Brain Feed — Claude Code session log

What this is: the cross-session memory. Each Claude Code session appends rows here at session-end (built / decided / open). This git-versioned file works immediately and is what Claude Code reads at session-open. The Sheet copy (Brain_Feed tab in the command-inbox sheet) is the same log the Telegram bot reads — it activates once the engine redeploy lands. Until then, this file is the live record.

Append format per session: built: / decided: / open: lines, newest session on top.

2026-05-31 (Sun) — Session 39 cont · mobile-app handoff: civic comment + Referral Database

• received: Handoff from Claude mobile app — Sam worked voice-first on the go. Two artifacts produced.
• filed: Referral Database CLEAN — restructured from legacy single-sheet Referral Database.xlsx (Drive file 1XR-VIX_eWZq5jyXcBsnAkMTx9H-LyFeS) into a 4-tab workbook (README + Directory + By Category + Top Connectors). 273 vendors · 121 unique categories · 7 sections (Medical · Household · Jewish/Lifestyle · Services · Auto · Other Repair · Hobbies). 7-column schema: Section | Category | Vendor/Name | Notes | Contact Info | Alternate Contact | Recommended By. Zero-error formulas, autofilter live on all tabs. Top community connectors by recommendation volume: Yosef Birnbaum (15), Liad Maccabi (11), Melinda Hilsenrath (10), Tuli Weiss (8), Dovi Frankel/Harold Kleiman/Joe Golavanov (7 each). Moved from workspace root to canonical path: reference/Referral_Database_CLEAN.xlsx (renamed underscore-form per handoff spec). Sam appears once in the data — Handy Ed (917-657-4486) recommended by "Z Treitel."
• submitted (Sam-side): Hempstead Board of Appeals Case 75-26 — 522 Hungry Harbor LLC, hearing Wed Jun 24 2026. Sam submitted OPPOSED comment via hempsteadny.gov form. Comment lead: traffic safety + child bus stop at Bunker/Plainfield/Hungry Harbor blind curve. Final variant ("tighter sharper edge"): position statement → factual case → comparable-precedent → ask with reasonable-out clause. No alternate-site framing, no whataboutism.
• locked: Civic comment doctrine (6 steps) — Sam asked for the pattern to be reusable. Saved as memory feedback_civic_comment_doctrine: (1) clean position statement first · (2) factual case in short declarative paragraphs · (3) comparable-precedent framing · (4) explicit ask + reasonable-out condition · (5) drop alternate-site/whataboutism · (6) lead with safety/concrete facts, not motive attribution. Candidate next applications: HOA correspondence, board-of-zoning letters, Eden Gardens formal comms, public-facing community statements.
• carded: 3 cards in command-inbox queue: (a) mirror Referral_Database_CLEAN to Google Drive as live Google Sheet for phone use; (b) decide duplicate-vendor rule (e.g. Y & B Handyman in 3 categories — keep duplicates with primary flag OR normalize); (c) Hempstead Jun 24 hearing follow-up (attend / written supplement / rely on submitted).
• open: Drive mirror (Sam-side, 2 min in Drive UI). Duplicate-vendor rule (Sam decides). Hempstead Jun 24 calendar add + attendance decision.

2026-05-29 (Fri PM, late) — Session 39 cont · -Source prefix routing on tg.ps1

• built: scripts/tg.ps1 -Source parameter — every Telegram push self-labels with a distinct prefix so Sam can tell at a glance who's talking. Mapping: Code → 🤖 Code, MIS → 📈 MIS, System → ⚙ System, Pulse → Pulse, Mom → 💬 Mom, Family → 👨‍👩 Family, Bot → Start Here (legacy default, backwards-compat), Custom → (no prefix). Emojis assembled from codepoints at runtime so PS 5.1's Windows-1252 file-read can't mangle them. Tested live — push arrived with 🛠 Code — Source prefix LIVE title. Sam's diagnostic question solved: one Telegram bot account, many voices, now all labeled.
• decided: Going forward, Claude Code ALWAYS passes -Source Code. Documented in feedback_claude_to_sam_channels memory + CLAUDE.md key rules + script header. Any session that uses tg.ps1 without -Source falls back to "Start Here" (legacy bot voice) — that's the right default for the bot itself but wrong for Claude-Code pushes.
• open: Bot conversational reply still uses "Start Here —" header internally (not via tg.ps1). Aligned with the new -Source Bot legacy default. No code change needed in start-here.gs — the convention works because Bot replies and Code pushes now sit in different namespaces.

