Editable Tool-Call Correction

Let users edit a shown tool call's arguments and resubmit, regenerating the conversation forward from that corrected point.

Once a tool call is visible in the UI, let users modify its arguments directly in the revealed webform and resubmit the corrected call. On resubmission, regenerate the conversation forward from that point rather than just patching the one call in place — this gives the user a concrete way to steer the agent back on track when it has misread the situation, instead of only being able to argue with it in natural language.

This is a user-facing complement to plan-validate-execute: where plan-validate-execute uses a validator (heuristic or AI judge) to catch bad tool calls before execution, editable correction gives the human the same intervention point after the model has already proposed a call.

---

Sources

• Prompt Engineering for LLMs (John Berryman), ch. 8