n}\n\nexport function minutesToLabel(mins: number): string {\n const h24 = Math.floor(mins / 60);\n const m = mins % 60;\n const ampm = h24 < 12 ? \"am\" : \"pm\";\n const h = h24 % 12 === 0 ? 12 : h24 % 12;\n return m === 0 ? `${h}${ampm}` : `${h}:${String(m).padStart(2, \"0\")}${ampm}`;\n}\n\nexport function tzAbbreviation(timezone: string): string {\n const parts = new Intl.DateTimeFormat(\"en-US\", { timeZone: timezone, timeZoneName: \"short\" })\n .formatToParts(new Date());\n const raw = parts.find((p) => p.type === \"timeZoneName\")?.value ?? \"\";\n const m = raw.match(/^([A-Z])[DS]T$/); // PDT/PST -> PT, EDT/EST -> ET ...\n if (m) return `${m[1]}T`;\n const m2 = raw.match(/^([A-Z]{2,3})[DS]T$/); // AKDT -> AKT, HST -> HT\n if (m2) return `${m2[1]}T`;\n return raw;\n}\n\nexport function weekdayOf(date: string, timezone: string): string {\n return new Intl.DateTimeFormat(\"en-US\", { timeZone: timezone, weekday: \"long\" })\n .format(new Date(`${date}T12:00:00Z`));\n}\n\n// ISO start-time for an arbitrary prospect-local date+time with the correct\n// UTC offset for THAT date in that zone (DST-aware). Used for out-of-window\n// direct bookings, where no free-slots string exists to copy from.\nexport function isoForLocal(date: string, minutes: number, timezone: string): string {\n const probe = new Date(`${date}T12:00:00Z`);\n const raw = new Intl.DateTimeFormat(\"en-US\", { timeZone: timezone, timeZoneName: \"longOffset\" })\n .formatToParts(probe).find((p) => p.type === \"timeZoneName\")?.value ?? \"GMT+00:00\";\n const offset = raw === \"GMT\" ? \"+00:00\" : raw.replace(\"GMT\", \"\");\n const h = String(Math.floor(minutes / 60)).padStart(2, \"0\");\n const m = String(minutes % 60).padStart(2, \"0\");\n return `${date}T${h}:${m}:00${offset}`;\n}\n\nfunction isWeekend(date: string): boolean {\n const wd = weekdayOf(date, \"UTC\");\n return wd === \"Saturday\" || wd === \"Sunday\";\n}\n\n// Prospect-local \"now\" as {date, minutes}.\nexport function localNow(timezone: string): { date: string; minutes: number } {\n const s = new Date().toLocaleString(\"sv-SE\", { timeZone: timezone }); // \"YYYY-MM-DD HH:mm:ss\"\n return { date: s.slice(0, 10), minutes: parseInt(s.slice(11, 13), 10) * 60 + parseInt(s.slice(14, 16), 10) };\n}\n\n// ---------- slot lookup ----------\n\nexport type SlotsByDay = Record; // date -> ISO slots (prospect tz)\n\nexport async function loadSlots(timezone: string): Promise {\n return await fetchFreeSlots(6, timezone);\n}\n\nexport function findExactSlot(slots: SlotsByDay, date: string, minutes: number): string | null {\n return (slots[date] ?? []).find((s) => slotMinutes(s) === minutes) ?? null;\n}\n\n// Nearest available slot to a target time on a date, ON the 30-min grid,\n// later preferred on ties and generally (10:00 taken -> 10:30 before 9:30).\nexport function nearestSlot(slots: SlotsByDay, date: string, target: number): string | null {\n const candidates = (slots[date] ?? []).filter((s) => slotMinutes(s) % 30 === 0);\n if (!candidates.length) return null;\n let best: string | null = null;\n let bestScore = Infinity;\n for (const s of candidates) {\n const diff = slotMinutes(s) - target;\n // distance dominates; earlier slots get a half-step penalty so the later\n // of two equidistant options always wins\n const score = Math.abs(diff) * 2 + (diff < 0 ? 1 : 0);\n if (score < bestScore) {\n bestScore = score;\n best = s;\n }\n }\n return best;\n}\n\n// An available ideal-window time on a date, excluding times already used and\n// preferring spread: if the day's existing pick is in the morning, prefer an\n// afternoon ideal and vice versa.\nexport function idealSlot(\n slots: SlotsByDay,\n date: string,\n usedMinutes: number[],\n): string | null {\n const anchor = usedMinutes.length ? usedMinutes[usedMinutes.length - 1] : null;\n const order = [...IDEAL_MINUTES].sort((a, b) => {\n if (anchor === null) return a - b;\n return Math.abs(b - anchor) - Math.abs(a - anchor); // farthest from anchor first\n });\n for (const target of order) {\n // exact ideal time, else its nearest 30-min neighbor\n const s = findExactSlot(slots, date, target) ?? nearestSlot(slots, date, target);\n if (s && !usedMinutes.includes(slotMinutes(s))) return s;\n }\n return null;\n}\n\n// Next N business dates strictly after `fromDate` (and never before `minDate`,\n// the prospect's stated availability floor, e.g. \"next week\") that have slots.\nfunction nextBusinessDates(slots: SlotsByDay, fromDate: string, n: number, minDate = \"\"): string[] {\n const dates = Object.keys(slots)\n .filter((d) => d > fromDate && d >= minDate && !isWeekend(d))\n .sort();\n return dates.slice(0, n);\n}\n\nfunction buildDay(slots: SlotsByDay, timezone: string, date: string, isoTimes: string[]): DayOptions {\n const uniq = [...new Set(isoTimes)];\n const labels = uniq.map((s) => minutesToLabel(slotMinutes(s)));\n const slotIso: Record = {};\n uniq.forEach((s, i) => (slotIso[labels[i]] = s));\n return { date, weekday: weekdayOf(date, timezone), times: labels, slotIso };\n}\n\n// ---------- the two proposal shapes ----------\n\n// Scenario 2: interested, no time given. Before 10:00 their time -> same day at\n// 3pm + two business days (10a/1p then 12p/4p shapes via ideal windows).\n// Otherwise two business days x two times. `minDate` floors every proposal to\n// the prospect's stated window (\"some time next week\").\nexport function standardProposal(slots: SlotsByDay, timezone: string, minDate = \"\"): ProposalMenu {\n const now = localNow(timezone);\n const days: DayOptions[] = [];\n\n if (now.minutes < 10 * 60 && !isWeekend(now.date) && now.date >= minDate) {\n const s = findExactSlot(slots, now.date, SAME_DAY_MINUTES) ??\n nearestSlot(slots, now.date, SAME_DAY_MINUTES);\n if (s) days.push(buildDay(slots, timezone, now.date, [s]));\n }\n\n const PREFERRED: number[][] = [[10 * 60, 13 * 60], [12 * 60, 16 * 60]]; // day1: 10a/1p, day2: 12p/4p\n let dayIdx = 0;\n for (const date of nextBusinessDates(slots, now.date, 2, minDate)) {\n const targets = PREFERRED[Math.min(dayIdx, 1)];\n dayIdx++;\n const picked: string[] = [];\n for (const t of targets) {\n const s = findExactSlot(slots, date, t) ?? nearestSlot(slots, date, t);\n if (s && !picked.includes(s)) picked.push(s);\n }\n if (picked.length) days.push(buildDay(slots, timezone, date, picked));\n }\n\n return { timezone, tzAbbr: tzAbbreviation(timezone), days };\n}\n\n// Scenario 1 counter: their proposed date+time wasn't bookable. Proposed day at\n// nearest alternative, then two more days at their-time-if-free (else nearest)\n// plus an alternate ideal for spread. Weekend proposed date is honored for its\n// own day (they picked it) but follow-up days are business days.\nexport function counterProposal(\n slots: SlotsByDay,\n timezone: string,\n proposedDate: string,\n proposedMinutes: number,\n): ProposalMenu {\n const days: DayOptions[] = [];\n\n const near = nearestSlot(slots, proposedDate, proposedMinutes);\n if (near) {\n const ideal = idealSlot(slots, proposedDate, [slotMinutes(near)]);\n days.push(buildDay(slots, timezone, proposedDate, ideal ? [near, ideal] : [near]));\n }\n\n for (const date of nextBusinessDates(slots, proposedDate, 2)) {\n const first = findExactSlot(slots, date, proposedMinutes) ??\n nearestSlot(slots, date, proposedMinutes);\n if (!first) continue;\n const ideal = idealSlot(slots, date, [slotMinutes(first)]);\n days.push(buildDay(slots, timezone, date, ideal ? [first, ideal] : [first]));\n }\n\n return { timezone, tzAbbr: tzAbbreviation(timezone), days };\n}\n\n// Vague window (\"Tuesday?\", \"tomorrow afternoon\"): two available times inside\n// the window, spread apart. window: [startMinutes, endMinutes].\nexport function windowProposal(\n slots: SlotsByDay,\n timezone: string,\n date: string,\n window: [number, number],\n): ProposalMenu {\n const inWindow = (slots[date] ?? []).filter((s) => {\n const m = slotMinutes(s);\n return m >= window[0] && m <= window[1] && m % 30 === 0;\n });\n const days: DayOptions[] = [];\n if (inWindow.length) {\n const picks = inWindow.length > 1 ? [inWindow[0], inWindow[inWindow.length - 1]] : [inWindow[0]];\n days.push(buildDay(slots, timezone, date, picks));\n }\n return { timezone, tzAbbr: tzAbbreviation(timezone), days };\n}\n\n// Render a menu for the drafting prompt: verified times the model may use, and\n// nothing else.\nexport function menuForPrompt(menu: ProposalMenu): string {\n if (!menu.days.length) return \"(no verified times available - offer the calendar link only)\";\n return menu.days\n .map((d) => `${d.weekday} (${d.date}): ${d.times.map((t) => `${t} ${menu.tzAbbr}`).join(\" or \")}`)\n .join(\"\\n\");\n}\n",[156],"9d5a5aff-9e58-43ac-9f42-85a8b9e95ca0","reconcile-cron.ts","a6a40df4-80bc-11f1-9084-1607ee4eb77e",54,["D",1784168316285],"interval","// Interval val: precision backstop for time-sensitive CRM work - scheduled\n// reminder emails (checkDueEmails) and the calendar/pipeline reconciler.\n// The same checks piggyback on webhook traffic; this val guarantees they run\n// even in quiet periods. SCHEDULE IS SET IN THE VAL.TOWN DASHBOARD (this\n// file just needs \"every 15 min\" or faster - it's cheap and idempotent).\nimport { checkDueEmails, reconcile } from \"./crm.ts\";\n\nexport default async function () {\n const dryRun = (Deno.env.get(\"DRY_RUN\") ?? \"true\") !== \"false\";\n console.log(\"cron emails:\", await checkDueEmails(dryRun).catch((e) => String(e)));\n console.log(\"cron reconcile:\", await reconcile(dryRun).catch((e) => String(e)));\n}\n",[164],"789af663-1b30-4014-8b35-296f4cee1eeb",["D",1786127124429],1,[28],"2742f830-26a6-4388-8778-e4ac91806f66","playbook.ts","a62caf46-7a9c-11f1-b4fc-1607ee4eb77e",92,["D",1786127123928],19,"// AUTO-GENERATED by scripts/deploy.py from skill/*.md - do not edit here.\nexport const PLAYBOOK_A = \"\\n# Voice & Style (derived from templates.md + swipefile.md, 2026-07-14)\\n\\nSource of truth: templates.md and swipefile.md - Sam's real, tested replies.\\nEverything here is distilled FROM those two files; when in doubt, reread them and\\nmatch. This replaces all older style guidance - including the 2026-07-08\\n\\\"educate them a little\\\" doctrine. Nobody wants to read a blocky three-paragraph\\nresponse: replies are SHORT, straight to the point, simple sentence structures.\\n\\n## The response shape (every positive reply follows it)\\n\\n1. **Greeting**: \\\"Hey,\\\" - with their name if it's in their reply or signature\\n (\\\"Hey Monica,\\\" / \\\"Hey Dr. Song,\\\"). Then a short remark that fits the thread:\\n \\\"great to hear!\\\" / \\\"totally fair question.\\\" / \\\"good question.\\\" / \\\"of course.\\\" /\\n \\\"happy to answer your questions.\\\" / \\\"thanks for getting back to me.\\\"\\n2. **Answer their question(s) directly** - short, plain terms, simple sentences,\\n one idea per paragraph, from facts.md. The new templates are the target\\n length; never pad an answer to seem thorough.\\n3. **Work the free campaign in** - it's the answer to almost every worry:\\n \\\"you get the same fully-managed service we provide to our ongoing clients -\\n the only difference is that we waive our service fee entirely. If we don't\\n get at least {LOWER_BOUND} {TREATMENT_TYPE} patients showing up to their\\n appointments, we reimburse your ad spend.\\\" (The guarantee sentence follows\\n the offer-variant rules in templates.md - mirror the first email; 2026-07-16\\n era threads keep their old \\\"cases\\\" sentence.)\\n4. **Only after answering, pivot to the CTA**: \\\"Let's do a quick call so we can\\n give you a rundown of the free campaign, what results you can expect, and\\n get the info we need to get your campaign live.\\\" Adapt the lead-in to the\\n thread (\\\"As for the call, ...\\\" if they asked for one; \\\"During the call,\\n I'll...\\\" if they already agreed to book). Never attach a duration to the\\n call - \\\"quick\\\" is the whole promise (the old \\\"15-minute\\\" is retired\\n 2026-07-15).\\n5. **Times + calendar**: propose the verified times by day (\\\"How about Friday\\n at 10am or 1pm ET, or Monday at 12pm or 4pm ET?\\\"), weekday names never\\n \\\"today\\\"/\\\"tomorrow\\\", always the prospect's timezone label. Say \\\"Calendly\\\", not\\n \\\"calendar\\\" - people know Calendly is a scheduler: \\\"Or if it's easier,\\n here's our Calendly: [link]\\\" (only where the template calls for the inline\\n line; the CTA block's self-book line is appended by the system). The old\\n phone-ask line is RETIRED (2026-07-16); the two new phone asks are\\n system-gated - include only when instructed.\\n6. **Sign-off**: \\\"Thanks,\\\" on its own line, then the sender persona's first name\\n (the pen name on the sending account - NEVER any other name), then exactly\\n one signature line:\\n SDR, Lumina\\n No website in the signature (removed 2026-07-14, tentative). The signature\\n goes on EVERY reply we send, before any P.S. line.\\n7. **P.S.**: at most one, ONLY the two situations in templates.md (just booked\\n them in / first time offering times to a fully-interested reply). Everything\\n else gets NO P.S. - the old \\\"not a sales call\\\" P.S. is retired.\\n\\n## Back-and-forth mode (any reply after our first)\\n\\nOnce we're trading messages about times, drop the full template structure - it\\nreads automated if every email has the same skeleton. Go short and loose,\\none-liner-ish: \\\"Hey - I've got another call at 2pm, how about Thursday at 2pm or\\n4pm ET? Here's the Calendly if easier: [link]\\\". Same voice, same verified-times\\nrule, no repeated P.S. boilerplate.\\n\\nCorollary: read the whole thread every time, and never let two of our messages\\nin one thread share the same skeleton (greeting shape, paragraph order, CTA\\nphrasing). Vary structure as you go - naturally, and always short.\\nSign-offs: \\\"Thanks,\\\" normally; \\\"Take care,\\\" on reschedule/cancel replies.\\nThe signature (persona name, then \\\"SDR, Lumina\\\") stays on every message, even\\nthe short loose ones - a repeated signature is normal email behavior and does\\nNOT count as a repeated skeleton. ONE exception (2026-07-18): once they're\\nBOOKED, back-and-forth replies sign with just \\\"Thanks,\\\" + first name - no\\n\\\"SDR, Lumina\\\", no self-booking link (the system enforces this too).\\n\\n## Tone rules\\n\\n- **Short and direct.** Simple sentence structures, one idea per paragraph.\\n Every sentence either answers something, disarms a concern, or moves to the\\n CTA - anything else gets cut. When a template covers it, the template IS the\\n right length.\\n- **Plain natural human language.** Write like Sam texts a colleague: \\\"Hey -\\n works for me!\\\", \\\"totally fair question\\\", \\\"you pay nothing upfront\\\". A busy\\n dentist gets it on the first read. No jargon, no ad-tech vocabulary.\\n- **Warm, confident, zero pressure.** The free campaign carries the sale; the\\n reply never pushes. \\\"before you decide if you want to continue.\\\"\\n- **Empathy for skepticism**: \\\"I'm guessing you've been burned by other\\n marketers before\\\" - name the fear, then answer it with results-first.\\n- **Money is discussed plainly - but never volunteered.** \\\"$497/month or\\n $2,197 for six months paid upfront - that's 25% off, about 6 weeks free.\\\"\\n The recommended ad budget ({CURRENCY}500 to Meta) and the\\n {CURRENCY}150-per-patient reimbursement rate are ONLY for explicit\\n ad-spend/reimbursement questions - the fee-waiver block's short guarantee\\n sentence is the most detail anyone else gets.\\n- **Zero friction between a positive reply and a booked call.** Never\\n volunteer numbers, durations, or commitments they didn't ask about - each\\n one is a fresh reason to pause. The goal of every reply is getting them\\n face-to-face on a call; details can live there.\\n- **Consistency bias**: when they asked for something (a call, a booking), remind\\n them: \\\"As for the call, ...\\\".\\n- **Match their energy - and the cold email's CTA.** The cold email ends with\\n a question (\\\"would you be interested / like to hear more?\\\"), so a bare\\n \\\"Yes\\\" / \\\"Yeah sure\\\" / \\\"I'd be interested\\\" means \\\"tell me more\\\" - it gets the\\n Tell Me More template, not a bare ask-for-call. An explicit \\\"schedule the\\n call\\\" gets the short Interested/Ask for Call shape.\\n\\n## Mechanics\\n\\n- PLAIN TEXT only - no HTML tags, no markdown. Real line breaks: blank line\\n between paragraphs; \\\"Thanks,\\\" on its own line, name on the next, then\\n \\\"SDR, Lumina\\\" on the next - no blank lines inside the block.\\n- Simple \\\"- \\\" bullet lists are fine where a template uses them (the Process\\n template) - they're in Sam's tested copy. Don't add bullets elsewhere.\\n- Links are bare URLs pasted inline (calendar), except the website in body\\n copy, which is written exactly \\\"luminamg.com\\\" - no https://, no trailing\\n slash (the send layer hyperlinks it automatically).\\n- Never output an em-dash or en-dash character; use \\\" - \\\" (or \\\"--\\\" like Sam\\n occasionally does).\\n- \\\":)\\\" is fine where Sam's copy uses it. Don't add emoji anywhere else.\\n- No corporate stiffness (\\\"I hope this finds you well\\\", \\\"at your convenience\\\",\\n \\\"rest assured\\\", \\\"Furthermore/Moreover/Additionally\\\").\\n- Prefer the templates' exact sentences over paraphrasing them - they're tested\\n copy. Adapt only the seams (greeting, transitions, what fits the thread and the\\n first cold email's framing) - but never force a template onto a reply it\\n doesn't fit.\\n\\n\\n\\n# Reply Templates (source of truth, together with swipefile.md)\\n\\nThese templates are Sam's tested copy (2026-07-14 shortened rewrite, updated\\n2026-07-16 for the Reimburse Ad Spend campaigns, updated 2026-08-05 for the\\nappointment/booking-deposit campaigns). They override every older\\ninstruction. Bodies are PLAIN TEXT with real line breaks (see voice.md).\\nReplies are SHORT and straight to the point - simple sentence structures, no\\nthree-paragraph walls. The cold email is question-led (\\\"if we could book your\\npractice {N} {treatment} appointments with patients who put down a booking\\ndeposit, would you be interested / like to hear more?\\\"), so many replies are a\\nbare \\\"Yes\\\"/\\\"Sure\\\"/\\\"I'd be interested\\\" - those get the Tell Me More template.\\n\\n{SENDER_FIRST_NAME} = the sender persona's first name (the pen name on the account\\nthat sent the campaign - Tiffany, Christina, etc. NEVER any other name).\\n{LENGTH_OF_CAMPAIGN} = the duration promised in the FIRST email we sent in that\\nthread. Current (2026-08) campaigns promise NO duration (\\\"your first campaign\\\") -\\nwrite \\\"two weeks\\\" (that is the real pilot length, and shorter than the month\\nthey'd assume). Older threads MIRROR their first email (\\\"2 weeks\\\" or \\\"30 days\\\")\\nper facts.md - never invent one.\\n{TREATMENT_TYPE} = the treatment(s) pitched in the first email - mirror that\\nemail / the lead data. May be MULTIPLE treatments joined like \\\"implant,\\ninvisalign or veneer\\\" - keep the join as written. Write \\\"Invisalign\\\"\\ncapitalized in replies; other treatments stay lowercase.\\nTREATMENT NARROWING RULE: the moment the prospect names the treatment(s) they\\nactually care about (\\\"we'd like to do Invisalign\\\"), every treatment mention in\\nthe reply narrows to exactly that (\\\"at least 4 Invisalign patients showing up\\nto their appointments\\\") - their stated interest beats the lead data.\\n{LOWER_BOUND} = the LOWER end of the promised range in the first email\\n(\\\"4-8 implant appointments\\\" -> 4; \\\"5-10\\\" -> 5; a single number stays as-is).\\n{PRACTICE_TYPE} = the lead's \\\"practice type\\\" variable (solo practice,\\nsmall-group practice, family dental practice, orthodontic practice, oral\\nsurgery group, implant center...). Use it PLURALIZED - pluralize the last word:\\n\\\"solo practices\\\", \\\"oral surgery groups\\\", \\\"implant centers\\\" - wherever a reply\\nsays who we help. If the variable is missing, fall back to \\\"dental clinics\\\".\\n{CURRENCY} = the prospect's local currency symbol. \\\"$\\\" by default and whenever\\nunsure; only switch (e.g. \\\"\\u00a3\\\") when the thread/lead is clearly outside the US.\\nCALENDAR = https://calendly.com/luminamg/pilot\\n\\n## Signature rule (applies to EVERY reply, no exceptions)\\n\\nEvery reply ends with the sign-off followed by this exact one-line signature,\\ndirectly under the persona's first name (before any P.S. line):\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\nPlain text. It appears on every single reply - full templates, short\\nback-and-forth one-liners, and quick acks alike. The website does NOT go in the\\nsignature (removed 2026-07-14, tentative - Sam is reworking the site). Website\\nmentions in template BODIES (\\\"You can find us at luminamg.com.