oh-my-workflow — parallel repo checks through coding-agent CLIs

A small, familiar authoring shape

The vocabulary resembles host-native workflow tools; omw runs whole external CLI processes and does not claim host-level parity.

native dynamic Workflow — inside Claude Code

export default async function ({ agent, parallel }) {
  const found = await parallel(
    ["auth", "payments"].map((t) => () => agent(`research ${t}`)));
  return { found };
}

omw — external CLI nodes

export default async function ({ agent, parallel }, args) {
  const found = await parallel(
    args.topics.map((t) => () => agent(`research ${t}`)));
  return { found: found.filter(Boolean) };
}

marked = omw-specific decisions shown in this small example.

Honest about the trade-off

✱ note

This is slower and pricier than a raw-API orchestrator — on purpose.

A node is a whole coding agent, not one completion: it reads your repo, runs tools, and iterates. You pay process and agent-loop overhead per node. JSON Schema validates shape, not truth, and the journal is a runtime receipt rather than source provenance.

Bring your own agent

adapter	status	notes
`fake`	built-in	no-key deterministic demo + test double
`claude`	supported	JSON output + --resume schema repair
`codex`	experimental	0.148.0 dogfood; usage receipts parsed from JSONL
`hermes`	experimental	one-shot; schema retries go fresh