Configuration, Ports, Quotas, and Storage Reference

This page lists local and Docker environment variables, default ports, workspace-related quotas, and storage locations. See Deployment for operational steps.

Environment variables (local mode)

cp .env.example .env
set -a && source .env && set +a
./scripts/run-local.sh

The Ascend NPU Broker is for deployments that need host Ascend devices, and administrators must enable it explicitly. Keep SCIENCE_AGENT_NPU_BROKER=0 when the host has no Ascend NPU, lacks CANN/MindSpore, or should not expose NPU jobs to the Agent; then run_npu_job is absent from the tool table. Enabling it does not change the normal local-mode startup command. Before enabling the Broker, create and verify at least one ScienceDiscovery managed scientific environment with Python that can import the required CANN/MindSpore stack. For built-in NPU workloads, including npu.smoke_test, select environment_id from environment_list; when omitted, the API resolves the Session-selected environment. The API selects that environment's latest Revision for the Broker's internal execution/audit receipt. The Agent cannot select a historical revision: environment_revision_id is rejected. When SCIENCE_AGENT_NPU_WORKLOAD_CONFIG is empty, the built-in allowlist currently contains npu.smoke_test and antibody.protenix.v1. Add models through a custom JSON allowlist with fixed entry points, not arbitrary Agent-supplied commands. SCIENCE_AGENT_NPU_PYTHON is kept only for custom allowlists that explicitly use ${python}; the built-in allowlist uses ${managedPython} and ignores it. Changing the allowlist is equivalent to changing executable host-code entry points and should be reviewed as a deployment change. Model weights, databases, HMMER, CANN, MindScience checkouts, and similar site assets stay outside the repository and are normally referenced through the environment variables or workload configuration above.

The browser stores only the local service access token in local storage. Model credentials stay in backend storage.

Quota levels

These defaults come from services/api/src/workspace-upload.ts, services/runner/src/executor.ts, and .env.example. They have different meanings and are not interchangeable:

In GET /health, workspace.maxFileBytes, maxRequestBytes, and maxWorkspaceBytes report the API file, API request, and runner workspace limits. The endpoint does not report the stdout/stderr limit.

Docker environment variables

Compose reads the root .env (template: .env.docker.example) and interpolates the keys below into docker-compose.yml. They form two layers: the orchestration layer only affects how Compose starts the container; the container layer is forwarded into the container, key by key, through the service's environment block, and an empty value means the built-in default. Procedures and the layering are explained under Docker deployment.

Orchestration layer

Container layer

Values fixed in the image are not changed through .env: SCIENCE_AGENT_DATA_DIR=/app/data, SCIENCE_AGENT_HOST=0.0.0.0, SCIENCE_AGENT_PORT=4310, SCIENCE_AGENT_RUNNER_HOST=127.0.0.1, SCIENCE_AGENT_RUNNER_PORT=4311, SCIENCE_AGENT_RUNNER_URL, and the paths of the baked Python environments, the model catalog snapshot, and the micromamba seed. In the default JiuwenSwarm mode, the adapter listens on 0.0.0.0:4310 inside the container and the API moves to internal :4410. In native mode the API listens on :4310 directly. Runner 4311 remains on container loopback; only 4310 is published. Local mode's SCIENCE_AGENT_MICROMAMBA_BASE_URL is not needed under Docker: the image carries and seeds the pinned micromamba, so nothing is downloaded at run time. Other local-mode variables (such as HTTP_PROXY) are not forwarded; add them to the service's environment block in a docker-compose.override.yml when needed.

Storage layout

Unless overridden, persistent application data is kept in the repository:

The data directory is the only runtime root. SCIENCE_DISCOVERY_DATA_DIR=/srv/science-discovery ./scripts/run-local.sh moves state and service environments together. The former SCIENCE_AGENT_DATA_DIR is still read as a compatibility fallback and produces a log; when both are set, SCIENCE_DISCOVERY_DATA_DIR wins and the choice is logged. For the repository launcher, an existing default data directory is moved once into .sciencediscovery-data. For the single-file launcher, an existing default ./science-discovery-data or the older ./science-agent-data is imported once into ./.sciencediscovery-data, newest first; an existing target is never overwritten and the skip is logged. Deleting the active data directory removes projects, sessions, credentials, and audit records. Under Docker deployment, it is the host ./data bind mount; only service envs/ live in the image. services/paper/.venv and services/gateway/.venv are used only by standalone development or smoke commands.

The single-file payload overrides follow the same naming and precedence rule: use SCIENCE_DISCOVERY_PAYLOAD_CACHE_DIR for the extraction cache or SCIENCE_DISCOVERY_PAYLOAD_DIR for a pre-extracted payload. The corresponding SCIENCE_AGENT_* names remain logged compatibility fallbacks.