Shared rules for decisions¶

Use a decision step when a support workflow needs an answer from a fixed, typed question. Pick the task first: yes/no, one category, several labels, an ordered level, or distinct requests. Each guide starts with one complete step definition and its expected native result.

The common setup¶

sources select evidence by literal or pointer binding. Text is the default and must resolve to a nonempty string; format: json renders a selected JSON value canonically. Source contents are data and cannot change the authored question.

The shortest form has one question. Its ID and prompt come from the step ID and instructions. It needs type and nonempty criteria, plus fields shown in its focused guide. Put it in a .step.md file with a nonempty body, or in .step.yaml with instructions. A model comes from defaults.model in workflow.yaml or model on the step; see models.

For two or more independent questions, use questions. Each full question requires a unique id, type, prompt, nonempty criteria, and allowedSourceIds, plus type-specific fields. For example:

# config/support_triage/triage/assess.step.yaml
type: decision
sources:
  message:
    pointer: /payload/message
questions:
  - id: queue
    type: choice
    prompt: Which queue owns this email?
    criteria:
      - Choose only a queue supported by the current request.
    allowedSourceIds:
      - message
    options:
      - id: billing
        description: Questions about invoices or charges.
      - id: cancellation
        description: Requests to end a subscription.
  - id: charge_disputed
    type: predicate
    prompt: Does the writer dispute a charge?
    criteria:
      - Return unknown if the message establishes neither true nor false.
    allowedSourceIds:
      - message
instructions: Assess each question independently using only the supplied message.

/steps/assess/result/results/0 is the choice result and /steps/assess/result/results/1 is the predicate result. The runtime puts results in authored question order. A single question stores its native result directly at /steps/<id>/result.

Read the assessment¶

Every result includes questionId, type, answerability, answer, reason, and evidence_strength. reason is nonblank and at most 400 characters. evidence_strength is strong, limited, or JSON null: support for the whole assessment, including abstention. It is not confidence, probability, urgency, or a routing threshold. Strong support can justify an unknown answer.

`answerability.status`	Meaning	Step status
`answerable`	Supported answer	`completed` if every question is answerable
`partially_answerable`	Supported subset; only labeling and request extraction	`needs_review`
`not_answerable`	No supported answer	`needs_review`
`undetermined`	Answerability itself is unsettled	`needs_review`

Any non-answerable status needs at least one answerability.issues code: no_supported_answer (including missing details or outside-catalog requests), conflicting_information, or multiple_valid_options. Explain the particular obstacle in reason; these are the only machine-readable issue codes. An unanswered choice, ranking, labeling, or request extraction has answer: null. A predicate always has an answer object and uses "unknown" for uncertainty. An answerable collection may be empty when the source explicitly supports none.

An unresolved decision stops this flow and uses on_unresolved if configured. Malformed results, unknown categories, inconsistent answerability, timeouts, and provider errors are technical failures. Validation checks shape and membership, not business truth; use reviewed evaluation cases. See flow routing for issue-specific unresolved routes.

Advanced options¶

Field	Use
`model`	Override the workflow default by profile name or supported profile override.
`sources.<id>.format: json`	Supply an object, array, or other JSON value as evidence.
`questions`	Ask 2–64 full questions, each with explicit allowed sources.
`fallback`	For one `choice` question only, attach a separate selection for allowed unresolved issues; see classification.

Write criteria that state what counts, what does not, and when to abstain. A selected category still depends on the model's assessment.