Examples

Review & auto-fix loop

Reviews uncommitted changes and auto-fixes in a loop.

id: review-uncommitted
steps:
  - id: remediation
    type: loop
    max: 5
    until: ${{ len(steps.review.result.findings) == 0 }}
    steps:
      - id: review
        type: codex
        with:
          kind: review
          target:
            type: uncommitted

      - id: fix
        if: ${{ len(steps.review.result.findings) > 0 }}
        retry:
          max: 3
          delay: 1s
        type: codex
        with:
          kind: turn
          prompt: |
            Address the accepted findings from this review.
            Review:
            ${{ toJSON(steps.review.result) }}

Plan, review, and refine

Drafts a plan, critiques it, refines it, then saves.

id: plan
inputs:
  requirements:
    type: string
  output_path:
    type: string
steps:
  - id: draft
    type: codex
    with:
      kind: turn
      collaboration_mode: plan
      prompt: |
        Clarify anything missing, then draft an implementation plan:
        ${{ inputs.requirements }}

  - id: critique
    type: codex
    with:
      kind: turn
      prompt: |
        Review for gaps:
        ${{ steps.draft.result }}

  - id: refine
    type: codex
    with:
      kind: turn
      prompt: |
        Improve this plan using the critique.
        Draft:
        ${{ steps.draft.result }}
        Critique:
        ${{ steps.critique.result }}

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

Implement with Claude

id: implement
inputs:
  requirements:
    type: string
steps:
  - id: build
    type: claude
    with:
      prompt: |
        Implement the following change in the current repository.

        Requirements:
        ${{ inputs.requirements }}

Parallel checks

Runs multiple shell checks concurrently, then summarizes them with OpenCode.

id: parallel-checks
steps:
  - id: all
    type: parallel
    branches:
      - id: unit
        steps:
          - id: run_unit
            type: shell
            with:
              command: npm test
              stdout:
                mode: text
      - id: lint
        steps:
          - id: run_lint
            type: shell
            with:
              command: npm run lint
              stdout:
                mode: text
    exports:
      unit: ${{ steps.run_unit.result }}
      lint: ${{ steps.run_lint.result }}

  - id: report
    type: opencode
    with:
      prompt: |
        Summarize:
        Unit: ${{ steps.all.result.unit }}
        Lint: ${{ steps.all.result.lint }}

Commit review

id: review-commit
inputs:
  commit_sha:
    type: string
steps:
  - id: review
    type: codex
    with:
      kind: review
      target:
        type: commit
        sha: ${{ inputs.commit_sha }}

Patterns

Review-driven loop

Use review findings as the loop condition:

until: ${{ len(steps.review.result.findings) == 0 }}

When until is omitted, the loop still succeeds and returns reason: max_reached.

Branch passthrough

- else:
  steps: []
  exports:
    value: ${{ steps.earlier.result }}