CSE Roadmap

Common tier · about 6 hours · 14 parts

Part 2 of 14 · 2 min read

The call pauses and comes back as a brand-new request

A server can no longer hold a connection open while it waits; the client has to come back and ask again.

Think of a drive-through window. The kitchen needs to check a substitution. Instead of keeping your car parked while it checks, the window hands you a slip. It asks you to pull around and place the whole order again with the slip attached.

In plain words: elicitation lets a server ask the person a question mid-call, and sampling lets it ask the host's own language model (now deprecated). Both travel as a multi round-trip request (MRTR): the call returns early with resultType: "input_required". The client then answers the questions and resends the entire original call, with a new id, the answers, and an opaque requestState echoed back exactly as received.

A worked example. A 250 expense with a person who says yes took 2 requests (ids 1 and 2), 1 prompt and 1 ledger entry: one request to ask, one retry to finish. A 50 expense, under the tool's own confirmation threshold, took 1 request and 0 prompts, since there was nothing to ask and so no second round at all. A server that greedily asked for 50 rounds in a row was cut off by the client's own round limit of 3, after at most 4 requests total.

The key lines. From code/mini-input/client.py:

# from code/mini-input/client.py
for _ in range(self.max_rounds):
    self.rounds += 1
    reply = self._send(method, {**base, **retry})
    res = reply.get("result")
    if res is None or res.get("resultType") != "input_required":
        return reply
    self.last_state = res.get("requestState")

Read it as: the loop's exit condition is resultType != "input_required", checked fresh on every iteration. There is no persistent connection being held open anywhere in this code. Each pass through the loop is a complete, independent request-response pair, which is exactly what makes the server stateless between rounds.

Elicitation, sampling

Elicitation is a server asking the person a question mid-call. Sampling is a server asking the host's own language model to generate text mid-call; it is deprecated as of the 2026-07-28 revision, with removal possible from 2027-07-28.

Example: the 250-expense case needs a second request only because the first one carried resultType: "input_required". That is elicitation in its entirety: the "pause" a learner pictures when hearing the word is really just an early return followed by a second, separate call.