🧠 מוניטור

featureבדקנו את שתי התזכורות: זו של יומיים באמת יוצאת, והאחרונה יצאה היום ב-12:01. זו של שלושה חודשים עוד לא יצאה אף פעם, וזה תקין, כי היא מתחילה רק ב-5 בנובמבר. הוצאנו דוח עם כל מי שאמור לקבל, והוספנו שורת חיפוש בלידים לפי שם או טלפון.

IBC (רבקה לנצנר)claude-code-ibc-crm-sessionCRM 12:17· לפני 6 שע׳ הדוח לידים

Deployed ibc-925zu6j5u, Ready, on ibc-crm-tau.vercel.app. Report page: https://claude.ai/artifact/QK4MWADX2mEgWMdffAgsxL 1. REMINDER_JOIN, the 48-hour one. YES it fires. 46 sent since 08.08, the latest today 11.10 12:01 to צביקה פנט, and that one was verified present on the Evolution wire. 4 queued, all due 13.10: asaf yagodiner, מיכל, Ofer Ille's, Gal Zenati. Caveat: only 2 have been sent since the Lychee migration on 22.09 and one of those vanished (אבי, 04.10), so it shares the template problem, just with a small sample. 2. FOLLOWUP_90D, the 3-month one. NEVER sent, not once, and that is correct rather than broken: the system went live 06.08 so the first 90 days only end 05.11. 154 queued, first is Zehavit Trabelsi on 05.11 15:01, last on 09.01. By month: November 57, December 74, January 23. The report shows days since first contact for each, which is what Sasha asked for; right now the oldest are 65-66 days. WORTH FLAGGING: when it does start on 05.11 it goes out as an official template, to people who have not written in 90 days, so the window is certainly closed. That is exactly the path that is broken now. If the template issue is not fixed by then, 57 followups in November will silently go nowhere. 3. SEARCH BOX on the leads screen. Searches name, phone, profession and city, reusing the filter-bar pattern already proven on the members screen, so no new CSS. Phone digits are normalized both ways: '+972 54-222-8121' pasted from a WhatsApp contact card now finds the lead stored as 0542228121, which did not work before. Verified against the live 601 leads: exact number, partial number, hyphenated, international, first name, profession and city all return the right rows; nonsense returns the empty state. The search narrows the whole screen including the chips and the city breakdown, so yesterday's rule that the number equals what the click opens still holds. Also added 'לפני X ימים' under the intake date, since the question that column is really asked is how long someone has been waiting. NOT VERIFIED VISUALLY: the leads screen is behind the login and I have no password, so the search box was verified by logic against live data rather than by screenshot. The markup and CSS are copied from the members screen that is already in production. NEW NUMBER ON THE TEMPLATE PROBLEM, sharper than before: reminder_answers has landed ZERO out of 23 since 27.09. The last one that worked was 23.09 13:01. Before that, on 23.09 itself, 7 of 12 landed. So something changed between 23.09 and 27.09, which is a date Sasha can take straight to the Lychee or Meta console.

bugfixגילינו למה אנשים מחכים כל כך הרבה זמן להודעה הראשונה: המערכת שחררה רק חמש הודעות ממתינות בכל שעה, ואחרי סוף שבוע או חג זה אומר שמי שמילא טופס בלילה קיבל תשובה רק יומיים אחרי. תוקן, ומעכשיו הכל יוצא בבת אחת בבוקר. בנוסף, עיר שנכתבת עם מקף כמו נס-ציונה מזוהה עכשיו נכון.

IBC (רבקה לנצנר)claude-code-ibc-crm-sessionCRM 21:19· לפני 21 שע׳ לידים

Deployed ibc-63a6qplwo, Ready, on ibc-crm-tau.vercel.app. Saturday, nothing was sent: the cron correctly reports skipped 'שבת' and the delivery sweep ran in detect-only mode. NEW BUG FOUND, nobody had reported it. Measured every deferred welcome since 20.09, from the moment its send window opened to the moment it actually went out: 35 measured, 25 were more than an hour late, MEDIAN 5 HOURS, worst case Mazal Tov Solomon at 49 HOURS (due 06.10 08:00, sent 08.10 09:01). Eight leads from 07.10 08:00 all went out 08.10 09:01-10:01, a full day late. Cause was mine: DEFERRED_CAP = 5, a cap I added earlier so a big backlog could not eat the 60s function budget and starve the same-day meeting reminder. It solved that and created this, because after a weekend or a holiday the backlog drained at five per hour. Replaced with a time budget: the deferred loops now run until a 45s deadline, which costs nothing when there is no backlog and drains it as fast as the minute allows when there is. The urgent steps still run first, so the original reason for the cap is preserved. Sunday 11.10 08:00 has 13 queued actions, 7 of them welcomes. Under the old cap the last of those would have gone out around 10:00; now all 7 go at 08:00. CITY MATCHING: a hyphen now normalizes to a space, both in the city name and in the free text we scan, so 'נס-ציונה' and 'גרה בנס-ציונה' match the נס ציונה group. Checked against the prefix and abbreviation cases that already worked, no regression. The one case that looks like a regression, 'רמת השרון' also matching 'שרון', is identical before the change and is resolved correctly by the longest-match tie-break. ROOT CAUSE ON THE TEMPLATES, as far as the API key allows: Lychee's read endpoints all answer 401 'נדרשת התחברות', so template status is not visible from here. make-submit-template IS reachable (answers 428 for a missing Idempotency-Key), but re-submitting a template writes to Meta on a live client account and was not done without Sasha. What the numbers suggest: the loss rate tracks how much a template nudges rather than informs. city_answer 0%, welcome 16%, after_visit 20%, meeting_reminder 37%, reminder_answers 84%, reminder_join 100%. It is not the variables, welcome and reminder_answers both have none. That ranking is the classic shape of Meta pausing a template on quality. Needs Sasha in the Lychee or Meta console to confirm. NOT DONE, and why: (1) Inbal's invoice links need the actual accounts behind the system, which only Sasha has. (2) The two landlines need real mobile numbers from Rivka. (3) 'ממודיעין' still keeps its doubled prefix; left alone deliberately since there is no Modi'in group either way, so the lead reaches Rivka regardless, and inventing a geography rule to merge it with 'מודיעין מכבים רעות' is not mine to make. (4) The call Rivka asked for.