WebFetch URL permission prompts can be shown one at a time with approval memory (tengu_orderly_lantern)

What

WebFetch is the tool Claude uses to read a web page. When it is not allowed to fetch a URL on its own, it asks you for permission. This prompt flow was reworked:

• Approved URLs are remembered while prompts are pending. Origins (the site part of a URL) are also remembered when your answer adds allow rules.
• A denial clears the remembered approvals.
• A request that waits behind another prompt and finds its URL or origin already approved is fetched again without asking (served_without_prompt).
• Earlier prompts left unanswered shorten the next prompt's timeout. A request can end as expired_in_line when waiting in the queue uses up its time, and the model is told: "The user did not answer an earlier permission request in time, so the request for this URL was not sent to them." The model is told to ask the user to include the URL.
• With tengu_orderly_lantern true, prompts are serialised through a lock (s.acquire), so only one is shown at a time. With it false, the new outcome handling exists but nothing is serialised. The compiled fallback is false, and nothing has been read about this gate yet.

Before, there was a single prompt with a fixed timeout, outcomes were limited to suppressed_*, timed_out, approved, denied and retry_denied, and nothing was remembered.

The stock WebFetch description captured under this release is unchanged.

Why

When Claude fetches several pages at once, you may see fewer stacked prompts and not be asked twice about a site you just approved. The one-at-a-time behaviour applies only where the flag is on.