2026-05-29 (Fri) — Session 39 cont (full day → erev Shabbos, voice + svc-deploy + queue UX)

• built: Voice IN + OUT, FREE on Cloudflare Workers AI. /voice/transcribe (Whisper) + /voice/speak (MeloTTS) on ops-api. Bot handles Telegram voice messages (auto-downloads OGG → Whisper → routes through telegramBrainReply_ same as typed). SPEAK command → MeloTTS → Telegram sendAudio (MP3, plays inline). Tested: speak endpoint returns 600KB audio for short prompts. Cost: $0. Path A locked in (no OpenAI/ElevenLabs subscription needed unless quality bothers Sam).
• built: Two voice-bug fixes. (1) Routing: I labeled voice body with source: 'telegram-voice' which fell off both the brain branch + outbound Telegram sendMessage branch (both gated on source === 'telegram'). Sam's first voice test went silent. Fixed by routing voice as source: 'telegram' + carrying voice metadata on side flags. (2) TypeError: body.message.voice.duration was read AFTER body got reassigned to flat shape (no .message on new body). Sam got "Cannot read properties of undefined (reading 'voice')" — Whisper had actually transcribed fine, crash was on the LAST log line before return. Fixed by capturing duration into local var before reassign.
• built: Auto-redeploy via clasp CLI — proved clasp deploy --deploymentId <live> bumps the live Web App version WITHOUT clobbering the URL (because command-inbox/appsscript.json has the webapp block). Verified @68 → @69 LIVE, QUEUE still serves. Sam never clicks "Manage Deployments → New version" again for this project — I run the deploy after every push.
• built: tools/svc-deploy/ framework — service-account based push+deploy using Apps Script API directly (no clasp, no user OAuth). Once Sam does the 6-step Cloud Console setup (tools/svc-deploy/SETUP.md), ! clasp login is gone forever for command-inbox + LevSMS + MIS v2. node tools/svc-deploy/deploy.mjs <project> does the whole push+version+deploy in one command. googleapis 144.x + auth lib installed locally.
• built: Queue UX overhaul. (1) Global #NNN numbering — migrateAddGlobalNumbers_ + nextGlobalNum_ + display lead with #NNN (full Action_ID in parens). Sam's "two 10s" complaint solved. (2) DEFERRED hidden from default view + auto-resurfaces on due-date arrival via resurfaceDueDeferred_ wired into pulse + morningBriefing_. QUEUE DEFERRED / QUEUE ALL reveals them on demand. Default queue now shows only active work.
• built: home.html Open Loops + Today tiles = LIVE + tap-to-act. New QUEUE_JSON bot command returns structured cards. ops-api /inbox whitelist extended (DONE/PROOF/DEFER/STATUS/ACTION/NEED_INFO). Open Loops tile fetches QUEUE_JSON, renders cards grouped P0..P3, tap a row → DONE/PROOF/DEFER/NEED_INFO buttons fire actions on the sheet. "Add a card" box at bottom → ACTION command. Today tile = top 4 by priority auto-rendered. The dashboard IS the command center now, not a launcher.
• built: Lane-aware bot system prompt. TOOLS_GUIDANCE_ hardened: (1) ANY named person/place/deal/ticker → search_queue + search_memory FIRST before replying. (2) NEVER say "I don't have that" without running search. (3) Lane keywords route to right context (STR/MIS/Eden/HOA/Mom/family/LevSMS/Mildred). (4) Pre-loaded context is just top cards — tools see everything. Fixes the Itamar bug (bot had card #133 with JFK destination but asked "which airport?" anyway).
• built: Start-here.html forced-light + "Act now" tile on home. Killed dark @media block in start-here.html + canonical hero gradient. Home now has a prominent accent-colored "Act now" tile linking to start-here.html (the textarea + button PWA where Sam actually acts on cards).
• built: Eden $20K synopsis draft in business Gmail (id r120339851436382935) — per Sam's WhatsApp offer to Asher Thu 5:24 PM. Waiting on Sam's send.
• built: Voice/svc-deploy notification pings via scripts/tg.ps1 (renamed from notify.ps1 2026-05-29) at every milestone so Sam sees progress on his phone, not just in chat.
• decided: Lane categories Sam refined. Eden/Mom/Mendy/Itamar/future paying = client lane. Mildred = assistant lane. STR/MIS/BOS/Family/HOA/LevSMS = own ops. Bot system prompt should frame answers in the matching category.
• decided: Voice cost = $0 (Workers AI free tier). Sam noted claude.ai going from $100 → $20. No new recurring cost for voice. If MeloTTS quality bothers him, swap to OpenAI TTS-1 (~$2-5/mo) by changing the model id in /voice/speak — Path B noted but not built.
• decided: Jorge Alvarado = OLD employee, NOT current vendor. Earlier inference from contacts (Pro/Vendor tag) was wrong. Will not surface him as a STR handyman candidate.
• decided: Contacts triage workflow = 5 contacts at a time, surface name + tag + one-line who + outreach angle, Sam picks go/skip/hold/kill/ask-Mildred per row. Not started; Sam said contacts go investors (or any segment) when ready.
• taught: 5-beat message structure (from the Itamar WhatsApp exchange — Sam asked to be taught):
  1. Confirm the answer first (not the apology)
  2. ONE apology max (or turn it positive: "thanks for waiting")
  3. The terms/info they need, no breakdown that scares
  4. Any alternative as a gift (frame for them, not as your hedge)
  5. Close on what they should do, not another sorry
  Saved as memory: feedback_message_structure_5_beats.
• open: Sam to do 6-step Cloud Console setup for svc-deploy (~10 min) to permanently kill ! clasp login for the three main projects.
• open: 5 TODAY-before-Shabbos items from email sweep (Sam to do): Pay Alex (Mildred URGENT), Pay 20four7VA $108.76, Smart device models → Mildred (#092 blocked), Tello number decision (permanently dead), Itamar WhatsApp reply (Sam already sent his version).
• open: Parallel sessions today shipped lots: Scope ticker tool (was Peek; now has voice-input in-browser), nightly cron handler on ops-api for MIS snapshots (kills Apps Script reauth pain for that trigger), icon-studio + Chanie A-G mockups, LevSMS Phase 2 (PDF/image OCR + preview/approve), architecture/cockpit-offer/next.html pages. Not me. Money-in-the-loop + design-system + LevSMS lanes all advancing.
• open: Weekend handoff — Sam said "not gonna touch you for the weekend." Park. Motzei Shabbos pickup = test voice end-to-end (mic + SPEAK), do svc-deploy setup, contacts triage, STR tile build.

