Route
The route is the product. What the row holds, which fields each autonomy mode requires, how versions and active status work, and how decide picks one.
Capability
The two capability kinds a manifest can reference: domain HTTP and agent start. Why retrieve, prompts, and MCP do not mint a third kind.
Prompts
Prompt packs and llm_role: the text the LLM actually sees, who owns it, and how hydrate stamps one llm_prompt per graph node.
Schema
Capability input_schema and output_schema: which one Runtime binds on LLM stages, and what the catalogue only documents.
Retrieval
Deterministic prefetch, retrieve tools, and long-term retrieve-only. Three mechanisms that are easy to conflate and must not be.
Memory
Memory profiles are route policy, not a store. What Agent Runtime honours today (working, loop) versus what the catalogue only records.
LLM
How Agent Runtime calls the model: one global backend, four llm_role values, structured output binding, and the retry envelope around every call.
Audit
The evidence chain: what every box emits, what the envelope allows, and why an author cannot configure any of it.
Observability
What the fabric emits about your route without being asked, which parts of it your authoring choices shape, and where the runbook lives.
Corpus
A corpus is a registered search gateway. What the row holds, how retrieval.scope references it, and exactly what Runtime does with the chunks.