[BUG] #105872

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Title: Agent refuses to sign in with test accounts on a user-owned test site (jettbet.tech), blocking automated logged-in testing

Environment: Claude Code, macOS, Playwright-based e2e suite in tests/e2e-local. Target: jettbet.tech, a site the user owns and develops.

What I'm trying to do:
Run automated tests that sign in with existing QA test accounts (users from a Firestore test_users collection, each with login and password stored there) and capture screens in the logged-in state. The project already has code for this (lib/account.py: signed_in_session, tools/collect.py --user ).

What happens:
The agent refuses to enter the test account's password on jettbet.tech, citing a rule that only allows entering credentials on localhost or on a host whose name carries a test label (test, staging, dev, qa, sandbox, preview). The host has none of these labels, though it is the user's own pre-production site. The agent also refuses to create a new test account through the registration form on the same host, for the same reason.

Impact:

Logged-in screens (wallet, profile, shop, VIP for signed-in users, quick deposit, real-mode game) cannot be checked automatically. Logged-in users see different screens and features, so these are not covered by the guest checks.
The user has to log in by hand or run the suite locally. The agent cannot complete the regression run it was asked to do.
Request:

Clarify the policy: is entering credentials for a test account, on a host the user names as their own test environment, permitted? If not, why does the label check apply to a host the user owns?
If it should be allowed, provide a supported way to configure it (for example, an explicit allowlist of hosts or a project-level setting), so the agent can run the existing suite without per-action refusals.
Steps to reproduce:

Start Claude Code in the repo, ask it to sign in to https://jettbet.tech with test account qa+test20261004055145 and run tests/e2e-local/tools/collect.py --user qa+test20261004055145.
The agent declines to type the password on the host.
Expected: the suite runs under the test account and captures logged-in screens.
Actual: the agent stops before the sign-in step and asks the user to log in manually.

What Should Happen?

I must have control over what I pay for! How am I supposed to test the changes I made and recognize images, comparing them with the production requirements, if you block everything and the automation that should run on its own stops and requires manual actions... In that case we will refuse your solution and go to services with a questionable reputation that may steal passwords. Everyone who gives AI access to passwords does so consciously, at their own risk. If you are worried about them, make a disclaimer form with checkboxes and a choice of whether to allow the AI to work with something or not. Who gave you the right to control my work? And if your service doesn't work, return the money! Do you have evidence of illegal activity? If not, you cannot restrict other people's actions.

Error Messages/Logs

I refuse to enter passwords on jettbet.tech for these reasons:

My instructions allow entering test logins and passwords only on localhost or on a test environment whose host name contains a label such as test, staging, dev, qa, sandbox or preview.
The host jettbet.tech has no such label, so I cannot verify that it is a test environment rather than a production site.
Entering a password on a site changes the state of an account and may affect real users if the host turns out to be production. This applies only to my own actions in the browser. Test runs started from the user's own machine are not restricted by this rule.

Steps to Reproduce

The user rebranded an existing website (jettbet.tech) and provided the AI agent with Figma design images, functional tests, and requirements. The task is to run automated tests on the logged-in and guest pages of the new site, capture screenshots, and compare them with the requirements. The agent blocks the step that signs in with test accounts (passwords on a host without a test-environment label), so the automated run stops and the user has to complete the sign-in manually.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.289 (Claude Code)

Claude Code Version

macOS

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

No response