openclaw - 💡(How to fix) Fix Proposal: pre-turn context-staleness check before treating last message as answer [1 participants]

There's a class of bug that shows up in assistants that do slow work mid-conversation.

The scenario:

1. Assistant asks: 'Which plan would you prefer - monthly or annual?'
2. Assistant kicks off a tool call that takes 8 seconds
3. User types 'I'll be on Telegram' while waiting
4. Tool finishes. Assistant reads latest message, treats 'I'll be on Telegram' as the answer
5. Wrong action follows

This happens because most turn-based assistants read 'the latest user message' and assume it replies to the last outbound question. That assumption holds in fast back-and-forth. It breaks whenever the assistant does anything slow.

A fix that works: I built a small Node module with timestamp + intent heuristics that catches this before the turn runs. Five rules in priority order: timestamp comparison, generic-ack detection, multiple-choice matching, domain word overlap, ambiguous fallback.

Repo: https://github.com/simonlpaige/response-context-check

Zero deps. Under 200 lines. Works as a module or CLI (pipe JSON in, get JSON out). Also includes an OpenClaw skill file at skills/response-context-check/SKILL.md that documents when to invoke it.

Where it would fit in OpenClaw: As middleware in the turn-processing pipeline, before the agent reads the inbound message. If the last outbound message was a question, and more than N seconds passed, run this check automatically. If recommendation is not 'treat-as-answer', either skip acting or ask for clarification.

Would you be open to a PR that wires this in as an optional middleware step? Or even just linking the standalone repo from the docs for users who want to add it themselves?