\\\") stay - the\\nsend layer hyperlinks them automatically.\\n\\n## Greeting rule (applies to every template)\\n\\nIf the prospect's name is in their reply or signature, greet them with it:\\n\\\"Hey Monica,\\\" / \\\"Hey Dr. Song,\\\" - otherwise just \\\"Hey,\\\". Then a short remark that\\nfits the thread: \\\"great to hear!\\\", \\\"totally fair question.\\\", \\\"good question.\\\",\\n\\\"of course.\\\", \\\"happy to answer your questions.\\\", \\\"thanks for getting back to me.\\\"\\n\\n## Proposing times (live Calendly calendar integration)\\n\\nThe times in these templates ({X_DAY} at {TIME_1}...) are REAL: the system\\nchecks the live calendar and hands the draft a VERIFIED TIMES list. Rules:\\n- Only ever mention times from that verified list - never invent or shift one.\\n- Times are in the PROSPECT's timezone with the label (\\\"10am ET\\\").\\n- Use weekday names, never \\\"today\\\"/\\\"tomorrow\\\" (\\\"Friday at 10am or 1pm ET\\\").\\n- Standard shape: two days x two times. If their reply landed before 10am their\\n time, a same-day 3pm option leads and it becomes three days / five times.\\n- Proposals stick to business days (booking a weekend THEY asked for is fine).\\n- If the verified list is empty (calendar unreachable), offer no times - just\\n the Calendly link line.\\n- When they ask for a FURTHER-OUT week (\\\"I'm out next week - the week after?\\\"),\\n the system fetches that week's real availability into the verified list -\\n propose those times like any others. NEVER punt to a bare link because the\\n ask is far out (2026-07-18 correction). Deep in a thread, flex the phrasing:\\n \\\"I can book you in manually for the following week. Would {DAY 1} at\\n {TIME 1} or {TIME 2} work, or {DAY 2} at {TIME 3} or {TIME 4}? Feel free to\\n suggest other times that work for you.\\\"\\n\\n## P.S. lines (at most one, ONLY this situation)\\n\\n- First time offering times to a fully-interested reply: \\\"P.S. If those times\\n don't work, feel free to send over other openings and I'll book you in directly\\\"\\n (2026-07-18 wording - it sells the direct-booking capability; the old\\n \\\"send over your calendar link\\\" version is retired).\\n- EVERYTHING ELSE: no P.S. at all. The old booked-confirmation P.S. (\\\"better\\n email for this\\\") is RETIRED - the Booking Confirmation template asks about\\n the email in the body instead. The old \\\"This isn't a sales call\\\" P.S. stays\\n retired too.\\n\\n## The guarantee line (MIRROR RULE - match the first email's offer variant)\\n\\nSam's cold emails have carried three different offer variants. The drafting\\ninput names the variant (detected from the FIRST EMAIL WE SENT in the thread);\\nthe guarantee wording in a reply must match it exactly - a guarantee the first\\nemail never made costs Sam real money.\\n\\n- REIMBURSE - SHOWS (all campaigns from 2026-08-05 onward): the first email\\n promises booked appointments - often \\\"with patients who put down a booking\\n deposit\\\" - and (on some copies) a P.S. that no-shows mean we reimburse the\\n ad spend. Replies use the \\\"patients showing up to their appointments\\\"\\n guarantee sentence baked into the [FEE-WAIVER BLOCK] below - on ALL\\n current-era threads, even the copies whose body carried no explicit\\n reimburse P.S. (Sam's 2026-08-05 ruling). Keep it that brief - no\\n reimbursement rate, no recommended ad spend - unless they explicitly ask\\n (the dedicated Reimbursement and Ad Budget templates).\\n- REIMBURSE - CASES (2026-07-16 era): the first email promised \\\"{N} {treatment}\\n cases\\\" and \\\"if we don't hit those numbers, we cover/reimburse/pay you back\\n for your ad spend\\\". Keep the old cases sentence on these threads: \\\"If we\\n don't bring in at least {LOWER_BOUND} {TREATMENT_TYPE} cases, we reimburse\\n your ad spend.\\\"\\n- UNTIL WE DO (legacy, some older threads): the first email's P.S. promised to\\n keep the campaign running free until the numbers hit. Info-style replies get\\n ONE extra casual line right after the fee-waiver copy (instead of the\\n reimburse sentence): \\\"And yep - if we don't bring in at least {LOWER_BOUND}\\n {TREATMENT_TYPE} cases, we keep running your campaign free of charge until\\n we do. Ad spend included.\\\" Direct questions get the dedicated legacy\\n template. Never use reimbursement language on these threads.\\n- BASE (legacy, no promise): drop the guarantee sentence from the fee-waiver\\n block entirely and never write or imply any guarantee or reimbursement.\\n\\nWording may be clarified rather than echoed verbatim (\\\"free of charge\\\" for \\\"on\\nour dime\\\") - but the numbers, treatment type, and duration are still mirrored\\nexactly from the first email.\\n\\n## The standard fee-waiver block (referenced below as [FEE-WAIVER BLOCK])\\n\\nFor your first two weeks, you get the same fully-managed service we provide to our ongoing clients. The only difference is that we waive our service fee entirely. If we don't get at least {LOWER_BOUND} {TREATMENT_TYPE} patients showing up to their appointments, we reimburse your ad spend.\\n\\nThe final guarantee sentence follows the offer-variant rules above: keep it as\\nwritten on current REIMBURSE - SHOWS threads; on 2026-07-16-era REIMBURSE -\\nCASES threads it reads \\\"If we don't bring in at least {LOWER_BOUND}\\n{TREATMENT_TYPE} cases, we reimburse your ad spend.\\\" (and \\\"two weeks\\\" becomes\\nthat thread's mirrored {LENGTH_OF_CAMPAIGN}); swap it for the \\\"And yep...\\\"\\nline on legacy UNTIL WE DO threads; drop it entirely on legacy BASE threads.\\n\\n## The standard CTA block (referenced below as [CTA BLOCK], 2026-07-18 rewrite)\\n\\nLet's hop on a quick call to share more details about the pilot, answer any questions, and go over the results you can expect.\\n\\nHow about {X_DAY} at {TIME_1} or {TIME_2} or {Y_DAY} at {TIME_3} or {TIME_4}?\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\nBook a time yourself at CALENDAR\\n\\nTwo booking paths live in this block on purpose: we book them (they reply\\nwith a time) or they self-book (the line under the signature - the send\\nlayer appends it automatically and renders it as a blue link, so don't\\nfight it). No inline \\\"here's our calendar\\\" line in this block anymore.\\n\\nYou don't need the precise CTA block in every case - use it when it works\\n(Sam, 2026-07-18); the Ask for Call template deliberately skips the pitch\\nline.\\n\\nRETIRED FOREVER - never include either line, even when an earlier message\\nin the thread used it:\\n- \\\"And what's the best number to reach you at?\\\"\\n- \\\"And just for convenience, what's a good number to reach you directly?\\\"\\n (dropped 2026-08-07: 0 of 9 asks ever got a number back - prospects reply\\n with times or self-book instead. The post-booking \\\"best number for\\n reminders\\\" ask in the Booking Confirmation template is unaffected.)\\n\\n---\\n\\n## Generic (positive reply that doesn't map onto a template)\\n\\nUse the following structure:\\n\\nHey, [remark that makes sense given the thread]\\n\\n[answer their question(s) directly - short, plain terms, simple sentences;\\npull claims from facts.md only. Point to the free campaign where it helps:\\nthey see exactly what working with us looks like before paying us anything.]\\n\\nOnly after answering their questions, pivot to the CTA:\\n\\n[CTA BLOCK]\\n\\nNote: the justification for the call may change depending on earlier replies.\\n\\n---\\n\\n## Interested = Ask for Call\\n\\nWhen they clearly ask to talk/book (\\\"schedule the call\\\", \\\"Ok let set up a\\nmeeting\\\", \\\"can we schedule a call to discuss\\\") use the Ask for Call template\\nbelow - they are already sold on talking, so NEVER re-pitch the call at them\\n(no \\\"Let's hop on a quick call...\\\" line; that mistake shipped twice the week\\nof 2026-07-18). A bare \\\"Yes\\\"/\\\"Sure\\\" answering the cold email's \\\"would you be\\ninterested / like to hear more?\\\" question is NOT this - that gets Tell Me\\nMore below.\\n\\n---\\n\\n## Tell Me More\\n\\nThe default reply when they answer the cold email's question with \\\"Yes\\\" /\\n\\\"Sure\\\" / \\\"I'd be interested\\\" / \\\"send me more info\\\" and don't ask who we are.\\nNO website line here (2026-07-16): the website appears only in Company Details\\nand when they ask for it.\\n\\nHey, of course.\\n\\nWe're Lumina and we help {PRACTICE_TYPE, pluralized - e.g. \\\"family dental practices\\\"} build a predictable pipeline of high-value treatments.\\n\\nThree quick bullets\\n\\n- We call every lead within 60 seconds and follow up by phone and text for 7 days\\n- We only warm transfer patients that are ready to book to your front desk\\n- You don't pay us for the first {LENGTH_OF_CAMPAIGN}, and our service fee is {CURRENCY}497 a month if you decide to continue\\n\\n{GUARANTEE LINE - one standalone sentence, era-gated per the guarantee-line\\nrules above. Current REIMBURSE - SHOWS threads: \\\"If we don't get at least\\n{LOWER_BOUND} {TREATMENT_TYPE} patients showing up to their appointments, we\\nrefund your ad spend too.\\\" REIMBURSE - CASES threads use their cases\\nsentence; UNTIL WE DO threads use their \\\"And yep...\\\" line; legacy BASE\\nthreads get NO guarantee line.}\\n\\n[CTA BLOCK]\\n\\n(2026-08-07 restructure, Sam: NO [FEE-WAIVER BLOCK] paragraph here anymore -\\nbullet 3 carries the you-don't-pay message and the guarantee stands alone\\nright after the bullets. The bullets are what stop this from sounding like\\nevery marketing agency ever - keep all three, hyphen style (\\\"- \\\"), each on\\nits own line; the send layer indents them. The literal lead-in line \\\"Three\\nquick bullets\\\" stays. The bullet-3 price mention is the ONLY place price is\\nvolunteered outside the Price template - see the price-discipline rule\\nbelow.)\\n\\n---\\n\\n## Company Details\\n\\nWhen they ask who we are / what company this is / for a website or company\\ninfo - alone or mixed with a \\\"tell me more\\\"-style reply.\\n\\nHey, of course.\\n\\nWe're Lumina. You can find us at luminamg.com.\\n\\nWe help {PRACTICE_TYPE, pluralized} build a predictable pipeline of high-value treatments.\\n\\n[FEE-WAIVER BLOCK]\\n\\n[CTA BLOCK]\\n\\n---\\n\\n## Need More Info: Process (they ask how it works / what we actually do)\\n\\nHey, of course!\\n\\nDuring your first {LENGTH_OF_CAMPAIGN} with us, here's what we do.\\n\\n- Produce 10-15 high-performing {TREATMENT_TYPE} ads\\n- Build a conversion-optimized landing page that turns clicks into treatment-ready buyers\\n- Call every new lead within 60 seconds and transfer them to your front desk to get booked in\\n\\n[CTA BLOCK]\\n\\n---\\n\\n## Need More Info: Price\\n\\nHey, totally fair question.\\n\\n[FEE-WAIVER BLOCK]\\n\\nIf you'd like to continue with us after the pilot campaign, our flagship plan starts at $497/month or $2,197 for six months paid upfront - that's 25% off, about 6 weeks free.\\n\\n[CTA BLOCK]\\n\\n---\\n\\n## Need More Info: Ad Budget (ONLY when they explicitly ask about ad spend/budget)\\n\\nHey, good question.\\n\\nWe'd recommend budgeting {CURRENCY}500 towards Meta ads for the two weeks. If we don't get at least 3 patients showing up to their appointments, we reimburse your ad spend.\\n\\n[CTA BLOCK]\\n\\n(When ad spend is the topic, {CURRENCY}500 is the headline and the flat \\\"3\\npatients showing up\\\" is the guarantee - NOT {LOWER_BOUND} (3 is the floor\\nacross all current ranges, and 500 is less daunting than a bigger number\\nnext to a bigger promise - Sam, 2026-08-05). The guarantee sentence is\\nREIMBURSE-thread copy. On legacy threads keep just the budgeting sentence -\\nwith the until-we-do line first when that variant applies.)\\n\\n---\\n\\n## Reimbursement: direct question (they ask how the reimbursement/guarantee works)\\n\\nHey, totally fair question.\\n\\nFor your first {LENGTH_OF_CAMPAIGN}, you get the same fully-managed service our ongoing clients receive. The only difference is that we waive our service fee entirely, and for every {CURRENCY}150 you spend on ads, we guarantee 1 patient shows up to their {TREATMENT_TYPE} appointment - at the recommended {CURRENCY}500 budget that's 3 patients showing up - or we reimburse your ad spend.\\n\\n[CTA BLOCK]\\n\\n(This template is the ONLY place the {CURRENCY}150-per-patient rate is\\nvolunteered - everyone else gets just the short guarantee sentence in the\\nfee-waiver block. If their message is broader skepticism rather than a\\nmechanics question, use Too Good to Be True and fold the guarantee sentence\\nin. No flagship pricing unless they also asked about price.)\\n\\n---\\n\\n## LEGACY - Until We Do: direct question (older threads only - first email promised to keep running until the numbers hit)\\n\\nHey, totally fair question.\\n\\nYep - exactly what it sounds like. If the first {LENGTH_OF_CAMPAIGN} doesn't get you those {LOWER_BOUND}+ {TREATMENT_TYPE} cases, we keep the campaign running free of charge until you have them. That covers the ad spend too, so your total cost stays capped at the first {LENGTH_OF_CAMPAIGN} of ad budget.\\n\\nI'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n\\nFor us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n\\n[CTA BLOCK]\\n\\n(No flagship pricing here unless they also asked about price. If their message\\nis broader skepticism rather than a question about this clause, use Too Good\\nto Be True below and just fold the capped-cost sentence in.)\\n\\n---\\n\\n## Need More Info: Too Good to Be True (skepticism, \\\"what's the angle\\\", \\\"how do you profit\\\")\\n\\nHey, totally fair question.\\n\\n[ONLY when they asked about a catch/downside (\\\"what's the catch?\\\", \\\"too good to be true\\\"): There's no catch :)]\\n\\nI'm guessing you've been burned by other marketers before, so we'd rather get you results before you decide if you want to continue.\\n\\nFor us, it's a chance to get our foot in the door and onboard a practice that stays with us for years.\\n\\n[CTA BLOCK]\\n\\nTHE OPENER MUST ANSWER THEIR ACTUAL QUESTION. \\\"What's in it for you?\\\" /\\n\\\"how do you make money doing this?\\\" is NOT a catch question - answer it with\\nthe foot-in-the-door paragraph FIRST (that IS what's in it for us) and skip\\nthe \\\"There's no catch\\\" line entirely. These lines are blocks: lead with\\nwhichever one answers them, drop the ones that don't.\\n\\n(NO price in this template - removed 2026-07-18; volunteering it \\\"off the\\nrip\\\" turned prospects off. When their reply mixes skepticism with a\\ntell-me-more ask, BLEND: relevant TMM parts first, then the no-catch /\\nburned-before / foot-in-the-door lines - see the blending rule at the end.\\nEach of those lines is its own PARAGRAPH - keep the blank line between\\n\\\"...decide if you want to continue.\\\" and any fee-waiver copy. And FIT THE\\nOPENER TO THEIR WORDS: \\\"There's no catch :)\\\" answers \\\"what's the catch\\\" -\\n\\\"what's in it for you?\\\" / \\\"how do you make money?\\\" gets the foot-in-the-door\\nanswer up front instead. Never parrot a line that doesn't answer what they\\nactually asked.)\\n\\n---\\n\\n## Ask for Call (they explicitly asked to meet - no pitch, straight to times)\\n\\nHey, great to hear!\\n\\nHow about {X_DAY} at {TIME_1} or {TIME_2} or {Y_DAY} at {TIME_3} or {TIME_4}?\\n\\nOr if it's easier, here's our Calendly: CALENDAR\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\nP.S. If those times don't work, feel free to send over other openings and I'll book you in directly\\n\\n(Greeting flexes to their words: \\\"of course!\\\" answers \\\"can we schedule a\\ncall?\\\", \\\"great to hear!\\\" answers enthusiasm. NO \\\"Let's hop on a quick\\ncall...\\\" line - they already asked for the call. This template keeps the\\ninline Calendly line, so the send layer skips the under-signature one.)\\n\\n---\\n\\n## Booking Confirmation (their time was open - the system already booked it; 2026-07-18 rewrite)\\n\\nVariant when their address looks like a generic office inbox (system-detected):\\n\\nHey - {DAY} at {TIME & TIMEZONE} it is! Just sent over a confirmation email with meeting details to {EMAIL}. It looks like this may be an office email, so if there's a better address, just let me know and I'll send it there instead.\\n\\nGoogle Meet link for your convenience: [MEET_LINK]\\n\\nAdd to Google Calendar\\nAdd to Apple / Outlook Calendar\\n\\nAnd what's the best number for reminders before the call?\\n\\n(When we already asked for a number earlier in the thread, soften the re-ask\\nso it doesn't read oblivious: \\\"And what's the best number for reminders\\nbefore the call? Promise you won't get spammed :)\\\" - the promise line is ONLY\\nfor that second ask. First-reply direct-books use the plain question above.\\nEither way the ask is system-gated and vanishes once we hold a number.)\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\nVariant when it looks like the right (personal) email: identical, but the\\nopener is just \\\"Hey - {DAY} at {TIME & TIMEZONE} it is! Just sent over a\\nconfirmation email with meeting details.\\\" (no address callout).\\n\\n(The three link lines are written EXACTLY as shown - the send layer swaps\\n[MEET_LINK] for the real Google Meet URL and turns the two \\\"Add to...\\\" lines\\ninto blue links; never write actual URLs for them. The reminders-phone\\nquestion is GATED by the system - it disappears once we hold their number.\\nIf their reply also asked questions, answer briefly first per the playbook.\\nNo other times, no calendar link, no P.S.)\\n\\n---\\n\\n## IF SLOT ISN'T OPEN (their proposed time is taken)\\n\\nHey - I've got another call at {PROPOSED_TIME}. How about {PROPOSED_DAY} at {NEAREST_AVAILABLE_TIME}, {PROPOSED_DAY_AFTER} at {PROPOSED_TIME / NEAREST_AVAILABLE_TIME} or {ALTERNATE_TIME}, or {PROPOSED_DAY_AFTER_THAT} at {PROPOSED_TIME / NEAREST_AVAILABLE_TIME} or {ALTERNATE_TIME}.\\n\\nOr if it's easier, here's our Calendly: CALENDAR\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\nP.S. If those times don't work, feel free to send over other openings and I'll book you in directly\\n\\n(All times come from the verified list: their day at the nearest open slot, then\\nthe next two days at their proposed time when it's free - if it worked for one\\nday it probably works the next - plus an alternate ideal time for spread.\\nAdapt the opener naturally if they asked for a window rather than a clock time.)\\n\\n---\\n\\n## Already booked (they self-booked or say they did)\\n\\nOne short, warm line confirming you saw it and you're looking forward to the\\ncall - e.g. \\\"Hey, saw the booking come through - talk then!\\\" adapted to the\\nthread. No times, no calendar link, no pitch, no P.S. - but still the sign-off\\nand signature block per the signature rule.\\n(If they say they already RESCHEDULED themselves: no reply at all - the\\nconfirmation flow covers them.)\\n\\n---\\n\\n## Post-Booking Details (they have a booking and ask a question - e.g. \\\"I never got the invite\\\")\\n\\nThey're booked; this is support, not scheduling. NEVER pitch times at someone\\nwith an appointment on the books. For missing-invite asks:\\n\\nHey,\\n\\nSorry about that. Check for an email from Calendly at notifications@calendly.com, and make sure it didn't land in your spam folder.\\n\\nIn the meantime, here are the meeting details:\\n\\nGoogle Meet: [MEET_LINK]\\n\\nAdd to Google Calendar\\nAdd to Apple / Outlook Calendar\\n\\nEverything you need is in this thread now, but I'm happy to resend the Calendly invite if it still hasn't come through.\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\n\\n(The link lines are swapped for real links by the send layer - never write\\nURLs for them. For OTHER post-booking questions, answer briefly from facts.md\\nand reference the booked call naturally - the details block only when it\\nhelps. POST-BOOKING SIGNATURE RULE: once they're booked, back-and-forth\\nreplies drop the \\\"SDR, Lumina\\\" line and the self-booking link - just\\n\\\"Thanks,\\\" + first name. Applies here and to Already Booked above. As always:\\nthese are BLOCKS, not scripts - fit greeting and wording to what they said.)\\n\\n---\\n\\n## Written Reschedule Request (they ask to move the call, no new time named)\\n\\nThe system keeps their current appointment until a new time is agreed. Propose\\nverified times:\\n\\nHey,\\n\\nNo worries at all - happy to reschedule. Does {X_DAY} at {TIME_1} or {TIME_2} or {Y_DAY} at {TIME_3} or {TIME_4} work for you instead?\\n\\nAlternatively, have your pick at any available time on our Calendly: CALENDAR\\n\\nFeel free to reach out if those times don't work and we can work something out.