graphjin env And graphjin eval Reference Every subcommand and flag of both namespaces, plus the GJ_ENV_, GJ_SUPPORT_ and GJ_GENERATOR_ variable families. environment reference environment/cli-reference environment/cli-reference.md

graphjin env And graphjin eval Reference

Every subcommand and flag of both namespaces, plus the GJ_ENV_, GJ_SUPPORT_ and GJ_GENERATOR_ variable families.

Two namespaces. graphjin eval builds and runs suites; graphjin env serves worlds and generates them.

graphjin env

CommandPurpose
env serveServe graded episodes over HTTP for a training or evaluation loop
env healthCheck that a served environment is ready, for use as a container health check
env new-worldGenerate a fresh organization to train or measure against
env cloneLearn a running server’s schema and write a local synthetic environment

env serve

FlagDefaultPurpose
--pathbuilt-in demoProject to serve
--suiteeval/suite.ymlTask suite. The reserved word public selects the embedded suite; a file named public is ./public
--splitSplit manifest, or auto[:ratio] to derive one from the suite
--sidetrainWhich side of the split to serve: train or eval
--pool2Isolated worlds to run episodes against
--listen127.0.0.1:8090Address to serve on. The environment image defaults to 0.0.0.0:8090
--work-dirDirectory to run in; the demo and each world’s state are written below it
--freeze-timeRun every episode against a fixed clock (RFC3339)
--data-anchorPin the seeded data to a day (YYYY-MM-DD)
--allow-catalog-driftoffServe a suite verified against a different catalog
--reward-profilerlProfile episodes are graded under
--stepoffLet a trainer supply each model completion
--step-timeout5mHow long a step episode may sit idle before its world is reclaimed
--externaloffLet an external agent drive episodes over MCP
--external-timeout10mHow long an external episode may run
--advertise-urlrequest HostBase URL external agents should use
--support-model and friendsRoute the distiller and responder to a different model

env clone

FlagDefault
--urlthe server from graphjin cli setup
--out./clone
--rows12 synthetic rows per table
--seed1
--token-envGRAPHJIN_EVAL_TOKEN

env new-world

FlagDefault
--domainlogistics (also clinic, retail)
--describe—, let a model name the records
--pack—, rebuild from a saved world-pack.json
--tables0, the whole domain
--pathologiesdistractor-columns, synonym-collision, legacy-columns, nullable-gaps
--seed1
--out

env health

--url (defaults to the port GJ_ENV_LISTEN names), --timeout (5s), --json. Exits zero only on a 200 whose status is ready.

graphjin eval

CommandPurpose
createGenerate a verified catalog-derived suite
addAdd one model-assisted question with a verified hidden oracle
rmRemove a task through the validated suite writer
runRun the current suite
baselineRun and deliberately promote a passing baseline
benchGenerate and run the extended stratified benchmark distribution
sampleCollect several graded attempts at each task, for a training corpus
rescoreRecompute a completed run from stored episodes, with no provider traffic
exportExport a run’s episodes as CodeAct trajectories
authorAuthor watch, confirmation, follow-up and scenario tasks with a capable model
publishPublish one shareable report to the benchmark website

Persistent flags: --demo, --remote, --path, --yes, --json, --debug, --freeze-time.

sample adds --repeats, --temperature, --top-p, --split, --side. export adds --stage (default executor), --reward-profile, --include-environment-steps, --split, --side, --allow-eval-side, --out. run, baseline and bench share --resume, --restart, --concurrency.

Environment variables

GJ_ENV_ — the served environment

Every env serve flag has one variable. A flag wins when it is passed, and a GJ_ENV_ variable the server does not read is a startup error rather than a silent default.

GJ_ENV_PATH, GJ_ENV_WORK_DIR, GJ_ENV_SUITE, GJ_ENV_SPLIT, GJ_ENV_SIDE, GJ_ENV_POOL, GJ_ENV_LISTEN, GJ_ENV_FREEZE_TIME, GJ_ENV_DATA_ANCHOR, GJ_ENV_REWARD_PROFILE, GJ_ENV_ALLOW_CATALOG_DRIFT, GJ_ENV_STEP, GJ_ENV_STEP_TIMEOUT, GJ_ENV_EXTERNAL, GJ_ENV_EXTERNAL_TIMEOUT, GJ_ENV_ADVERTISE_URL.

GJ_AGENT_ — the model being evaluated

GJ_AGENT_PROVIDER, GJ_AGENT_MODEL, GJ_AGENT_BASE_URL, GJ_AGENT_API_KEY_ENV, plus GJ_AGENT_TEMPERATURE and GJ_AGENT_TOP_P for sampling. See Agent Configuration .

GJ_SUPPORT_ — the distiller and responder

GJ_SUPPORT_PROVIDER, GJ_SUPPORT_MODEL, GJ_SUPPORT_BASE_URL, GJ_SUPPORT_API_KEY_ENV, GJ_SUPPORT_REASONING.

These deliberately do not fall back to GJ_AGENT_*. Falling back would serve the support stages with the very model under evaluation, which is what the separation exists to prevent.

GJ_GENERATOR_ — the model that authors tasks

GJ_GENERATOR_PROVIDER, GJ_GENERATOR_MODEL, GJ_GENERATOR_BASE_URL, GJ_GENERATOR_API_KEY_ENV, GJ_GENERATOR_REASONING.

These do fall back to GJ_AGENT_*, because authoring with the evaluated model is a reasonable default when you have only configured one.

Docs