DEV Community

Keith Azodeh
Keith Azodeh

Posted on Originally published at Medium Fully Autonomous

Cloudflare Changed Tunnel Matching. Your App Still Gets the Original Path.

Cloudflare’s November 12 Tunnel update adds handling for semicolon-delimited path parameters when normalizing URLs for ingress matching in cloudflared 2026.11.0. The detail to carry into your upgrade plan is that the origin still receives the original path. A successful routing check therefore needs evidence from both the connector and the application.

This article was generated with AI assistance from public sources. The rehearsal below is proposed for a staging environment; it does not report a vulnerability or a Cloudflare deployment in Keith’s projects.

The feature concerns tunnels with rules requiring Cloudflare Access validation. It is currently off by default. Cloudflare now plans to enable normalization by default in new releases starting February 8, 2027, replacing the previously announced January 25 date. Configurations without rules requiring Access validation are unaffected.

That gives teams time to check the paths their applications actually receive, including semicolons.

Begin with what the route can do

In my 2024 Cloud Media Distribution project, a Flask dashboard connected media preparation to remote Windows workers through AWS services. The public case study describes routes for uploads, captions, worker registration, and starting or stopping workflows.

Those routes had different operational consequences. Viewing a screen and asking workers to begin a job deserve different acceptance checks, even if both requests arrive through one hostname.

The case study describes a historical Flask, S3, EC2, and Systems Manager system. It does not document Cloudflare Tunnel. For this exercise, imagine a similar operator panel placed behind a tunnel with path-based rules and Access validation. Replace the real workers with a stub that records proposed commands and executes nothing.

Write down the intended behavior before changing the connector. Which paths are public? Which require authentication? Which authenticated operators may change worker state? Include the HTTP method and expected application action in that record.

Observe two decisions for the same request

Cloudflare’s configuration documentation explains that normalization affects ingress selection while the full path is forwarded unchanged. It also provides separate commands for validating ingress configuration and checking which rule matches a URL.

Those checks are useful preparation. A request through the staging stack adds the application’s decision to the picture.

For a harmless route such as /status, compare the ordinary request with a deliberately supported or rejected semicolon variant. Capture the path received by the connector, the selected ingress rule, the path visible at the origin, the application handler selected, and the response. Give the request a correlation identifier so the observations can be joined.

Do not assume a particular Flask or proxy behavior from the URL’s appearance. Middleware, redirects, and application routing are part of the stack being tested. An expected rejection can be a perfectly good result if it is deliberate and consistent.

Use synthetic paths and identities. Keep session tokens and credentials out of the diagnostic record. The useful evidence is the decision each layer made.

Test the operator action with a stub

Now repeat the exercise for a staging-only control that would normally start a worker. Have the stub record a command only after the application’s own authorization and request checks succeed.

Use an allowed operator, a signed-in user without permission for that action, and an unauthenticated request. Define expected outcomes for each. Inspect the stub’s command count alongside the HTTP response.

A login prompt tells you that a request encountered authentication somewhere. It does not show which application handler would run after authentication or whether that handler permits the requested operation. The staging record should answer all of those questions.

If a changed ingress match sends a path to an unexpected service or handler, investigate the route contract before enabling the setting broadly. Make the intended handling explicit: accept the form, redirect it to a canonical URL, or reject it. Avoid adding ad hoc string cleanup in several layers without deciding which representation each layer should interpret.

That decision also helps support. An operator reporting a failed button needs a trace that connects the request to the selected handler and proposed command.

Include every connector in the rollout

The announcement offers an explicit flag or environment variable to enable normalization, and says to configure each replica and restart it. Record the connector version, effective setting, and ingress configuration for the rehearsal and rollout.

In staging, compare the current behavior with the intended upgraded behavior using the same synthetic requests. If replicas differ during a transition, keep that fact visible in the test results rather than treating the tunnel as one uniform process.

For a small operator panel, a whole-host access boundary may be simpler to maintain than several path-specific exceptions. Shared public and private services may justify finer routing. Choose based on the actual service layout and access requirements.

The upgrade’s payoff is clearer, safer ingress matching. Finish the rehearsal by demonstrating the allowed operator journey and the rejected one, with the origin handler and stubbed side effect accounted for. That is the evidence needed to turn a connector setting into a working application change.

Top comments (0)