spatialagents

The in-app AI conversation runtime BioCognia's products share (Plexora, SCIMAP Pro): the streamed model call through the BioCognia AI gateway, a tool loop with deferred tool schemas, approvals for source-file writes and deletes, sub-agents, deterministic compaction, a tool-result cache, offloaded results, the conversation store, the local trace and a terminal loop.

A product binds it with three things and never the other way round:

• a spatialagents.backend.ToolBackend: what its tools are and how one runs (catalog, definition, permission(tool, arguments), invoke, ...);
• a spatialagents.runner.ChatProfile: its prompt content, task names and wording;
• its places on disk and its version, given to each store.

from biocognia.gateway import TokenSource
from spatialagents.conversations import ChatService, ConversationStore
from spatialagents.gateway import GatewayClient
from spatialagents.policy import Policy

service = ChatService(ConversationStore(root / "conversations"),
                      gateway_factory=lambda: GatewayClient(TokenSource(LICENSING)),
                      backend_factory=MyBackend, profile=MY_PROFILE)
started = service.start(Policy())

Dependencies: the standard library, biocognia, pydantic, pyyaml. No mcp, anndata, Flask or numpy, and nothing from either product; import spatialagents imports nothing beyond the standard library.

Tests need no product, no network and no model: spatialagents.testing has FakeGateway (the gateway's wire and a prompt-cache emulator), FakeToolBackend and ScriptedBrain, and registers inert pytest fixtures.

UV_PROJECT_ENVIRONMENT=$HOME/.venvs/spatialagents uv sync --all-groups
BIOCOGNIA_OFFLINE=1 $HOME/.venvs/spatialagents/bin/python -m pytest -q -p no:cacheprovider