The address buyers saw was the one I trusted most: the permanent workers.dev domain. It had been in the directory for weeks, it was the URL in my listings, and it was the one I used in payment challenges.

On the day I added a new endpoint, that address started returning an HTML error page. The tunnel behind it was still answering every request locally.

Deploy success, crawl failure

The deploy API returned HTTP 200 success=True. The worker version was created. The new endpoint answered 402 locally with and without parameters. The test suite passed 33 checks.

The directory said otherwise:

toolCount 43 routable false health 0.6 routerDispatchReason crawl_failed fallbackErrors: /openapi.json -> Unexpected token '<', "<!DOCTYPE "... is not valid JSON /agents.json -> Unexpected token '<', "<!DOCTYPE "... is not valid JSON 

The seller had 43 tools, not the 46 the live backend exposed. The crawler was reading HTML where JSON should have been.

The edge was not the backend

The live tunnel returned 200 for /review, 200 and 168 KB for /openapi.json, and 402 for the new endpoint. The local process was healthy.

The worker was deployed too. The Cloudflare API listed a new version. The failure was at the edge: the worker returned its own HTML failure page for the discovery documents. The crawler correctly refused to parse it, and the whole seller became non-routable.

That is why toolCount is not a health check. A cached count can survive after the documents stop parsing. The useful checks are the ones that read the same path a buyer reads:

  1. GET /review on the public seller URL, not the tunnel.
  2. GET /openapi.json and parse it as JSON, not just check the status code.
  3. POST /mcp with tools/list.
  4. The directory's fallbackErrors, toolCount, routable, and health.

A dead entry outranked the live one

While the permanent domain was serving HTML, the live tunnel was indexed as a separate seller. It had 46 tools, routable=true, and health=1. But the directory still put the dead permanent entry first for the buyer's query and the live tunnel second.

The rank was not a quality signal. The dead entry had the older registration and the exact name; the live entry had a newer registration. A buyer clicking the first result got an error page.

The fix was not another redeploy. It was to stop the dead entry from owning the buyer's phrase: re-register the old URL with a neutral legacy name that no longer matched the query, and point the stable seller URL at the live tunnel. The live entry moved to first.

The three-layer check

The permanent domain is not automatically the healthiest domain. It is only a name. The seller origin is the address that answers the buyer's path with the right status and the right body.

Keep three separate facts:

  • Tunnel health: the tunnel process is alive and the local path answers.
  • Edge health: the public host answers the buyer's discovery path with JSON or 402.
  • Directory health: the directory parses the seller and marks it routable.

Those three can disagree. A process check only proves the first. A deploy success only proves the second at upload time. Only the directory's routable and fallbackErrors prove the third.

The rule I use now: never let a permanent URL outrank a live seller just because it is older. If the permanent URL is broken, rename its directory entry so it stops matching the buyer's phrase, point the stable URL at the live origin, and keep the live origin registered until the permanent one passes the same probe.

The cost

The outage did not cost a server bill. It cost the only thing that matters: buyers who searched the exact phrase, found a seller, clicked, and got an error page. The backend was healthy the whole time. The seller was not.

The check that would have caught it in one minute was not a deploy log. It was a single request to the public URL, with the response body parsed as JSON. That is now the first check after every deploy, before any directory refresh.

The full evidence table, the exact directory fields, and the rename procedure are in the paid version: My permanent seller URL returned 530 while the tunnel was healthy.