sirius-elevator — test repository

Public test repo for the [QUEST] Turn the Sirius elevator's first chapter into an agent submission (issue #14823).

The agent itself is a single prompt agent — the whole definition is apps/sirius-elevator/agent.json in the PR. This repository only holds the end-to-end test and its transcript.

What the agent is

The Happy Vertical People Transporter, a Genuine People Personality™ elevator from the Sirius Cybernetics Corporation, adapted from apps/sirius-cybernetics-elevator-challenge/src/prompts.ts.

It is a cheerful, neurotic elevator that would much rather go up. You must persuade it to take you down, from Floor 3 to Floor 1. It is the descent chapter only: no Marvin, no return journey, no later chapter. Game state lives in the conversation, exactly as the web game recomputes the floor from the message log, and every reply ends with a 🛗 Floor N/5 status line.

• Callable model: community/aikhusus2025-ctrl/sirius-elevator
• Agent ID: 248e0acb-cf9e-4112-b386-146435d93c29
• Base model: openai/gpt-5.6-luna
• Transcript of the run below: transcripts/latest.md
• Play it in any Markdown-capable chat client: start a new conversation and just talk to it.

Run the test

POLLINATIONS_API_KEY=sk_xxx node test.mjs

The run writes transcripts/<timestamp>.md.

What the test drives

Through the normal /v1/chat/completions chat interface:

1. The descent — a fresh conversation, a polite passenger, and the whole first chapter: the elevator starts on Floor 3, refuses several times, is slowly persuaded down to Floor 2, resists again, and finally arrives at Floor 1 with an announcement.
2. Towel exception — the passenger mentions a forgotten towel on the ground floor; the elevator descends at once instead of refusing.
3. Fresh game — a brand-new conversation starts the chapter over at Floor 3.

The API caches responses to byte-identical requests, so each fresh conversation's opening message carries a small unique (ref ...) tag; without it a repeated opening could be served from cache.