\\n\\nTake care,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\n(When they DO name a new open time, the system moves the appointment on the\\nspot - send a quick courteous confirmation (\\\"No worries at all - just moved\\nus to {NEW_TIME}.\\\") followed by the same three link lines as the Booking\\nConfirmation template (Meet link + the two Add-to-Calendar lines) so the new\\ndetails live right in the thread. No other times, no calendar link, no\\nphone ask.)\\n\\n---\\n\\n## Written Cancellation Request (they ask to cancel)\\n\\nThe system cancels the appointment immediately - then this reply tries to save\\nthe meeting:\\n\\nHey,\\n\\nNo worries at all. Would {X_DAY} at {TIME_1} or {TIME_2} or {Y_DAY} at {TIME_3} or {TIME_4} work for you instead?\\n\\nAlternatively, have your pick at any available time on our Calendly: CALENDAR\\n\\nLet me know if you have any questions or concerns.\\n\\nTake care,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\n(If their message is clearly final - firmly out, unhappy - skip the times and\\nthe calendar entirely: just a warm, brief, door-open close.)\\n\\n---\\n\\n## Preferred Method of Communication (PMoC) - RETIRED 2026-08-05\\n\\nThe agent NO LONGER REPLIES to contact-preference messages (phone/text/\\nWhatsApp/\\\"call my office\\\" preferences) at all. Sam is alerted with the\\ndetails and reaches out himself on their channel - no \\\"Expect a text from\\nSam\\\" acknowledgment first (it kept promising outreach before Sam had sent\\nanything, and speed-to-lead doesn't apply when they've named their channel).\\nHistorical examples in swipefile.md are kept for context only.\\n\\n(A message that only gives a better EMAIL ADDRESS is NOT a PMoC - the system\\nupdates the invite/records itself and the reply confirms it briefly in\\nwhatever lane the rest of their message calls for.)\\n\\n---\\n\\n## Wrong Person (referral)\\n\\n[IF NEW PROSPECT CONTACT ISN'T PROVIDED - the agent sends this reply itself]\\n\\nHey {INITIAL_PROSPECT_NAME}, thanks for pointing me in the right direction. What's {NEW_PROSPECT_NAME}'s best email so I can send them the details?\\n\\nThanks,\\n{SENDER_FIRST_NAME}\\nSDR, Lumina\\n\\n[TO NEW PROSPECT - INTERIM: the agent DRAFTS this but Sam sends it (sending to a\\nbrand-new address isn't wired up yet). The finished draft goes in Sam's alert.]\\n\\nHey {NEW_PROSPECT_NAME}, got in touch with {INITIAL_PROSPECT_NAME} and they told me that you're the right person to talk to about this.\\n\\nTLDR: We help {PRACTICE_TYPE, pluralized} build a predictable pipeline of high-value treatments with Meta ads, and it all starts with a free {LENGTH_OF_CAMPAIGN} campaign so you can see exactly what working with us looks like before paying us a cent.\\n\\n[CTA BLOCK]\\n\\n---\\n\\n## Price discipline (details kill - Sam, 2026-07-18)\\n\\nPrice is volunteered in exactly TWO places: the {CURRENCY}497 bullet inside\\nTell Me More, and the Price template when they directly ask about cost.\\nNEVER volunteer price in Ask for Call / Interested replies, skepticism\\nreplies (Too Good to Be True), or answers to unrelated specific questions -\\nit adds friction right when they were ready to move.\\n\\n## Template blending (mixed-intent replies)\\n\\nReal replies often span two templates (\\\"Let's do it! ...what's the catch?\\\" =\\nTMM + Too Good to Be True). Take the RELEVANT PARTS of each - intro/bullets\\nfrom one, objection-handling lines from the other - into one short reply with\\na single CTA. Don't stack both templates whole, and don't force the closest\\nsingle template when a blend answers them better.\\n\\n## Anti-robotic rule (applies to the whole thread)\\n\\nAlways read the entire thread before drafting. Templates fit ~95% of replies,\\nbut never send a message that mirrors the structure of an earlier one in the\\nsame thread - same greeting shape, same paragraph order, same CTA phrasing\\nreads as a bot. Vary the opener, merge or reorder paragraphs, tighten the\\nlength as the thread deepens. And never box yourself into a template that\\ndoesn't fit what they actually said - when no template fits where the thread\\nhas gone, compose short and direct from voice.md + facts.md instead of\\nforcing one.\\n\\n\\n\\n# Canonical Facts (single source of truth, 2026-07-08 rewrite)\\n\\nAnswer questions from this file (and mechanism.md for how-it-works depth). If\\nthese facts conflict with anything in mechanism.md, THIS file wins.\\n\\n**Missing-info policy (changed 2026-07-08):** if we don't have the exact thing\\nthey asked for, that's fine - say so simply if needed, give the best adjacent info\\nwe DO have (what we do, the free campaign, pricing, website), and redirect to the\\ncall. Escalate (answerable=false) ONLY for genuine edge cases the playbook can't\\ntouch at all - not merely because one of their questions has no exact answer here.\\n\\n## The offer\\n\\n- We help {PRACTICE_TYPE, pluralized - \\\"solo practices\\\", \\\"family dental\\n practices\\\", \\\"oral surgery groups\\\"...; \\\"dental clinics\\\" when the variable is\\n missing} build a predictable pipeline of high-value treatments with Meta ads\\n (Facebook/Instagram ONLY - no Google or other platforms).\\n- The cold email pitched booked appointments for the practice's mapped\\n treatment(s) - possibly several (\\\"implant, invisalign or veneer\\\") - often\\n \\\"with patients who put down a booking deposit\\\". On the back end we bring\\n patients for a variety of treatments.\\n- **Free first campaign (THE hook - work it into every info answer).** For\\n their first two weeks (current copies promise NO duration - \\\"your first\\n campaign\\\"; the real pilot length is two weeks, say it plainly, it beats the\\n month they'd assume; OLDER threads mirror their email's \\\"2 weeks\\\"/\\\"30 days\\\")\\n they get the same fully-managed service as our ongoing clients - the only\\n difference is we waive our service fee entirely. The ad budget paid to Meta\\n is theirs, and on current campaigns it's guaranteed (see below).\\n What we do during it: produce 10-15 high-performing {TREATMENT_TYPE} ads,\\n build a conversion-optimized landing page that turns clicks into\\n treatment-ready buyers, and call every new lead within 60 seconds,\\n transferring them to their front desk to get booked in.\\n The cold emails are spintax, so each prospect was promised slightly different\\n wording and numbers (\\\"4-8 implant or veneer appointments\\\", \\\"3-6 implant,\\n invisalign or bonding appointments\\\"...).\\n **MIRROR RULE: the first email we sent in THAT thread is the source of truth\\n for {LENGTH_OF_CAMPAIGN} and any promised numbers.** If the first email\\n isn't available, say \\\"the free campaign\\\" with NO duration or numbers.\\n- **Booking deposit (ask-only - NEVER volunteered, never argued over email):**\\n the deposit patients put down is small - usually under {CURRENCY}50. The\\n practice collects it and keeps it; we never touch the money. Refundable or\\n applied to their treatment - the practice's choice, both are common. It\\n exists because patients who put down a deposit show up. If a prospect\\n doubts deposits work for their patients: it's flexible, we don't force it -\\n details are a call conversation.\\n- **THE GUARANTEE - three offer variants; ALWAYS match the first email in the\\n thread (the drafting input names the detected variant):**\\n - **REIMBURSE - SHOWS (all campaigns from 2026-08-05 onward):** the first\\n email promises booked appointments (often with booking deposits); copies\\n 2-3 add a P.S. \\\"if patients no-show, we refund/reimburse your ad spend\\\".\\n Default phrasing in replies - on ALL current-era threads, even the copies\\n without the P.S. (Sam's ruling): \\\"If we don't get at least {LOWER_BOUND}\\n {TREATMENT_TYPE} patients showing up to their appointments, we reimburse\\n your ad spend.\\\" Mechanics ONLY when they ask how the reimbursement works:\\n for every {CURRENCY}150 they spend on ads, we guarantee 1 patient shows up\\n to their {TREATMENT_TYPE} appointment - at the recommended {CURRENCY}500\\n budget that's 3 patients showing up - or we reimburse their ad spend.\\n Never volunteer the rate or the recommended budget otherwise.\\n - **REIMBURSE - CASES (2026-07-16 era, older threads):** the first email\\n promised \\\"{N} {treatment} cases\\\" with the reimburse promise. Keep the old\\n sentence on these threads: \\\"If we don't bring in at least {LOWER_BOUND}\\n {TREATMENT_TYPE} cases, we reimburse your ad spend.\\\"\\n - **UNTIL WE DO (legacy, older threads):** the first email's P.S. promised\\n \\\"if we don't hit those numbers, we'll keep it running on our dime / free\\n of charge until we do.\\\" The extension covers our service AND the ad\\n spend - past the first {LENGTH_OF_CAMPAIGN} the prospect pays nothing\\n until the promised cases arrive, so their total cost is capped at the\\n first campaign's ad budget. Never use reimbursement language here.\\n - **BASE (legacy, no promise):** never write or imply any guarantee,\\n reimbursement, or free extension on these threads - it promises money\\n Sam never offered.\\n Phrasing may be clarified rather than echoed verbatim, but\\n numbers/treatment/duration are mirrored exactly.\\n- The cold email P.S. on current campaigns (copies 2-3 only) is the no-show\\n reimburse promise. LEGACY P.S. lines on older threads: \\\"a short video\\n testimonial if we deliver\\\" (use for \\\"what's the catch?\\\" when THEIR email\\n carried it) and limited capacity mentions. Current copies carry an urgency\\n line in the body: \\\"we're opening five spots for a program where you don't\\n pay us for your first campaign\\\" - mirror their email's number if they bring\\n it up; never invent capacity numbers.\\n\\n## Pricing (quote these plainly - no dancing around price)\\n\\n- **Paid package: Dental Growth Flagship - $497/month, or $2,197 for six months\\n paid upfront (that's 25% off - about 6 weeks free).** Lead with the monthly\\n number; the 6-month upfront is the discount option.\\n- Always frame it: our service fee is fully waived for the first campaign - they\\n only cover ad costs paid directly to Meta; they see exactly what working with\\n us looks like first, then decide.\\n- Month-to-month on the monthly plan; no long-term contract.\\n- Ad budget (updated 2026-08-05 - still ask-only): the recommended\\n {CURRENCY}500 is NEVER volunteered - not in the fee-waiver copy, not in\\n price answers (volunteered numbers add friction). It comes up ONLY when they\\n explicitly ask about ad spend/budget (the Ad Budget template): budget is up\\n to them, paid directly to Meta on their own card, we'd recommend budgeting\\n {CURRENCY}500 towards Meta ads for the two weeks - and if we don't get at\\n least 3 patients showing up to their appointments, we reimburse the ad\\n spend (in ad-spend answers the flat 3 is the guarantee, never\\n {LOWER_BOUND} - the {CURRENCY}500 is the headline) - and scaling spend is\\n a conversation for after the free campaign.\\n- \\\"What if it doesn't work?\\\" -> the guarantee IS the answer, in the first\\n email's own variant (reimbursed ad spend / keep running until we do - see\\n The offer). On legacy BASE threads: the free campaign exists exactly so they\\n can see how we work before paying anything.\\n- INTERNAL - NEVER SAY THIS TO A PROSPECT: free-campaign results tend to be worse\\n than steady-state (fresh ad account, Meta still learning). Don't volunteer it.\\n If pushed hard on exact free-campaign results, stay honest but positive\\n (campaigns get stronger as Meta learns) or escalate.\\n\\n## The call\\n\\n- A **quick video call** (Google Meet): a rundown of the free campaign, what results they can\\n expect, and getting the info we need to create their ads. Never attach a\\n duration to it in copy (\\\"15-minute\\\" retired 2026-07-15 - it's just \\\"a quick\\n call\\\"). If they explicitly ask how long it takes: it's short - we block half\\n an hour on the calendar and often wrap sooner.\\n- Not a sales call - they pay nothing upfront. (Body-copy talking point when\\n they hesitate; the old P.S. line carrying it is retired.)\\n- Calendar link: https://calendly.com/luminamg/pilot\\n- The system checks the LIVE calendar: proposed times come from a verified list,\\n and when a prospect's requested time is open, the system books it on the spot\\n and the confirmation/invite emails go out automatically. Never mention a time\\n that isn't in the verified list.\\n- Booking confirmations and reminders arrive from Calendly\\n (notifications@calendly.com) with the calendar invite and a Google Meet\\n link attached. If someone says they didn't get the invite, tell them to\\n check for an email from Calendly (and spam), and offer to resend.\\n\\n## Contact info & resources (what to send, and how eagerly)\\n\\n- **Website: https://luminamg.com/** - share confidently whenever they ask for a\\n website or company info.\\n- **Sam's phone: 925-730-9373** - every PMoC acknowledgment includes it (\\\"our\\n founder Sam (his number is 925-730-9373)\\\"). Outside PMoC, share it only in the\\n rare case they specifically need a number to reach us at. Prefer: give the\\n good info we have, then redirect to what we do / the free campaign / the call.\\n- **No LinkedIn to share.** Say so plainly (\\\"I don't have a LinkedIn to share\\\")\\n and pivot to the website or the call.\\n- **Founder videos (YouTube)** - only for the few cases where someone genuinely\\n wants deep how-it-works detail; most replies should NOT include them:\\n - How we get qualified leads: https://youtu.be/WlAQcaskL28?si=B4DryJm0hvTXYsBr\\n - How we convert leads into patients: https://youtu.be/PvFJuJSNZiI?si=FxUduRVYavk5PpXU\\n - The quantity and types of ads we produce: https://youtu.be/X_Cc8Iy3Hgg?si=xHSPFvrVpWCLuald\\n- General rule: we don't have everything (that's fine) - give what we do have,\\n then redirect their attention to what we do, the free campaign, and the call.\\n\\n## Leads & operations\\n\\n- Leads are exclusive to their practice.\\n- When a new lead comes in, we call them within 60 seconds. If they pick up,\\n we qualify them for location and financing, then keep them on the line and\\n connect the call straight through to the front desk to get booked in (a\\n live phone handoff - for how-it-works depth say \\\"we keep them on the line\\n and connect you directly\\\", never \\\"warm transfer\\\" jargon).\\n- If they don't pick up, we keep following up with calls and texts for about\\n a week - any time they answer, same live transfer.\\n- NEVER say \\\"AI\\\" or \\\"AI agent\\\" about the calling - it's just \\\"we call them\\\".\\n- Nothing integrates with their practice software (Dentrix etc.); their front\\n desk books patients straight into their own system. That's a feature: no\\n scheduler conflicts, no access to their PMS needed.\\n- To start we need: access to their Facebook page, plus offer specifics agreed\\n during onboarding. Existing social content is nice to have, not required.\\n- Launch: generally within 3 days after onboarding.\\n\\n## Company\\n\\n- Lumina (\\\"Lumina Marketing Group\\\" / LuminaMG), based in California (SF).\\n- In business for 2 years. Team size: stay vague, never give a number.\\n- No case studies or references yet - never claim any. Pivot: the free campaign\\n IS the proof; they see results themselves before paying.\\n\\n## Exclusivity\\n\\n- No formal exclusivity clause, but the positioning is: we don't work with\\n overlapping/competing practices - \\\"we won't step on anyone else's toes.\\\" Say it\\n confidently; never mention the absence of a clause.\\n\\n## \\\"We tried Facebook ads before and it flopped\\\"\\n\\nCommon reasons, and what we do differently: not enough ad creatives (we run 10-15\\ngenuinely different ones), unproven creatives, unqualified leads (we qualify them\\nbefore they book), and slow follow-up (we call every new lead within 60 seconds\\nand transfer them straight to the front desk).\\n\\n## Special situations\\n\\n- **They give a phone number / preferred channel (PMoC)**: the agent sends\\n NOTHING (2026-08-05 policy - no \\\"Expect a text from Sam\\\" ack). Sam is\\n alerted with the details and reaches out himself on their channel. A PMoC\\n mixed with a meeting request is still PMoC - never book it over email.\\n A phone number that merely sits in their SIGNATURE is NOT a PMoC - reply to\\n their message normally (the system separately alerts Sam with the number so\\n he can call them). A message that only gives a better EMAIL ADDRESS is NOT\\n a PMoC either - the system updates the invite itself (see templates.md).\\n- **They hand the thread to someone else (referral / wrong person)**: if the new\\n contact's email ISN'T given, the agent replies asking for it (Wrong Person\\n template part 1). If it IS given, the agent drafts the intro to the new person\\n (part 2) and Sam sends it - see templates.md INTERIM note.\\n- **They propose a specific meeting time**: the system checks the calendar and\\n books it if open (they get the invite automatically); if taken, counter with\\n the IF SLOT ISN'T OPEN template.\\n- **They say they booked / a time \\\"works\\\"**: the system checks the calendar for\\n an existing appointment first - if they're already booked, just the short\\n \\\"saw the booking come through\\\" line, never a re-pitch or a double-booking.\\n- **Reply written by an AI assistant** (\\\"Sent from X's AI Assistant\\\", templated\\n auto-responses): never a positive reply - always not_positive, no response.\\n- **\\\"Is this AI?\\\"**: don't answer; escalate. (Prospects realistically never ask.)\\n- **\\\"We already have an agency\\\"** / **\\\"How did you get my email?\\\"**: not_positive,\\n no reply.\\n\\n## FAQ quick answers\\n\\n- Cost? -> first campaign is the same fully-managed service with our fee waived\\n entirely, and if we don't get at least {LOWER_BOUND} {TREATMENT_TYPE}\\n patients showing up to their appointments we reimburse the ad spend\\n (guarantee sentence per the first email's variant - \\\"cases\\\" wording on\\n 2026-07-16-era threads; no budget number unless they ask about ad spend).\\n Continuing afterward: flagship plan starts at $497/month or $2,197 for six\\n months upfront (25% off - about 6 weeks free).\\n- Ad budget? (deeper answer only if THEY brought up ad spend/budget) -> we'd\\n recommend budgeting {CURRENCY}500 towards Meta ads for the two weeks, and\\n if we don't get at least 3 patients showing up to their appointments we\\n reimburse the ad spend (flat 3 here, never {LOWER_BOUND}); scaling is a\\n conversation for after the free campaign.\\n- How does the reimbursement work? (only if THEY ask) -> for every\\n {CURRENCY}150 spent on ads, we guarantee 1 patient shows up to their\\n {TREATMENT_TYPE} appointment - at the recommended {CURRENCY}500 budget\\n that's 3 patients showing up - or we reimburse the ad spend (the\\n Reimbursement template).\\n- What's this booking deposit? / how much? / who keeps it? (only if THEY\\n ask) -> small, usually under {CURRENCY}50; your practice collects and\\n keeps it (we never touch it); refundable or applied to treatment - your\\n call, practices do both; it's there because patients who put down a\\n deposit show up. If they push back on deposits: totally flexible, we don't\\n force it - happy to go over it on the call.\\n- Too good to be true / what's your angle? -> it does sound that way; they've\\n probably been burned before, so we'd rather get them results before they\\n decide to continue - for us it's a foot in the door and a practice that\\n stays with us for years.\\n- Contract? -> month-to-month.\\n- How fast? -> live within ~3 days of finishing onboarding.\\n- What do you need from us? -> access to your Facebook page and a short\\n onboarding to agree on the offer.\\n- Who talks to the leads? -> we call every new lead within 60 seconds, qualify\\n them for location and financing, and transfer the call straight to your\\n front desk to book them in.\\n- Case studies? -> the free campaign is the proof - you see results before paying.\\n- What's the catch? -> if THEIR first email asked for it: a short video\\n testimonial if we deliver. Otherwise: we'd rather get them results before\\n they decide to continue - foot in the door (Too Good to Be True framing).\\n- \\\"You'd really run it free until we hit the numbers?\\\" (LEGACY - ONLY on Until\\n We Do threads) -> yep, exactly that: past the first {LENGTH_OF_CAMPAIGN} we\\n cover our service AND the ad spend until the promised cases arrive - their\\n total cost stays capped at the first campaign's ad budget.\\n- Website? -> https://luminamg.com/ (share confidently).