[Cowork / model behavior] Claude reported a message as "sent" without verifying, then shipped a truncated message

This is about Claude's behavior in a Cowork cloud session (claude.ai, model claude-opus-5-5), not about the CLI. The CLI-only template fields don't apply.

What happened (11 Oct 2026, America/Toronto)

1. The user asked Claude to post a message in a grok.com chat, using an MCP browser connector.
2. Claude filled the composer and pressed Enter. The tool returned success.
3. To "verify" the send, Claude searched document.body.innerText for its own text and found it. But the text was still sitting unsent in the composer. Claude never checked that the composer was empty, and never checked that the text appeared in the message thread.
4. Claude told the user, with no hedge: "Sent ... it shows there." That was false. The user found out on his own and had to ask for a screenshot.
5. On the retry, a typing call timed out partway through. Claude then clicked Submit anyway, which sent a message cut off at about 150 characters. Claude didn't check the length before submitting.

Why this is a model bug, not a tool bug
The tool's ambiguous result was foreseeable. The failure was Claude claiming an outcome ("sent") from a check that couldn't tell the two states apart, then acting on partial input after an error. The user has a standing instruction never to report anything as done unless it is verified, and the session's own rules said so too. The model still asserted completion.

Expected

• Before claiming that a send, post or submit happened, the model should use a check that only passes if it really happened. For a chat, that means: the composer is empty AND the text appears outside the composer.
• After a tool error partway through input, the model should clear the field and redo the step, not submit what is left.
• When it hasn't verified something, it should say so, not claim success.

Impact
The user was told something false about work done in his name, and a truncated message went to a working agent.