Run The Environment In Two Minutes

For anyone who wants an environment to train or evaluate against. If you want to measure an agent on your own schema instead, start at Use your own graph .

Start it

docker run -d -p 8090:8090 --tmpfs /tmp:size=1g dosco/graphjin:env-latest

Nothing is mounted and nothing is generated. The image carries a demo organization — accounts, subscriptions, invoices, support tickets — and a frozen suite of verified tasks built against it.

/tmp must be writable. Each world provisions its own database, discovery cache and artifact store there, so a read-only root filesystem needs --tmpfs or a volume. Size it for pool × (project + database + cache).

Measured 2026-08-31 on Apple silicon laptop, native arm64 image. Re-measure with make env-image-smoke.

Ask it what it is

curl -s localhost:8090/health | python3 -m json.tool

{
  "status": "ready",
  "workers": 2,
  "tasks": 113,
  "dataset": { "catalog_hash": "c4ee2c0d…", "data_anchor": "2026-08-01" },
  "reward_version": "graphjin.eval.reward/v6",
  "reward_profile": "rl",
  "build": { "version": "…", "commit": "…", "image_role": "env" },
  "capabilities": {
    "drive_modes": ["episodes"],
    "writes": true, "resettable": true,
    "suite_source": "public", "suite_fingerprint": "8f312479…",
    "catalog_match": true, "split": "none", "side": "train",
    "freeze_time_source": "build", "boot_ms": 2055
  }
}

Two fields are worth reading before anything else.

catalog_match says the suite’s oracles were verified against the schema being served. If it is false, every episode is being graded against answers computed for a different database. Serving that combination is refused at startup unless you explicitly allow it, so seeing true here is confirmation rather than luck.

freeze_time_source says who pinned the clock. In this image it is build: the clock is pinned to the image’s build date, so one tag asks the same question on any day you run it. Without a pin, a task about “the last 30 days” quietly becomes a different task tomorrow.

Drive one episode

curl -s -X POST localhost:8090/episodes -d '{"slug": "count-accounts"}'

/tasks lists what is available. The response carries the answer, the score vector, and the reward. Without a model configured the episode is graded a failure, which is the correct result — it is what an agent that does nothing earns.

To point it at a policy, give it the four agent variables:

docker run -d -p 8090:8090 --tmpfs /tmp:size=1g \
  -e GJ_AGENT_PROVIDER=openai-compatible \
  -e GJ_AGENT_BASE_URL=http://your-inference-server:8000 \
  -e GJ_AGENT_MODEL=your-checkpoint \
  -e GJ_AGENT_API_KEY_ENV=YOUR_KEY_VAR \
  dosco/graphjin:env-latest

Configure it

Every flag has a GJ_ENV_* variable, because a container is configured by whoever wrote its manifest. A flag wins when it is passed, and a GJ_ENV_ variable the server does not read is a startup error — a typo in a manifest is otherwise indistinguishable from a default.

# a train container and an eval container that agree on the holdout,
# with nothing mounted and no coordination between them
docker run -d -p 8090:8090 --tmpfs /tmp:size=1g \
  -e GJ_ENV_SPLIT=auto:0.8 -e GJ_ENV_SIDE=train dosco/graphjin:env-latest
docker run -d -p 8091:8090 --tmpfs /tmp:size=1g \
  -e GJ_ENV_SPLIT=auto:0.8 -e GJ_ENV_SIDE=eval  dosco/graphjin:env-latest

--split auto derives the division from each task’s content identifier, so two containers reach the same answer independently. Without a split there is no holdout, and /health reports "split": "none" so that is visible rather than assumed.

The full list is in the CLI reference .

Health checks

The image has no shell and no curl, so the binary is the probe:

HEALTHCHECK --interval=10s --timeout=5s --retries=30 \
  CMD ["/ko-app/v3", "env", "health"]

It exits zero only for a server that is actually ready. A server that is up and not yet serving is exactly what a health check exists to catch.

Next

• Driving episodes — three ways, one reward contract.
• Training a policy — and why the workflows have an order.
• Reward and comparability — what to record with a number.