step

name	step
user-invocable	false
description	Run ONE prompt once and capture its structured output — the single LLM/work unit between /agentflow:reduce (one agent over N inputs) and /agentflow:foreach (one agent per item). Pick where it runs with `--runtime`: `main` (inline), `subagent` (one Agent), or sessionless CLIs `claude-cli` (`claude -p`) / `codex-cli` (`codex exec`) that the engine executes itself. This is how a workflow step can be an arbitrary skill, an MCP-using agent, `claude -p`, or a different model (e.g. cross-model). USE for a single ad-hoc step: "run this one prompt", "ask codex to …", "have a second model review this", "do this one thing and capture the result". For per-item work use /agentflow:foreach; to synthesize many inputs use /agentflow:reduce.
allowed-tools	Bash, Read, Write, Edit, Agent
argument-hint	--prompt "<text>" [--prompt-file <path>] --runtime main|subagent|claude-cli|codex-cli [--model …] [--subagent-type <name>] [--input <file>]

/agentflow:step

Make it visible: say in one line that you're running a step (id + runtime).

1. Init

node "${CLAUDE_PLUGIN_ROOT}/dist/state/step.js" init <id> \
  (--prompt "<text>" | --prompt-file <path>) --runtime <main|subagent|claude-cli|codex-cli> \
  [--model <…>] [--subagent-type <name>] [--input <file>] [--force]

2. Execute — branch on config.runtime

• claude-cli / codex-cli — the engine runs it sessionlessly; you don't dispatch anything:

  node "${CLAUDE_PLUGIN_ROOT}/dist/state/step.js" run <id>
  

  It spawns claude -p <prompt> --output-format json (or codex exec <prompt> --json), captures the result, and marks the run done/failed. Binaries must be installed + logged in; override with $STEP_CLAUDE_BIN / $STEP_CODEX_BIN. (Validate the exact flags against your installed CLIs.)
• subagent — dispatch one Agent (subagent_type + model from config) with the step's prompt (and --input file if any); tell it to return the result. Then record it:

  node "${CLAUDE_PLUGIN_ROOT}/dist/state/step.js" complete <id> --output "<agent result>"
  # or, if the agent wrote a file: ... complete <id> --output-path <file>
  

• main — do the work inline in this thread, then complete <id> --output "<result>".

On failure: node "${CLAUDE_PLUGIN_ROOT}/dist/state/step.js" fail <id> --error "…".

3. Result

status <id> reports output_pointer — a file holding the step's output, consumable by a later stage ({{stages.<name>.result_pointer}} in a workflow) or a /agentflow:reduce.

In a workflow (WORKFLOW.md)

A ## name · step stage runs a step; bullets become flags:

## critique · step
- runtime: claude-cli
- model: opus
- prompt: Critique the draft at {{stages.draft.result_pointer}}; list concrete weaknesses.

Two step stages with different --model inside a repeat/until loop give you a cross-model conversation (model A proposes, model B critiques) — adversarial or cooperative — with the loop predicate as the deterministic convergence check.