[BUG] Web (cloud) session, Auto mode: classifier denies an explicitly user-requested add_repo(push); UI/mode guidance from the assistant was wrong

Preflight Checklist

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

What's Wrong?

Title: Web (cloud) session, Auto mode: classifier denies an explicitly user-requested add_repo(push); UI/mode guidance from the assistant was wrong

Environment

• Claude Code on the web (claude.ai/code), cloud session
• Permission mode: Auto (the mode menu offers only Auto / Accept edits / Plan)
• Model shown in the UI: Sonnet 5.5

What happened

1. I explicitly asked the assistant, in my own chat message, to request access to several GitHub repositories one after another (add_repo with access: "push").
2. The auto-mode classifier denied the call with the reason "[Permission Grant]", although my own message asked for exactly this action.
3. The denial text said the user can "add a Bash permission rule to their settings". The tool involved is an MCP tool, not Bash, so this hint does not apply and there was no clear way to grant the permission.
4. I asked the assistant how to allow it. It told me to switch to the default mode and mentioned bypassPermissions. Neither exists in the mode menu of the web UI (see the screenshot: only Auto, Accept edits, Plan). The assistant answered from a list of internal mode names instead of what the product actually shows, and stated it with confidence.

Expected behavior

• An action the user explicitly requested in chat is either allowed by the classifier or turned into a one-click approval prompt for the user. Auto mode should not leave the user with no way to approve.
• The denial message names the mechanism that really applies (here: MCP tool permission rules, or the approval prompt) and where the user can change it in the web UI.
• The assistant has accurate, current documentation of the UI it runs in (mode names, where permissions are configured) and does not guess.

Actual behavior

• The explicit request was denied and the user could not find a way to allow it.
• The assistant's advice on how to fix it pointed to options that do not exist in the UI, which cost the user time.

Impact

The session was stuck: the task needed push access to an accounting repository, and neither the user nor the assistant could grant it from inside the session. The user had to spend several turns finding out that the suggested modes were not available.

Claude Model

Sonnet (default)