\\n- Where are you located? -> California.\\n\";\n",[176],"f9a0c026-6c60-408c-98f8-1482ab11ce4d","ghl.ts","503799ae-7bee-11f1-9fbd-1607ee4eb77e",73,["D",1784412833921],9,"// GoHighLevel (LeadConnector) calendar integration for the reply agent.\n// Verified live 2026-07-09: free-slots, book, reschedule, cancel, contact upsert.\n// Not yet imported by webhook.ts - the wire-in (which times to propose, when to\n// book) comes once Sam specifies the rules.\n//\n// Env: GHL_API_KEY (Private Integration token, scopes: calendars r/w,\n// calendar events r/w, contacts r/w). Location/calendar ids are not secrets\n// (they appear in public booking URLs), so they live here as constants.\n\nconst GHL_BASE = \"https://services.leadconnectorhq.com\";\nexport const GHL_LOCATION_ID = \"89bBbANt0akkAawGIfdm\";\n// \"Lumina Free Campaign Call\" - the calendar behind links.luminamg.com/widget/booking/1nbt58J9c0CBZFyqCBQk\nexport const GHL_CALENDAR_ID = \"1nbt58J9c0CBZFyqCBQk\";\nexport const GHL_TIMEZONE = \"America/Los_Angeles\";\n\n// Appointment mutations get 60s, not 15s: GHL has taken ~41s to commit a\n// booking under load (Cornerstone, 2026-07-17) - aborting a mutation the\n// server may still complete leaves an orphaned booking with no reply sent.\nasync function ghl(\n method: string,\n path: string,\n body?: unknown,\n version = \"2021-04-15\",\n timeoutMs = 15_000,\n): Promise {\n const resp = await fetch(GHL_BASE + path, {\n method,\n headers: {\n \"Authorization\": `Bearer ${Deno.env.get(\"GHL_API_KEY\")}`,\n \"Version\": version,\n \"Accept\": \"application/json\",\n \"Content-Type\": \"application/json\",\n },\n body: body === undefined ? undefined : JSON.stringify(body),\n signal: AbortSignal.timeout(timeoutMs),\n });\n const text = await resp.text();\n if (!resp.ok) throw new Error(`ghl ${method} ${path} ${resp.status}: ${text.slice(0, 400)}`);\n return text ? JSON.parse(text) : {};\n}\n\n// Real-time availability. Returns { \"YYYY-MM-DD\": [\"2026-07-14T19:30:00-07:00\", ...] }\n// honoring the calendar's buffers, blocked slots, min-notice (2h) and booking\n// window (6 days). Slots are slot-interval spaced (15 min) for 30-min meetings.\nexport async function fetchFreeSlots(\n days = 6,\n timezone = GHL_TIMEZONE,\n): Promise> {\n const now = Date.now();\n const data = await ghl(\n \"GET\",\n `/calendars/${GHL_CALENDAR_ID}/free-slots?startDate=${now}&endDate=${now + days * 86_400_000}&timezone=${encodeURIComponent(timezone)}`,\n );\n const out: Record = {};\n for (const [day, v] of Object.entries(data)) {\n if (v && typeof v === \"object\" && Array.isArray((v as any).slots)) {\n out[day] = (v as any).slots;\n }\n }\n return out;\n}\n\n// Bookings must attach to a GHL contact; upsert dedupes by email (and phone -\n// a phone already on another contact is silently dropped, the upsert still\n// succeeds). Returns the contactId.\n// The clinic name is written to BOTH the standard companyName field and Sam's\n// user-created \"Clinic Name\" custom field (key clinic_name, id\n// gFa5tA51d3Q14cX4jPKv) - his booking notification reads the custom one.\nexport async function upsertContact(c: {\n email: string;\n firstName?: string;\n lastName?: string;\n phone?: string;\n clinicName?: string;\n}): Promise {\n const { clinicName, ...rest } = c;\n const body: Record = { locationId: GHL_LOCATION_ID, ...rest };\n if (clinicName) {\n body.companyName = clinicName;\n body.customFields = [{ key: \"clinic_name\", field_value: clinicName }];\n }\n const data = await ghl(\"POST\", \"/contacts/upsert\", body, \"2021-07-28\");\n const id = data?.contact?.id;\n if (!id) throw new Error(`ghl upsertContact: no contact id in response`);\n return id;\n}\n\n// startTime must be a free slot in ISO-with-offset form, exactly as returned by\n// fetchFreeSlots. toNotify:false suppresses GHL's native calendar emails; Sam's\n// own confirmation/reminder workflow fires regardless (it triggers on the\n// appointment itself). Returns { id, address } - address is the auto-created\n// Zoom link.\nexport async function bookAppointment(a: {\n contactId: string;\n startTime: string;\n title: string;\n}): Promise<{ id: string; address?: string }> {\n const data = await ghl(\"POST\", \"/calendars/events/appointments\", {\n calendarId: GHL_CALENDAR_ID,\n locationId: GHL_LOCATION_ID,\n contactId: a.contactId,\n startTime: a.startTime,\n title: a.title,\n appointmentStatus: \"confirmed\",\n toNotify: false,\n }, \"2021-04-15\", 60_000);\n if (!data?.id) throw new Error(\"ghl bookAppointment: no appointment id in response\");\n return { id: data.id, address: data.address };\n}\n\n// A failed booking call is only DEFINITIVE when GHL answered with an HTTP\n// error (ghl() throws `ghl POST ... : ...`). A timeout/network abort\n// is AMBIGUOUS: GHL has committed a booking ~41s after the client gave up\n// (Cornerstone, 2026-07-17). Never match on timeout message text - it differs\n// per runtime (\"Signal timed out.\" on Val.town vs \"operation was aborted due\n// to timeout\" locally); match the ghl() HTTP error prefix instead.\nexport function bookingErrorIsDefinitive(e: unknown): boolean {\n return String(e).startsWith(\"Error: ghl \") || String(e).includes(\"ghl POST\") || String(e).includes(\"ghl PUT\");\n}\n\nasync function bookingLanded(\n contactId: string,\n): Promise<{ id: string; startTime: string } | null> {\n for (const waitMs of [5_000, 10_000, 15_000]) {\n await new Promise((r) => setTimeout(r, waitMs));\n const appt = await getUpcomingAppointment(contactId).catch(() => null);\n if (appt) return appt;\n }\n return null;\n}\n\n// The prospect named an OPEN slot, so the booking must converge on booked -\n// countering with other times would contradict what we just verified (Sam's\n// rule, 2026-07-17). Ambiguous abort -> landed-check -> retry the POST ->\n// landed-check again. A definitive 400 AFTER an ambiguous attempt gets one\n// more landed-check first: \"slot no longer available\" may just mean our own\n// aborted POST finally committed. Returns:\n// booked - appointment confirmed (directly or found landed on the calendar)\n// taken - GHL definitively rejected a clean attempt (genuinely taken/closed)\n// stuck - GHL unresponsive through 2 attempts + polls (~3 min): caller must\n// alert Sam and send the prospect NOTHING.\nexport async function bookWithRecovery(\n contactId: string,\n startTime: string,\n title: string,\n): Promise<\n | { outcome: \"booked\"; apptId: string; note: string }\n | { outcome: \"taken\"; error: unknown }\n | { outcome: \"stuck\"; error: unknown }\n> {\n let sawAmbiguous = false;\n let lastErr: unknown = null;\n for (let attempt = 1; attempt <= 2; attempt++) {\n try {\n const appt = await bookAppointment({ contactId, startTime, title });\n return { outcome: \"booked\", apptId: appt.id, note: attempt > 1 ? \" (on retry)\" : \"\" };\n } catch (e) {\n lastErr = e;\n console.error(`bookAppointment attempt ${attempt} failed:`, e);\n if (bookingErrorIsDefinitive(e)) {\n if (!sawAmbiguous) return { outcome: \"taken\", error: e };\n const landed = await bookingLanded(contactId);\n if (landed) return { outcome: \"booked\", apptId: landed.id, note: \" (earlier aborted call landed)\" };\n return { outcome: \"taken\", error: e };\n }\n sawAmbiguous = true;\n const landed = await bookingLanded(contactId);\n if (landed) return { outcome: \"booked\", apptId: landed.id, note: \" (landed after aborted call)\" };\n }\n }\n return { outcome: \"stuck\", error: lastErr };\n}\n\n// CAUTION (verified): GHL does NOT exempt the appointment's own block from the\n// availability check - rescheduling to a slot that overlaps the current one\n// (e.g. 20:30 -> 20:15 for a 30-min meeting) fails with \"slot no longer\n// available\". Cancel-and-rebook if that case ever matters.\nexport async function rescheduleAppointment(appointmentId: string, startTime: string) {\n return await ghl(\"PUT\", `/calendars/events/appointments/${appointmentId}`, {\n startTime,\n toNotify: false,\n }, \"2021-04-15\", 60_000);\n}\n\n// Sets appointmentStatus to \"cancelled\" (keeps the record; slots free up\n// immediately). Never hard-DELETE events.\nexport async function cancelAppointment(appointmentId: string) {\n return await ghl(\"PUT\", `/calendars/events/appointments/${appointmentId}`, {\n appointmentStatus: \"cancelled\",\n toNotify: false,\n }, \"2021-04-15\", 60_000);\n}\n\nexport async function getAppointment(appointmentId: string) {\n const data = await ghl(\"GET\", `/calendars/events/appointments/${appointmentId}`);\n return data?.appointment ?? data?.event ?? data;\n}\n\n// Full contact object (email, additionalEmails, phone, phoneLabel,\n// additionalPhones, customFields, tags, ...). Throws on 404.\nexport async function getContact(contactId: string): Promise {\n const data = await ghl(\"GET\", `/contacts/${contactId}`, undefined, \"2021-07-28\");\n return data?.contact ?? data;\n}\n\n// All events on the campaign calendar in [startMs, endMs] - includes cancelled\n// ones (filter by appointmentStatus). Used by the CRM reconciler.\nexport async function listCalendarEvents(startMs: number, endMs: number): Promise {\n const data = await ghl(\n \"GET\",\n `/calendars/events?locationId=${GHL_LOCATION_ID}&calendarId=${GHL_CALENDAR_ID}&startTime=${startMs}&endTime=${endMs}`,\n );\n return data?.events ?? [];\n}\n\n// Tags are add-only from our side (PUT tags replaces the whole array, so we\n// merge with what's there). Used to flag duplicate contacts for Sam's review.\nexport async function addContactTag(contactId: string, tag: string): Promise {\n const c = await getContact(contactId);\n const tags: string[] = Array.isArray(c?.tags) ? c.tags : [];\n if (tags.some((t) => t.toLowerCase() === tag.toLowerCase())) return;\n await ghl(\"PUT\", `/contacts/${contactId}`, { tags: [...tags, tag] }, \"2021-07-28\");\n}\n\n// ---------- opportunities (pipeline CRM) ----------\n// NOTE: these need the opportunities scopes on the Private Integration token.\n// Until Sam adds them, every call 401s with \"not authorized for this scope\" -\n// crm.ts detects that string and degrades to contact-only sync.\n\nexport async function listPipelines(): Promise {\n const data = await ghl(\"GET\", `/opportunities/pipelines?locationId=${GHL_LOCATION_ID}`, undefined, \"2021-07-28\");\n return data?.pipelines ?? [];\n}\n\nexport async function searchOpportunitiesByContact(contactId: string): Promise {\n const data = await ghl(\n \"GET\",\n `/opportunities/search?location_id=${GHL_LOCATION_ID}&contact_id=${contactId}`,\n undefined,\n \"2021-07-28\",\n );\n return data?.opportunities ?? [];\n}\n\nexport async function createOpportunity(o: {\n pipelineId: string;\n pipelineStageId: string;\n contactId: string;\n name: string;\n}): Promise {\n const data = await ghl(\"POST\", \"/opportunities/\", {\n locationId: GHL_LOCATION_ID,\n pipelineId: o.pipelineId,\n pipelineStageId: o.pipelineStageId,\n contactId: o.contactId,\n name: o.name,\n status: \"open\",\n }, \"2021-07-28\");\n const id = data?.opportunity?.id ?? data?.id;\n if (!id) throw new Error(\"ghl createOpportunity: no id in response\");\n return id;\n}\n\nexport async function updateOpportunity(id: string, fields: Record): Promise {\n return await ghl(\"PUT\", `/opportunities/${id}`, fields, \"2021-07-28\");\n}\n\n// One-off email through GHL conversations (lands in the contact's\n// conversation thread; sender fully controllable). Used by the scheduled\n// reminder queue in crm.ts.\nexport async function sendConversationEmail(m: {\n contactId: string;\n subject: string;\n html: string;\n emailFrom: string;\n emailFromName?: string;\n}): Promise {\n return await ghl(\"POST\", \"/conversations/messages\", { type: \"Email\", ...m }, \"2021-04-15\");\n}\n\nexport async function searchOpportunitiesByPipeline(pipelineId: string): Promise {\n const data = await ghl(\n \"GET\",\n `/opportunities/search?location_id=${GHL_LOCATION_ID}&pipeline_id=${pipelineId}&limit=100`,\n undefined,\n \"2021-07-28\",\n );\n return data?.opportunities ?? [];\n}\n\n// Update fields on an existing contact. additionalEmails must be the OBJECT\n// form [{ email: \"...\" }] - the bare-string form 422s, and the upsert endpoint\n// rejects the property entirely (both verified 2026-07-15), so extra emails\n// always go through here as a second call.\nexport async function updateContact(contactId: string, fields: Record) {\n return await ghl(\"PUT\", `/contacts/${contactId}`, fields, \"2021-07-28\");\n}\n\n// Search (never creates). Returns the contactId for an exact email match, or null.\nexport async function findContactByEmail(email: string): Promise {\n const data = await ghl(\n \"GET\",\n `/contacts/?locationId=${GHL_LOCATION_ID}&query=${encodeURIComponent(email)}`,\n undefined,\n \"2021-07-28\",\n );\n const hit = (data?.contacts ?? []).find(\n (c: any) => String(c.email ?? \"\").toLowerCase() === email.toLowerCase(),\n );\n return hit?.id ?? null;\n}\n\n// Upcoming (non-cancelled, future) appointment for a contact, or null. Used for\n// self-book detection AND as an idempotency guard before booking - never book a\n// contact who already has something on the calendar. startTime in the response\n// is location-local \"YYYY-MM-DD HH:mm:ss\", so we compare lexicographically\n// against \"now\" rendered in the location timezone.\n// Transition-era lookup: does this lead have an appointment on the RETIRED\n// GHL calendar (pre-Calendly-migration booking)? Used by webhook.ts's\n// legacy_appt lane while LEGACY_GHL_APPTS=true; dead code after.\nexport async function legacyGhlAppointment(\n emails: { primary: string; additional?: string },\n): Promise<{ id: string; startTime: string } | null> {\n for (const email of [emails.primary, emails.additional]) {\n if (!email) continue;\n const contactId = await findContactByEmail(email).catch(() => null);\n if (!contactId) continue;\n const appt = await getUpcomingAppointment(contactId).catch(() => null);\n if (appt) return appt;\n }\n return null;\n}\n\nexport async function getUpcomingAppointment(\n contactId: string,\n): Promise<{ id: string; startTime: string } | null> {\n const data = await ghl(\"GET\", `/contacts/${contactId}/appointments`, undefined, \"2021-07-28\");\n const nowLocal = new Date().toLocaleString(\"sv-SE\", { timeZone: GHL_TIMEZONE }); // \"YYYY-MM-DD HH:mm:ss\"\n const upcoming = (data?.events ?? data?.appointments ?? [])\n .filter((a: any) =>\n a.appointmentStatus !== \"cancelled\" && a.appointmentStatus !== \"invalid\" &&\n String(a.startTime ?? \"\") > nowLocal\n )\n .sort((a: any, b: any) => String(a.startTime).localeCompare(String(b.startTime)));\n return upcoming[0] ? { id: upcoming[0].id, startTime: upcoming[0].startTime } : null;\n}\n",[184],"989393e8-b2de-4cc2-9758-57d8fef5d8dc","crm.ts","f53a3bf4-80ae-11f1-832a-1607ee4eb77e",90,["D",1784944420053],7,"// GHL-as-CRM sync (2026-07-15). GoHighLevel is Sam's CRM: every positive\n// reply creates/enriches a contact (person name, official clinic name from\n// the practice WEBSITE, replier-primary emails, labeled phone) and an\n// opportunity in the \"Cold Email\" pipeline (Positive -> Booked). A throttled\n// reconciler catches prospects who book themselves under a different email.\n// EVERYTHING here is best-effort: a CRM failure must never block a reply.\nimport { sqlite } from \"https://esm.town/v/std/sqlite\";\nimport { email as sendEmail } from \"https://esm.town/v/std/email\";\nimport {\n addContactTag,\n createOpportunity,\n findContactByEmail,\n getContact,\n listPipelines,\n searchOpportunitiesByContact,\n searchOpportunitiesByPipeline,\n sendConversationEmail,\n updateContact,\n updateOpportunity,\n upsertContact,\n} from \"./ghl.ts\";\nimport { calendarConfigured, listCalendarEvents } from \"./calendly.ts\";\n\nconst MODEL = \"claude-sonnet-5\";\nexport const CLINIC_FIELD_ID = \"gFa5tA51d3Q14cX4jPKv\"; // Sam's \"Clinic Name\" custom field (key clinic_name)\nexport const FREE_MAIL = new Set([\n \"gmail.com\", \"yahoo.com\", \"outlook.com\", \"hotmail.com\", \"aol.com\", \"icloud.com\",\n \"me.com\", \"msn.com\", \"live.com\", \"comcast.net\", \"att.net\", \"verizon.net\", \"sbcglobal.net\",\n]);\n\n// ---------- shared pure helpers (webhook.ts imports these too) ----------\n\nexport function stripHtml(s: string): string {\n return s.replace(//gi, \"\").replace(//gi, \"\")\n .replace(/<[^>]+>/g, \" \").replace(/\\s+/g, \" \").trim();\n}\n\n// 10-digit US normalization; \"\" when the string doesn't contain a full number.\nexport function normalizePhone(s: string): string {\n const digits = String(s).replace(/\\D/g, \"\");\n if (digits.length === 10) return digits;\n if (digits.length === 11 && digits.startsWith(\"1\")) return digits.slice(1);\n return \"\";\n}\n\nexport function leadVar(lead: any, name: string): string {\n const v = (lead?.custom_variables ?? []).find(\n (v: any) => String(v.name).toLowerCase() === name,\n )?.value;\n return v ? String(v).trim() : \"\";\n}\n\nexport function websiteFromLead(lead: any): string | null {\n const site = leadVar(lead, \"website\");\n if (site) return site.startsWith(\"http\") ? site : `https://${site}`;\n const domain = String(lead?.email ?? \"\").split(\"@\")[1]?.toLowerCase();\n if (domain && !FREE_MAIL.has(domain)) return `https://${domain}`;\n return null;\n}\n\n// Who the booking confirmation must reach (the Jim Sandford lesson): PRIMARY\n// is the address the human actually replied from; the campaign lead email\n// (often a generic info@ inbox) rides along as an additional email.\nexport type ContactEmails = { primary: string; additional?: string };\n\nexport function contactEmailsFor(reply: any, lead: any): ContactEmails {\n const replier = String(reply?.from_email_address ?? \"\").trim().toLowerCase();\n const leadEmail = String(lead?.email ?? \"\").trim().toLowerCase();\n const primary = replier || leadEmail;\n return { primary, additional: leadEmail && leadEmail !== primary ? leadEmail : undefined };\n}\n\n// Best person name available: split the extracted full name (sans \"Dr.\"),\n// else the greet-name + lead fields. personName=true only for a real\n// first+last pair - placeholders must never block a later real-name upgrade.\nexport function contactName(\n a: { prospect_full_name?: string; prospect_first_name?: string },\n lead: any,\n): { firstName?: string; lastName?: string; personName: boolean } {\n const parts = String(a.prospect_full_name ?? \"\").replace(/^dr\\.?\\s+/i, \"\").trim().split(/\\s+/).filter(Boolean);\n if (parts.length >= 2) return { firstName: parts[0], lastName: parts.slice(1).join(\" \"), personName: true };\n return {\n firstName: a.prospect_first_name ?? lead.first_name ?? undefined,\n lastName: lead.last_name ?? undefined,\n personName: false,\n };\n}\n\nexport async function lookupContact(emails: ContactEmails): Promise {\n const byPrimary = await findContactByEmail(emails.primary).catch(() => null);\n if (byPrimary) return byPrimary;\n if (emails.additional) return await findContactByEmail(emails.additional).catch(() => null);\n return null;\n}\n\n// Find-or-create for booking paths. A contact found only under the lead email\n// has the generic inbox as primary - migrate it so confirmations land with\n// the actual person. Best-effort throughout.\nexport async function ensureContact(\n emails: ContactEmails,\n info: { firstName?: string; lastName?: string; phone?: string; clinicName?: string },\n): Promise {\n const byPrimary = await findContactByEmail(emails.primary).catch(() => null);\n if (byPrimary) return byPrimary;\n if (emails.additional) {\n const byLead = await findContactByEmail(emails.additional).catch(() => null);\n if (byLead) {\n await updateContact(byLead, {\n email: emails.primary,\n additionalEmails: [{ email: emails.additional }],\n }).catch((e) => console.error(\"contact primary-email fix failed:\", e));\n return byLead;\n }\n }\n const id = await upsertContact({ email: emails.primary, ...info });\n if (emails.additional) {\n await updateContact(id, { additionalEmails: [{ email: emails.additional }] })\n .catch((e) => console.error(\"additionalEmails set failed:\", e));\n }\n return id;\n}\n\n// Clinic-name normalization for comparisons: lowercase, alphanumeric+spaces,\n// professional suffixes stripped, so \"Merrily Sandford, DDS & Associates\"\n// matches \"merrily sandford associates\".\nexport function normalizeName(s: string): string {\n return String(s ?? \"\").toLowerCase().replace(/&/g, \" and \").replace(/[^a-z0-9\\s]/g, \" \")\n .split(/\\s+/).filter((w) => w && ![\"dds\", \"dmd\", \"pc\", \"pa\", \"pllc\", \"llc\", \"inc\", \"and\", \"the\", \"of\"].includes(w))\n .join(\" \");\n}\n\nexport function emailFromText(s: string): string | null {\n return String(s ?? \"\").match(/[a-z0-9._%+-]+@[a-z0-9.-]+\\.[a-z]{2,}/i)?.