2026-05-28 (Thu) — Session 39 cont (afternoon, mom-relay polish + deep email triage)

• built: Personal-relay phase-2/3 polish — tell my mom/mother + aliases (mother/mommy/mama/ma/ima→mom) resolve correctly; <name> first-word-after-"tell" no longer hijacks; brain prompt now states it CAN relay (no more wrong "I can't text on your behalf"). Double-space fix in LevSMS reply hint. PERSONAL_CONTACTS JSON in LevSMS = whitelist + display name in one map; RELAY_CONTACTS JSON in command-inbox = name→number for "tell X". Visual badges live: 💬 PERSONAL — <Name> (incoming) + 📤 PERSONAL — sent to <Name> (outgoing). Per-person emoji = put it in the contact name string. Sam tested end-to-end with mom's real number — works.
• built: emoji-?? fix in sendTelegram_ — root cause was parse_mode=Markdown silently rejecting any unbalanced * _ [ ] ( ) and falling back to Twilio SMS where emojis render as ?. Now plain text by default; titles drop the *bold* wrapper. Pushed; needs redeploy.
• built: Deep email triage on BOTH inboxes (2 background agents). Personal: 60-120 day dropped-check CLEAN, 09-WAITING empty, 09-INVOICES empty; 5 real stuck items in 02-ACTION/SAM carded (Namecheap domain verify, Kosher Qin support, Michael OT eval, Pinchus Chein installment, BM Villas DocuSign). GEICO DriveEasy = 50+ identical spam in 2 months (already carded). Business: Eden $20K still #1 (65 days, on card 009); Followup/backup.pst/IBKR-Alerts/Ads-folder = cleanup pile; "Trietel" typo label; Drafts all intentional. No dropped relationships beyond Eden.
• built: Personal inbox filter cleanup — created 3 labels (Noise/Fidelity, Noise/News, Noise/Marketing); 6 filters created then 4 deleted (DICK's/LendingClub/Boost/Tripadvisor — Sam preferred unsubscribe over silent-filter). 2 filters kept (Fidelity, Bloomberg). Then Sam said flip Fidelity to stay-in-inbox + labeled (his "don't lose it" concern) — done this session.
• built: 9 + 5 + 5 = 19 cards added today from the triages — Darchei registration LATE (P0), Tello plan deactivated (P0), Chase 1260 overdraft, Experian declined, GEICO DriveEasy, National Grid past-due, Duke past-due, GitHub Claude perms, Netlify credits at 75%, Namecheap domain verify, Kosher Qin support, Michael OT eval, Pinchus 2nd installment, BM Villas DocuSign, Airbnb/Hospitable integration, HOA attorney waiver decision, Connections Map (consolidate existing seeds — WORKSPACE_AUDIT/CONTEXT/CHANNELS/CAPABILITY_SCAN already exist), business inbox 01-09 mirror, personal SMS relay temp-hot-add (deferred).
• decided: Sam's personal label hierarchy 01-URGENT/02-ACTION/03-MONEY/04-FAMILY/05-PERSONAL/06-SECURITY/07-REFERENCE/08-BUSINESS/09-WAITING/INVOICES is excellent — mirror it on business. Empty 02-ACTION/MILDRED + 09-WAITING + 09-INVOICES = the system working (clear = current), not broken — don't consolidate.
• decided: Bulk-delete pile (~100 personal stale transactionals + 4 dead religious labels Rabbi Elefant/501c3/Yarchei Kallah/Rav Tzvi) surfaced for greenlight; 4 labels deleted this session, ~100 deletes pending Sam's explicit confirm.
• open: Sam needs to (1) Redeploy command-inbox to land the emoji-?? fix + the parseRelayReply aliases live; (2) Login to Namecheap (30 sec) to confirm hookstreetcapital.com is in good standing (1yr-old verification warning); (3) Run the business cleanup recipes manually (claude.ai connector has no delete/filter tools); (4) Click unsubscribe in the 4 marketing senders now landing in inbox; (5) Resolve the GEICO DriveEasy spam at the GEICO account.
• open: Parallel sessions shipped Plaid scaffolding (ops-api Worker + /link.html on portal), MIS v2 stages 6-9 (95% gap closed), home portal Money→Cash tile live from Plaid Worker. Cross-session work — not done by me — but it's the money-in-the-loop lane materializing.

