Facility Knowledge#
For the OSPREY agent to be useful at your facility, it has to know your facility: what the subsystems are, how devices are named, which procedures apply, and where any of it sits on the control system. OSPREY keeps that knowledge in three tiers, split by when the agent needs it and what shape the knowledge has.
Always in context – the facility’s core operating context, carried as Markdown rules under the project’s
.claude/rules/directory. These load into the main agent’s context at the start of every session, the same wayCLAUDE.mddoes.Fetched on demand – the Open Knowledge Format (OKF) bundle, a directory of Markdown concept documents served by the
osprey_facility_knowledgeMCP server. Subsystem descriptions, device specs, procedures and physics notes stay out of context until a task calls for them.Queried – the facility graph, which holds the machine’s structure rather than prose: one node per device, the sections devices sit in, the classes they belong to, and the control-system addresses bound to each one. A specialist agent (
facility-knowledge-graph) writes the Cypher, so you ask in plain language.
Each tier is documented separately:
The always-in-context tier – what the demo preset ships
in .claude/rules/ and how to change it.
The on-demand tier – authoring, configuring, and serving the Open Knowledge Format bundle of concept documents.
The queryable tier – the questions the graph answers, how the specialist agent reaches it, and how its answers differ from the channel finder’s.