Replicate Run a Prediction with Bounded Wait and Cancel

arazzo: 1.0.1
info:
  title: Replicate Run a Prediction with Bounded Wait and Cancel
  summary: Create a prediction, poll a bounded number of times, and cancel it if it has not finished.
  description: >-
    A guardrail pattern for cost and latency control. The workflow creates a
    prediction and polls it a bounded number of times; if the prediction reaches
    a terminal state it ends, but if it is still running when the poll budget is
    exhausted it cancels the prediction to stop billing for a stuck run. Every
    step spells out its request inline so the flow can be read and executed
    without opening the underlying OpenAPI description.
  version: 1.0.0
sourceDescriptions:
- name: cancelApi
  url: ../openapi/replicate-cancel-api-openapi.yml
  type: openapi
- name: predictionsApi
  url: ../openapi/replicate-predictions-api-openapi.yml
  type: openapi
workflows:
- workflowId: predict-with-timeout-cancel
  summary: Create a prediction, poll within a bounded budget, and cancel it if still running.
  description: >-
    Submits a prediction, polls it a limited number of times, and on exhausting
    the poll budget while still starting or processing, cancels the prediction
    and reports the canceled state.
  inputs:
    type: object
    required:
    - apiToken
    - version
    - input
    properties:
      apiToken:
        type: string
        description: Replicate API token used as a Bearer credential.
      version:
        type: string
        description: The ID of the model version that you want to run.
      input:
        type: object
        description: The model's input as a JSON object matching the version's input schema.
  steps:
  - stepId: createPrediction
    description: >-
      Create a prediction for the supplied model version and input.
    operationId: predictions.create
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    requestBody:
      contentType: application/json
      payload:
        version: $inputs.version
        input: $inputs.input
    successCriteria:
    - condition: $statusCode == 201
    outputs:
      predictionId: $response.body#/id
  - stepId: getPrediction
    description: >-
      Retrieve the prediction state, retrying a bounded number of times. If the
      budget is exhausted while still running, control falls through to the
      cancel step.
    operationId: predictions.get
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    - name: prediction_id
      in: path
      value: $steps.createPrediction.outputs.predictionId
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      status: $response.body#/status
      output: $response.body#/output
    onSuccess:
    - name: predictionDone
      type: end
      criteria:
      - context: $response.body
        condition: $.status == "succeeded" || $.status == "failed" || $.status == "canceled"
        type: jsonpath
    - name: keepPolling
      type: retry
      retryAfter: 2
      retryLimit: 10
      stepId: getPrediction
      criteria:
      - context: $response.body
        condition: $.status == "starting" || $.status == "processing"
        type: jsonpath
  - stepId: cancelPrediction
    description: >-
      Cancel the prediction because it did not finish within the bounded poll
      budget, stopping further compute usage.
    operationId: predictions.cancel
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiToken
    - name: prediction_id
      in: path
      value: $steps.createPrediction.outputs.predictionId
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      canceledStatus: $response.body#/status
  outputs:
    predictionId: $steps.createPrediction.outputs.predictionId
    finalStatus: $steps.getPrediction.outputs.status
    output: $steps.getPrediction.outputs.output
    canceledStatus: $steps.cancelPrediction.outputs.canceledStatus