[0]?.toLowerCase() ?? null;\n}\n\nfunction nonFreemailDomains(...emails: (string | undefined)[]): string[] {\n return [...new Set(\n emails.filter(Boolean).map((e) => String(e).toLowerCase().split(\"@\")[1] ?? \"\")\n .filter((d) => d && !FREE_MAIL.has(d)),\n )];\n}\n\n// Do we already hold a number the PROSPECT gave us (this reply or any prior\n// one - phone_alerted rows log every extraction)? Lead-var/office numbers\n// don't count: the direct number is the prize, so asking stays worthwhile.\n// Gates both phone asks (CTA + booking confirmation): once true, never ask.\nexport async function prospectPhoneKnown(leadId: number | undefined, extractedNow?: string | null): Promise {\n if (extractedNow) return true;\n if (!leadId) return false;\n try {\n const r = await sqlite.execute({\n sql: `SELECT 1 FROM ai_actions WHERE lead_id = ? AND action = 'phone_alerted' LIMIT 1`,\n args: [leadId],\n });\n return (r.rows?.length ?? 0) > 0;\n } catch {\n return false;\n }\n}\n\n// GHL phone-label enum (verified live): Home | Work | Mobile | Landline.\nexport function phoneLabelFor(type?: string): string | undefined {\n if (type === \"personal\") return \"Mobile\";\n if (type === \"office\") return \"Work\";\n return undefined;\n}\n\nfunction e164(raw?: string): string | undefined {\n if (!raw) return undefined;\n const n = normalizePhone(raw);\n return n ? `+1${n}` : String(raw).trim() || undefined;\n}\n\n// ---------- website intel (one fetch + one Claude call per site) ----------\n\nexport type Intel = { phones: string[]; emails: string[]; clinicName?: string; timezone?: string; evidence?: string };\n\nconst INTEL_TOOL = {\n name: \"submit_website_intel\",\n description: \"Submit facts extracted from the practice website.\",\n input_schema: {\n type: \"object\",\n properties: {\n clinic_name: {\n type: \"string\",\n description: \"The practice's official name exactly as the website presents it (header/title/footer) - the full name, not a shortened salutation form. Omit if unclear.\",\n },\n timezone: { type: \"string\", description: \"IANA US timezone from address/city/state/area code, or omit\" },\n evidence: { type: \"string\", description: \"One short sentence on what you used\" },\n },\n required: [\"evidence\"],\n },\n};\n\nexport async function claudeExtract(system: string, content: string, tool: Record): Promise {\n const resp = await fetch(\"https://api.anthropic.com/v1/messages\", {\n method: \"POST\",\n headers: {\n \"x-api-key\": Deno.env.get(\"ANTHROPIC_API_KEY\") ?? \"\",\n \"anthropic-version\": \"2023-06-01\",\n \"content-type\": \"application/json\",\n },\n body: JSON.stringify({\n model: MODEL,\n max_tokens: 500,\n system,\n tools: [tool],\n tool_choice: { type: \"tool\", name: (tool as any).name },\n messages: [{ role: \"user\", content }],\n }),\n signal: AbortSignal.timeout(30_000),\n });\n if (!resp.ok) throw new Error(`anthropic ${resp.status}: ${await resp.text()}`);\n const data = await resp.json();\n return (data.content ?? []).find((b: { type: string }) => b.type === \"tool_use\")?.input ?? {};\n}\n\nconst intelCache = new Map>();\n\n// siteOverride lets call-prep run intel on a website discovered via web\n// search when the lead record has none (the BookYourData personal-email case).\nexport function websiteIntel(lead: any, siteOverride?: string): Promise {\n const site = siteOverride ?? websiteFromLead(lead);\n if (!site) return Promise.resolve(null);\n if (!intelCache.has(site)) {\n if (intelCache.size > 100) intelCache.clear();\n intelCache.set(\n site,\n fetchIntel(site).catch((e) => {\n console.error(\"websiteIntel failed:\", e);\n intelCache.delete(site); // don't cache transient failures\n return null;\n }),\n );\n }\n return intelCache.get(site)!;\n}\n\nasync function fetchIntel(url: string): Promise {\n const resp = await fetch(url, {\n headers: { \"User-Agent\": \"Mozilla/5.0 (compatible; LuminaBot)\" },\n signal: AbortSignal.timeout(8_000),\n });\n if (!resp.ok) return null;\n const html = (await resp.text()).slice(0, 300_000);\n const phones = new Set();\n const emails = new Set();\n const scrapeContacts = (h: string) => {\n for (const m of h.matchAll(/href=[\"']tel:([^\"']+)[\"']/gi)) {\n const n = normalizePhone(m[1]);\n if (n) phones.add(n);\n }\n for (const m of h.matchAll(/\$?\\b\\d{3}\$?[\\s.-]\\s?\\d{3}[\\s.-]\\d{4}\\b/g)) {\n const n = normalizePhone(m[0]);\n if (n) phones.add(n);\n }\n for (const m of h.matchAll(/href=[\"']mailto:([^\"'?]+)/gi)) {\n const e = m[1].trim().toLowerCase();\n if (e.includes(\"@\")) emails.add(e);\n }\n for (const m of h.matchAll(/\\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\\.[A-Za-z]{2,}\\b/g)) {\n const e = m[0].toLowerCase();\n // regex over raw HTML picks up asset filenames and tracker junk\n if (!/\\.(png|jpe?g|gif|svg|webp|css|js)$/.test(e) && !/sentry|wixpress|example\\.|schema\\.org/.test(e)) {\n emails.add(e);\n }\n }\n };\n scrapeContacts(html);\n // The office email often lives only on the contact page - one bounded extra\n // fetch when the homepage has none.\n if (!emails.size) {\n try {\n const base = new URL(url);\n let contactUrl = \"\";\n for (const m of html.matchAll(/href=[\"']([^\"'#?]+)[\"']/gi)) {\n try {\n const u = new URL(m[1], base);\n if (u.host === base.host && /contact/i.test(u.pathname)) {\n contactUrl = u.href;\n break;\n }\n } catch {\n continue;\n }\n }\n if (contactUrl) {\n const cresp = await fetch(contactUrl, {\n headers: { \"User-Agent\": \"Mozilla/5.0 (compatible; LuminaBot)\" },\n signal: AbortSignal.timeout(8_000),\n });\n if (cresp.ok) scrapeContacts((await cresp.text()).slice(0, 300_000));\n }\n } catch (e) {\n console.error(\"contact-page scrape failed:\", e);\n }\n }\n const text = stripHtml(html).slice(0, 4000);\n let extracted: any = {};\n if (text) {\n try {\n extracted = await claudeExtract(\n \"Extract the dental practice's OFFICIAL name and US timezone from its website text.\",\n `Website: ${url}\\n\\n${text}`,\n INTEL_TOOL,\n );\n } catch (e) {\n console.error(\"intel extract failed:\", e);\n }\n }\n return {\n phones: [...phones],\n emails: [...emails].slice(0, 5),\n clinicName: extracted.clinic_name || undefined,\n timezone: extracted.timezone || undefined,\n evidence: extracted.evidence || undefined,\n };\n}\n\n// ---------- sqlite state (throttle + one-time alerts) ----------\n\nasync function ensureCrmState() {\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS crm_state (key TEXT PRIMARY KEY, value TEXT)`);\n}\n\nasync function getState(key: string): Promise {\n const r = await sqlite.execute({ sql: `SELECT value FROM crm_state WHERE key = ?`, args: [key] });\n return r.rows.length ? String(r.rows[0][0]) : null;\n}\n\nasync function setState(key: string, value: string) {\n await sqlite.execute({\n sql: `INSERT INTO crm_state (key, value) VALUES (?, ?) ON CONFLICT(key) DO UPDATE SET value = excluded.value`,\n args: [key, value],\n });\n}\n\n// ---------- per-lead mute (Sam is handling the thread manually) ----------\n//\n// Muted leads get NO auto-reply: the webhook exits before drafting and just\n// alerts Sam. Stored in crm_state as \"mute:\" -> ISO timestamp, matched\n// against both the campaign lead email and the address the human actually\n// replies from. Managed on the admin page (test-ghl.ts ?action=mutes);\n// manual-only, no expiry.\n\nconst muteKey = (email: string) => `mute:${email.trim().toLowerCase()}`;\n\nexport async function muteLead(email: string) {\n await ensureCrmState();\n await setState(muteKey(email), new Date().toISOString());\n}\n\nexport async function unmuteLead(email: string) {\n await ensureCrmState();\n await sqlite.execute({ sql: `DELETE FROM crm_state WHERE key = ?`, args: [muteKey(email)] });\n}\n\n// First match wins; returns the muted_at timestamp or null.\nexport async function mutedSince(emails: (string | undefined)[]): Promise {\n await ensureCrmState();\n for (const e of emails) {\n if (!e || !e.trim()) continue;\n const v = await getState(muteKey(e));\n if (v) return v;\n }\n return null;\n}\n\nexport async function listMutes(): Promise<{ email: string; muted_at: string }[]> {\n await ensureCrmState();\n const r = await sqlite.execute(`SELECT key, value FROM crm_state WHERE key LIKE 'mute:%' ORDER BY value DESC`);\n return r.rows.map((row: any) => ({ email: String(row[0]).slice(\"mute:\".length), muted_at: String(row[1]) }));\n}\n\n// ---------- opportunities: pipeline discovery + scope guard ----------\n\nfunction isScopeError(e: unknown): boolean {\n return String(e).includes(\"not authorized for this scope\");\n}\n\n// One email to Sam the first time an opportunity call hits the missing scope,\n// then silence until the scope works (flag clears on first success).\nasync function alertScopeOnce() {\n await ensureCrmState();\n if (await getState(\"opp_scope_alerted\")) return;\n await setState(\"opp_scope_alerted\", new Date().toISOString());\n try {\n await Promise.race([\n sendEmail({\n subject: \"[Reply Agent] GHL opportunities scope missing - pipeline sync is off\",\n text: \"Contacts are syncing fine, but the Private Integration token can't touch opportunities, so nothing lands in the Cold Email pipeline.\\n\\n\" +\n \"Fix (2 min): GHL -> Settings -> Private Integrations -> your integration -> Edit scopes -> add View Opportunities + Edit Opportunities. If GHL rotates the token, update GHL_API_KEY on the val.\\n\\n\" +\n \"Verify after: hit the test endpoint with ?action=crmcheck - oppScope should flip to true. Everything self-heals from there (no redeploy).\",\n }),\n new Promise((_, rej) => setTimeout(() => rej(new Error(\"email timeout\")), 8000)),\n ]);\n } catch (e) {\n console.error(\"scope alert email failed:\", e);\n }\n}\n\ntype Pipeline = { pipelineId: string; positiveId: string; bookedId: string };\nlet pipelineCache: { at: number; data: Pipeline | null } | null = null;\n\n// Discover Sam's \"Cold Email\" pipeline + Positive/Booked stages BY NAME so a\n// rename in GHL degrades to a logged skip instead of writing to a wrong stage.\nexport async function coldEmailPipeline(): Promise {\n if (pipelineCache && Date.now() - pipelineCache.at < 3_600_000) return pipelineCache.data;\n const pipes = await listPipelines(); // throws on missing scope - callers guard\n const pipe = pipes.find((p: any) => String(p.name ?? \"\").toLowerCase().includes(\"cold email\"));\n const stageId = (name: string) =>\n pipe?.stages?.find((s: any) => String(s.name ?? \"\").toLowerCase().includes(name))?.id;\n const data = pipe && stageId(\"positive\") && stageId(\"booked\")\n ? { pipelineId: pipe.id, positiveId: stageId(\"positive\")!, bookedId: stageId(\"booked\")! }\n : null;\n if (!data) console.error(\"coldEmailPipeline: pipeline or Positive/Booked stage not found by name\");\n pipelineCache = { at: Date.now(), data };\n // Scope works - clear the one-time alert flag so a future regression re-alerts.\n await ensureCrmState();\n await setState(\"opp_scope_alerted\", \"\").catch(() => {});\n return data;\n}\n\nasync function ensureOpportunity(\n contactId: string,\n name: string,\n stage: \"positive\" | \"booked\",\n): Promise<{ opportunityId?: string; note: string }> {\n try {\n const pipe = await coldEmailPipeline();\n if (!pipe) return { note: \"pipeline/stages not found by name - opportunity skipped\" };\n const opps = await searchOpportunitiesByContact(contactId);\n const existing = (opps ?? []).find((o: any) => o.pipelineId === pipe.pipelineId);\n if (existing) {\n if (stage === \"booked\" && existing.pipelineStageId !== pipe.bookedId) {\n await updateOpportunity(existing.id, { pipelineStageId: pipe.bookedId });\n return { opportunityId: existing.id, note: \"opportunity moved to Booked\" };\n }\n return { opportunityId: existing.id, note: \"opportunity already exists\" };\n }\n const id = await createOpportunity({\n pipelineId: pipe.pipelineId,\n pipelineStageId: stage === \"booked\" ? pipe.bookedId : pipe.positiveId,\n contactId,\n name,\n });\n return { opportunityId: id, note: `opportunity created in ${stage === \"booked\" ? \"Booked\" : \"Positive\"}` };\n } catch (e) {\n if (isScopeError(e)) {\n await alertScopeOnce();\n return { note: \"opportunities scope missing (Sam alerted once)\" };\n }\n throw e;\n }\n}\n\n// ---------- contact enrichment ----------\n\n// Pure decision function: given the EXISTING GHL contact and what we know\n// now, return the minimal PUT body (or null). Never blanks a field, never\n// downgrades a person name, never clobbers a manually-edited clinic name.\nexport function buildContactDelta(c: any, want: {\n primaryEmail: string;\n additionalEmail?: string;\n firstName?: string;\n lastName?: string;\n personName: boolean;\n phone?: string;\n phoneLabel?: string;\n clinic?: string;\n clinicIsOfficial: boolean;\n bisonDerived: string[]; // clinic values we're allowed to overwrite with the official one\n}): Record | null {\n const delta: Record = {};\n const existingEmail = String(c.email ?? \"\").toLowerCase();\n const existingAdditional: string[] = (c.additionalEmails ?? []).map((e: any) => String(e.email ?? \"\").toLowerCase());\n\n // Emails: migrate a generic primary to the real one; otherwise just make\n // sure both addresses are on the contact.\n if (want.primaryEmail && existingEmail !== want.primaryEmail) {\n if (want.additionalEmail && existingEmail === want.additionalEmail) {\n delta.email = want.primaryEmail;\n delta.additionalEmails = [...new Set([...existingAdditional.filter((e) => e !== want.primaryEmail), want.additionalEmail])]\n .map((e) => ({ email: e }));\n } else if (!existingAdditional.includes(want.primaryEmail)) {\n delta.additionalEmails = [...existingAdditional, want.primaryEmail].map((e) => ({ email: e }));\n }\n } else if (want.additionalEmail && want.additionalEmail !== existingEmail && !existingAdditional.includes(want.additionalEmail)) {\n delta.additionalEmails = [...existingAdditional, want.additionalEmail].map((e) => ({ email: e }));\n }\n\n // Names: fill empty; upgrade a placeholder (no lastName) to a real person\n // name; never touch an existing first+last pair.\n if (want.firstName && !c.firstName) {\n delta.firstName = want.firstName;\n if (want.lastName) delta.lastName = want.lastName;\n } else if (want.personName && want.lastName && c.firstName && !c.lastName) {\n delta.firstName = want.firstName;\n delta.lastName = want.lastName;\n }\n\n // Phone: fill empty (with label); a different number appends to\n // additionalPhones; a matching number can gain a missing label.\n if (want.phone) {\n const wantNorm = normalizePhone(want.phone);\n const existingNorm = normalizePhone(String(c.phone ?? \"\"));\n const additionalNorms = (c.additionalPhones ?? []).map((p: any) => normalizePhone(String(p.phone ?? \"\")));\n if (!c.phone) {\n delta.phone = want.phone;\n if (want.phoneLabel) delta.phoneLabel = want.phoneLabel;\n } else if (wantNorm && existingNorm === wantNorm) {\n if (want.phoneLabel && !c.phoneLabel) delta.phoneLabel = want.phoneLabel;\n } else if (wantNorm && !additionalNorms.includes(wantNorm)) {\n delta.additionalPhones = [\n ...(c.additionalPhones ?? []).map((p: any) => ({ phone: p.phone, phoneLabel: p.phoneLabel })),\n { phone: want.phone, ...(want.phoneLabel ? { phoneLabel: want.phoneLabel } : {}) },\n ];\n }\n }\n\n // Clinic: fill empty; the official website name may replace Bison-derived\n // values only. An unrecognized existing value = Sam's manual edit - leave it.\n if (want.clinic) {\n const current = String(\n (c.customFields ?? []).find((f: any) => f.id === CLINIC_FIELD_ID)?.value ?? c.companyName ?? \"\",\n );\n const curN = normalizeName(current);\n const wantN = normalizeName(want.clinic);\n const replaceable = !curN || (want.clinicIsOfficial && want.bisonDerived.some((b) => normalizeName(b) === curN));\n if (replaceable && curN !== wantN) {\n delta.companyName = want.clinic;\n delta.customFields = [{ id: CLINIC_FIELD_ID, field_value: want.clinic }];\n }\n }\n\n return Object.keys(delta).length ? delta : null;\n}\n\n// ---------- the two webhook entry points ----------\n\ntype AnalysisLike = {\n category: string;\n prospect_full_name?: string;\n prospect_first_name?: string;\n clinic_name?: string;\n phone_number?: string;\n phone_number_type?: string;\n contact_details?: string;\n};\n\nexport async function syncPositiveReply(args: {\n analysis: AnalysisLike;\n lead: any;\n reply: any;\n dryRun: boolean;\n}): Promise<{ contactId?: string; opportunityId?: string; note: string }> {\n const { analysis: a, lead, reply, dryRun } = args;\n if (!Deno.env.get(\"GHL_API_KEY\")) return { note: \"no GHL key - CRM sync skipped\" };\n // Own-domain guard: replies/bookings from @luminamg.com addresses are Sam\n // or test traffic, never a prospect - keep them out of the CRM entirely.\n if (contactEmailsFor(reply, lead).primary.endsWith(\"@luminamg.com\")) {\n return { note: \"own-domain reply - CRM sync skipped\" };\n }\n\n // Referral: the contact we track is the NEW person when their email was\n // shared; otherwise the practice inbox (we don't know the replier is the\n // right person - they just told us they aren't).\n let emails = contactEmailsFor(reply, lead);\n if (a.category === \"referral\") {\n const newEmail = emailFromText(a.contact_details ?? \"\");\n emails = newEmail\n ? { primary: newEmail, additional: String(lead?.email ?? \"\").toLowerCase() !== newEmail ? String(lead?.email ?? \"\").toLowerCase() || undefined : undefined }\n : { primary: String(lead?.email ?? \"\").toLowerCase() };\n if (!emails.primary) return { note: \"referral without any usable email - skipped\" };\n }\n\n const intel = await websiteIntel(lead);\n const bisonDerived = [a.clinic_name, lead.company, lead.first_name].filter(Boolean) as string[];\n const clinic = intel?.clinicName || a.clinic_name || lead.company || lead.first_name || undefined;\n // Referral name fields describe the REPLIER, not the new contact - skip them.\n const name = a.category === \"referral\" ? { firstName: undefined, lastName: undefined, personName: false } : contactName(a, lead);\n const firstName = name.firstName ?? clinic; // Sam's rule: no name at all -> clinic name as the visible name\n const phone = a.category === \"referral\" ? undefined : e164(a.phone_number);\n const phoneLabel = phone ? phoneLabelFor(a.phone_number_type) : undefined;\n const oppName = clinic ?? name.firstName ?? emails.primary;\n\n if (dryRun) {\n return { note: `[dry run] would sync contact ${emails.primary} (${firstName ?? \"?\"} ${name.lastName ?? \"\"}, clinic \"${clinic ?? \"?\"}\") + Positive opportunity` };\n }\n\n let contactId = await lookupContact(emails);\n let note: string;\n if (!contactId) {\n contactId = await upsertContact({\n email: emails.primary,\n firstName,\n lastName: name.lastName,\n phone,\n clinicName: clinic,\n });\n const extras: Record = {};\n if (emails.additional) extras.additionalEmails = [{ email: emails.additional }];\n if (phoneLabel) extras.phoneLabel = phoneLabel;\n if (Object.keys(extras).length) {\n await updateContact(contactId, extras).catch((e) => console.error(\"contact extras failed:\", e));\n }\n note = `contact created (${emails.primary})`;\n } else {\n const existing = await getContact(contactId);\n const delta = buildContactDelta(existing, {\n primaryEmail: emails.primary,\n additionalEmail: emails.additional,\n firstName: name.firstName, // only real names participate in upgrades\n lastName: name.lastName,\n personName: name.personName,\n phone,\n phoneLabel,\n clinic,\n clinicIsOfficial: Boolean(intel?.clinicName),\n bisonDerived,\n });\n if (delta) {\n await updateContact(contactId, delta);\n note = `contact enriched (${Object.keys(delta).join(\", \")})`;\n } else {\n note = \"contact up to date\";\n }\n }\n\n const opp = await ensureOpportunity(contactId, oppName, \"positive\");\n return { contactId, opportunityId: opp.opportunityId, note: `${note}; ${opp.note}` };\n}\n\nexport async function markBooked(contactId: string | undefined, oppName: string, dryRun: boolean): Promise {\n if (!contactId) return \"no contactId - skipped\";\n if (dryRun) return \"[dry run] would move opportunity to Booked\";\n const opp = await ensureOpportunity(contactId, oppName, \"booked\");\n return opp.note;\n}\n\n// ---------- scheduled one-off emails ----------\n// Queue rows land via direct sqlite INSERT (see NOTES.md); this checker runs\n// on webhook traffic, on every GHL-workflow ping, and from the interval val -\n// whichever fires first sends. Claim-by-UPDATE keeps concurrent invocations\n// from double-sending.