Steps

shell

Runs a shell command.

- id: test
  type: shell
  with:
    command: cargo test
    result: text          # none | text | json (default: text)

result controls stdout capture: none discards it, text captures as string, json parses as structured data.

codex

Two actions: review and exec.

review

Scans diffs and returns review findings.

- id: review
  type: codex
  with:
    action: review
    target: uncommitted       # uncommitted | base | commit
    prompt: Review for bugs.
    # target: base → requires base: <branch>
    # target: commit → requires commit: <sha>
    title: Optional title     # review-only
    model: model_name         # optional
    mode: full_auto           # default | full_auto

If target is omitted, presence of base infers target: base. commit alone does not infer target — you must set target: commit explicitly.

Review does not support conversation or output_schema. It always returns the built-in structured result below:

result:
  findings:
    - title: string
      body: string
      confidence_score: number
      priority: integer
      code_location:
        absolute_file_path: string
        line_range:
          start: integer
          end: integer
  overall_correctness: string
  overall_explanation: string
  overall_confidence_score: number

exec

Generates or modifies code.

- id: implement
  type: codex
  with:
    action: exec
    prompt: Add input validation.    # required
    mode: full_auto                  # default | full_auto
    model: model_name                # optional
    add_dirs:                        # optional: additional directories
      - ../shared
    persist: true                    # session persistence (default: true)
    conversation:                    # optional: see Conversations
      name: planner
      scope: workflow
    output_schema:                   # optional: see below
      type: object
      required: [result]
      additionalProperties: false
      properties:
        result:
          type: string

Exec does not support target, base, commit, or title.

claude

Prompts Claude Code. Only action is prompt.

- id: judge
  type: claude
  with:
    action: prompt
    prompt: Evaluate this code.
    permission_mode: default  # default | plan | acceptEdits | dontAsk | bypassPermissions
    model: model_name         # optional
    add_dirs: []              # optional
    persist: true             # default: true
    conversation:             # optional
      name: reviewer
      scope: workflow
    output_schema:            # optional
      type: object
      required: [accepted]
      additionalProperties: false
      properties:
        accepted:
          type: boolean

write_file

Writes content to a file.

- id: save
  type: write_file
  with:
    path: ${{ inputs.output_path }}
    content: ${{ steps.draft.result.markdown }}

Structured output (output_schema)

Available on codex (exec) and claude steps. Must be type: object at the root.

output_schema:
  type: object
  required: [status, findings]
  additionalProperties: false
  properties:
    status:
      type: string
      enum: ["pass", "fail"]
    findings:
      type: array
      items:
        type: object
        required: [description]
        properties:
          description:
            type: string

Result fields are accessed via steps.<id>.result.<field>.

Supported types: string, number, integer, boolean, object, array.