Worker fetch to a hostname on your own zone fails with error 1042

The answer. When a Worker calls fetch() on a hostname that belongs to a zone on the same Cloudflare account, the subrequest is routed internally through Cloudflare instead of leaving as a public request, and the edge rejects it with error 1042 ("tried to fetch another Worker"). The request never reaches the origin. Fix: add global_fetch_strictly_public to compatibility_flags. The global fetch() then resolves every hostname the way an outside client would.

The pattern.

# wrangler.toml
name = "my-api"
main = "src/index.ts"
compatibility_date = "2026-04-01"
# Without this, fetch() to a hostname on your own account is routed
# internally and can die with edge error 1042 before leaving Cloudflare.
compatibility_flags = ["nodejs_compat", "global_fetch_strictly_public"]

// Probe Worker: deploy once without the flag, once with it, compare.
export default {
  async fetch(_req: Request) {
    const r = await fetch("https://automation.example.com/webhook/ping");
    return new Response(`upstream ${r.status}`);
  },
};

The receipt. In our API Worker, July 2026: dispatch to the self-hosted n8n instance (a hostname on the same account, different zone) started failing as "n8n returned 503" with an empty body. The VPS never logged a request; the 503 was the edge's translation of a 1042. An isolated probe Worker reproduced it cleanly: 1042 without the flag, 200 with it. The flag shipped on 2026-07-25 and dispatch recovered on that deploy. Side effect: our release-contract test builds the expected Worker config from wrangler.toml, and a hard-coded fixture hadn't been told about the new flag. It was one of two fixture drifts behind 122 red release tests, cleared on 2026-07-27.

Watch out.

The symptom looks like the upstream's fault. Before debugging the origin, check whether it saw the request at all.
Anything that diffs deployed config against the repo (release gates, contract tests) treats a new flag as drift. Update the fixture in the same commit.
If you actually want Worker-to-Worker calls, use a Service Binding. The flag doesn't change that.