Agent SDK

OpenAI SDK

You run the agent loop with the OpenAI SDK; one call turns the model's function calls into the outputs to send back, each one shrunk the moment the tool returns. Both the Responses API and Chat Completions.

Responses API#

Add the expand tool to your tools, and replace the code that runs function calls and builds function_call_output items with one call:

import OpenAI from "openai";
import { createContextwaise } from "contextwaise";
import { expandToolResponses, functionOutputs } from "contextwaise/openai";

const client = new OpenAI();
// Reads CONTEXTWAISE_API_KEY and CONTEXTWAISE_URL from the environment.
const cw = createContextwaise({ agent: "support-triage" });

const input: OpenAI.Responses.ResponseInputItem[] = [{ role: "user", content: question }];
while (true) {
  const response = await client.responses.create({
    model,
    tools: [...myTools, expandToolResponses],
    input,
  });
  input.push(...(response.output as OpenAI.Responses.ResponseInputItem[]));
  if (!response.output.some((item) => item.type === "function_call")) break;

  // Runs your tools, shrinks their outputs, answers expand.
  input.push(...(await functionOutputs(cw, response, runTool, { request: question })));
}

Chat Completions#

The same, with tool messages:

import { expandTool, toolMessages } from "contextwaise/openai";

const messages: OpenAI.ChatCompletionMessageParam[] = [{ role: "user", content: question }];
while (true) {
  const completion = await client.chat.completions.create({
    model,
    tools: [...myTools, expandTool],
    messages,
  });
  const message = completion.choices[0].message;
  messages.push(message);
  if (!message.tool_calls?.length) break;

  // A message doesn't carry usage: pass it so the dashboard can count what each expand cost.
  messages.push(...(await toolMessages(cw, message, runTool, { request: question, promptTokens: completion.usage?.prompt_tokens })));
}

Your tools stay as they are: runTool(name, args) returns a tool's raw output, text or any JSON value. The arguments arrive parsed.

What the call does#

For every function call in the response, in order:

Runs your tool and keeps the raw output on your side.
Shrinks the output for the model: noise removed, the relevant parts kept within the tool's budget. Small results, and results where a cut would save too little (under 1,000 tokens or 30%), pass whole.
Answers expand itself, from the original output. It never reaches your tools.
Reports failures as Error: … text: a tool that throws, or arguments that aren't valid JSON.

Pass request, the user's request for the turn: with the text the model wrote before its calls, it tells relevance what the agent is after.

Good to know#

The Responses API's expand tool is not strict, because its question is optional. Use expandToolResponses there, and expandTool for Chat Completions.
Several function calls in one response run in parallel.
Some models accept function tools on Chat Completions only with reasoning off: if the API refuses your tools, set reasoning_effort to "none" or use the Responses API.

Note

The adapter reads the SDK's response objects (or plain objects); the OpenAI SDK isn't a dependency of contextwaise. All options are on Options; for other loops, see the Core API.