Claude Count Tokens Then Create Message

arazzo: 1.0.1
info:
  title: Claude Count Tokens Then Create Message
  summary: Pre-flight a prompt through token counting before sending it to the Messages API.
  description: >-
    A budget-aware completion flow. The workflow first counts the number of
    input tokens a prompt would consume using the count_tokens endpoint, then
    sends the same prompt to the Messages API to generate a response. Counting
    first lets a caller verify the request fits within a model's context window
    and estimate cost before incurring generation charges. Every step spells
    out its request inline — including the required x-api-key and
    anthropic-version headers — so the flow can be read and executed without
    opening the underlying OpenAPI description.
  version: 1.0.0
sourceDescriptions:
- name: claudeApi
  url: ../openapi/claude-messages-api.yml
  type: openapi
workflows:
- workflowId: count-tokens-then-create-message
  summary: Count input tokens for a prompt, then generate a message with the same prompt.
  description: >-
    Counts the input tokens for a model and prompt, then sends the prompt to the
    Messages API, returning both the pre-flight token count and the generated
    message usage.
  inputs:
    type: object
    required:
    - apiKey
    - model
    - prompt
    properties:
      apiKey:
        type: string
        description: Anthropic API key sent in the x-api-key header.
      anthropicVersion:
        type: string
        description: Value for the required anthropic-version header.
        default: '2023-06-01'
      model:
        type: string
        description: The model to count tokens for and to generate the message with.
      prompt:
        type: string
        description: The user prompt to count and send.
      maxTokens:
        type: integer
        description: Maximum number of tokens to generate in the response.
        default: 1024
  steps:
  - stepId: countTokens
    description: >-
      Count the number of input tokens the prompt will consume for the chosen
      model, without creating a message.
    operationId: countMessageTokens
    parameters:
    - name: x-api-key
      in: header
      value: $inputs.apiKey
    - name: anthropic-version
      in: header
      value: $inputs.anthropicVersion
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.model
        messages:
        - role: user
          content: $inputs.prompt
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      inputTokens: $response.body#/input_tokens
  - stepId: createMessage
    description: >-
      Send the same prompt to the Messages API now that its token footprint is
      known, capturing the generated content and usage.
    operationId: createMessage
    parameters:
    - name: x-api-key
      in: header
      value: $inputs.apiKey
    - name: anthropic-version
      in: header
      value: $inputs.anthropicVersion
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.model
        max_tokens: $inputs.maxTokens
        messages:
        - role: user
          content: $inputs.prompt
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      messageId: $response.body#/id
      content: $response.body#/content
      stopReason: $response.body#/stop_reason
      usage: $response.body#/usage
  outputs:
    estimatedInputTokens: $steps.countTokens.outputs.inputTokens
    messageId: $steps.createMessage.outputs.messageId
    content: $steps.createMessage.outputs.content
    usage: $steps.createMessage.outputs.usage