Expose client-facing chat sanitization hooks separate from persistence sanitization

[arthurfiorette]

Today Think & AiChatAgent allows to sanitize persisted messages via sanitizeMessageForPersistence(), but doesn't expose a hook to sanitize what is exposed to the client. For example, I might want to hide from the frontend a tool output. I know this does not 100% guarantee and the user might just ask like "what was that tool output", but for that we can add more system prompts and other anti-jailbreaks.

Something like:

protected sanitizeMessageForClient(message: UIMessage): UIMessage

// and

protected sanitizeChatResponseChunkForClient(
  chunk: UIMessageChunk
): UIMessageChunk | null

or just a

protected sanitizeOutgoingChatFrame(
  frame: OutgoingMessage
): OutgoingMessage | null

Ideally they should cover CF_AGENT_USE_CHAT_RESPONSE, CF_AGENT_CHAT_MESSAGES, /get-messages...

A more declarative option like:

clientTranscriptPolicy = {
  stripToolOutputs: true,
  stripToolInputs: true,
  stripProviderMetadata: true,
}

would also be welcomed :)

[threepointone]

We use the entire messages array, so stripping on the way out to the client isn't a good idea, probably better to just have client side logic to not render, or to have your tool calls itself not return sensitive data (and use other storage on the server).

[arthurfiorette]

Could you expand a bit on "We use the entire messages array, so stripping on the way out to the client isn't a good idea"? Not sure if I understood.