> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trycarhub.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Inference completion callback (internal)

> **Not part of the public API.** Called by the inference plane when a submitted unit of
work settles. Authenticated by an HMAC-SHA256 of the raw body in `X-Modal-Signature`,
keyed with a shared secret; unsigned or stale (> 5 min skew) calls are rejected with
`401`.

The callback is what closes the billing loop: `units_actual` is the real billable count
(the per-detection meter is only knowable here), and the gateway commits that amount and
releases the rest of the reservation in the same transaction. Deliveries are at-least-once
and must be treated as idempotent on `(job_id, status)`.




## OpenAPI

````yaml /openapi.yaml post /internal/callbacks/modal
openapi: 3.1.0
info:
  title: CarHub API
  version: '2026-08-01'
  summary: >-
    Automotive AI as a service — 24 inference endpoints, one asynchronous
    contract.
  description: >
    CarHub exposes automotive computer-vision and pricing models as plain HTTP
    endpoints.

    Every model is priced and rate-limited on its own — nothing requires
    adopting the rest.


    ## Two API surfaces


    | Surface | Prefix | Authentication |

    |---|---|---|

    | Inference (the product) | `/v1/…` | API key — `Authorization: Bearer
    chk_live_…` / `chk_test_…` |

    | Management (the dashboard) | `/mgmt/v1/…` | Short-lived JWT —
    `Authorization: Bearer <jwt>` |


    ## Everything is a job


    Every inference endpoint is `POST` and **asynchronous by default**: it
    answers `202` with a

    [`Job`](#/components/schemas/Job) in status `queued`. Collect the result by
    polling

    `GET /v1/jobs/{id}` or by subscribing to the `job.succeeded` / `job.failed`
    webhooks.


    Fast endpoints additionally accept `?wait=true`, which holds the connection
    until the job

    settles (30 s ceiling) and answers `200` with the same `Job` object in
    status `succeeded`.

    Video, 360 and bundle endpoints reject `wait` with `400 wait_not_supported`
    — their work

    outlives any reasonable request timeout.


    ## Inputs


    Every endpoint accepts two request encodings for the same operation:


    * `multipart/form-data` — the binary directly in the request (`image`,
    `audio`, `video` or
      `images[]` depending on the endpoint). Simplest path, capped at 10 MB per file.
    * `application/json` — `{"input_ref": "upl_…"}` (or `input_refs` for
    multi-asset endpoints)
      referencing an upload created through `POST /v1/uploads`. Required for videos and batches.

    Endpoint parameters are identical in both encodings.


    ## Idempotency


    Send `Idempotency-Key` on any `POST`. A replayed key returns the original
    response and never

    bills twice; reusing a key with a different payload is a `409
    idempotency_key_reuse`.


    ## Test mode


    Keys prefixed `chk_test_` run the entire pipeline — auth, rate limiting, job
    lifecycle,

    ledger, signed webhooks — against a fixture backend instead of a GPU. Test
    jobs carry

    `livemode: false` and draw on a separate, freely rechargeable test credit
    balance.

    Drive the fixture backend per request with `X-Carhub-Test-Scenario`.
  termsOfService: https://trycarhub.com/terms
  contact:
    name: CarHub API support
    url: https://docs.trycarhub.com
    email: api@trycarhub.com
  license:
    name: Proprietary
    identifier: LicenseRef-CarHub-Commercial
servers:
  - url: https://api.trycarhub.com
    description: Production (EU)
security:
  - ApiKeyAuth: []
tags:
  - name: A — Understand the vehicle
    description: >-
      Identity, attributes and a clean capture — everything downstream reads
      from here.
  - name: B — Damage & condition
    description: Find it, then grade it, so your own rules can act on the result.
  - name: C — Parts
    description: >-
      Turns "there is damage" into "there is damage on this component" — what
      makes every downstream estimate defensible.
  - name: D — Money
    description: Valuation and repair cost, computed from what was actually observed.
  - name: E — Image production
    description: >-
      Exposed as raw infrastructure, not a batch retouching service. Call it per
      asset, in your own pipeline.
  - name: Inspections
    description: The eight-stage bundle, billed as a single blended inspection.
  - name: Jobs
    description: Status, results and history of asynchronous work.
  - name: Uploads
    description: Presigned storage for videos and batches.
  - name: Management
    description: >-
      Organisation, API keys, usage, billing and outgoing webhooks. JWT
      authenticated.
