04 Schema

One schema, every language

The wire is validated by a single JSON Schema (Draft 2020-12): contextgraph-envelope.schema.json. The root validates one envelope; every payload shape is also exposed under $defs for granular validation.

There is no separate IDL — the contextgraph-types Rust structs serialized by serde are the protocol, and the schema mirrors them exactly. It validates with ajv, Python jsonschema, Rust jsonschema, Go gojsonschema — any Draft 2020-12 validator. Find it in the repository under schema/.

ContextFrame

Table 2 — ContextFrame fields. ● = required (what the reference serializer always emits).
Field	Type	Notes
● id	string	Non-empty, provider-scoped.
● kind	FrameKind	`snippet \| symbol \| fact \| doc \| memory \| episode \| graph`
● title	string	Non-empty.
● score	number	Relevance in [0, 1].
● token_cost	u32	Canonical: `ceil(utf8_bytes(content)/4)`.
content	string?	Untrusted data. Governed per representation — omitted entirely for `reference` frames.
representation	enum?	`full \| compact \| reference`; absent ⇒ `full`.
content_digest	string?	`sha256:` + 64 hex. Omitting it makes the frame unverifiable.
uri	string?	The frame’s own URI. A query anchor equal to it anchors the frame.
content_ref	object?	`{ provider_id, uri, expires_at? }` — opaque resolver handle.
canonical_content_hash	string?	Hash of the full canonical source.
content_fidelity	enum?	`exact \| normalized \| summarized \| omitted`
transform	object?	`{ method, implementation, version }` for `compact` renderings.
minimum_content_fidelity	enum?	The lowest `content_fidelity` accepted where the frame is used.
inline_content_requirement	enum?	`required \| resolvable_reference_allowed`: whether the frame must carry its content inline.
canonical_token_cost	u32?	Cost of the canonical source; `tokenizer_ref` names the tokenizer (e.g. `openai:o200k_base`).
valid_from / valid_to	timestamp?	Bi-temporal validity bounds.
recorded_at	timestamp?	When the frame was captured.
provenance	Provenance[]	Omitted when empty.
citation_label	string?	Non-empty when present.
embedding	object?	`{ fingerprint, vector? }`
relations	Relation[]	The graph: labelled edges. Omitted when empty.

A real conforming frame — from the golden fixturesjson

{
  "id": "frame:minimal",
  "kind": "fact",
  "title": "Minimal conforming frame",
  "content": "Default arrays are omitted on the wire.",
  "score": 1.0,
  "token_cost": 10,
  "citation_label": "minimal fixture"
}

ContextQuery

Table 3 — ContextQuery fields.
Field	Type	Notes
● goal	string	The driving task — verbatim intent, not keywords.
● max_frames	u32	Hard cap on returned frames.
● max_tokens	u32	A hard contract, not a hint.
kinds	FrameKind[]	Absent or empty means “your best frames of any kind.”
anchors	string[]	Focal URIs, such as open files and mentioned symbols. Absent means none.
query_text	string?	Optional keyword rendering of the goal.
embedding	number[]?	Only sendable when host and provider fingerprints match exactly (E1).
as_of	timestamp?	Pin retrieval to an instant.
representation_preferences	enum[]?	Ordered; absent ⇒ `[full]`.

Provenance and Relation

A Provenance record requires a type (the Rust field kind serializes to type on the wire) and may carry uri, range, digest (pattern ^sha256:[0-9a-f]{64}$), method, and by. A Relation requires rel and target_uri, with an optional display_name.

Structural rules the schema enforces: required fields, score ∈ [0,1], non-empty title/citation_label/id, the frame-kind enum, the contextgraph/MAJOR.MINOR(-draft)? version pattern, u32 ranges, and the per-representation content/content_ref/transform conditionals.