generated: '2026-08-11'
method: derived
source: openapi/infer-by-flow7-public-api-openapi.yml (components.schemas, 22 schemas)
description: >-
  The entity graph behind Infer's four operations. It is small and deliberately shallow — there are no
  persistent, addressable resources in the public contract at all. No entity has a GET-by-id
  operation; a Response and its Receipt are returned once and never re-fetchable through the public
  API. Everything else is a read-only projection of pricing and health.
entity_count: 8
persistent_addressable_entities: 0
id_prefixes:
  - prefix: resp_
    entity: Response
    source: example in https://infer.flow7.org/docs
  - prefix: rcpt_
    entity: Receipt (relay.receipt_id)
    source: >-
      https://infer.flow7.org/docs example and the real value rcpt_d32aefa9cbc630fa67934a4b published
      in Route Note 005
  - prefix: pv_
    entity: PriceVersion
    source: >-
      Documented example in /docs. NOTE — the live catalog returns the far longer form
      live-beta-mode-<hex>-<tier>, so pv_ appears to be an illustrative shorthand rather than the
      production format.
  - prefix: req_
    entity: RequestId (X-Request-ID)
    source: observed live on GET /api/public/status — req_6c40bad6e6d044e6b58ce18f
entities:
  - name: PublicCatalog
    schema: PublicCatalog
    operation: getPublicCatalog
    kind: document
    key: none (singleton document)
    fields: [currency, unit, environment, simulated, live_available, models, disclosure]
  - name: CatalogModel
    schema: CatalogModel
    kind: value
    key: slug
    fields: [slug, name, kind, environment, tiers, runtime]
    note: kind is "exact" or "dynamic"; 22 models present at probe time.
  - name: CatalogTier
    schema: CatalogTier
    kind: value
    key: selector
    fields:
      - selector
      - available
      - privacy_modes
      - input_per_million_usd
      - cached_input_per_million_usd
      - cache_write_per_million_usd
      - output_per_million_usd
      - request_usd
      - minimum_margin
      - target_contribution_margin
      - retry_budget
      - price_version
  - name: CatalogRuntime
    schema: CatalogRuntime
    kind: value
    note: >-
      Per-model capability block observed live — status, p95_latency_ms, max_context_tokens,
      max_output_tokens, tools, structured_outputs, vision, streaming, responses_api. At probe time
      22/22 models advertised tools, streaming and responses_api; 21/22 advertised structured_outputs.
  - name: PublicStatus
    schema: PublicStatus
    operation: getPublicStatus
    kind: document
    key: none (singleton document)
    fields: [environment, simulated, updated_at, status, services, note]
  - name: PublicServiceStatus
    schema: PublicServiceStatus
    kind: value
    key: service
    note: One row per published model service, with status and p95_latency_ms.
  - name: Model
    schema: Model
    operation: listModels
    kind: resource projection
    key: id (the selector)
    fields: [id, object, owned_by, relay]
    note: owned_by is a const "infer" — the upstream owner is never named.
  - name: Response
    schema: ResponseObject
    operation: createResponse
    kind: transient resource
    key: id (resp_...)
    retrievable: false
    fields: [id, object, created_at, status, model, output, output_text, usage, relay]
relationships:
  - from: PublicCatalog
    to: CatalogModel
    type: has_many
    via: models[]
  - from: CatalogModel
    to: CatalogTier
    type: has_many
    via: tiers{} (keyed by wire tier — economy, balanced, priority, official)
  - from: CatalogModel
    to: CatalogRuntime
    type: has_one
    via: runtime
  - from: PublicStatus
    to: PublicServiceStatus
    type: has_many
    via: services[]
  - from: PublicServiceStatus
    to: CatalogModel
    type: belongs_to
    via: service == CatalogModel.slug
    confidence: high
    evidence: >-
      All 22 service names in the live status response match the 22 model slugs in the live catalog
      exactly. This join is not declared in the spec but is the join a client must make to answer
      "is the selector I want callable right now".
  - from: Model
    to: ModelPrice
    type: has_one
    via: relay
  - from: Model
    to: CatalogTier
    type: belongs_to
    via: Model.id == CatalogTier.selector
    confidence: high
  - from: CreateResponseRequest
    to: CatalogTier
    type: belongs_to
    via: model (selector) == CatalogTier.selector
    confidence: high
  - from: CreateResponseRequest
    to: RelayRequestOptions
    type: has_one
    via: relay
  - from: Response
    to: ResponseUsage
    type: has_one
    via: usage
  - from: Response
    to: RelayResponseMetadata
    type: has_one
    via: relay
  - from: RelayResponseMetadata
    to: Receipt
    type: belongs_to
    via: receipt_id
    confidence: medium
    note: >-
      The Receipt is referenced by id but has no schema and no operation in the public contract. It is
      addressable only inside the console and through the local receipt inspector at
      /tools/receipt-inspector.
  - from: RelayResponseMetadata
    to: CatalogTier
    type: belongs_to
    via: price_version
    confidence: high
    note: The locked price that governs the charge; the same string appears on the catalog tier.
  - from: Response
    to: Session
    type: belongs_to
    via: relay.session_id / X-Session-Affinity
    confidence: medium
    note: >-
      A Session is not a schema — it is a client-supplied opaque string that pins route affinity. It
      has no lifecycle, no expiry and no retrieval operation in the public contract.
  - from: ErrorEnvelope
    to: ErrorDetail
    type: has_one
    via: error
observations:
  - >-
    The public data model has no customer, workspace, key, wallet, transaction or receipt entity —
    all of which demonstrably exist in the product (they appear in the docs, the Terms and Route Note
    005's ledger reconciliation). Everything commercial lives behind the console, not the API.
  - >-
    The requested `model` and the served class are deliberately different fields. Response.model
    echoes the requested selector; relay.resolved_model_class names what ran. For a dynamic selector
    like infer/auto/code:stable these differ by design.
  - >-
    relay.tier uses the wire vocabulary (economy | balanced | priority | official) while the selector
    uses the customer vocabulary (:low-cost | :balanced | :stable | :official). Two of the four do
    not match by string.
render: none