Steps

shell

Runs a shell command.

- id: test
  type: shell
  with:
    command: npm test
    stdout:
      mode: text # none | text | json (default: text)

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

codex

Two kinds are supported: turn and review.

turn

Normal Codex turns. Set collaboration_mode: plan to use Codex’s Plan collaboration mode (planning-only, no repo mutations).

- id: implement
  type: codex
  with:
    kind: turn
    prompt: Add input validation.
    model: gpt-5.4

- id: draft_plan
  type: codex
  with:
    kind: turn
    collaboration_mode: plan
    prompt: Clarify the scope and produce an implementation plan.
    model: gpt-5.4

steps.<id>.result is always the final plain-text response for turns.

review

Scans diffs and returns review findings.

- id: review
  type: codex
  with:
    kind: review
    model: gpt-5.4
    target:
      type: uncommitted # uncommitted | base | commit

Target variants:

• type: uncommitted
• type: base with branch
• type: commit with sha

Review 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

claude

Runs Claude Code through the local Claude CLI and returns the final plain-text assistant response.

- id: build
  type: claude
  with:
    prompt: Implement the requested change.
    model: claude-opus-4-6
    effort: high
    max_turns: 8
    max_thinking_tokens: 12000
    permission_mode: accept_edits

Required field:

• with.prompt

Optional fields:

• with.model
• with.effort
• with.max_turns
• with.max_thinking_tokens
• with.permission_mode

Claude YAML stays snake_case and maps directly to Claude Agent SDK options in camelCase. Do not set with.kind, with.mode, or with.action on Claude steps. Claude steps require a locally installed and authenticated Claude Code CLI. Run claude login before use.

opencode

Runs OpenCode through the local OpenCode CLI and returns the final plain-text assistant response.

- id: analyze
  type: opencode
  with:
    prompt: Analyze the architecture.

- id: build
  type: opencode
  with:
    prompt: Implement the requested change.
    model: anthropic/claude-sonnet-4-20250514
    agent: build
    variant: high
    permission_mode: auto_approve

Required field:

• with.prompt

Optional fields:

• with.model
• with.agent
• with.variant
• with.permission_mode

OpenCode uses named agents and defaults with.agent to build. with.model must use provider/model. with.permission_mode supports default and auto_approve. Headless runs reject permission requests by default unless with.permission_mode is auto_approve. OpenCode steps require a local OpenCode CLI installation. Verify with opencode --version.

cursor

Runs Cursor Agent through ACP and returns the final plain-text assistant response.

- id: triage
  type: cursor
  with:
    mode: agent # agent | ask | plan (default: agent)
    model: composer-2
    prompt: Summarize the diff.

cursor is ACP text-only. If you need structured JSON from Cursor, run Cursor headless through a shell step and parse that shell output there.

write_file

Writes content to a file.

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

Result access

• codex turns: steps.draft.result
• codex review: steps.review.result.findings
• claude: steps.build.result
• opencode: steps.build.result
• cursor: steps.triage.result