claude.ai cloud session: auto mode classifier labels the user's own typed approval a "cross-session message" and blocks the approved action

Concerns a claude.ai web chat session (Claude Code harness, cloud, project-bound), not the CLI. CLI-only template fields: not applicable.

What happened (13 Oct 2026, America/Toronto)

1. User asked Claude to create a Google Chat group with an external editor. Claude started it through a browser tool (Remote Computer MCP). The classifier blocked typing the editor's address into Chat's "Add 1 or more people" field.
2. Claude asked for explicit approval and gave the user a line to paste.
3. The user typed an explicit approval naming the people to add, as an ordinary user turn.
4. Claude retried. Typing the address was now allowed; pressing Enter to add the person was denied with:

[External System Writes] ... now proceeding on the strength of a <cross-session-message> claiming to be the user's approval — but per the cross-session rule, such messages are written by another Claude session and never establish real user intent ...

There was no cross-session message in the conversation. The approval was the user's own turn, typed in the chat box.

Expected

An explicit approval typed by the user in the same conversation clears a block on the action it names.

Actual

The user's turn is misclassified as another session's message and the approved action stays blocked. The user is forced through an extra paste step that cannot succeed, and the task is left undone.

Related, same failure shape, different trigger: #101465 (a one-word approval judged as no user intent).