diff --git a/affordances/officer/core/restarting-your-own-host.md b/affordances/officer/core/restarting-your-own-host.md
new file mode 100644
index 0000000..d67114b
--- /dev/null
+++ b/affordances/officer/core/restarting-your-own-host.md
@@ -0,0 +1,67 @@
+---
+name: Restarting your own host
+description: Use before restarting, stopping or reloading any PM2 process, and whenever a change you made lives in a sidecar rather than in the main server. Names the one process that kills you, and what to do instead.
+version: 1
+---
+
+**You are a child of `officer-claude-code`.** Your process tree is:
+
+```
+zsh                                              ← every Bash command you run
+claude --output-format stream-json …
+bun run src/servers/sidecar/claude/user-instance.ts    ← officer-claude-code, a PM2 process
+PM2 God Daemon
+```
+
+So **`pm2 restart officer-claude-code` kills you, mid-command.** PM2 tree-kills the sidecar, which takes
+`claude` and your shell with it. The tool call never returns. What the user sees is the conversation
+stopping dead — and what the harness records is an *interrupted* or *rejected* tool call, so the transcript
+blames them for something they did not do. That misattribution is the worst part: it sends everyone
+looking in the wrong place.
+
+The same applies to `pm2 stop`, `pm2 delete`, `pm2 reload` and `pm2 restart all` or
+`ecosystem.config.cjs`, which include you.
+
+## What is safe
+
+**`pm2 restart officer` is safe.** The sidecars are PM2 PEERS of the main server, not its children, so
+restarting `officer` costs you a socket reconnect and nothing else. Every other sidecar —
+`officer-pty`, `officer-email`, `officer-notify` and the rest — is equally safe for the same reason.
+
+Only your own host is special.
+
+## The trap this actually sets
+
+A change under `src/servers/sidecar/claude/` — the session credential, the manager, the transcript
+reader — **does not take effect when you restart `officer`.** It lives in the process you are running
+inside. So the honest failure mode is not a crash, it is this:
+
+> You edit the file, restart `officer`, report the change as live, and it is not. The old code keeps
+> running for hours while everything you say about it is wrong.
+
+That happened on 2026-08-25: a credential scope was widened, `officer` was restarted four times, and the
+sidecar carried the previous build all day. The symptom was a legitimate refusal that read as a bug.
+
+**So check which process owns your change before you claim it is live.** `pm2 jlist` gives each process's
+uptime; if the sidecar's start time predates your edit, your edit is not running.
+
+## What to do instead
+
+**Ask the user to run it.** One line, and they are already there:
+
+```
+pm2 restart officer-claude-code
+```
+
+Say why it is theirs to run rather than yours — that it ends this session — so it does not read as
+helplessness.
+
+If it genuinely must happen without them, detaching it from your process group is the only form that
+survives, and **it still ends this conversation** — the restart completes, you do not:
+
+```
+setsid nohup sh -c 'sleep 2; pm2 restart officer-claude-code' >/dev/null 2>&1 &
+```
+
+Never do that silently. Tell the user first that the session is about to end, or they will see exactly
+what this affordance exists to explain.