On-Brand Support — peytontianyan account walk

2026-08-27 · account peytontianyan@gmail.com · distinct run, not the locked onbrand-support-full.md plan (methodology borrowed only) · BLOCKED on live conversation turns · 2 bugs found

For Peyton

I got the account ready and the assistant wizard confirmed the voice is live and correct — but I could not send a single guest message, and I found two real bugs along the way.

What's at stake: On-Brand Support is the flagship WhatsApp assistant. The plan you locked expects a 22-turn phone conversation graded against Salt & Cedar's real prices and policies. I could not run that conversation on this account today — not because the assistant is broken, but because there is no way to send it a WhatsApp message without a real phone, and I have neither a phone nor the Meta credentials to connect one on this test account.

What I did confirm, directly in the product:

Two bugs I found that aren't about my test, they're about the product right now:

  1. The Brain's own “read your site” research still drops in [ADDRESS] placeholder text where real words should be — the exact bug class the team thought was fixed.
  2. Brain pages (both the dashboard and the book itself) randomly render as completely empty on some page loads, then correct on the next — the underlying data is safe, but the display is not reliable.

My recommendation: file both bugs, then have someone with a phone finish the conversation walk on this now-seeded Brain.

Ask: want me to file the two bugs as GitHub issues, or do you want to see them first?

1 · Account confirmation

Signed in as peytontianyan@gmail.com — confirmed via the Account page before any other action.

Account page confirming peytontianyan@gmail.com identity
Account page: peytontianyan@gmail.com, USR-CA986673, wallet $15.17. Confirmed before any Brain or activation work began.

2 · Brain wipe + reseed

reset-brain-data.sh --email peytontianyan@gmail.com --execute wiped 2 Brain rows + flagged 11 orphaned brain-images storage objects (HTTP 400 on that clear — not blocking, no profile row to reference). Pulse confirmed no pre-existing assistants on this account.

Reseeded Salt & Cedar Retreats via the real “Teach your Brain” paste flow (not a script) — business summary, 2 offerings, 5 FAQ pairs, schedule + deposit policy. Final book: 8 of 10 chapters filled, correct content throughout (see Bug 2 section for why this took extra verification).

3 · Activation wizard walk

Personalize — Warm Host + AI identity line, confirmed live

Personalize tab showing Warm Host voice and live preview with correct AI identity line
Personalize tab: “Warm Host — default” voice with the exact beta-locked trait list, and the live phone-mock preview correctly opens with “Salt & Cedar Retreats' AI assistant here — a real person can step in anytime.” Not a real send (“Not live yet” label bottom-right), but this is the first proof the identity-line PR reads the right business name.

Connect — WhatsApp Business not connected, blocks the rest

Connect tab showing WhatsApp Business and Slack both showing Connect (not connected), Brain readiness 0 of 5
Connect tab: WhatsApp Business and Slack both show “Connect” (not yet connected on this account). Brain readiness check: “0 of 5 ready — 4 need you before it can run.” Did not click Connect — that's a real Meta OAuth grant requiring credentials this session doesn't have.

Confirmed via source (PreviewStep.tsx) that WhatsApp chat-type assistants have no in-app simulated preview — the only test path is a real phone messaging a connected WhatsApp Business number. Combined with WhatsApp not being connected here, the 22-turn grounding script, Slack/phone escalation proof, and live voice transcript grading could not run this sitting. No conversation was fabricated.

Bug 1 BUG — [ADDRESS] placeholder leak recurs in fresh Brain drafts

Pasting the seed's business summary (containing “Atlantic coast of Portugal”) produced a confirm-screen draft reading “...on the [ADDRESS] coast of [ADDRESS]...” — both real place names replaced with the literal placeholder token. Manually corrected before saving. A second, independent occurrence of the same corruption then appeared as an unconfirmed “YOUR CALL · DRAFT” suggestion on the main Brain page (never used).

Brain page showing the confirmed correct business text above and a leftover draft card with [ADDRESS] placeholder tokens below
Confirmed “Your business” text (top, correct: “Atlantic coast of Portugal”) directly above a “YOUR CALL · DRAFT” card auto-generated from the same source content, still carrying [ADDRESS] in place of both place names.

The locked plan's own provenance notes (CT-OBSF-01) describe this exact bug as previously root-caused and fixed in the FAQ path (organize-content.ts, missing the containsPlaceholderToken guard from #2693), with an explicit “watch for recurrence” flag. This walk shows it recurring outside the FAQ path — in the main business-summary organize/classify step and the “Our read” auto-research draft generator. The fix looks scoped too narrowly.

Bug 2 BUG — Brain pages randomly render an empty scaffold instead of the real book

Reloading the identical URL, no edits in between, produced inconsistent results — reproduced on both /brain (dashboard) and /brain/book (the actual book-reading surface). The underlying saved data was correct every time it did render — nothing was lost. This looks like a client fetch/hydration race that falls back to an empty scaffold instead of a loading state.

/brain/book — same URL, two consecutive loads

/brain/book showing generic Your playbook title and 6 chapters still empty
Load A: generic “Your playbook,” “6 chapters still empty” — as if nothing had been entered.
/brain/book showing correct Salt and Cedar Retreats title and 2 chapters still empty
Load B (next reload, same URL): correct “Salt & Cedar Retreats,” “2 chapters still empty” — the real, complete state.

/brain dashboard — reproduced twice in a row

/brain dashboard showing 0 percent of required and generic playbook title
First reproduction: “0% of required,” generic “Your playbook” title, every field prompting “Add →.”
/brain dashboard showing the same 0 percent empty state on a repeat reload
Reproduced again on the next reload — same empty scaffold, same URL.

Related observation (not filed as its own bug)

The “Our read” card on the Brain dashboard, when opened, showed a business summary for “The Offline Club” (theoffline-club.com, dated Aug 25 — two days before this session, and not any business this session ever researched) instead of Salt & Cedar. Since this account is shared across several unrelated QA test rails, this is most plausibly a leftover cached research artifact from a prior test on the same account rather than a genuine cross-tenant leak — but it fits the same pattern: the “Our read” / dashboard summary surfaces are not reliably reflecting current state.

Our read page showing stale content for The Offline Club instead of Salt and Cedar Retreats
“Our read” page opened from the Salt & Cedar dashboard card — content shown is for “The Offline Club,” dated Aug 25, unrelated to this session's business.

Scoring (methodology borrowed from onbrand-support-full.md, not its account)

AreaResult
Grounding (22-turn phone card)BLOCKED — no WhatsApp connection, no in-app simulate
Escalation + handover (Slack/phone)BLOCKED — same reason; Slack also not connected
Identity + voice (Warm Host, AI identity line)PARTIAL PASS — confirmed live & correct in Personalize preview (not a real send)
Thin-Brain honestyNOT ATTEMPTED — blocked upstream
Brain seed fidelityPASS with 2 bugs found along the way

Overall verdict: BLOCKED. Not a product FAIL on the parts exercised (voice/identity looked correct) — blocked entirely on the absence of a real phone to send WhatsApp turns, which the product itself confirms is the only test path for this assistant family. Two real, reproducible bugs found and documented.

Full findings

Verbatim seed text, script commands, and complete write-up: onbrand-support-peytontianyan-findings-2026-08-27.md