diff --git a/skills/deploy/SKILL.md b/skills/deploy/SKILL.md
index 9d36756..5a97eb5 100644
--- a/skills/deploy/SKILL.md
+++ b/skills/deploy/SKILL.md
@@ -66,6 +66,38 @@ money. That is not hypothetical.
 - **A dirty tree means TRACKED changes.** `git status --porcelain` counts untracked files; a stray script a
   fast-forward could never touch is not a reason to skip a host. Use `--untracked-files=no`.
 
+# `db:push` can offer data loss at a prompt, and it waits for an answer
+
+A schema push on a host that is many commits behind can stop and ask whether a table was RENAMED. Both
+answers destroy something, so neither is safe to guess and neither is safe to automate.
+
+Seen 2026-08-27 catching `edge-pertento-bridge` up by 215 commits:
+
+```
+Is command_hq_spawned table created or renamed from another table?
+❯ + command_hq_spawned                     create table
+  ~ dav_app_passwords › command_hq_spawned rename table
+```
+
+drizzle-kit saw one table leaving the platform schema and another arriving, and could not tell a rename
+from a coincidence. `rename` would have turned the DAV app passwords into the spawned-commander table;
+`create` would have dropped them.
+
+**The table had not gone anywhere — it had moved to a PLUGIN.** Caldav was extracted that day, so
+`dav_app_passwords` belongs to `plugins/caldav/db/schema.ts` and reaches the barrel through
+`gen-plugin-schemas`. Installing the plugin FIRST declares it, and the question disappears.
+
+So, on a host far enough behind that a plugin extraction is in the gap:
+
+1. Install the plugins whose tables are involved, **before** pushing.
+2. Then `bun db:push < /dev/null` — stdin closed, so it can never hang waiting on a prompt in an
+   unattended run. A push that needs an answer fails instead, which is the outcome you want.
+3. Expect ONE failure on a `role_permissions` foreign key, then `seed-roles.ts`, then push again. That
+   sequence is normal on a host predating the roles table.
+
+**Back up the database first.** `pg_dump` against the URL in `.env` — and read the variable name rather
+than assuming: several hosts use `OFFICERDEV_POSTGRES_URL`, not `POSTGRES_URL`.
+
 # The report is an artefact, the lock is the record
 
 Write the local record FIRST, then commit and push the report. If the push fails, say so and stop: the