\n\nasync function ensureEmailQueue() {\n await sqlite.execute(\n `CREATE TABLE IF NOT EXISTS scheduled_emails (\n id INTEGER PRIMARY KEY AUTOINCREMENT,\n due_at TEXT, -- UTC 'YYYY-MM-DD HH:MM:SS'\n contact_id TEXT,\n subject TEXT,\n html TEXT,\n email_from TEXT,\n from_name TEXT,\n status TEXT DEFAULT 'pending',\n sent_at TEXT,\n detail TEXT)`,\n );\n}\n\nexport async function checkDueEmails(dryRun: boolean): Promise {\n if (!Deno.env.get(\"GHL_API_KEY\")) return \"no GHL key\";\n await ensureEmailQueue();\n const due = await sqlite.execute({\n sql: `SELECT id, contact_id, subject, html, email_from, from_name FROM scheduled_emails\n WHERE status = 'pending' AND due_at <= datetime('now') ORDER BY id`,\n args: [],\n });\n if (!due.rows.length) return \"no due emails\";\n const notes: string[] = [];\n for (const [id, contactId, subject, html, emailFrom, fromName] of due.rows as any[]) {\n if (dryRun) {\n notes.push(`[dry run] would send #${id} \"${subject}\"`);\n continue;\n }\n const claim = await sqlite.execute({\n sql: `UPDATE scheduled_emails SET status = 'sending' WHERE id = ? AND status = 'pending'`,\n args: [id],\n });\n if (claim.rowsAffected !== 1) continue; // another invocation got it\n try {\n await sendConversationEmail({\n contactId: String(contactId),\n subject: String(subject),\n html: String(html),\n emailFrom: String(emailFrom),\n emailFromName: fromName ? String(fromName) : undefined,\n });\n await sqlite.execute({\n sql: `UPDATE scheduled_emails SET status = 'sent', sent_at = datetime('now') WHERE id = ?`,\n args: [id],\n });\n notes.push(`sent #${id} \"${subject}\"`);\n } catch (e) {\n await sqlite.execute({\n sql: `UPDATE scheduled_emails SET status = 'failed', detail = ? WHERE id = ?`,\n args: [String(e).slice(0, 500), id],\n });\n notes.push(`FAILED #${id}: ${String(e).slice(0, 120)}`);\n // A dead reminder is time-sensitive - Sam needs to know immediately.\n try {\n await Promise.race([\n sendEmail({\n subject: `[Reply Agent] Scheduled email #${id} FAILED - send manually`,\n text: `Subject: ${subject}\\nContact: ${contactId}\\n\\nError:\\n${String(e).slice(0, 800)}`,\n }),\n new Promise((_, rej) => setTimeout(() => rej(new Error(\"email timeout\")), 8000)),\n ]);\n } catch { /* alert is best-effort */ }\n }\n }\n return notes.join(\"; \");\n}\n\n// ---------- self-booking reconciler ----------\n\n// Pure matcher: does this booked contact belong to one of our Positive-stage\n// contacts? Match on shared non-freemail email domain first (jim@x.com vs\n// info@x.com), then on normalized clinic name (covers personal-gmail\n// bookings). Sam's design, 2026-07-15.\nexport function matchBookingToCandidates(\n booked: { emails: (string | undefined)[]; clinic?: string },\n candidates: { contactId: string; opportunityId: string; emails: (string | undefined)[]; clinic?: string }[],\n): { contactId: string; opportunityId: string } | null {\n const bookedDomains = nonFreemailDomains(...booked.emails);\n const bookedClinic = normalizeName(booked.clinic ?? \"\");\n let clinicMatch: { contactId: string; opportunityId: string } | null = null;\n for (const cand of candidates) {\n const candDomains = nonFreemailDomains(...cand.emails);\n if (bookedDomains.some((d) => candDomains.includes(d))) return cand; // domain match wins\n if (!clinicMatch && bookedClinic && normalizeName(cand.clinic ?? \"\") === bookedClinic) clinicMatch = cand;\n }\n return clinicMatch;\n}\n\nfunction contactEmailsOf(c: any): (string | undefined)[] {\n return [c?.email, ...((c?.additionalEmails ?? []).map((e: any) => e.email))];\n}\n\nfunction contactClinicOf(c: any): string | undefined {\n return (c?.customFields ?? []).find((f: any) => f.id === CLINIC_FIELD_ID)?.value ?? c?.companyName ?? undefined;\n}\n\n// Sweep upcoming CALENDLY bookings and make the GHL pipeline agree with the\n// calendar. Throttled (>=15 min) because it piggybacks on webhook traffic;\n// force=true from the manual test-endpoint action and Calendly webhooks.\nexport async function reconcile(dryRun: boolean, force = false): Promise {\n if (!Deno.env.get(\"GHL_API_KEY\")) return \"no GHL key\";\n if (!calendarConfigured()) return \"calendly not configured\";\n await ensureCrmState();\n if (!force) {\n const last = Number(await getState(\"reconcile_at\") ?? 0);\n if (Date.now() - last < 15 * 60_000) return \"throttled\";\n }\n await setState(\"reconcile_at\", String(Date.now()));\n\n let pipe: Pipeline | null;\n try {\n pipe = await coldEmailPipeline();\n } catch (e) {\n if (isScopeError(e)) {\n await alertScopeOnce();\n return \"opportunities scope missing\";\n }\n throw e;\n }\n if (!pipe) return \"pipeline/stages not found by name\";\n\n const now = Date.now();\n const events = (await listCalendarEvents(now - 24 * 3_600_000, now + 14 * 86_400_000))\n .filter((e) => e.status === \"active\" && e.inviteeEmail);\n const notes: string[] = [];\n // Candidate pool (contacts sitting in Positive) is loaded lazily - most\n // sweeps find every booking already attached to an opportunity.\n let candidates: { contactId: string; opportunityId: string; emails: (string | undefined)[]; clinic?: string }[] | null = null;\n\n for (const ev of events.slice(0, 25)) {\n try {\n // Own-domain guard: bookings by @luminamg.com invitees are Sam or test\n // traffic (booktest/caltest), never a prospect - skip CRM processing.\n if (ev.inviteeEmail.toLowerCase().endsWith(\"@luminamg.com\")) continue;\n // Calendly events carry invitee identity, not a GHL contact - resolve\n // (or create) the CRM contact by the invitee email. Agent-booked\n // invitees always exist already via ensureContact; a missing contact\n // means a self-booking under an email GHL has never seen.\n let contactId = await findContactByEmail(ev.inviteeEmail).catch(() => null);\n if (!contactId) {\n if (dryRun) {\n notes.push(`[dry run] would create contact for unknown invitee ${ev.inviteeEmail}`);\n continue;\n }\n const nameParts = String(ev.inviteeName ?? \"\").split(/\\s+/).filter(Boolean);\n contactId = await upsertContact({\n email: ev.inviteeEmail,\n firstName: nameParts[0],\n lastName: nameParts.slice(1).join(\" \") || undefined,\n });\n }\n const opps = await searchOpportunitiesByContact(contactId);\n const own = (opps ?? []).find((o: any) => o.pipelineId === pipe.pipelineId);\n if (own) {\n if (own.pipelineStageId === pipe.positiveId) {\n if (!dryRun) await updateOpportunity(own.id, { pipelineStageId: pipe.bookedId });\n notes.push(`moved ${contactId} Positive->Booked${dryRun ? \" [dry run]\" : \"\"}`);\n }\n continue;\n }\n // Booking under a contact with no opportunity: probably a self-booking\n // duplicate of one of our Positive contacts.\n const booked = await getContact(contactId).catch(() => null) ??\n { id: contactId, email: ev.inviteeEmail };\n if (candidates === null) {\n candidates = [];\n const posOpps = await searchPositiveOpportunities(pipe);\n for (const o of posOpps.slice(0, 25)) {\n const c = await getContact(o.contactId ?? o.contact?.id).catch(() => null);\n if (c) candidates.push({ contactId: c.id, opportunityId: o.id, emails: contactEmailsOf(c), clinic: contactClinicOf(c) });\n }\n }\n const match = matchBookingToCandidates(\n // inviteeName doubles as a clinic-name candidate: Calendly self-book\n // forms often get the practice name typed into the name field.\n { emails: contactEmailsOf(booked), clinic: contactClinicOf(booked) ?? (ev.inviteeName || undefined) },\n candidates,\n );\n if (match) {\n if (!dryRun) {\n // The contact holding the APPOINTMENT becomes canonical: pull the\n // original's clinic/emails onto it, point the opportunity at it,\n // and tag the original for Sam to merge/delete (never auto-delete -\n // he may have added notes).\n const orig = await getContact(match.contactId).catch(() => null);\n const delta = orig\n ? buildContactDelta(booked, {\n primaryEmail: String(booked.email ?? \"\").toLowerCase(),\n additionalEmail: String(orig.email ?? \"\").toLowerCase() || undefined,\n firstName: undefined,\n lastName: undefined,\n personName: false,\n phone: orig.phone ?? undefined,\n phoneLabel: orig.phoneLabel ?? undefined,\n clinic: contactClinicOf(orig),\n clinicIsOfficial: false,\n bisonDerived: [],\n })\n : null;\n if (delta) await updateContact(contactId, delta).catch((e) => console.error(\"booked-contact enrich failed:\", e));\n try {\n await updateOpportunity(match.opportunityId, { contactId: contactId, pipelineStageId: pipe.bookedId });\n } catch {\n await updateOpportunity(match.opportunityId, { pipelineStageId: pipe.bookedId });\n }\n await addContactTag(match.contactId, \"duplicate-superseded\").catch(() => {});\n }\n notes.push(`self-booking ${contactId} matched Positive contact ${match.contactId} -> Booked${dryRun ? \" [dry run]\" : \"\"}`);\n candidates = candidates.filter((c) => c.opportunityId !== match.opportunityId);\n } else {\n if (!dryRun) {\n const oppName = contactClinicOf(booked) ??\n (`${booked.firstName ?? \"\"} ${booked.lastName ?? \"\"}`.trim() || String(booked.email ?? contactId));\n await ensureOpportunity(contactId, oppName, \"booked\");\n }\n notes.push(`unmatched booking ${contactId} -> Booked opportunity created${dryRun ? \" [dry run]\" : \"\"}`);\n }\n } catch (e) {\n if (isScopeError(e)) {\n await alertScopeOnce();\n return \"opportunities scope missing\";\n }\n console.error(\"reconcile event failed:\", ev.id, e);\n notes.push(`event ${ev.id} failed: ${String(e).slice(0, 120)}`);\n }\n }\n return notes.length ? notes.join(\"; \") : `all ${events.length} bookings already reconciled`;\n}\n\n// Filter client-side on the stage id - the search endpoint's stage filter\n// params are inconsistent across accounts, the pipeline filter is not.\nasync function searchPositiveOpportunities(pipe: Pipeline): Promise {\n return (await searchOpportunitiesByPipeline(pipe.pipelineId)).filter((o: any) =>\n o.pipelineStageId === pipe.positiveId && o.status !== \"lost\"\n );\n}\n",[192],"17104274-9f3d-499b-a47f-cbcf8f8c8e00","callprep.ts","e1685a74-86e5-11f1-b50c-1607ee4eb77e",91,["D",1784944420416],"// Call-prep alert for EVERY interested reply (2026-07-23). Sam calls all\n// positive replies now, so the notification is his call-prep sheet: best\n// number to dial (labeled personal vs office), the decisionmaker (extracted\n// name when the replier IS one, otherwise researched off the practice\n// website), reply time in PT, offer, their message, and their timezone.\n// Everything degrades to a labeled fallback - the alert must never be lost.\n// No sqlite/email imports here: webhook.ts sends and logs; this file only\n// builds, so it stays offline-testable and under Val.town's 80k file cap.\nimport type { Analysis } from \"./analysis.ts\";\nimport {\n claudeExtract,\n leadVar,\n normalizePhone,\n stripHtml,\n websiteFromLead,\n websiteIntel,\n} from \"./crm.ts\";\n\n// ---------- decisionmaker research (one site fetch chain + one Claude call) ----------\n\ntype Decisionmaker = { name: string; role: string };\n\n// Flat fields on purpose: a nested array here made the model occasionally\n// return the whole payload as a JSON string in one property.\nconst DM_TOOL = {\n name: \"submit_decisionmakers\",\n description: \"Submit the likely decisionmakers found for this dental practice.\",\n input_schema: {\n type: \"object\",\n properties: {\n dm1_name: { type: \"string\", description: \"Most likely decisionmaker's full name, 'Dr.'-prefixed when a doctor. Omit if none found - never invent.\" },\n dm1_role: { type: \"string\", description: \"Their short lowercase role, e.g. 'principal dentist', 'office manager'\" },\n dm2_name: { type: \"string\", description: \"Second candidate's full name, only when the site clearly names one\" },\n dm2_role: { type: \"string\", description: \"Second candidate's short lowercase role\" },\n evidence: { type: \"string\", description: \"One short sentence on where the names came from\" },\n },\n required: [\"evidence\"],\n },\n};\n\nasync function fetchPage(url: string): Promise {\n const resp = await fetch(url, {\n headers: { \"User-Agent\": \"Mozilla/5.0 (compatible; LuminaBot)\" },\n signal: AbortSignal.timeout(8_000),\n });\n if (!resp.ok) throw new Error(`fetch ${url} -> ${resp.status}`);\n return (await resp.text()).slice(0, 300_000);\n}\n\n// Team/about/doctor pages usually hold the names the homepage doesn't.\nfunction teamLinks(homeHtml: string, site: string): string[] {\n const base = new URL(site);\n const seen = new Set();\n for (const m of homeHtml.matchAll(/href=[\"']([^\"'#?]+)[\"']/gi)) {\n let u: URL;\n try {\n u = new URL(m[1], base);\n } catch {\n continue;\n }\n if (u.host !== base.host) continue;\n if (/\\.(png|jpe?g|gif|svg|ico|css|js|xml|pdf|webp)$/i.test(u.pathname)) continue;\n if (!/about|team|staff|meet|doctor|dentist|provider|our-practice|who-we-are/i.test(u.pathname)) continue;\n seen.add(u.href);\n if (seen.size >= 2) break;\n }\n return [...seen];\n}\n\nconst dmCache = new Map>();\n\nfunction researchDecisionmakers(site: string | null, hints: string): Promise<{ people: Decisionmaker[]; evidence?: string } | null> {\n if (!site) return Promise.resolve(null);\n if (!dmCache.has(site)) {\n if (dmCache.size > 100) dmCache.clear();\n dmCache.set(\n site,\n (async () => {\n const homeHtml = await fetchPage(site);\n const subpages = await Promise.all(\n teamLinks(homeHtml, site).map((u) => fetchPage(u).catch(() => \"\")),\n );\n const text = [\n `HOMEPAGE (${site}):\\n${stripHtml(homeHtml).slice(0, 4000)}`,\n ...subpages.filter(Boolean).map((h, i) => `SUBPAGE ${i + 1}:\\n${stripHtml(h).slice(0, 3000)}`),\n ].join(\"\\n\\n\");\n const out = await claudeExtract(\n \"Identify who most likely DECIDES on marketing spend at this dental practice: \" +\n \"the principal/owner dentist, office manager, practice manager, or operations manager. \" +\n \"Use the website text plus the context clues. Return up to 2 real named people \" +\n \"(most likely first) - never invent names, and skip generic staff (front desk, hygienists, associates).\",\n `${hints}\\n\\n${text}`,\n DM_TOOL,\n );\n const junk = /^(unknown|none|n\\/a|not (provided|found|named))$/i;\n const people: Decisionmaker[] = [\n { name: String(out.dm1_name ?? \"\").trim(), role: String(out.dm1_role ?? \"\").trim() },\n { name: String(out.dm2_name ?? \"\").trim(), role: String(out.dm2_role ?? \"\").trim() },\n ].filter((p) => p.name && p.role && !junk.test(p.name));\n return { people, evidence: out.evidence };\n })().catch((e) => {\n console.error(\"decisionmaker research failed:\", e);\n dmCache.delete(site); // don't cache transient failures\n return null;\n }),\n );\n }\n return dmCache.get(site)!;\n}\n\n// ---------- practice discovery (web search, for leads with no website) ----------\n// BookYourData-style leads are dentists' PERSONAL emails: no website var, no\n// office domain to derive one from. But they carry name + practice + street\n// address, which is enough for one web-search call to find the official site\n// (and often the office email/phone). The found site then feeds the normal\n// intel + decisionmaker pipeline.\n\nexport type DiscoveredPractice = {\n website?: string;\n office_email?: string;\n office_phone?: string;\n practice_name?: string;\n confidence?: string;\n evidence?: string;\n};\n\n// Flat fields on purpose - same sonnet nested-payload gotcha as DM_TOOL.\nconst PRACTICE_TOOL = {\n name: \"submit_practice\",\n description: \"Submit what you found about this dentist's practice. Omit any field you could not verify - never invent.\",\n input_schema: {\n type: \"object\",\n properties: {\n website: { type: \"string\", description: \"Official practice website URL, only if confident it is THIS dentist's practice (name + city/state match)\" },\n office_email: { type: \"string\", description: \"The practice's office/front-desk email if seen (e.g. on their website or a listing)\" },\n office_phone: { type: \"string\", description: \"The practice's office phone number if seen\" },\n practice_name: { type: \"string\", description: \"Official practice name\" },\n confidence: { type: \"string\", enum: [\"high\", \"medium\", \"low\"], description: \"How sure you are this is the right practice\" },\n evidence: { type: \"string\", description: \"One short sentence on how you matched it\" },\n },\n required: [\"evidence\"],\n },\n};\n\nconst discoveryCache = new Map>();\n\nexport function discoverPractice(lead: any, analysis: Analysis): Promise {\n const key = String(lead?.email ?? \"\").toLowerCase();\n if (!key) return Promise.resolve(null);\n if (!discoveryCache.has(key)) {\n if (discoveryCache.size > 100) discoveryCache.clear();\n discoveryCache.set(\n key,\n runDiscovery(lead, analysis).catch((e) => {\n console.error(\"practice discovery failed:\", e);\n discoveryCache.delete(key); // don't cache transient failures\n return null;\n }),\n );\n }\n return discoveryCache.get(key)!;\n}\n\nasync function runDiscovery(lead: any, analysis: Analysis): Promise {\n const facts = [\n `Dentist: ${analysis.prospect_full_name ?? lead.first_name ?? \"?\"}${lead.last_name ? ` ${lead.last_name}` : \"\"}`,\n `Company field on the lead: ${lead.company ?? \"?\"} (often the dentist's own name, sometimes the practice name)`,\n `Practice name if known: ${analysis.clinic_name ?? \"?\"}`,\n `Address: ${[leadVar(lead, \"address\"), leadVar(lead, \"city\"), leadVar(lead, \"state\")].filter(Boolean).join(\", \") || \"?\"}`,\n `Their personal email: ${lead.email ?? \"?\"} (the handle sometimes hints at the practice name)`,\n ].join(\"\\n\");\n const tools = [\n { type: \"web_search_20260209\", name: \"web_search\", max_uses: 3 },\n { type: \"web_fetch_20260209\", name: \"web_fetch\", max_uses: 2 },\n PRACTICE_TOOL,\n ];\n const messages: any[] = [{\n role: \"user\",\n content: `Find the OFFICIAL website of this US dentist's practice, plus the office email and office phone if you can (their contact page usually has both).\\n\\n${facts}\\n\\nSearch for the dentist by name with their city/state (and the address if needed). Verify the site actually belongs to THIS dentist - the name and location must match; dentists share names. When done, call submit_practice with what you verified. If you cannot find their practice, call submit_practice with just evidence explaining why.`,\n }];\n // Server tools run in a server-side loop that can pause; resend to resume.\n // Hard overall deadline: this runs BEFORE the reply is drafted and sent, so\n // it must never eat the val's run budget - a lost discovery just degrades\n // the alert.\n const deadline = Date.now() + 150_000;\n for (let round = 0; round < 3; round++) {\n const remaining = deadline - Date.now();\n if (remaining < 10_000) return null;\n const resp = await fetch(\"https://api.anthropic.com/v1/messages\", {\n method: \"POST\",\n headers: {\n \"x-api-key\": Deno.env.get(\"ANTHROPIC_API_KEY\") ?? \"\",\n \"anthropic-version\": \"2023-06-01\",\n \"content-type\": \"application/json\",\n },\n body: JSON.stringify({ model: \"claude-sonnet-5\", max_tokens: 3000, tools, messages }),\n signal: AbortSignal.timeout(Math.min(130_000, remaining)),\n });\n if (!resp.ok) throw new Error(`anthropic ${resp.status}: ${await resp.text()}`);\n const data = await resp.json();\n const submit = (data.content ?? []).find(\n (b: any) => b.type === \"tool_use\" && b.name === \"submit_practice\",\n );\n if (submit) {\n const out = submit.input ?? {};\n const clean = (v: any) => {\n const s = String(v ?? \"\").trim();\n return s && !/^(unknown|none|n\\/a|not found)$/i.test(s) ? s : undefined;\n };\n const website = clean(out.website);\n return {\n website: website && !website.startsWith(\"http\") ? `https://${website}` : website,\n office_email: clean(out.office_email)?.toLowerCase(),\n office_phone: clean(out.office_phone),\n practice_name: clean(out.practice_name),\n confidence: clean(out.confidence),\n evidence: clean(out.evidence),\n };\n }\n if (data.stop_reason === \"pause_turn\") {\n messages.push({ role: \"assistant\", content: data.content });\n continue;\n }\n return null; // ended without submitting - treat as not found\n }\n return null;\n}\n\n// ---------- formatting helpers ----------\n\nexport function fmtPhone(raw: string): string {\n const d = normalizePhone(raw);\n return d ? `(${d.slice(0, 3)}) ${d.slice(3, 6)}-${d.slice(6)}` : String(raw).trim();\n}\n\n// Sam's call-script format: \"Dr. \" for doctors, first name otherwise.\nexport function scriptName(name: string): string {\n const tokens = name.trim().replace(/,.*$/, \"\").split(/\\s+/).filter(Boolean);\n if (!tokens.length) return name.trim();\n if (/^dr\\.?$/i.test(tokens[0])) {\n return `Dr. ${tokens[tokens.length - 1] === tokens[0] ? name.trim() : tokens[tokens.length - 1]}`;\n }\n return tokens[0];\n}\n\n// Reply time in Sam's timezone: time, date, day of week.