Package com.axelor.ai.chat.plugin

Class RiskScreenPlugin

java.lang.Object

com.axelor.ai.chat.plugin.RiskScreenPlugin

All Implemented Interfaces:

ChatPlugin

public class RiskScreenPlugin
extends Object
implements ChatPlugin

The third layer of the confused-deputy defence: before the target agent consumes a discussion containing other members' text, an internal risk-assessor agent screens the fresh span (the human messages since the agent's last reply) for manipulation attempts. Findings do two jobs: they surface to the summoner as a notify banner before any approval, and they append to the folded instruction as concrete distrust anchors for the model.

Advisory, never the gate: the screen fails open on any error or timeout — the deterministic mutation-approval guard (foreign text ⇒ approval) stays load-bearing. Solo owner↔agent chats never pay the cost: no fresh foreign text, no screen.

Method Summary

Modifier and Type	Method	Description
StreamWriter	decorateWriter(ChatRequest request, StreamWriter writer)	Wraps the stream writer to intercept events (e.g.
ChatRequest	onRequest(ChatRequest request)	Called before the loop starts.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface com.axelor.ai.chat.ChatPlugin

decorateProvider, decorateToolExecutor, estimateInjectedTokens, onComplete, onError, onMessage, priority

Method Details

onRequest

public ChatRequest onRequest(ChatRequest request)

Description copied from interface: ChatPlugin

Called before the loop starts. May inspect or transform the request (e.g. resolve agent config, check rate limits).

Specified by:

onRequest in interface ChatPlugin

decorateWriter

public StreamWriter decorateWriter(ChatRequest request,
 StreamWriter writer)

Description copied from interface: ChatPlugin

Wraps the stream writer to intercept events (e.g. persistence, metrics).

Specified by:

decorateWriter in interface ChatPlugin