Crew Comms harness · MSG-36

One press live-fires a real message through the whole machine — send → Kindness Gate → store → delivery — plus the AI judge, the media scanner and a WS canary. Traffic lands only in the crew-only bmsg-harness org; every gate decision is a real, kept log row. (MSG-36 · MSG-33)

Run all checks now What do these checks actually do?

The checks, in plain English close ✕

a — round-trip	Sends a real test message and then reads it back: send → store → catch-up. Proves the core post-office works — a message you send is saved and can be fetched again, in order.
e — WS canary	Opens a live connection (the same one the chat screen uses), sends a message, and checks it ARRIVES over that live wire — then reconnects part-way through and checks nothing was lost. Proves real-time delivery and the "catch up after a dropped connection" promise.
b — Kindness Gate + AI judge	Feeds the gate one friendly sentence (must be let through) and one nasty one (must be flagged). Proves the machine that reads every message before anyone sees it is awake and judging correctly.
c — media scanner	Uploads a real tiny photo the whole proper way — ask for an upload ticket, upload the bytes, wait for the scanner to check them — and confirms the photo only becomes visible AFTER the scan passes. Proves no picture ever appears unscanned.
d — deployed build	Confirms the running service is the exact version that passed the full durability test suite in the deploy pipeline — nothing hand-patched, nothing stale.

All test traffic stays inside the crew-only bmsg-harness org — real users never see it — and every gate decision it triggers is a real, kept log row.

Last nightly run

The same checks run automatically each night — a broken organ shows up here, not in a user report.

Persona runs

After every deploy and every night, Crew Comms is used as each kind of person — member, owner, Crew Support, a BMSG Admin in and out of scope, the Super Admin, a verified child, a stranger — and every answer is checked against what it should be and the record it should leave. A red line is a finding for a person to read before the next build.