Step gate

The step gate is a PreToolUse hook that stops every tool call (the main agent's and its subagents') until a person answers. It's what interactive mode means. Tools that ask the person themselves (ask_human, present_plan, the approval and manual-intervention tools, request_file, retire_source) and the task list (TodoWrite) go through without asking first.

When it's active

buildSessionOptions() registers it in guided and interactive sessions and checks the current mode on each call: while the mode is interactive it asks, otherwise it gives no decision and the call proceeds as if it weren't there. That's what lets a guided session switch to interactive and back (Modes). An autonomous session doesn't have it.

What it asks

A decision checkpoint titled Proposed action, with the tool's name and its parameters as formatted JSON:

Proposed action
Tool: WebFetch
Parameters: {
  "url": "https://example.com/news",
  "prompt": "Summarize the headlines"
}

What each answer does

Answer	Values	Effect
Yes	y, yes, Enter ("")	The call proceeds.
No	n, anything else	The call is denied. The model is told "Action rejected by the user in step-by-step mode." and decides what to do next.
Stop	q	The call is denied and the turn stops (continue: false): "Execution stopped manually by the user."

The texts for the model are always in English, whatever the kit's language; the panel is in the kit's language.

Using it in your own hook chain

createStepGate(runDir, isActive?) is exported to build the same hook yourself, for instance in a session you assemble by hand:

import { createStepGate } from "@falkenslab/agent-kit";

const stepGate = createStepGate(runDir, () => supervised); // asks only while `supervised` is true

const options: Options = {
  // …
  hooks: { PreToolUse: [{ hooks: [stepGate] }] },
};