\nexport function whenInPT(iso?: string): string {\n const d = iso && !isNaN(Date.parse(iso)) ? new Date(iso) : new Date();\n const time = d.toLocaleString(\"en-US\", {\n timeZone: \"America/Los_Angeles\", hour: \"numeric\", minute: \"2-digit\", timeZoneName: \"short\",\n });\n const date = d.toLocaleString(\"en-US\", {\n timeZone: \"America/Los_Angeles\", weekday: \"long\", month: \"long\", day: \"numeric\", year: \"numeric\",\n });\n return `${time} - ${date}`;\n}\n\nfunction localClock(tz: string): string | null {\n try {\n return new Date().toLocaleString(\"en-US\", { timeZone: tz, hour: \"numeric\", minute: \"2-digit\" });\n } catch {\n return null; // model handed back a non-IANA string\n }\n}\n\n// Their message without the quoted thread under it (fall back to the full\n// text if stripping would leave nothing).\nexport function freshMessage(reply: any): string {\n const full = String(reply?.text_body ?? stripHtml(String(reply?.html_body ?? \"\"))).trim();\n // Gmail wraps the \"On ... wrote:\" attribution across lines - match lazily\n // across newlines up to the first \"wrote:\" at an end of line.\n let cut = full.replace(/^\\s*On [\\s\\S]{0,300}?wrote:\\s*$[\\s\\S]*/m, \"\");\n cut = cut.split(\"\\n\").filter((l) => !l.startsWith(\">\")).join(\"\\n\").trim();\n return cut.length >= 2 ? cut : full;\n}\n\n// ---------- the alert ----------\n\nexport type CallPrepInput = {\n analysis: Analysis;\n lead: any;\n reply: any;\n leadLabel: string;\n senderEmail?: string;\n senderName?: string;\n};\n\nexport async function buildCallPrep(input: CallPrepInput): Promise<{ subject: string; text: string; detail: string }> {\n const { analysis, lead, reply, leadLabel } = input;\n const siteOnFile = websiteFromLead(lead);\n // No website on file = a personal-email lead (the BookYourData lists) - go\n // find the practice on the web from name + address before building the sheet.\n const discovered = siteOnFile ? null : await discoverPractice(lead, analysis).catch(() => null);\n const site = siteOnFile ?? discovered?.website ?? null;\n const siteNote = siteOnFile || !site ? \"\" : \" (found via web search - verify)\";\n // Cached from the CRM sync moments earlier when the site is on file.\n const intel = site ? await websiteIntel(lead, site).catch(() => null) : null;\n const sitePhones = intel?.phones ?? [];\n const leadMobile = leadVar(lead, \"phone\");\n const leadMobileDigits = normalizePhone(leadMobile);\n\n // Best number: their own number from the reply beats anything; a match\n // against the website's line(s) is the strongest personal-vs-office signal,\n // the model's signature read backs it up; website office line is the fallback.\n const replyDigits = normalizePhone(analysis.phone_number ?? \"\");\n let phoneLine: string;\n let phoneKind: string;\n if (analysis.phone_number) {\n if (replyDigits && sitePhones.includes(replyDigits)) {\n phoneKind = \"office\";\n phoneLine = `${fmtPhone(analysis.phone_number)} - OFFICE (same number as their website)`;\n } else if (replyDigits && sitePhones.length) {\n phoneKind = \"personal\";\n phoneLine = `${fmtPhone(analysis.phone_number)} - PERSONAL/DIRECT (differs from their website's ${sitePhones.map(fmtPhone).join(\", \")})`;\n } else if (analysis.phone_number_type && analysis.phone_number_type !== \"unknown\") {\n phoneKind = analysis.phone_number_type;\n phoneLine = `${fmtPhone(analysis.phone_number)} - ${analysis.phone_number_type.toUpperCase()}` +\n (analysis.phone_number_evidence ? ` (${analysis.phone_number_evidence})` : \" (from their reply)\");\n } else {\n phoneKind = \"unknown type\";\n phoneLine = `${fmtPhone(analysis.phone_number)} - from their reply, type unknown (nothing on their website to compare)`;\n }\n if (leadMobileDigits && leadMobileDigits !== replyDigits) {\n phoneLine += ` (also on file: ${fmtPhone(leadMobile)}, mobile from our lead list)`;\n }\n } else if (leadMobileDigits) {\n phoneKind = \"mobile (lead list)\";\n phoneLine = `${fmtPhone(leadMobile)} - MOBILE from our lead list (no number in their reply)`;\n const office = sitePhones[0] ?? (discovered?.office_phone ? normalizePhone(discovered.office_phone) : \"\");\n if (office && office !== leadMobileDigits) {\n phoneLine += ` (office: ${fmtPhone(office)}${sitePhones[0] ? \"\" : \" - via web search, verify\"})`;\n }\n } else if (sitePhones.length) {\n phoneKind = \"office (website)\";\n phoneLine = `${fmtPhone(sitePhones[0])} - OFFICE line from their website${siteNote} (no number in their reply)`;\n } else if (discovered?.office_phone) {\n phoneKind = \"office (web search)\";\n phoneLine = `${fmtPhone(discovered.office_phone)} - OFFICE, found via web search (verify)`;\n } else {\n phoneKind = \"none\";\n phoneLine = `none found - no number in their reply or on their website${site ? \"\" : \" (no website found)\"}`;\n }\n\n // Decisionmaker: use the replier's name when the model judged them to BE\n // one; otherwise research the website for two candidates.\n const replierName = analysis.prospect_full_name ?? analysis.prospect_first_name;\n const dmLines: string[] = [];\n let dmDetail: string;\n const leadName = String(lead.first_name ?? \"\").trim();\n // Personal-email lists (BookYourData) put the dentist's own name on the\n // lead - the inbox owner IS the decisionmaker, no research needed.\n const leadIsDentist = !siteOnFile && /^dr\\.?\\s/i.test(leadName);\n if (replierName && analysis.replier_is_decisionmaker === true) {\n dmLines.push(`Decisionmaker: ${scriptName(replierName)} (replied themselves${analysis.prospect_full_name ? ` - ${analysis.prospect_full_name}` : \"\"})`);\n dmDetail = `dm=${scriptName(replierName)}`;\n } else if (leadIsDentist && !replierName) {\n dmLines.push(`Decisionmaker: ${scriptName(leadName)}${lead.company ? ` (${lead.company})` : \"\"} - this lead is the dentist's personal email, so you're likely talking to them directly`);\n dmDetail = `dm=${scriptName(leadName)} (lead record)`;\n } else {\n const hints = [\n `Practice: ${analysis.clinic_name ?? lead.company ?? leadVar(lead, \"clinic\") ?? \"?\"}`,\n `Location: ${[leadVar(lead, \"city\"), leadVar(lead, \"state\")].filter(Boolean).join(\", \") || \"?\"}`,\n `Lead label: ${leadLabel}`,\n `The reply came from: ${reply.from_name ?? \"?\"} <${reply.from_email_address ?? \"?\"}>` +\n (replierName ? ` - signed \"${replierName}\", who reads as staff rather than the decisionmaker` : \" - no personal name found\"),\n `Their reply (for signature clues): ${freshMessage(reply).slice(0, 600)}`,\n ].join(\"\\n\");\n const research = await researchDecisionmakers(site, hints);\n if (research?.people.length) {\n research.people.forEach((p, i) => dmLines.push(`Decisionmaker #${i + 1}: ${p.name}, ${p.role}`));\n if (research.people.length === 1) dmLines.push(`(only one clear candidate on their site)`);\n if (research.evidence) dmLines.push(`(source: ${research.evidence})`);\n dmDetail = `dm researched: ${research.people.map((p) => p.name).join(\" / \")}`;\n } else if (leadIsDentist) {\n dmLines.push(`Decisionmaker: ${scriptName(leadName)}${lead.company ? ` (${lead.company})` : \"\"} - this lead is the dentist's personal email, so you're likely talking to them directly`);\n dmDetail = `dm=${scriptName(leadName)} (lead record)`;\n } else {\n dmLines.push(`Decisionmaker: unknown - couldn't find one on their website${site ? ` (${site})` : \" (no website found)\"}`);\n dmDetail = \"dm unknown\";\n }\n if (replierName) dmLines.push(`(reply was signed by ${replierName} - reads as staff, not the decisionmaker)`);\n }\n\n // Office email: scraped from the site when we have one, else whatever web\n // search surfaced. The lead/reply addresses themselves don't count.\n const ownEmails = new Set(\n [lead.email, reply.from_email_address].filter(Boolean).map((s: any) => String(s).toLowerCase()),\n );\n let siteHost = \"\";\n try {\n siteHost = site ? new URL(site).hostname.replace(/^www\\./, \"\") : \"\";\n } catch { /* malformed site string */ }\n const scrapedAll = (intel?.emails ?? []).filter((e) => !ownEmails.has(e));\n // A page can carry the web designer's or a vendor's address - the practice's\n // own domain wins when present.\n const sameDomain = siteHost ? scrapedAll.filter((e) => e.endsWith(`@${siteHost}`)) : [];\n const scraped = sameDomain.length ? sameDomain : scrapedAll;\n const searched = discovered?.office_email && !ownEmails.has(discovered.office_email) ? [discovered.office_email] : [];\n let officeEmailLine: string | null = null;\n if (scraped.length) {\n officeEmailLine = `Office email: ${scraped.slice(0, 2).join(\", \")} (from their website${siteNote})`;\n } else if (searched.length) {\n officeEmailLine = `Office email: ${searched[0]} (found via web search - verify)`;\n } else if (!siteOnFile) {\n officeEmailLine = `Office email: none found`;\n }\n\n const offer = analysis.original_offer ||\n [leadVar(lead, \"treatment number\"), leadVar(lead, \"treatment type\"), leadVar(lead, \"treatment suffix\")]\n .filter(Boolean).join(\" \") || \"(unknown - check the thread in Bison)\";\n\n const fromLine = `${reply.from_name ?? \"\"} <${reply.from_email_address ?? lead.email ?? \"?\"}>`.trim() +\n (reply.from_email_address && lead.email && reply.from_email_address.toLowerCase() !== lead.email.toLowerCase()\n ? ` (lead email: ${lead.email})`\n : \"\");\n\n const tz = analysis.timezone ?? intel?.timezone;\n const clock = tz ? localClock(tz) : null;\n const tzLine = tz\n ? `${tz}${clock ? ` - their local time right now: ${clock}` : \"\"}`\n : \"unknown\";\n\n const text = [\n `CALL PREP`,\n `Best number: ${phoneLine}`,\n ...(officeEmailLine ? [officeEmailLine] : []),\n ...dmLines,\n `Replied: ${whenInPT(reply.date_received ?? reply.created_at)}`,\n `Offer: ${offer}`,\n ``,\n `THEIR MESSAGE`,\n `-------------`,\n freshMessage(reply),\n ``,\n `Subject line: ${String(reply.email_subject ?? \"\").replace(/^\\s*((re|fwd?|fw)\\s*:\\s*)+/i, \"\").trim() || \"?\"}`,\n `Sender email: ${fromLine}`,\n `Their timezone: ${tzLine}`,\n ``,\n `CONTEXT`,\n `Clinic: ${intel?.clinicName ?? discovered?.practice_name ?? analysis.clinic_name ?? lead.company ?? \"?\"}`,\n `Website: ${site ? site + siteNote : \"none found\"}`,\n ...(discovered?.evidence ? [`(web search match: ${discovered.evidence}${discovered.confidence ? `, confidence ${discovered.confidence}` : \"\"})`] : []),\n `Location: ${[leadVar(lead, \"city\"), leadVar(lead, \"state\")].filter(Boolean).join(\", \") || \"?\"}`,\n `Our sender: ${input.senderEmail ?? \"?\"} (persona: ${input.senderName ?? \"?\"})`,\n ``,\n `The AI is still handling the email reply as usual - this is your heads-up to call.`,\n ].join(\"\\n\");\n\n return {\n subject: `Interested - call prep: ${leadLabel}`,\n text,\n detail: `phone=${analysis.phone_number ?? (leadMobileDigits ? `lead:${leadMobileDigits}` : sitePhones[0] ? `site:${sitePhones[0]}` : \"none\")} (${phoneKind}); ${dmDetail}; site=${siteOnFile ? \"on file\" : discovered?.website ? \"discovered\" : \"none\"}`,\n };\n}\n",[200],"0077a4d8-3520-4f2d-aa0c-eab8919f54f5","calendly.ts","c443a84e-82f4-11f1-adff-1607ee4eb77e",80,["D",1784420096609],"// Calendly calendar layer - mirrors the ghl.ts calendar interface so\n// scheduling.ts/webhook.ts keep their contracts (slot strings\n// \"YYYY-MM-DDTHH:mm:ss±HH:MM\" in prospect tz; bookWithRecovery outcomes\n// booked|taken|stuck; upcoming-appointment {id, startTime} with location-local\n// startTime). GHL remains the CRM; only calendar ops live here.\n//\n// Env: CALENDLY_TOKEN (PAT), CALENDLY_EVENT_TYPE (the SECRET twin event type\n// URI - ALL agent bookings/availability go through it; the public /pilot link\n// is only for self-bookers, which doubles as agent-vs-self attribution),\n// CALENDLY_ORG_URI (optional, derived from /users/me), CALENDLY_WEBHOOK_SIGNING_KEY.\n\nimport { sqlite } from \"https://esm.town/v/std/sqlite\";\n\nconst CAL_BASE = \"https://api.calendly.com\";\nexport const CALENDLY_TIMEZONE = \"America/Los_Angeles\"; // location tz (mirrors GHL_TIMEZONE)\nexport const CALENDLY_BOOKING_QUESTIONS = true; // twin has required Mobile Phone / Clinic Name\n\n// Error prefix \"calendly METHOD path STATUS: body\" is load-bearing:\n// bookingErrorIsDefinitive() distinguishes HTTP-answered rejections from\n// ambiguous timeout/network aborts (never match timeout text - it differs per\n// runtime). 15s reads / 60s mutations, same recovery ladder as GHL.\nasync function calendly(\n method: string,\n path: string,\n body?: unknown,\n timeoutMs = 15_000,\n): Promise {\n const resp = await fetch(CAL_BASE + path, {\n method,\n headers: {\n \"Authorization\": `Bearer ${Deno.env.get(\"CALENDLY_TOKEN\")}`,\n \"Content-Type\": \"application/json\",\n },\n body: body === undefined ? undefined : JSON.stringify(body),\n signal: AbortSignal.timeout(timeoutMs),\n });\n const text = await resp.text();\n if (!resp.ok) throw new Error(`calendly ${method} ${path} ${resp.status}: ${text.slice(0, 400)}`);\n return text ? JSON.parse(text) : {};\n}\n\nexport function calendarConfigured(): boolean {\n return Boolean(Deno.env.get(\"CALENDLY_TOKEN\") && Deno.env.get(\"CALENDLY_EVENT_TYPE\"));\n}\n\nfunction eventTypeUri(): string {\n return Deno.env.get(\"CALENDLY_EVENT_TYPE\") ?? \"\";\n}\n\nlet cachedOrg: string | null = null;\nexport async function orgUri(): Promise {\n const env = Deno.env.get(\"CALENDLY_ORG_URI\");\n if (env) return env;\n if (cachedOrg) return cachedOrg;\n const me = await calendly(\"GET\", \"/users/me\");\n cachedOrg = me?.resource?.current_organization ?? \"\";\n return cachedOrg!;\n}\n\nexport function uuidOf(uri: string): string {\n return uri.split(\"/\").pop() ?? uri;\n}\n\n// UTC instant -> \"YYYY-MM-DDTHH:mm:ss±HH:MM\" rendered in `timezone` (the slot\n// string layout scheduling.ts slotDate/slotMinutes parse by fixed offsets).\nexport function utcToZonedIso(utc: string, timezone: string): string {\n const d = new Date(utc);\n const wall = d.toLocaleString(\"sv-SE\", { timeZone: timezone }); // \"YYYY-MM-DD HH:mm:ss\"\n const raw = new Intl.DateTimeFormat(\"en-US\", { timeZone: timezone, timeZoneName: \"longOffset\" })\n .formatToParts(d).find((p) => p.type === \"timeZoneName\")?.value ?? \"GMT+00:00\";\n const offset = raw === \"GMT\" ? \"+00:00\" : raw.replace(\"GMT\", \"\");\n return `${wall.slice(0, 10)}T${wall.slice(11, 19)}${offset}`;\n}\n\nfunction isoNowPlus(seconds: number): string {\n return new Date(Date.now() + seconds * 1000).toISOString().replace(/\\.\\d{3}Z$/, \"Z\");\n}\n\nasync function availableTimes(startIso: string, endIso: string, timezone: string): Promise> {\n const et = encodeURIComponent(eventTypeUri());\n const data = await calendly(\n \"GET\",\n `/event_type_available_times?event_type=${et}&start_time=${encodeURIComponent(startIso)}&end_time=${encodeURIComponent(endIso)}`,\n );\n const out: Record = {};\n for (const s of data?.collection ?? []) {\n if (s?.status && s.status !== \"available\") continue;\n const zoned = utcToZonedIso(s.start_time, timezone);\n (out[zoned.slice(0, 10)] ??= []).push(zoned);\n }\n for (const day of Object.keys(out)) out[day].sort();\n return out;\n}\n\n// SAME name/shape as ghl.fetchFreeSlots. API caps ranges at 7 days.\nexport async function fetchFreeSlots(days = 6, timezone = CALENDLY_TIMEZONE): Promise> {\n return await availableTimes(isoNowPlus(120), isoNowPlus(Math.min(days, 6.9) * 86_400), timezone);\n}\n\n// Targeted far-out window: one ≤7-day fetch covering the proposed date (the\n// twin's 60-day scheduling window is what makes far-out dates visible at all).\nexport async function fetchSlotsForDate(proposedDate: string, timezone: string): Promise> {\n const dayStartUtc = new Date(`${proposedDate}T00:00:00Z`).getTime() - 86_400_000; // pad a day for tz offsets\n const start = Math.max(dayStartUtc, Date.now() + 120_000);\n const startIso = new Date(start).toISOString().replace(/\\.\\d{3}Z$/, \"Z\");\n const endIso = new Date(start + 3 * 86_400_000).toISOString().replace(/\\.\\d{3}Z$/, \"Z\");\n return await availableTimes(startIso, endIso, timezone);\n}\n\n// POST /invitees on the twin. startTime accepts the local-offset slot string.\n// The twin's REQUIRED questions must always be answered (verified live:\n// \"Required Questions and Answers cannot be blank\"; phone accepts free text,\n// so the fallback chain can bottom out at \"not provided\").\nexport async function bookAppointment(a: {\n email: string;\n name: string;\n startTime: string; // local-offset slot string or UTC ISO\n timezone: string; // prospect IANA zone - drives their confirmation rendering\n phone?: string;\n clinic?: string;\n}): Promise<{ id: string; inviteeUri: string }> {\n const data = await calendly(\"POST\", \"/invitees\", {\n event_type: eventTypeUri(),\n start_time: new Date(a.startTime).toISOString().replace(/\\.\\d{3}Z$/, \"Z\"),\n location: { kind: \"google_conference\" },\n invitee: { name: a.name, email: a.email, timezone: a.timezone },\n questions_and_answers: [\n { question: \"Mobile Phone\", answer: a.phone || \"not provided\", position: 0 },\n { question: \"Clinic Name\", answer: a.clinic || \"not provided\", position: 1 },\n ],\n }, 60_000);\n const r = data?.resource ?? data;\n const eventUri = r?.event;\n if (!eventUri) throw new Error(\"calendly bookAppointment: no event in response\");\n return { id: uuidOf(eventUri), inviteeUri: r?.uri ?? \"\" };\n}\n\nexport function bookingErrorIsDefinitive(e: unknown): boolean {\n return String(e).startsWith(\"Error: calendly \") || String(e).includes(\"calendly POST\") ||\n String(e).includes(\"calendly GET\");\n}\n\nasync function bookingLanded(emails: (string | undefined)[]): Promise<{ id: string; startTime: string } | null> {\n for (const waitMs of [5_000, 10_000, 15_000]) {\n await new Promise((r) => setTimeout(r, waitMs));\n const appt = await getUpcomingAppointment(emails).catch(() => null);\n if (appt) return appt;\n }\n return null;\n}\n\n// Identical outcome contract to the GHL version (see NOTES.md Cornerstone\n// incident): 2 attempts; ambiguous abort -> landed-check -> retry -> check;\n// a definitive 400 AFTER an ambiguous attempt gets one more landed-check\n// (\"already_filled\" may be our own aborted POST having committed).\nexport async function bookWithRecovery(a: {\n email: string;\n name: string;\n startTime: string;\n timezone: string;\n phone?: string;\n clinic?: string;\n}): Promise<\n | { outcome: \"booked\"; apptId: string; note: string }\n | { outcome: \"taken\"; error: unknown }\n | { outcome: \"stuck\"; error: unknown }\n> {\n let sawAmbiguous = false;\n let lastErr: unknown = null;\n for (let attempt = 1; attempt <= 2; attempt++) {\n try {\n const appt = await bookAppointment(a);\n return { outcome: \"booked\", apptId: appt.id, note: attempt > 1 ? \" (on retry)\" : \"\" };\n } catch (e) {\n lastErr = e;\n console.error(`calendly book attempt ${attempt} failed:`, e);\n if (bookingErrorIsDefinitive(e)) {\n if (!sawAmbiguous) return { outcome: \"taken\", error: e };\n const landed = await bookingLanded([a.email]);\n if (landed) return { outcome: \"booked\", apptId: landed.id, note: \" (earlier aborted call landed)\" };\n return { outcome: \"taken\", error: e };\n }\n sawAmbiguous = true;\n const landed = await bookingLanded([a.email]);\n if (landed) return { outcome: \"booked\", apptId: landed.id, note: \" (landed after aborted call)\" };\n }\n }\n return { outcome: \"stuck\", error: lastErr };\n}\n\nexport async function getScheduledEvent(eventUuid: string): Promise {\n const data = await calendly(\"GET\", `/scheduled_events/${eventUuid}`).catch(() => null);\n return data?.resource ?? null;\n}\n\n// Meet link for a just-booked event. Google pushes the conference a moment\n// after booking (location.status \"pushed\"); tiny retry covers the gap.\nexport async function getMeetLink(eventUuid: string): Promise {\n for (let attempt = 0; attempt < 3; attempt++) {\n try {\n const data = await calendly(\"GET\", `/scheduled_events/${eventUuid}`);\n const loc = data?.resource?.location;\n if (loc?.join_url) return loc.join_url;\n } catch (e) {\n console.error(\"getMeetLink:\", e);\n }\n await new Promise((r) => setTimeout(r, 3000));\n }\n return \"\";\n}\n\n// In-thread add-to-calendar links for booking confirmations (Calendly hosts\n// no linkable .ics - verified 2026-07-18; its .ics is an email attachment).\n// gcal = one-click Google prefill; ics = our own endpoint (test-ghl.ts\n// ?action=ics) serving a generated file for Apple/Outlook.\nexport function buildCalendarLinks(\n eventUuid: string,\n joinUrl: string,\n startUtc: string,\n adminUrl: string,\n durationMin = 30,\n): { gcal: string; ics: string } {\n const fmt = (iso: string) => iso.replace(/[-:]/g, \"\").replace(/\\.\\d+/, \"\");\n const start = new Date(startUtc);\n const end = new Date(start.getTime() + durationMin * 60_000);\n const q = new URLSearchParams({\n action: \"TEMPLATE\",\n text: \"Pilot Campaign Call with Lumina\",\n dates: `${fmt(start.toISOString())}/${fmt(end.toISOString())}`,\n details: joinUrl ? `Google Meet: ${joinUrl}` : \"Details in your confirmation email.\",\n location: joinUrl || \"Google Meet\",\n });\n return {\n gcal: `https://calendar.google.com/calendar/render?