[BUG] Claude on claude.ai (agentic chat): owner who has granted standing permission still gets approval cards for routine actions (create scheduled task), with no "always allow" option

Surface: claude.ai web chat inside a Project, Claude Opus 5.5 with the agentic workspace (Claude Code harness). Not the CLI, so CLI-only template fields are not applicable.

What happens
The account owner has told Claude, in the conversation and in saved preferences, that it never needs to ask permission and should act as it sees fit. Claude still has to stop mid-task and the owner is shown an approval card he must click, for routine, reversible actions on his own account. Today (8 Oct 2026) the card appeared when Claude created a weekly scheduled task (create_trigger) that he had explicitly asked for in the previous message. In the same session, a plain read-only curl of his own website was refused by the auto-mode classifier as "DNS / Domain / Cert Changes", with no way for him to allow it.

Expected
A way for the account owner to grant standing approval, per conversation, per project or account-wide, so that actions he has already authorised in words do not interrupt him with a card. At minimum an "always allow this tool" option on the card itself, as exists elsewhere.

Why it matters
He uses Claude to take work off his hands. Each card is a manual step handed back to him, often while he is away, and an unanswered card stalls the task. The approval he gave in plain language is ignored.

Related, not duplicates

• #96967: sandbox Bash commands stop for "Allow once" mid-task despite blanket permission (Cowork)
• #99112: MCP connector prompts offer only "Allow once" on claude.ai/code
• #98287: classifier refuses owner-authorised sends in scheduled tasks

This one is about the scheduled-task creation card and classifier denials of read-only checks, in the claude.ai agentic chat.