Let an HTML render be reused with fresh data instead of rewritten

[jieyuexing]

Note

Drafted by Claude Opus 5.5 in Claude Code (via T3 Code); reviewed and posted by @jieyuexing.

Problem / use case

An HTML render is a snapshot: the agent bakes the data into the page. To show the same view again with newer data, the agent has to write the whole page again.

I had an agent render a usage chart in a thread: about 7 KB of markup around less than 2 KB of data. If my weekly scheduled usage report is to show the same chart every week, the agent has to regenerate the whole page on every run, and the layout, colors and labels will drift between runs, which makes weeks hard to compare.

This is different from #16943, which asks when to choose an interactive reply. This is about reusing a view once it exists.

Proposed behavior

Either of these would solve it; the first looks smaller:

1. Reuse a published render with new data. html_render accepts an earlier render's id plus a JSON data value. T3 republishes that page with the data available to its script, for example as window.t3RenderData. The agent sends data only, and the layout stays identical between runs.
2. Let a render fetch its own data. htmlRender.ts notes that the render bridge already speaks the MCP Apps protocol so it can later drive MCP apps. Allowing tools/call from a render, limited to read-only tools under the same rules McpAppFrame already applies, would let a page refresh itself.

Expected result

A scheduled task or a follow-up turn can show "the same chart as last week, with this week's data" without regenerating the page. Its output stays consistent, and it costs far fewer tokens.