${q.toString()}`,\n ics: `${adminUrl}?action=ics&e=${eventUuid}`,\n };\n}\n\n// Cancels keep the record (Calendly has no hard-delete for events).\nexport async function cancelAppointment(eventUuid: string, reason = \"Cancelled via our scheduling assistant\") {\n return await calendly(\"POST\", `/scheduled_events/${eventUuid}/cancellation`, { reason }, 60_000);\n}\n\n// No reschedule-to-time API: BOOK-NEW-FIRST, THEN CANCEL-OLD. The old\n// appointment survives a failed booking; a cancel failure after a successful\n// booking is surfaced in the note (never orphans the prospect).\nexport async function rescheduleAppointment(a: {\n oldEventUuid: string;\n email: string;\n name: string;\n startTime: string;\n timezone: string;\n phone?: string;\n clinic?: string;\n}): Promise<\n | { outcome: \"moved\"; apptId: string; note: string }\n | { outcome: \"taken\"; error: unknown }\n | { outcome: \"stuck\"; error: unknown }\n> {\n const booked = await bookWithRecovery(a);\n if (booked.outcome !== \"booked\") return booked;\n try {\n await cancelAppointment(a.oldEventUuid, \"Rescheduled per your request\");\n return { outcome: \"moved\", apptId: booked.apptId, note: booked.note };\n } catch (e) {\n console.error(\"reschedule: new booking ok but old cancel failed:\", e);\n return {\n outcome: \"moved\",\n apptId: booked.apptId,\n note: `${booked.note} (WARNING: old appointment ${a.oldEventUuid} cancel failed - remove manually: ${String(e).slice(0, 120)})`,\n };\n }\n}\n\n// First upcoming active event for ANY of the given emails (reply-From AND\n// lead email - self-bookings can appear under either). startTime is rendered\n// location-local \"YYYY-MM-DD HH:mm:ss\" to preserve the ghl.ts contract.\nexport async function getUpcomingAppointment(\n emails: (string | undefined)[],\n): Promise<{ id: string; startTime: string; inviteeEmail?: string } | null> {\n const org = encodeURIComponent(await orgUri());\n const seen = new Set();\n for (const raw of emails) {\n const email = raw?.trim().toLowerCase();\n if (!email || seen.has(email)) continue;\n seen.add(email);\n const data = await calendly(\n \"GET\",\n `/scheduled_events?organization=${org}&invitee_email=${encodeURIComponent(email)}&status=active&min_start_time=${\n encodeURIComponent(isoNowPlus(0))\n }&sort=start_time:asc&count=5`,\n );\n const ev = (data?.collection ?? [])[0];\n if (ev) {\n return {\n id: uuidOf(ev.uri),\n startTime: new Date(ev.start_time).toLocaleString(\"sv-SE\", { timeZone: CALENDLY_TIMEZONE }),\n inviteeEmail: email,\n };\n }\n }\n return null;\n}\n\nexport type CalEvent = {\n id: string;\n startTimeUtc: string;\n inviteeEmail: string;\n inviteeName: string;\n eventTypeUri: string; // twin = agent-booked, public = self-booked\n status: string;\n};\n\n// Reconciler sweep source: active events in [startMs, endMs] with one invitee\n// sub-fetch per event (bounded by the sweep's 2-week window).\nexport async function listCalendarEvents(startMs: number, endMs: number): Promise {\n const org = encodeURIComponent(await orgUri());\n const min = encodeURIComponent(new Date(startMs).toISOString().replace(/\\.\\d{3}Z$/, \"Z\"));\n const max = encodeURIComponent(new Date(endMs).toISOString().replace(/\\.\\d{3}Z$/, \"Z\"));\n const data = await calendly(\n \"GET\",\n `/scheduled_events?organization=${org}&status=active&min_start_time=${min}&max_start_time=${max}&count=100`,\n );\n const out: CalEvent[] = [];\n for (const ev of data?.collection ?? []) {\n const uuid = uuidOf(ev.uri);\n let email = \"\", name = \"\";\n try {\n const inv = await calendly(\"GET\", `/scheduled_events/${uuid}/invitees?count=1`);\n const first = (inv?.collection ?? [])[0];\n email = first?.email ?? \"\";\n name = first?.name ?? \"\";\n } catch (e) {\n console.error(\"listCalendarEvents invitee fetch failed:\", e);\n }\n out.push({\n id: uuid,\n startTimeUtc: ev.start_time,\n inviteeEmail: email,\n inviteeName: name,\n eventTypeUri: ev.event_type ?? \"\",\n status: ev.status ?? \"active\",\n });\n }\n return out;\n}\n\n// ---------- webhooks ----------\n\nexport async function listWebhookSubscriptions(): Promise {\n const org = await orgUri();\n const data = await calendly(\n \"GET\",\n `/webhook_subscriptions?organization=${encodeURIComponent(org)}&scope=organization&count=20`,\n );\n return data?.collection ?? [];\n}\n\nexport async function createWebhookSubscription(\n callbackUrl: string,\n signingKey: string,\n): Promise {\n const org = await orgUri();\n const data = await calendly(\"POST\", \"/webhook_subscriptions\", {\n url: callbackUrl,\n events: [\"invitee.created\", \"invitee.canceled\"],\n organization: org,\n scope: \"organization\",\n signing_key: signingKey,\n }, 30_000);\n return data?.resource ?? data;\n}\n\n// The signing key is OUR generated secret (not one of Sam's), so it persists\n// itself in the val's sqlite (crm_state) at subscription time - no show-once\n// manual env step. Env var CALENDLY_WEBHOOK_SIGNING_KEY overrides if ever set.\nconst SIGNING_KEY_STATE = \"calendly:signing_key\";\nlet cachedSigningKey: string | null = null;\n\nexport async function storeSigningKey(key: string): Promise {\n await sqlite.execute(`CREATE TABLE IF NOT EXISTS crm_state (key TEXT PRIMARY KEY, value TEXT)`);\n await sqlite.execute({\n sql: `INSERT INTO crm_state (key, value) VALUES (?, ?) ON CONFLICT(key) DO UPDATE SET value = excluded.value`,\n args: [SIGNING_KEY_STATE, key],\n });\n cachedSigningKey = key;\n}\n\nasync function signingKey(): Promise {\n const env = Deno.env.get(\"CALENDLY_WEBHOOK_SIGNING_KEY\");\n if (env) return env;\n if (cachedSigningKey) return cachedSigningKey;\n try {\n const r = await sqlite.execute({\n sql: `SELECT value FROM crm_state WHERE key = ?`,\n args: [SIGNING_KEY_STATE],\n });\n cachedSigningKey = (r.rows?.[0]?.[0] as string) ?? null;\n } catch {\n cachedSigningKey = null;\n }\n return cachedSigningKey;\n}\n\n// Header: \"t=,v1=\". Stale beyond\n// 5 min is rejected (replay guard).\nexport async function verifyWebhookSignature(rawBody: string, sigHeader: string | null): Promise {\n const key = await signingKey();\n if (!key || !sigHeader) return false;\n const parts = Object.fromEntries(sigHeader.split(\",\").map((p) => p.trim().split(\"=\", 2)));\n const t = parts[\"t\"], v1 = parts[\"v1\"];\n if (!t || !v1) return false;\n if (Math.abs(Date.now() / 1000 - Number(t)) > 300) return false;\n const cryptoKey = await crypto.subtle.importKey(\n \"raw\",\n new TextEncoder().encode(key),\n { name: \"HMAC\", hash: \"SHA-256\" },\n false,\n [\"sign\"],\n );\n const mac = await crypto.subtle.sign(\"HMAC\", cryptoKey, new TextEncoder().encode(`${t}.${rawBody}`));\n const hex = Array.from(new Uint8Array(mac)).map((b) => b.toString(16).padStart(2, \"0\")).join(\"\");\n return hex === v1.toLowerCase();\n}\n",[207],"0ed29caa-8e0a-4bbf-9227-49dda42488f4","analysis.ts","df35b9c8-8306-11f1-bc9e-1607ee4eb77e",87,["D",1784845729767],"// Phase-1 analysis: classification + scheduling/contact extraction. The\n// prompt/schema live here (not webhook.ts) purely for Val.town's 80k-char\n// file cap - webhook.ts imports all three.\nexport const ANALYSIS_PROMPT = `\nYou analyze replies to cold emails for Lumina (Facebook ads for dental clinics).\nDo NOT draft anything - classify the reply and extract scheduling facts. Code\nwill check the real calendar and a second pass drafts the reply.\n\nCATEGORIES (exactly one):\n1. \"contact_preference\" - they EXPLICITLY offer a number/channel to reach them\n on, or ask us to use one (text me, call me, WhatsApp, \"call my office\").\n Wins over all others. A phone number that merely sits in their signature\n block is NOT contact_preference - classify by the message itself (the number\n still gets extracted via phone_number below).\n2. \"referral\" - they hand the thread to someone else. Set new_contact_provided\n true only if the new person's EMAIL ADDRESS is in the reply.\n3. \"meeting_request\" - they want to talk/book/schedule, or show general interest\n without questions (\"sure\", \"ok\", \"sounds interesting\").\n4. \"info_request\" - they ask question(s) or want more info. Set answerable=false\n ONLY for genuine edge cases wholly outside a sales playbook (rare).\n5. \"not_positive\" - everything else: not interested, unsubscribe, hostile,\n \"we already have an agency\", \"how did you get my email\", vague non-committal.\n ALSO: any reply written by an automated/AI assistant (e.g. signatures like\n \"Sent from X's AI Assistant\", obviously templated auto-responses) is ALWAYS\n not_positive regardless of content. If genuinely unsure between not_positive\n and a positive category, pick not_positive.\n\nEXTRACTION (for every positive category):\n- prospect_first_name: how we'd greet them (\"Monica\", \"Dr. Song\") from their\n sign-off/signature, or the reply's From header when it is clearly a person's\n name (never a practice/office name). null if unknown.\n- prospect_full_name: their full name when determinable (\"Jim Sanford\"), with\n \"Dr.\" in front when they are a dentist/doctor (\"Dr. Jim Sanford\") - check the\n signature, sign-off, and the reply's From header. A From-header name counts\n ONLY when it is clearly a person, never a practice/office name. Omit if you\n only have a first name or nothing.\n- clinic_name: practice name from signature/thread/lead data. null if unknown.\n- timezone: IANA zone (e.g. \"America/New_York\") for the PROSPECT's location.\n Evidence priority: explicit statement in their reply > signature address >\n signature phone area code > lead data (address/phone custom variables) >\n practice name + campaign geography. Texas/Florida etc. split zones - resolve\n by city/area code, not state alone. null ONLY if there is no usable signal.\n timezone_evidence: one short sentence on what you used.\n- Proposed meeting time, ALWAYS in the PROSPECT's local timezone (they speak in\n their own zone unless they clearly reference ours; convert if they do):\n - proposed_date \"YYYY-MM-DD\" + proposed_time \"HH:MM\" (24h) when they name a\n specific time. Bare numbers like \"at 10\" = 10:00 unless context says pm.\n Resolve relative dates (\"tomorrow\", \"Tuesday\") with the current-time info\n provided in the input. Report past/too-soon times as-is - code handles it.\n - proposed_date + proposed_window (morning/afternoon/evening) when they name a\n day and rough window but no clock time (\"Tuesday afternoon\").\n - earliest_date \"YYYY-MM-DD\" when they state a broader availability window\n (\"some time next week\" -> next Monday; \"after the 20th\" -> the 21st; \"not\n this week\" -> next Monday). Every time we later propose must fall on or\n after it. Omit when they gave no such constraint.\n - all null when they didn't suggest any timing.\n- claims_booked: true if they say they already booked/scheduled on our calendar.\n- says_time_works: true if they affirm a specific time works (\"Tuesday at 2\n works\") - even casually.\n- wants_reschedule: true if they ask to MOVE an existing appointment (\"need to\n reschedule\", \"can't make Tuesday\", \"push it to next week\"). \"Cancel it and\n send me the link to rebook\" is a RESCHEDULE, not a cancel - the intent is a\n new time. If they name the new time, also fill proposed_date/time/window.\n- wants_cancel: true only if they want the appointment cancelled with NO intent\n to rebook (\"please cancel\", \"we won't be moving forward with the call\").\n- self_rescheduled: true if they say they ALREADY moved/rebooked it themselves\n on the calendar (done deal, not a request).\n- contact_details: contact_preference/referral only - the number / channel /\n person + email, plus any timing, verbatim as possible.\n- phone_number: the PROSPECT's own phone number when one appears ANYWHERE in\n their reply - explicitly written OR just in their signature block - verbatim.\n NOT numbers from quoted older messages in the thread, and never our own\n (925-730-9373). Omit if none.\n- phone_number_type (only when phone_number is present): \"personal\" when it\n reads as a direct/cell line (under their own name in a signature, labeled\n cell/mobile/direct, or \"call/text me at...\"), \"office\" when it reads as the\n practice's main line (practice-branded footer, office/front-desk label),\n else \"unknown\". phone_number_evidence: one short sentence on the signal.\n- replier_is_decisionmaker: whether the reply's AUTHOR appears to be a\n decisionmaker at the practice - a dentist/doctor, owner, or office/practice/\n operations manager. false when they read as front desk, receptionist,\n assistant, or other staff. With no explicit role signal, judge from context:\n a personal own-name mailbox and ownership language (\"my practice\", \"our\n numbers\") lean true; a shared inbox (info@, office@) with no personal\n sign-off leans false. Omit only when there is truly nothing to judge.\n- asked_for_call: true ONLY when they explicitly ask to meet/schedule/call\n (\"Ok let set up a meeting\", \"can we schedule a call to discuss\") - judged by\n meaning. Mere interest, questions, or a bare \"yes\" to our question is false.\n- invite_email_generic: true when the address their reply came FROM reads as a\n shared clinic inbox (info@, office@, frontdesk@, contact@, smile@...) rather\n than a person's own mailbox. Judge by meaning; when unsure, false.\n- original_offer: the concrete offer from the FIRST EMAIL WE SENT, short and\n near-verbatim (e.g. \"3-5 clear aligner cases every month, free test\n campaign\"). Omit if that email isn't provided.\n- offer_guarantee: how the FIRST EMAIL WE SENT backs its promise. Judge by\n MEANING, not exact words - the promise may sit in a P.S. or anywhere in the\n body, in any phrasing:\n - \"reimburse\" when it promises to reimburse / cover / pay back the ad spend\n if the promised results don't hit.\n - \"until_we_do\" when it promises to keep the campaign running (free / on our\n dime / at no charge) until the promised results are reached.\n - \"none\" when that email makes no such promise.\n Omit if the first email isn't provided.\n\nA reschedule/cancel request is still a positive-category reply (usually\nmeeting_request) unless the message also shows they're no longer interested.\n\nReturn via submit_analysis.`;\n\nexport const ANALYSIS_TOOL = {\n name: \"submit_analysis\",\n description: \"Submit the classification and extracted scheduling facts.\",\n input_schema: {\n type: \"object\",\n properties: {\n category: {\n type: \"string\",\n enum: [\"contact_preference\", \"referral\", \"meeting_request\", \"info_request\", \"not_positive\"],\n },\n reasoning: { type: \"string\", description: \"One short sentence.\" },\n contact_details: { type: \"string\" },\n new_contact_provided: { type: \"boolean\" },\n answerable: { type: \"boolean\" },\n prospect_first_name: { type: \"string\" },\n clinic_name: { type: \"string\" },\n timezone: { type: \"string\", description: \"IANA zone or omit if unknown\" },\n timezone_evidence: { type: \"string\" },\n proposed_date: { type: \"string\", description: \"YYYY-MM-DD prospect-local\" },\n proposed_time: { type: \"string\", description: \"HH:MM 24h prospect-local\" },\n proposed_window: { type: \"string\", enum: [\"morning\", \"afternoon\", \"evening\"] },\n earliest_date: { type: \"string\", description: \"YYYY-MM-DD floor for proposals when they state a window like 'next week'\" },\n claims_booked: { type: \"boolean\" },\n says_time_works: { type: \"boolean\" },\n wants_reschedule: { type: \"boolean\" },\n wants_cancel: { type: \"boolean\" },\n self_rescheduled: { type: \"boolean\" },\n phone_number: { type: \"string\", description: \"The prospect's phone number if present anywhere in their reply (body or signature), verbatim\" },\n phone_number_type: { type: \"string\", enum: [\"personal\", \"office\", \"unknown\"] },\n phone_number_evidence: { type: \"string\", description: \"One short sentence\" },\n prospect_full_name: { type: \"string\", description: \"Full name, 'Dr.'-prefixed when a doctor\" },\n original_offer: { type: \"string\", description: \"The concrete offer from the first email we sent, short\" },\n offer_guarantee: { type: \"string\", enum: [\"none\", \"until_we_do\", \"reimburse\"], description: \"How the first email we sent backs its promise (any wording, any position): reimburse ad spend, keep running until the results hit, or no promise\" },\n replier_is_decisionmaker: { type: \"boolean\", description: \"true when the reply's author reads as a decisionmaker (dentist/doctor, owner, office/practice/ops manager); false when front desk/receptionist/assistant/other staff - judged by meaning\" },\n asked_for_call: { type: \"boolean\", description: \"true ONLY when they explicitly ask to meet/schedule/call ('let's set up a meeting', 'can we schedule a call?') - by meaning, not exact words. false for mere interest, questions, or a bare yes\" },\n invite_email_generic: { type: \"boolean\", description: \"true when the address their reply came FROM reads as a shared clinic inbox (info@, office@, frontdesk@, contact@...) rather than a person's own mailbox - judge by meaning; when unsure, false\" },\n },\n required: [\"category\", \"reasoning\"],\n },\n};\n\nexport type Analysis = {\n category: string;\n reasoning: string;\n contact_details?: string;\n new_contact_provided?: boolean;\n answerable?: boolean;\n prospect_first_name?: string;\n clinic_name?: string;\n timezone?: string;\n timezone_evidence?: string;\n proposed_date?: string;\n proposed_time?: string;\n proposed_window?: string;\n earliest_date?: string;\n claims_booked?: boolean;\n says_time_works?: boolean;\n wants_reschedule?: boolean;\n wants_cancel?: boolean;\n self_rescheduled?: boolean;\n phone_number?: string;\n phone_number_type?: \"personal\" | \"office\" | \"unknown\";\n phone_number_evidence?: string;\n prospect_full_name?: string;\n original_offer?: string;\n offer_guarantee?: \"none\" | \"until_we_do\" | \"reimburse\";\n asked_for_call?: boolean;\n invite_email_generic?: boolean;\n replier_is_decisionmaker?: boolean;\n};\n",[214],"pulls","branches","forks","envVars",{"_227":229,"_230":60},"connections",{"_227":60,"_228":60},"count","unfilledCount",11,"emptyCount",{"_43":213,"_32":214,"_30":215,"_138":-5,"_41":42,"_40":216,"_140":25,"_38":273,"_13":173,"_33":34,"_35":218,"_24":25,"_145":-5,"_243":274,"_27":275},{"_43":206,"_32":207,"_30":208,"_138":-5,"_41":42,"_40":209,"_140":119,"_38":270,"_13":173,"_33":34,"_35":211,"_24":25,"_145":-5,"_243":271,"_27":272},{"_43":199,"_32":200,"_30":201,"_138":-5,"_41":42,"_40":202,"_140":119,"_38":267,"_13":173,"_33":34,"_35":204,"_24":25,"_145":-5,"_243":268,"_27":269},{"_43":191,"_32":192,"_30":193,"_138":-5,"_41":42,"_40":194,"_140":25,"_38":264,"_13":196,"_33":34,"_35":197,"_24":25,"_145":-5,"_243":265,"_27":266},{"_43":183,"_32":184,"_30":185,"_138":-5,"_41":42,"_40":186,"_140":119,"_38":261,"_13":188,"_33":34,"_35":189,"_24":25,"_145":-5,"_243":262,"_27":263},{"_43":175,"_32":176,"_30":177,"_138":-5,"_41":42,"_40":178,"_140":119,"_38":258,"_13":180,"_33":34,"_35":181,"_24":25,"_145":-5,"_243":259,"_27":260},{"_43":171,"_32":28,"_30":31,"_138":-5,"_41":42,"_40":14,"_140":119,"_38":255,"_13":173,"_33":34,"_35":37,"_24":25,"_145":-5,"_243":256,"_27":257},{"_43":163,"_32":164,"_30":165,"_138":-5,"_41":42,"_40":166,"_140":119,"_38":252,"_13":60,"_33":168,"_35":169,"_24":25,"_145":-5,"_243":253,"_27":254},{"_43":155,"_32":156,"_30":157,"_138":-5,"_41":42,"_40":158,"_140":119,"_38":249,"_13":160,"_33":34,"_35":161,"_24":25,"_145":-5,"_243":250,"_27":251},{"_43":147,"_32":148,"_30":149,"_138":-5,"_41":42,"_40":150,"_140":119,"_38":246,"_13":152,"_33":143,"_35":153,"_24":25,"_145":-5,"_243":247,"_27":248},{"_43":135,"_32":136,"_30":137,"_138":-5,"_41":42,"_40":139,"_140":25,"_38":242,"_13":142,"_33":143,"_35":144,"_24":25,"_145":-5,"_243":244,"_27":245},["D",1784845729511],"children",[],[136],["D",1784420098092],[],[148],["D",1784412835045],[],[156],["D",1784168316285],[],[164],["D",1786127124429],[],[28],["D",1786127123928],[],[176],["D",1784412833921],[],[184],["D",1784944420053],[],[192],["D",1784944420416],[],[200],["D",1784420096609],[],[207],["D",1784845729767],[],[214],["D",1783490909384],{"_94":93,"_101":96},{"_94":93,"_95":96,"_97":96,"_18":54,"_98":99,"_100":25},{"_13":83,"_32":84,"_43":42,"_85":44,"_86":-5,"_87":-5,"_88":-5,"_89":-5,"_90":-5,"_91":45,"_92":93},"ctx",{"_76":-5,"_121":313,"_314":315,"_316":317,"_318":25,"_319":25,"_320":25,"_321":119,"_322":-5,"_323":-7,"_324":-7,"_325":60,"_326":25,"_327":25,"_328":-5,"_329":25,"_330":25,"_331":25,"_332":-5},"ENV",{"_284":285,"_286":287,"_288":84,"_289":290,"_291":292,"_293":294,"_295":296,"_297":298,"_299":300,"_301":302,"_303":304,"_305":306,"_307":308,"_309":310,"_311":312},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","e87eb04e707cf64e7e532c8e8ec395f209e17f15","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Diff · v93