Human-in-the-loop

Pause the agent mid-response to ask the user a clarifying question, then resume with their answer.

Some turns need to stop and ask the user something before they can finish — picking between options, confirming a destructive action, or clarifying an ambiguous request. The AI SDK calls this human-in-the-loop (HITL), and the building block is a tool with no execute function.

When the LLM calls a tool that has no execute, streamText ends with the tool call still pending. The turn completes cleanly, the frontend renders UI to collect the answer, and when the user responds, a new turn resumes with the answer merged into the same assistant message.

How it works

Turn N:
  User message → run()
  LLM streams text → calls askUser tool (no execute)
  streamText ends with tool-call in `input-available` state
  onTurnComplete fires (finishReason = "tool-calls")
  Agent idle

Frontend:
  Renders question + option buttons from tool input
  User clicks → addToolOutput({ tool, toolCallId, output })
  sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls
  → sendMessage() fires next turn

Turn N+1:
  hydrateMessages / accumulator sees the updated assistant message
  run() is called, LLM continues from the tool result
  onTurnComplete fires (finishReason = "stop", responseMessage is the FULL merged message)

The AI SDK’s toUIMessageStream automatically reuses the assistant message ID across the pause (we pass originalMessages internally), so responseMessage in the post-resume onTurnComplete is the full merged message — the original text, the completed tool call, and any follow-up content — not just the new parts.

Backend: define the tool

A HITL tool has an inputSchema describing what the model can ask, but no execute function. When the LLM calls it, streamText returns control to your agent.

trigger/my-chat.ts

import { chat } from "@trigger.dev/sdk/ai";
import { streamText, tool } from "ai";
import { openai } from "@ai-sdk/openai";
import { z } from "zod";

const askUser = tool({
  description:
    "Ask the user a clarifying question when you need their input. " +
    "Present 2-4 options for them to pick from.",
  inputSchema: z.object({
    question: z.string(),
    options: z
      .array(
        z.object({
          id: z.string(),
          label: z.string(),
          description: z.string().optional(),
        })
      )
      .min(2)
      .max(4),
  }),
  // No execute function — streamText ends, the frontend supplies the output
  // via addToolOutput, and the next turn continues from the result.
});

export const myChat = chat.agent({
  id: "my-chat",
  run: async ({ messages, signal }) => {
    return streamText({
      model: openai("gpt-4o"),
      messages,
      tools: { askUser },
      abortSignal: signal,
    });
  },
});

Frontend: render the question and collect the answer

Two pieces on the client:

1. UI for the pending tool call — render when the tool part is in input-available state, i.e. the LLM has called the tool but there’s no output yet.
2. Auto-send on resolution — use sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls so answering kicks off the next turn without the user having to hit “send.”

import { useChat, lastAssistantMessageIsCompleteWithToolCalls } from "@ai-sdk/react";
import { useTriggerChatTransport } from "@trigger.dev/react-hooks";

function ChatView({ chatId, accessToken }: { chatId: string; accessToken: string }) {
  const transport = useTriggerChatTransport({ task: "my-chat", accessToken });
  const { messages, sendMessage, addToolOutput } = useChat({
    id: chatId,
    transport,
    sendAutomaticallyWhen: lastAssistantMessageIsCompleteWithToolCalls,
  });

  return (
    <>
      {messages.map((m) =>
        m.parts.map((part, i) => {
          if (part.type === "tool-askUser" && part.state === "input-available") {
            return (
              <AskUserCard
                key={i}
                question={part.input.question}
                options={part.input.options}
                onAnswer={(opt) =>
                  addToolOutput({
                    tool: "askUser",
                    toolCallId: part.toolCallId,
                    output: { optionId: opt.id, label: opt.label },
                  })
                }
              />
            );
          }
          if (part.type === "text") return <Markdown key={i}>{part.text}</Markdown>;
          return null;
        })
      )}
    </>
  );
}

addToolOutput patches the assistant message locally with state: "output-available" and fills in output. lastAssistantMessageIsCompleteWithToolCalls detects that every pending tool call now has a result, and useChat fires a new sendMessage — the backend picks it up as the next turn.

