Graceful Ctrl-C shutdown for mise dev / dev-all

[habdelra]

Summary

Ctrl-C through the mise run dev / mise run dev-all stack now lands as a clean WIFEXITED(0) across mise, pnpm, run-p, and every service, with no leaked file watchers, no orphaned vite/worker processes, and no spurious error lines from intermediate wrapper layers.

Orchestrators (mise tasks)

• mise-tasks/dev and mise-tasks/dev-all: signal trap does a fast_shutdown_kick — SIGTERM every recorded pgroup and return immediately. The slow TERM→KILL escalation + sweep_orphaned_services runs in the background under the cleanup guardian after this bash exits, so mise records WIFEXITED(0) instead of WIFSIGNALED when its per-task grace runs out mid-cleanup.
• The trailing wait "$SAT_PID"'s 128+signal return is normalized to 0. With set -m, Ctrl-C is delivered to SAT's pgroup — not this bash — so the INT/TERM/HUP trap doesn't fire; the EXIT trap still kicks the guardian and the script exits 0.
• Per-service mise-tasks/services/* scripts trap INT/TERM and exit 0 so the 143-on-Ctrl-C from the ts-node … | dev-log-tee.sh pipeline isn't reported by mise / run-p as a task failure. The two services with existing icon-server EXIT cleanup (realm-server, test-realms) preserve that path; INT/TERM additionally call cleanup once and exit 0.
• mise-tasks/lib/dev-common.sh sweep regexes:
  • VITE_SERVE_RE drops the ${REPO_ROOT_RE}/… anchor — pnpm invokes the wrapper as node scripts/vite-serve.js (relative argv), so the absolute pattern never matched and the sweep silently skipped it.
  • VITE_BIN_RE drops the trailing --port 4200 — in local-HTTPS dev mode the wrapper puts vite on a dynamic internal port, so pinning to 4200 missed every vite process.

Realm-server

• packages/realm-server/main.ts wires SIGTERM/SIGINT to the same stopRealmServer path used by the IPC stop message. Process-group sweeps from mise dev previously had no signal handler and had to escalate to SIGKILL.
• stopRealmServer iterates the mounted realms and calls realm.unsubscribe() on each, closing the underlying sane → fs.watch FSWatcher handles. Without this, each realm pinned a watcher and the process couldn't exit naturally — wtfnode dumps showed hundreds of FSWatcher handles after shutdown signal.

Host vite wrapper

packages/host/scripts/vite-with-traefik.js:

• Spawns vite with stdio: ['ignore', 'inherit', 'inherit'] so process.stdin.isTTY is false inside vite, suppressing the bindCLIShortcuts readline that produced the read EIO stack trace when the parent TTY tore down.
• Spawns without shell: true so child is the npx process, not an intermediate sh -c that would absorb our forwarded signal alone.
• Forwards SIGTERM/SIGINT/SIGHUP to vite, then exits 0 in the same tick. Waiting for the child to acknowledge made the orchestrator's ~2s SIGKILL grace expire mid-wait, surfacing as [ERR_PNPM_RECURSIVE_RUN_FIRST_FAIL] from pnpm.
• Non-INT/TERM signal exits from the child translate to 128 + signum instead of silently exit 0, so SIGKILL/SIGSEGV/SIGABRT on vite aren't masked as a clean shutdown.

packages/host/scripts/vite-serve.js runs ensure-boxel-ui inline via execFileSync, and packages/host/package.json start collapses to a single node scripts/vite-serve.js. The previous pnpm ensure-boxel-ui && node scripts/vite-serve.js chain forced pnpm to invoke the script through sh -c, and the shell — having no SIGTERM handler — died via signal on Ctrl-C even though Node exited 0, surfacing as [ERR_PNPM_RECURSIVE_RUN_FIRST_FAIL] and Command failed with signal "SIGTERM". Removing the && keeps Node as pnpm's direct child.

Observability for future hang investigations

New BOXEL_WTFNODE=1 opt-in helper (packages/realm-server/lib/wtfnode-on-signal.ts and packages/host/scripts/wtfnode-on-signal.js) dumps the active handles on SIGINT/SIGTERM and again 5 seconds later. Wired into every node entry point (realm-server main, worker-manager, worker children, prerender-server, prerender manager-server, vite wrapper) so future shutdown-hang investigations have evidence without ad-hoc edits. Disabled by default; the runtime cost is one process.on listener.

