What lands in hello@ that isn't a customer

An inventory of the internal and test mail that goes to hello@propflowai.co: what it is, how often it arrives, who sends it, and what to do about it. For Gera.

2026-10-08 · source: the hello@ inbox (searched over 7 and 14 days) plus origin/main code · counts are messages, not threads

The answer

The test that set this off doesn't repeat. It was a one-off. No synthetic test sends to hello@ every day. The only email that does arrive daily is the Daily Summary, which is a report, not a test. The volume comes from internal alerts and from bursts of test traffic whenever a session is testing.

The TEST-904 email came from one coding session on the mini, the "email-live-tests" lane, at 14:11Z on Oct 8. Nothing in the job list (config/automations.toml) or the GitHub workflows runs that test residents script on a schedule.
Correction (2026-10-08, later the same day): the 79 alert emails were not in hello@. Every one is addressed To: gera@ (live BAKE_ALERT_EMAIL = gera@). This inventory read hello@ through the google-workspace MCP, which runs as gera@ only, so it counted Gera's mailbox. Mail that truly reaches hello@ carries List-Id: hello.propflowai.co. Identity-refusal alerts are now Slack-only anyway (#17954).
Over 14 days there were about 27 synthetic test emails, in 5 or 6 bursts that all came from sessions. They were not scheduled.

Full breakdown: every kind of mail, three fixes, open questions

What arrives, by kind

Kind	Example subject	How often	Sent by	Is it a test?
Internal alerts	`[PropFlow bake alert] Identity refusal: phone-name-mismatch…`	79 / 7d	`sendBakeAlert`, addressed to gera@, not hello@ (corrected; see above). Identity refusals Slack-only since #17954	Not in hello@
Daily Summary	`PropFlow AI Agents - Daily Summary - Oct 07, 2026`	1 / day, ~00:10Z	The `/whats-shipping` skill's `~/.claude/daily-summary` sender (`cli.py install-schedule`), via SendGrid, `To: hello@`. Not installed on the mini, Gera's MacBook or AWS, so most likely on Fede's Mac (not confirmed)	No, it's a report
Test-building replies	`Re: Kitchen cabinet door hinge broken - unit TEST-904`	4 / 14d	Clara at the Yale sandbox, replying to fake residents C and D (`hello+yale-resident-c/d@`, created by `scripts/seed-email-live-test-residents.ts`)	Yes, a synthetic test
Test-company team notices	`Pat H. — leaking kitchen sink`, `Five — The Willows`, `Tour rescheduled: Jordan Ellis — 118 Quillwood Lane`	~23 / 14d (13 of them on Oct 4 between 03:30 and 04:25Z)	Clara, sending notices for the test companies (Harness Way, The Willows, Fairhaven) to a team inbox that lands in hello@. One of those addresses is `hello+team-notify-e2e@` (`scripts/team-notify-e2e/scenarios.ts`)	Yes, a synthetic test
Copies of real customer events	`Tour confirmed: … — Camellia`, `… — asked us something we could not answer`, `Mike — callback`	several / day	The hello@ copy that the team asked for (`OBSERVER_BCC_ADDRESS`, `src/lib/domain/notifications/observer-copy.ts`)	No, by design
Manual tests by people	`Routing test 3: partnerships@ to hello@ (ignore)`, `tester`, `gera+morgan-engine1@` prospects	~7 / 2d	Gera, by hand	Yes, done by hand
Vendor mail	Vercel "Preview deployment failed", GitHub, Zoom, LinkedIn	~15 / 2d	Outside services	No, just noise

Three things worth fixing

Move synthetic tests off hello@. DONE 2026-10-08. test@propflowai.co did not exist until Gera renamed the Test Client Workspace user to it today. Until then SendGrid had it on its bounce list (since April) and dropped every send there; that entry is cleared. PR #17939 moved fake residents, hello+team-notify-e2e@ and the staff-answer test's renters to test+<tag>@. Residents C and D are archived, and new residents E and F (test+yale-resident-e/f@) replace them. The Fairhaven (TEST) and Sandbox companies' team-notice inbox (escalationOwnerEmail) now reads test+team-notify-e2e@. A test building configured with the old hello+ address is now refused.
Identity-refusal alerts: DONE 2026-10-08. Slack-only since #17954. Each one still triggers an Agent Smith investigation, because Fede ruled every refusal a bug to reproduce. They had been going to gera@, not hello@.
The repeated "could not answer" emails are not a repeat-send bug. Corrected 2026-10-08; this line first called it one. The Thomas Rossi notices went to Western Slope's real leasing inbox, with hello@ on the intended copy. Each one quotes a different message: Thomas was texting live, and Clara sent the team one email per message, 11 in about an hour. Several were not questions ("You can rent it to me. Then I would be very happy"), and one question went out twice, 23 seconds apart (raw, then summarized). The fix is a product call on grouping a live conversation's notices and not flagging non-questions. Fede owns it and has already told the customer.

Not checked yet

Answered: bake alerts never reached hello@; they go to gera@ (see the correction above).
Which machine runs the Daily Summary schedule. It is the /whats-shipping sender, but its install is not on the mini, Gera's MacBook or AWS.

Method: Gmail searches on hello@propflowai.co (newer_than:7d and newer_than:14d, filtered by sender and subject) and git grep on a freshly fetched origin/main for hello@ addresses and schedules (config/automations.toml, .github/workflows/harnesses.yml, nightly.yml; the two weekly harness runs fire Sundays at 08:00 and 10:00Z, which doesn't match the Oct 4 burst).