feat(examples): keep alchemy-cloudflare Durable Objects running on pending I/O

[K-Mistele]

Stacked on #52 (Effect 4.0.0 / Alchemy beta.80).

Cloudflare's pending I/O keep-alive change: from compatibility date 2026-10-01, pending calls to another Durable Object, service bindings, container.monitor(), waitUntil and timers keep a Durable Object in memory with no client connected. Each call is covered for up to 15 minutes. Outbound fetch() already was.

Changes

• Worker.ts and computer/ComputerWorker.ts move from compatibility date 2026-08-31 to 2026-10-01. The only other change in that range is Python 3.14 for Python Workers.
• Keepalive stays, with corrected comments. Pending I/O keeps a running turn in memory, but a deploy or crash still ends it. Keepalive's alarm is what wakes a cut-off session so it continues the turn. Its comment also claimed model fetches didn't count as activity, which was wrong: the real gap was the RPC call to the Computer.

Deploy-only for now

Alchemy beta.80 passes the date straight through, so deploys work. Local alchemy dev runs workerd 1.20260918.1, which probably refuses a later compatibility date until Alchemy pins a newer workerd.

Watch out: Alchemy beta.80's default stage changed

The default stage changed from dev_${USER} to live_${USER}. A plain alchemy deploy on beta.80 therefore planned a whole new stack. It created a second ChatWorker, re-uploaded the fixed-name Computer Worker, and failed on the container app. I cleaned up as follows:

• removed the live stage's record of the shared Computer Worker, without deleting the Worker;
• destroyed the live stage, which deleted its ChatWorker and its stray container app;
• redeployed with --stage dev_kylemistele.

Existing deployments need --stage dev_<user> or ALCHEMY_STAGE from now on.

Tested on Cloudflare

I deployed the dev_kylemistele stage and sent a message whose bash container command was sleep 240; echo finished-sleep. The HTTP client disconnected after 20 seconds. The command returned finished-sleep about four minutes later and the turn completed, with no restart nudge in the log. Caveat: I polled the log every 30 seconds and Keepalive's alarm still runs, so this shows the turn survives a disconnected client, not the new rule on its own.

Typecheck, lint, format and the example's 25 tests pass.

🤖 Generated with Claude Code