Anthropic Count Tokens Then Create Message

arazzo: 1.0.1
info:
  title: Anthropic Count Tokens Then Create Message
  summary: Estimate input token usage for a prompt, then send the message only when it fits a budget.
  description: >-
    A cost-aware generation flow that pairs the Token Counting API with the
    Messages API. The workflow first counts the input tokens for a prospective
    request without charging output tokens, branches on whether the count is
    within the caller's token budget, and only then sends the real message to
    create a response. Every step spells out its request inline — including the
    inline x-api-key and anthropic-version headers the specs define — so the
    flow can be read and executed without opening the underlying OpenAPI
    description.
  version: 1.0.0
sourceDescriptions:
- name: tokenCountingApi
  url: ../openapi/anthropic-token-counting-api-openapi.yml
  type: openapi
- name: messagesApi
  url: ../openapi/anthropic-messages-api-openapi.yml
  type: openapi
workflows:
- workflowId: count-tokens-then-message
  summary: Count input tokens for a prompt and conditionally create the message.
  description: >-
    Calls Count Message Tokens to estimate the prospective input token total,
    and only proceeds to Create Message when that total is at or below the
    supplied token budget.
  inputs:
    type: object
    required:
    - apiKey
    - model
    - prompt
    - tokenBudget
    properties:
      apiKey:
        type: string
        description: A valid Anthropic API key sent in the x-api-key header.
      model:
        type: string
        description: The Claude model id to count against and generate with.
      prompt:
        type: string
        description: The user prompt text to evaluate and send.
      tokenBudget:
        type: integer
        description: The maximum acceptable input token count before sending the message.
      maxTokens:
        type: integer
        description: Maximum number of output tokens to generate.
        default: 1024
  steps:
  - stepId: countTokens
    description: >-
      Count the input tokens for the prospective message request without
      creating the message or charging output tokens.
    operationId: countMessageTokens
    parameters:
    - name: x-api-key
      in: header
      value: $inputs.apiKey
    - name: anthropic-version
      in: header
      value: "2023-06-01"
    - name: content-type
      in: header
      value: application/json
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.model
        messages:
        - role: user
          content: $inputs.prompt
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      inputTokens: $response.body#/input_tokens
    onSuccess:
    - name: withinBudget
      type: goto
      stepId: createMessage
      criteria:
      - context: $response.body
        condition: $.input_tokens <= $inputs.tokenBudget
        type: jsonpath
    - name: overBudget
      type: end
      criteria:
      - context: $response.body
        condition: $.input_tokens > $inputs.tokenBudget
        type: jsonpath
  - stepId: createMessage
    description: >-
      Send the prompt as a real message now that the input token total is known
      to be within budget, and capture the generated response.
    operationId: createMessage
    parameters:
    - name: x-api-key
      in: header
      value: $inputs.apiKey
    - name: anthropic-version
      in: header
      value: "2023-06-01"
    - name: Content-Type
      in: header
      value: application/json
    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
      usage: $response.body#/usage
  outputs:
    inputTokens: $steps.countTokens.outputs.inputTokens
    messageId: $steps.createMessage.outputs.messageId
    content: $steps.createMessage.outputs.content
    usage: $steps.createMessage.outputs.usage