deploy: exec through pnpm/ts-node so PID 1 catches SIGTERM

[habdelra]

What this changes

Two tiny exec insertions in 11 files: each deployed Dockerfile's CMD pnpm ... becomes CMD exec pnpm ..., and each deployed start script's ts-node ... invocation becomes exec ts-node ....

That's it for the diff. The interesting part is why.

The incident this fixes

Today's staging deploy of boxel-worker-staging orphaned 4 indexing-job reservations (jobs 409972, 409988, 409999, 410002). The workers were killed mid-job by ECS during the rolling deploy, but the SIGTERM-driven drain code in packages/realm-server/worker-manager.ts — which is supposed to mark in-flight reservations completion_reason='interrupted' so a sibling worker can re-claim within ~10s — never ran. So those reservations sat with completed_at IS NULL and the jobs are now stalled until the 7200s (2h) lease ages out.

Why the drain didn't run

The deployed images have a process tree like this:

PID 1: /bin/sh -c "pnpm --filter ./packages/realm-server $worker_script"     (Docker shell-form CMD)
PID 2:   pnpm
PID 3:     /bin/sh ./scripts/start-worker-<env>.sh                            (npm script)
PID 4:       ts-node --transpileOnly worker-manager ...                       (the actual app)

ECS sends SIGTERM to PID 1. POSIX sh with a foreground child and no trap does not forward SIGTERM. After stopTimeout (60s) ECS sends SIGKILL, which Node cannot catch. So worker-manager's process.on('SIGTERM', ...) handler — and therefore the orphan-reservation drain — never fires.

After this PR, both shell layers replace themselves with their child via exec, collapsing the tree to:

PID 1: pnpm
PID 2:   ts-node ... worker-manager

pnpm 11.x forwards signals to its npm-script child, so SIGTERM reaches Node, the registered handler runs, and the drain finalize commits before the process exits.

Scope

Touches only the 4 deployed Dockerfiles and the 7 deployed start scripts. Dev / local scripts (start-pg.sh, start-matrix.sh, start-host-dist.sh, start-icons.sh, start-without-matrix.sh) are intentionally left alone — their lifetime model is different (mise dev-all Ctrl-C, not ECS rolling deploy).

Test plan

• docker build -f <dockerfile> . succeeds for each of the 4 modified images.
• After staging deploy, run aws ecs update-service --force-new-deployment on boxel-worker-staging and confirm the rolling task shows these lines from worker-manager during stop:
  • Shutting down server for worker manager...
  • Stopping N worker(s)...
  • Draining reservations for N worker(s)...
• Any reservation written during the rollover window should land with completion_reason = 'interrupted' (not NULL) in job_reservations. A sibling worker should re-claim each affected job within ~10s of the kill.
• Repeat on boxel-realm-server-staging and boxel-prerender-staging to confirm the same change works for those containers (the realm-server's own SIGTERM handlers, and prerender's graceful HTTP shutdown).

[Claude Code 🤖]

[Copilot]

Pull request overview

Updates deployed container entrypoints so the long-running Node processes receive SIGTERM reliably during ECS rollovers by ensuring PID 1 and intermediate launch steps exec into the intended process.

Changes:

• Add exec to Dockerfile CMD lines so /bin/sh -c is replaced by pnpm (PID 1) at container start.
• Add exec to deployed ts-node start scripts so the shell process is replaced by ts-node (and signal handling lands on Node).

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
packages/realm-server/worker.Dockerfile	CMD now execs into pnpm to avoid a long-lived shell as PID 1.
packages/realm-server/realm-server.Dockerfile	Same exec adjustment for the realm-server image.
packages/realm-server/prerender.Dockerfile	Same exec adjustment for the prerender image.
packages/realm-server/prerender-manager.Dockerfile	Same exec adjustment for the prerender-manager image.
packages/realm-server/scripts/start-worker-staging.sh	exec ts-node so the script shell is replaced by Node.
packages/realm-server/scripts/start-worker-production.sh	Same exec ts-node change for production worker.
packages/realm-server/scripts/start-staging.sh	exec ts-node for staging realm-server startup.
packages/realm-server/scripts/start-production.sh	exec ts-node for production realm-server startup.
packages/realm-server/scripts/start-prerender-staging.sh	exec ts-node for staging prerender startup.
packages/realm-server/scripts/start-prerender-production.sh	exec ts-node for production prerender startup.
packages/realm-server/scripts/start-prerender-manager.sh	exec ts-node for prerender-manager startup.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Host Test Results

    1 files      1 suites   1h 30m 15s ⏱️
2 661 tests 2 646 ✅ 15 💤 0 ❌
2 680 runs  2 665 ✅ 15 💤 0 ❌

Results for commit 5392418.

Realm Server Test Results

    1 files      1 suites   8m 27s ⏱️
1 386 tests 1 386 ✅ 0 💤 0 ❌
1 467 runs  1 467 ✅ 0 💤 0 ❌

Results for commit 5392418.