Real failures still surface

Only INT/TERM/HUP are translated to exit 0. Any non-signal child failure (ts-node faulting on its own, an indexer crash, vite OOM, etc.) still flows through pipefail and surfaces as a real error.

Test plan

• mise run dev, wait for the stack to come up, hit a card preview URL to kick off indexing, then Ctrl-C. Expect:
  • No exited with 143 / ERROR task failed / read EIO / [ERR_PNPM_RECURSIVE_RUN_FIRST_FAIL] / Command failed with signal "SIGTERM" / no exit status.
  • Shell prompt returns in well under 5 seconds; background cleanup continues silently after.
• mise run dev-all — same expectations.
• After shutdown, ports 4200/4201/4202/4210/4211/4221/4222 should be free (lsof -i:4200,4201,4202,4210,4211,4221,4222).
• With BOXEL_WTFNODE=1 mise run dev, Ctrl-C and confirm the 5s-later dump shows no FSWatcher / unexpected timer handles for realm-server, worker-manager, etc.
• Sanity: induce a fake crash (e.g. kill -KILL a ts-node child mid-run) and confirm the orchestrator still surfaces it as a real failure rather than swallowing.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 91a24422bc

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

Preview deployments

• Host staging preview

• Host production preview

Host Test Results

    1 files  ±0      1 suites  ±0   1h 44m 58s ⏱️ - 1m 12s
2 661 tests +2  2 646 ✅ +2  15 💤 ±0  0 ❌ ±0 
2 680 runs  +2  2 665 ✅ +2  15 💤 ±0  0 ❌ ±0 

Results for commit e324125. ± Comparison against earlier commit d9f833e.

Realm Server Test Results

    1 files  ±0      1 suites  ±0   8m 13s ⏱️ -33s
1 386 tests ±0  1 386 ✅ +1  0 💤 ±0  0 ❌  - 1 
1 467 runs  ±0  1 467 ✅ +1  0 💤 ±0  0 ❌  - 1 

Results for commit e324125. ± Comparison against earlier commit d9f833e.

[Copilot]

Pull request overview

This PR improves the developer experience when stopping mise run dev / mise run dev-all (and related service tasks) by making Ctrl-C / signal-driven shutdowns clean and non-erroring, while still allowing genuine failures to surface.

Changes:

• Add SIGINT/SIGTERM traps to service-level mise tasks so signal-driven exits don’t show up as task failures (143/130) in mise/run-p.
• Update mise-tasks/dev and mise-tasks/dev-all to run cleanup on INT/TERM/HUP and then exit 0, avoiding propagation of signal exit codes from the underlying wait.
• Adjust the host’s Vite launcher to ignore stdin (avoids Vite readline read EIO noise) and forward shutdown signals to the Vite child, translating signal exits to 0 to prevent pnpm recursive-run failures during shutdown.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
packages/host/scripts/vite-with-traefik.js	Detaches stdin for Vite and forwards shutdown signals; maps signal exits to 0 for clean pnpm start shutdown.
mise-tasks/services/worker-test	Trap INT/TERM to exit 0 to avoid 143/130 shutdown noise.
mise-tasks/services/worker-base	Trap INT/TERM to exit 0 to avoid 143/130 shutdown noise.
mise-tasks/services/worker	Trap INT/TERM to exit 0 for pipefail pipeline shutdown behavior under the orchestrators.
mise-tasks/services/test-realms	Split EXIT cleanup from INT/TERM handling to ensure cleanup runs once while exiting 0 on shutdown signals.
mise-tasks/services/realm-server-base	Trap INT/TERM to exit 0 to avoid 143/130 shutdown noise.
mise-tasks/services/realm-server	Split EXIT cleanup from INT/TERM handling to ensure cleanup runs once while exiting 0 on shutdown signals.
mise-tasks/services/prerender-mgr	Trap INT/TERM to exit 0 to avoid 143/130 shutdown noise.
mise-tasks/services/prerender	Trap INT/TERM to exit 0 to avoid 143/130 shutdown noise.
mise-tasks/dev-all	Split EXIT cleanup from INT/TERM/HUP handling; cleanup + exit 0 on signal-driven shutdown.
mise-tasks/dev	Split EXIT cleanup from INT/TERM/HUP handling; cleanup + exit 0 on signal-driven shutdown.

---

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