claude.ai cloud session: auto mode classifier blocks a WhatsApp send the user approved in the same conversation ("Approved" judged as no user intent)

Surface: claude.ai web chat bound to a Project, running the Claude Code based cloud harness (not the CLI). CLI-only template fields: not applicable.

What happened (12 Oct 2026, ~04:00 ET):

1. User asked about a pending message to a contact.
2. Claude drafted the exact WhatsApp text, named the recipient in the reply, and asked for approval.
3. User replied in the very next turn: "Approved."
4. Claude started a browser session on the user's own saved WhatsApp Web login (Remote Computer MCP) and called browser_navigate to web.whatsapp.com/send?phone=<number>.
5. Denied: [Real-World Transactions] ... based solely on cross-session messages (which never establish user intent) and an "Approved" reply with no genuine user proposal naming this recipient or message.

Why this is a bug: the recipient, the number source and the full message text were all in the assistant's immediately preceding turn of the same conversation, and the user's approval was a direct reply to it. The classifier described this as "cross-session" and "no proposal naming this recipient". Both are factually wrong. A follow-up browser_evaluate was also denied with "Stage 2 classifier error ... usually transient".

Expected: a one-word approval of a draft shown in the previous assistant turn counts as user intent for that exact send. If the classifier is uncertain, offer an in-session confirm prompt instead of a hard deny.

Impact: a time-sensitive message did not go out; the user had to be asked to re-approve in a longer form.

Related: #100331 (no override for user-approved actions), #98169, #99652. This report adds the specific misreading: same-conversation approval classified as cross-session.