On this page

Projection · 3 levels

The mechanism that lets an agent decide how much of an atom to load. Three levels, kind-aware splitting, deterministic chunker.

The three levels

Level	Token budget	Loaded when	Contains
`summary`	~30 tok	Always — already in `_index.xml`	`description` + tags + 1-line claim
`core`	~150 tok	Retrieved as adjacent / supporting	+ key body fields (claim, applies-to, severity)
`full`	~380 tok	Retrieved as direct hit / focal atom	+ sources, examples, remediation, exceptions

The agent picks level per query via the MCP tool:

prime_query({
  scope: "focus:input-validation",
  level: "core",        // ← summary | core | full
  tags: ["api-endpoint"],
  budget_tokens: 800,
})

Projection model — existence is not content; agent picks summary, core, or full

The chunker is kind-aware

What goes in summary depends on the kind. A rule's summary is its claim; a method's summary is its goal; a fact's summary is its statement. The chunker (packages/compiler/src/chunker.ts) walks the AST per-atom:

// packages/compiler/src/chunker.ts (signature)
export interface ChunkLevels {
  /** ~30 tok: description + tags + 1-line claim */
  summary: string;
  /** ~150 tok: + core body fields */
  core: string;
  /** ~380 tok: + sources + examples + relations + notes */
  full: string;
}

export function chunk(atom: AtomDeclaration): ChunkLevels { ... }

Per-kind splitting rules

Kind	summary	core	full
`rule`	The claim	+ applies-to, severity	+ remediation, exceptions, validates-with
`pattern`	Problem statement	+ solution + structure	+ examples, behaviours
`fact`	Statement	+ confidence + applies-to	+ sources, counter-conditions
`method`	Goal	+ key steps	+ contract, tools, examples
`persona`	Posture	+ voice + style hooks	+ examples, anti-patterns

Compiled artifact layout

compiled/
└── @scope/
    └── kind-name/
        ├── atom.yaml             # metadata only
        ├── graph.yaml            # resolved edges
        └── chunks/
            ├── summary.md        # ~30 tok
            ├── core.md           # ~150 tok
            └── full.md           # ~380 tok

The runtime never reads chunks/. The MCP server returns paths; the agent reads them via its own Read tool.

Why exactly three levels?

Two is too few — agents need a "summary in the index" + "core for retrieval-hits" + "full for focal atoms." Four is unnecessary — the third tier already handles everything that doesn't fit in core. Empirically (~12 corpora compiled across the reference and example Primes) the three-level split has not needed extension.