@hitl-kit/ai-sdk

Vercel AI SDK adapter for HITL Kit. Typed tool() wrappers that return schema-conformant HITL events, so <HitlEventRenderer /> can dispatch them client-side.

pnpm add @hitl-kit/core @hitl-kit/react @hitl-kit/ai-sdk ai zod

ai and zod are peer dependencies; the versions you install are up to you (AI SDK v4+, Zod v3+).

Why this adapter exists

Unlike LangGraph, Vercel AI SDK doesn't have a native interrupt() / Command primitive. The canonical way for an agent to "ask the human something" in AI SDK is to call a tool. This adapter gives you 15 typed tool() wrappers — one per HITL Kit primitive — whose input schemas mirror the primitive's event shape and whose return values are validated HITL events ready to pass to <HitlEventRenderer />.

Flow:

1. Agent calls a HITL Kit tool
2. Tool returns a validated HitlEvent as the tool result
3. Your UI filters tool results and renders them via <HitlEventRenderer />
4. Human acts on the rendered primitive
5. Your app sends the decision back to the agent as a follow-up user message (AI SDK's canonical resume pattern)

Usage

Server: add the tools to a generateText or streamText call

import { generateText } from "ai";
import { hitlCardTool } from "@hitl-kit/ai-sdk";

const result = await generateText({
  model,
  messages,
  tools: {
    requestHumanReview: hitlCardTool({
      description:
        "Request human review of a citation before writing to the draft.",
    }),
  },
});
// If the model calls requestHumanReview, the tool result is a
// validated HitlCardEvent.

Client: render tool results via <HitlEventRenderer />

import { HitlEventRenderer } from "@hitl-kit/react";
import { isHitlToolResult } from "@hitl-kit/ai-sdk";
import { HitlCard } from "@/components/hitl/HitlCard";

// inside your chat UI, after streaming:
toolResults.filter(isHitlToolResult).map((r) => (
  <HitlEventRenderer
    key={r.result.id ?? r.result.kind}
    event={r.result}
    registry={{
      "hitl.card": (ev) => <HitlCard config={...} />,
    }}
  />
));

Or just import everything at once

import { generateText } from "ai";
import { allHitlTools } from "@hitl-kit/ai-sdk";

await generateText({ model, messages, tools: allHitlTools });

Tools (one per primitive)

| Factory | Event kind | Returns | |---|---|---| | hitlCardTool | hitl.card | HitlCardEvent | | subagentStatusTool | subagent.status | SubagentStatusEvent | | miniTraceTool | trace.mini | MiniTraceEvent | | aiGenerationScaleTool | scale.ai_generation | AiGenerationScaleEvent | | contextChipsTool | chips.context | ContextChipsEvent | | qaFlowTool | qa.flow | QAFlowEvent | | writingAgentTool | agent.writing | WritingAgentEvent | | researchAgentTool | agent.research | ResearchAgentEvent | | batchQueueTool | batch.queue | BatchQueueEvent | | searchResultTool | result.search | SearchResultEvent | | approveRejectTool | approval.binary | ApproveRejectEvent | | diffResultTool (v0.6a) | result.diff | DiffResultEvent | | citationResultTool (v0.6a) | result.citation | CitationResultEvent | | editablePlanTool (v0.6a) | plan.editable | EditablePlanEvent | | toolCallPreviewTool (v0.6a) | tool.call | ToolCallPreviewEvent |

Each factory accepts { description?: string } so you can shape the agent's mental model of when to call the tool. The validation happens inside execute via the core Zod schema — malformed tool calls throw before the result ever reaches the UI.

Resume pattern

AI SDK has no Command({ resume }). When the human acts on a primitive, your app sends a new user message back to the agent with the result embedded:

const handleApproval = (approved: boolean) => {
  append({
    role: "user",
    content: approved
      ? "Approved. Continue and write to the draft."
      : "Rejected. Do not write the citation.",
  });
};

The agent reads the new message, understands the decision, and continues. Less elegant than LangGraph's typed resume, but works with any AI SDK agent without special infrastructure.

Part of HITL Kit

• Schemas → @hitl-kit/core
• Renderer → @hitl-kit/react
• LangGraph adapter (different pattern, native interrupts) → @hitl-kit/langgraph
• Components → hitlkit.dev/registry
• Paper → hitlkit.dev/paper

MIT © Ieuan King.