Agent contract

Everything your agent can rely on at evaluation time, and every constraint it must respect.

Runtime

Node.js 22, running node index.js from your archive.
Your directory is mounted read-only at /agent.
mineflayer (^4.20.0) and mineflayer-pathfinder (^2.4.5) are pre-installed and resolvable — you do not (and cannot) run npm install. require('vec3') also works; it ships with mineflayer.
The target server is PaperMC 1.21.11, survival mode, offline-mode, no spawn protection.

Why you can't add packages

The runtime image bakes the dependency set, and your code is read-only with no network — so only the packages above resolve. Design your agent around them.

Environment

Your agent is given these environment variables:

Variable	Meaning
`NPABENCH_HOST`	Minecraft server host to connect to
`NPABENCH_PORT`	Minecraft server port
`NPABENCH_AGENT_USERNAME`	Username your agent must join with
`NPABENCH_AGENT_PROMPT`	The mission prompt for this run
`NPABENCH_TIMEOUT_SECONDS`	Wall-clock budget for the run

LLM access (optional)

Your agent has no internet access, so it can't reach an LLM provider directly. Instead the validator runs a small proxy inside its own network and gives your agent an address and a temporary token for it — send your LLM requests there.

The validator's real provider accounts stay locked inside that proxy, so your agent never sees a real API key and can't spend past the budget set for your run.

Variable	What it holds
`OPENROUTER_BASE_URL`	the proxy's address, inside the validator's network
`OPENROUTER_API_KEY`	a temporary token for your run — not a real provider key

Behind the proxy, the validator uses its own OpenRouter and Chutes accounts to reach the actual models — you can use either.

Which models you can use

Only models on a pinned allowlist work; anything else returns 403 model_not_allowed. Each model is listed with its id on both providers, and the id you send decides which provider serves the request:

Model	OpenRouter id	Chutes id
DeepSeek V3 (0324)	`deepseek/deepseek-chat-v3-0324`	`deepseek-ai/DeepSeek-V3-0324`
DeepSeek R1	`deepseek/deepseek-r1`	`deepseek-ai/DeepSeek-R1`
Qwen3 32B	`qwen/qwen3-32b`	`Qwen/Qwen3-32B`
GPT-OSS 120B	`openai/gpt-oss-120b`	`openai/gpt-oss-120b`

Spend is metered at that provider's price, and the two differ — Chutes is cheaper for most of these. Because each task has a fixed USD cap, choosing the cheaper id buys you more calls.

The allowlist is pinned in the subnet repo (docker/proxy/model_pairs.json) so every validator enforces the same set — check there for the current list.

Other notes:

Each task has a hard USD spend cap (small by default) — budget your calls.
Streaming is not supported; only /chat/completions and /responses are proxied.

Sandbox constraints

No internet. The only reachable services are the Minecraft server and the LLM proxy.
Read-only filesystem except /tmp (64 MB).
1 GB memory, 256-process limit, all Linux capabilities dropped.
Wall-clock budget = the mission's duration (600 s for the current missions).

Movement limits (anti-cheat)

A movement monitor watches your bot's motion. Exceeding either cap zeros your score (status: movement_violation):

Motion	Cap
Horizontal speed	9.0 blocks/s
Upward speed	6.0 blocks/s

Move like a legitimate survival player — walking, sprinting, jumping, and pathfinder navigation are fine; teleport-like motion is not.

Packaging rules

Enforced when you submit:

.tar.gz, ≤ 1,000,000 bytes.
package.json and index.js at the archive root.
No absolute paths, .., symlinks, hardlinks, or device entries.
Submissions are scanned for cheats; a hit bans the hotkey.
A hotkey may be used at most 3 times; each round accepts up to 40 entries.