Detecting a paused turn in onTurnComplete

Two ways to detect “this turn paused for user input” vs “this turn finished normally”:

Via finishReason (recommended)

The AI SDK’s finish reason is surfaced on every onTurnComplete event. If the model stopped on tool calls, it’s "tool-calls":

onTurnComplete: async ({ finishReason, responseMessage }) => {
  if (finishReason === "tool-calls") {
    // Turn paused — assistant message has pending tool call(s)
    const pending = responseMessage?.parts.filter(
      (p) => p.type.startsWith("tool-") && p.state === "input-available"
    );
    // Persist as a checkpoint / partial turn
  } else {
    // finishReason === "stop" — normal completion
    // Persist as a completed turn
  }
};

finishReason is only undefined for manual chat.pipe() flows or aborted streams. For the common run() → return streamText(...) pattern it’s always populated.

Via response parts

If you need more nuance (e.g. which specific tool is pending), inspect the parts directly:

function pendingToolCalls(message: UIMessage): string[] {
  return message.parts
    .filter((p) => p.type.startsWith("tool-") && p.state === "input-available")
    .map((p) => p.toolCallId);
}

Both finishReason === "tool-calls" and pendingToolCalls(responseMessage).length > 0 are equivalent in practice. Use finishReason for dispatch, parts for detail.

Persistence: one message vs one record per pause

Because the AI SDK reuses the assistant message ID across the pause, the “same turn” from the user’s perspective maps to two onTurnComplete firings on the server — but both receive a responseMessage with the same id, and the second firing’s responseMessage contains the fully merged content.

Two common persistence patterns:

Overwrite on every turn (simplest)

Just store the latest uiMessages array on every onTurnComplete. The paused-turn write is overwritten by the resume-turn write; the final DB state has the full merged message.

onTurnComplete: async ({ chatId, uiMessages }) => {
  await db.chat.update({
    where: { id: chatId },
    data: { messages: uiMessages },
  });
},

Use this unless you specifically need an audit trail.

Checkpoint nodes (immutable history)

For apps that want every pause point recorded as its own immutable snapshot (branching, replay, diff review), save a checkpoint when paused and a sibling when complete:

onTurnComplete: async ({ chatId, responseMessage, finishReason, uiMessages }) => {
  if (!responseMessage) return;

  if (finishReason === "tool-calls") {
    // Paused — save a checkpoint
    await db.turnCheckpoint.create({
      data: {
        chatId,
        messageId: responseMessage.id,
        parts: responseMessage.parts,
        kind: "partial",
      },
    });
  } else {
    // Completed — save a sibling with the merged full message
    await db.turnCheckpoint.create({
      data: {
        chatId,
        messageId: responseMessage.id,
        parts: responseMessage.parts,
        kind: "final",
      },
    });
  }

  // Always update the canonical chat record for `hydrateMessages` to load
  await db.chat.update({
    where: { id: chatId },
    data: { messages: uiMessages },
  });
};

Both writes see responseMessage.id as the same value — they’re checkpoints of the same logical message. Grouping by messageId + ordering by createdAt gives you the progression.

Multi-pause turns

A single logical turn can pause more than once — the LLM asks question A, gets the answer, thinks, then asks question B before finishing. Each pause fires its own onTurnComplete with finishReason === "tool-calls"; only the last firing has finishReason === "stop". The checkpoint pattern above handles this naturally — each pause adds a new checkpoint sharing the same responseMessage.id.

Gotchas

• Don’t set an execute function on the HITL tool. If it has one, streamText will call it immediately instead of handing control back.
• The frontend must use sendAutomaticallyWhen. Without it, the user has to press Enter after answering — addToolOutput updates local state but doesn’t fire a new turn by itself.
• Don’t mutate responseMessage in onTurnComplete. It’s the captured snapshot. To add custom parts, use chat.response.append() in onBeforeTurnComplete (while the stream is open).
• Stop handling. If the user stops the run while a pause is active (chat.stop() on the transport), onTurnComplete fires with stopped: true and finishReason reflecting the last successful step. Treat stopped paused turns the same as stopped normal turns.