v0.4.0

Two things a visitor could not do before: check what the assistant sends before using it, and get anywhere useful when the page does not carry the answer.

See what would be sent

A new per-content-element switch, Show configuration, renders a collapsed block naming the system prompt, the editor's supplemental instruction, the page area used as the source and the context limit — plus a sentence for the two things it cannot list verbatim, the page text itself and the automatic answer-language instruction.

The values come from the same normalized settings the assistant uses, not a second copy, so the disclosure cannot drift from the behaviour it describes. It sits outside the block that stays hidden until the browser reports a usable model: what would be sent is worth reading precisely in a browser that cannot run the assistant, which is what most visitors have.

An answer for the questions the page cannot answer

Until now an unanswerable question ended in a refusal, which is correct and useless — the visitor came with a question and left with nothing, while the site usually does have somewhere to send them.

notFoundMode set to Content element shows an editor-selected element in place of the model's refusal; notFoundContent picks it under the same rules as fallbackContent (same page, enabled, no cycles). Put a link to the search, a contact address, a related page — whatever actually helps.

Only the model knows whether the page answered, so it is asked to signal an unanswerable question with a marker, and the interface swaps in the prepared element. Two properties make that safe:

• A model that ignores the marker answers as before. The failure mode is the previous behaviour, not a broken page.
• The bare marker can never reach a visitor. The instruction is added only when the selected element actually renders, so a missing, hidden or cyclic reference leaves the prompt untouched.

The beginning of a streamed reply is withheld for as long as it could still become the marker, so the token never flashes on screen before the editor's content replaces it — and is released the instant the reply turns out to be an answer.

Upgrading

No migration, and nothing changes for an existing plugin: both settings are off by default.

Full changelog: https://github.com/netresearch/t3x-nr-browser-ai/blob/main/CHANGELOG.md