  - name: Internal
    description: Inbound callbacks from the inference plane. Never called by customers.
externalDocs:
  description: CarHub developer documentation
  url: https://docs.trycarhub.com
paths:
  /internal/callbacks/modal:
    post:
      tags:
        - Internal
      summary: Inference completion callback (internal)
      description: >
        **Not part of the public API.** Called by the inference plane when a
        submitted unit of

        work settles. Authenticated by an HMAC-SHA256 of the raw body in
        `X-Modal-Signature`,

        keyed with a shared secret; unsigned or stale (> 5 min skew) calls are
        rejected with

        `401`.


        The callback is what closes the billing loop: `units_actual` is the real
        billable count

        (the per-detection meter is only knowable here), and the gateway commits
        that amount and

        releases the rest of the reservation in the same transaction. Deliveries
        are at-least-once

        and must be treated as idempotent on `(job_id, status)`.
      operationId: modalCallback
      parameters:
        - name: X-Modal-Signature
          in: header
          required: true
          description: '`t=<epoch>,v1=<hex hmac-sha256 of "<t>.<raw body>">`.'
          schema:
            type: string
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/ModalCallback'
      responses:
        '204':
          description: Callback accepted.
        '400':
          $ref: '#/components/responses/BadRequest'
        '401':
          $ref: '#/components/responses/Unauthorized'
        '404':
          $ref: '#/components/responses/NotFound'
        '409':
          $ref: '#/components/responses/Conflict'
        '500':
          $ref: '#/components/responses/InternalError'
      security: []
components:
  schemas:
    ModalCallback:
      type: object
      title: Modal completion callback
      description: >-
        Payload posted by the inference plane when a unit of work settles.
        Internal contract, not part of the public API surface.
      required:
        - job_id
        - status
      properties:
        job_id:
          $ref: '#/components/schemas/JobIdString'
        status:
          type: string
          description: Terminal state reached by the inference call.
          enum:
            - succeeded
            - failed
        output_ref:
          type:
            - string
            - 'null'
          description: >-
            Storage key of the serialised result written by the worker. Required
            when `status` is `succeeded`.
          examples:
            - s3://carhub-out/2026/08/16/job_3ZxK9pQr7TnW.json
        units_actual:
          type:
            - integer
            - 'null'
          minimum: 0
          description: >-
            Billable units really consumed. Only the worker knows this for the
            per-detection meter; the gateway commits exactly this amount and
            releases the rest.
        model_version:
          type:
            - string
            - 'null'
          description: Build that produced the output, surfaced on the job.
          examples:
            - plate-reader@2026-08-01
        error:
          description: Populated when `status` is `failed`.
          oneOf:
            - $ref: '#/components/schemas/JobError'
            - type: 'null'
        provider_job_id:
          type:
            - string
            - 'null'
          description: >-
            Identifier on the inference provider's side, for cross-system
            tracing.
        duration_ms:
          type:
            - integer
            - 'null'
          minimum: 0
          description: Wall-clock time of the inference call, fed into latency metrics.
    JobIdString:
      type: string
      description: Job identifier — `job_` followed by a base58 string.
      pattern: ^job_[1-9A-HJ-NP-Za-km-z]{8,32}$
      examples:
        - job_3ZxK9pQr7TnW
    JobError:
      type: object
      description: Why a job settled in `failed` or `expired`.
      required:
        - type
        - code
        - message
      properties:
        type:
          type: string
          enum:
            - processing_error
            - invalid_request
            - internal
        code:
          type: string
          examples:
            - model_inference_failed
            - unreadable_input
            - upstream_timeout
            - job_expired
        message:
          type: string
    ErrorResponse:
      type: object
      title: Error response
      description: >-
        The single error envelope of the API. Every non-2xx response on every
        route group has exactly this shape.
      required:
        - error
      additionalProperties: false
      properties:
        error:
          $ref: '#/components/schemas/Error'
    Error:
      type: object
      required:
        - type
        - code
        - message
        - param
        - doc_url
        - request_id
      properties:
        type:
          type: string
          description: >-
            Broad class of the failure — enough to branch on without matching
            codes.
          enum:
            - invalid_request
            - authentication
            - permission
            - rate_limit
            - insufficient_credits
            - not_found
            - conflict
            - processing_error
            - internal
        code:
          type: string
          description: >-
            Stable machine-readable identifier, narrower than `type`. New codes
            may appear; treat an unknown code as its `type`.
          examples:
            - missing_input
            - invalid_api_key
            - rate_limit_exceeded
        message:
          type: string
          description: >-
            English, human-readable, safe to log. Never localised, never for end
            users.
        param:
          type:
            - string
            - 'null'
          description: >-
            The offending field or header, when the failure can be attributed to
            one.
