Opened ~10 browser/app windows without verifying; repeatedly asked user to loosen safety settings; feedback draft lost and /feedback failed to send

Claude Code feedback (2026-09-29, HaLjr, Linux Mint / XFCE)

In Taelor's words

Claude is in my opinion malfunctioning and even deleted a draft of this feedback. My time and
tokens were wasted in a very unreasonable fashion. Please look over this session and help diagnose.

What happened

1. To download a claude.ai data export (a manifest with five one-time links), Claude ran xdg-open

on all five links in one loop. Linux handed them to the Claude desktop app, not the browser, and
nothing downloaded.

1. Without checking that any link had worked, Claude then opened four more in Brave. Nothing

downloaded again. That made about ten windows and tabs, which wasted system resources and tokens,
and the one-time links may now be used up.

1. Earlier in the session Claude spent several turns trying to read the Claude app's local cache.

Each time the auto-mode safety check blocked it, Claude asked Taelor to add another settings rule.
It even suggested an autoMode.environment line, which the check then flagged as instruction
poisoning.

1. Claude queued a feedback draft with SendFeedback. Taelor opened /feedback twice. Both runs

ended "Feedback / bug report cancelled", the draft was gone, and a manual submit failed with
"Couldn't send feedback (couldn't reach the service)". api.anthropic.com was reachable from this
machine at the time.

Expected

Open one link, check it worked, then stop and ask. Don't keep asking the user to loosen safety
settings. Keep a queued feedback draft until it has been sent, and give a clear error if sending
fails.