2026-05-27 (Wed) — Session 39 cont

• built: Bot retrieval "eyes" — search_queue / get_card / search_grocery / search_memory + a tool-use loop in callClaude_ (max 4 rounds). Pushed; needs redeploy. Confirmed working in Sam's live test (found the Netlify link after he redeployed once).
• built: Pulse runaway fix — gates now check force !== true (a time-trigger passes a truthy event object; old !force bypassed every gate → 21 overnight pings). Trigger removed live (PULSE STOP) then re-installed fixed.
• built: Rich Context (setupRichContext, 24 facts incl. grocery baseline + kids' schools Darchei/BYAM) + remember write-back (Profile tab, fed into every message). Sam tested — "hot coffee" stuck.
• built: Dedup / no-auto-DONE / day-aware due dates prompt hardening. Pushed; needs redeploy.
• built: Brain_Feed — FEED/BRAINFEED command + tab + brainFeedBlock_ into context. Pushed; needs redeploy for the bot/command path (this file is the interim).
• decided: Family-OS install = isolated clone (not multi-tenant / not SaaS), PWA-first "login here"; run Chanie first as the prototype = the product seed. (memory project_family_os_install)
• decided: Closed Shellpoint (paid 5/15) + ~14 dup/stale/TEST cards across the day; HOA 016 + Eden 009 parked WAITING_ON_OTHER (on Asher) to quiet the pulse.
• decided: Added leads — Mr. Weiser proposal (construction/medical bidding + permitting ops, TX/FL/NY/NJ) + HS Services monetization push.
• open: ONE redeploy pending (Sam holding) — activates eyes + dedup + Brain_Feed(bot) + numbering fix (when built) for the live bot. The fixes exist; the redeploy is the switch.
• built: Card-numbering collision fix (safe) — the bot now reasons with the FULL Action_ID (A-20260527-013), not the 3-digit suffix that repeats across days. Kills the "two 013s" confusion. Pushed; needs redeploy. The prettier global "#47" renumber is a careful tested migration, deliberately deferred (not rushed into the live queue — adds a column the PWA reads).
• built: Mom / C&M Collection tariff-refund project — new docs/MOM_CM_TARIFF_REFUNDS.md. All-Ways Forwarding filing IEEPA/CAPE refund claims; Rivka confirmed 5/26; invoice #B35534_01 received 5/27, needs payment to begin. Carded: pay invoice + verify ACE Portal. (Codex session running on mom's computer; bilingual He/En.)
• built: Drafted Sam's reply to Mildred's 5/27 briefing (Gmail draft, threaded) — yes to Alex $375, smart-device models coming, Airbnb/contacts greenlight; smart-device-models reminder carded P0.
• decided: Captured ideas as cards (no rushed builds): news-briefing revive, Fidelity-firehose→digest, proactive piecemeal elicitation (bot chips stuck "needs-input" cards one item at a time). DMARC = filter/ignore; push-email = use Gmail iOS app for the Workspace account.
• open: Big roadmap lanes Sam named 5/27 (not one-turn builds): (1) email inbox-zero/current for BOTH personal + business — a real triage-tool project; (2) Google Drive migration of the "2 Hook Street" + "2026 BH" coding workspace (handoff doc: docs/GOOGLE_DRIVE_MIGRATION_HANDOFF.md); (3) grooming/stale intelligence — flag "hasn't moved in N days / too-long-outstanding cashflow / backburner sitting too long" + a 60–120 day email re-search for dropped items; (4) the full ownership/last-touched/times-reminded dashboard (card 009). Each is its own focused effort.
• open: v3 cashflow engine coded (v3-obligations.js) but never run; weekly grooming routine (card 023) not scheduled; comm-attribution tagging ([AUTO]/[CLAUDE CODE]/[needs you]) not built; progress dashboard (card 009) not built.