        doc_url:
          type: string
          format: uri
          description: Documentation page for this exact code.
          examples:
            - https://docs.trycarhub.com/errors/missing_input
        request_id:
          type: string
          description: >-
            Identifier of the failing request, mirrored in the `Request-Id`
            header.
          examples:
            - req_2Kd8xQmPvL4T
  responses:
    BadRequest:
      description: >
        Malformed or unusable request — bad JSON, missing or oversized file,
        unknown parameter,

        `wait=true` on an endpoint that does not support it.
      headers:
        Request-Id:
          $ref: '#/components/headers/RequestId'
      content:
        application/json:
          schema:
            $ref: '#/components/schemas/ErrorResponse'
          examples:
            missingInput:
              summary: No input supplied
              value:
                error:
                  type: invalid_request
                  code: missing_input
                  message: Provide either a multipart `image` field or an `input_ref`.
                  param: image
                  doc_url: https://docs.trycarhub.com/errors/missing_input
                  request_id: req_2Kd8xQmPvL4T
            waitUnsupported:
              summary: wait=true on a long-running endpoint
              value:
                error:
                  type: invalid_request
                  code: wait_not_supported
                  message: >-
                    This endpoint is long-running; poll the job or use a webhook
                    instead.
                  param: wait
                  doc_url: https://docs.trycarhub.com/errors/wait_not_supported
                  request_id: req_2Kd8xQmPvL4T
    Unauthorized:
      description: Missing, malformed, revoked or expired credential.
      headers:
        Request-Id:
          $ref: '#/components/headers/RequestId'
      content:
        application/json:
          schema:
            $ref: '#/components/schemas/ErrorResponse'
          examples:
            invalidKey:
              value:
                error:
                  type: authentication
                  code: invalid_api_key
                  message: The provided API key is invalid or has been revoked.
                  param: null
                  doc_url: https://docs.trycarhub.com/errors/invalid_api_key
                  request_id: req_2Kd8xQmPvL4T
    NotFound:
      description: No such resource, or it is not visible to this credential.
      headers:
        Request-Id:
          $ref: '#/components/headers/RequestId'
      content:
        application/json:
          schema:
            $ref: '#/components/schemas/ErrorResponse'
          examples:
            unknownJob:
              value:
                error:
                  type: not_found
                  code: job_not_found
                  message: No job with id `job_3ZxK9pQr7TnW`.
                  param: id
                  doc_url: https://docs.trycarhub.com/errors/job_not_found
                  request_id: req_2Kd8xQmPvL4T
    Conflict:
      description: >-
        The request contradicts the current state — an idempotency key replayed
        with a different body, or an upload reference already consumed.
      headers:
        Request-Id:
          $ref: '#/components/headers/RequestId'
      content:
        application/json:
          schema:
            $ref: '#/components/schemas/ErrorResponse'
          examples:
            idempotencyReuse:
              value:
                error:
                  type: conflict
                  code: idempotency_key_reuse
                  message: >-
                    This Idempotency-Key was already used with a different
                    request body.
                  param: Idempotency-Key
                  doc_url: https://docs.trycarhub.com/errors/idempotency_key_reuse
                  request_id: req_2Kd8xQmPvL4T
    InternalError:
      description: >-
        Something broke on our side. The request was not billed. Retry with the
        same `Idempotency-Key`; if it persists, quote `request_id`.
      headers:
        Request-Id:
          $ref: '#/components/headers/RequestId'
      content:
        application/json:
          schema:
            $ref: '#/components/schemas/ErrorResponse'
          examples:
            internal:
              value:
                error:
                  type: internal
                  code: internal_error
                  message: An unexpected error occurred.
                  param: null
                  doc_url: https://docs.trycarhub.com/errors/internal_error
                  request_id: req_2Kd8xQmPvL4T
  headers:
    RequestId:
      description: Unique identifier of this request. Quote it in any support conversation.
      schema:
        type: string
        examples:
          - req_2Kd8xQmPvL4T
  securitySchemes:
    ApiKeyAuth:
      type: http
      scheme: bearer
      bearerFormat: CarHub API key
      description: >
        Organisation API key, sent as `Authorization: Bearer chk_live_…`.


        * `chk_live_…` — real inference, real credits, `livemode: true`.

        * `chk_test_…` — identical pipeline against the fixture backend, test
        credits,
          `livemode: false`.

        Keys are stored hashed; only the prefix (`chk_live_a1b2…`) is ever
        displayed again. A key

        may be scoped to a subset of endpoint families — calling outside its
        scopes is a `403`.

````