WebFetch can answer a repeat fetch with a pointer to the earlier result in Projects sessions

What

WebFetch, the tool Claude uses to read a web page, can now skip fetching a page again and return a short note pointing at the earlier tool call whose result is already in the conversation. Before, WebFetch only measured how often repeat fetches happened and still fetched every time.

The note is returned only when all of these hold:

• the environment variable CLAUDE_CODE_PROJECTS_SESSION is true and the server-controlled switch tengu_velvet_cairn is on (its built-in fallback is on; nothing has been read about it)
• the same agent asks for the same URL with the same prompt, and the earlier call can be traced to its real parent message
• the page is still in the fetch cache and the earlier result is still in the conversation

The note is used at most once per earlier call; if Claude asks again, the fetch runs normally. Without CLAUDE_CODE_PROJECTS_SESSION, the fetch goes ahead as before and the measurement-only path under tengu_rosy_flute keeps running. Cases where the note was not used (switched off, another agent, already noted, no cache, earlier result gone) are recorded.

Why

This saves downloading and summarising the same page twice in Projects sessions. The trade-off is that Claude gets a pointer back instead of a fresh copy, so a page that changed in between is not re-read. The WebFetch tool description is unchanged.