Saved Workflow

Saved one-shot workflows live in the target repository:

.orca/workflows/<name>.ts

They are self-executing flow scripts. Run them through the Orcats CLI:

bunx -p @twelvehart/orcats orcats --backend codex .orca/workflows/my-workflow.ts -- "task input"

Use flowArgs() for user input:

import { flow, flowArgs } from "@twelvehart/orcats";

const task = flowArgs().join(" ");

await flow()(async () => {
  if (task.length === 0) {
    throw new Error("Pass a task after --");
  }
});

For mutating work, a saved workflow should gate changes with the target repo’s real test and lint commands. The orcats-author skill refuses to emit an ungated mutating workflow.

Generated mutating workflows default to baseline policy repair: they require a clean worktree, run the baseline gates, and repair red gates before the main workflow stage. Use --baseline=strict to fail immediately on red baseline gates, or --baseline=accept-dirty to explicitly accept a dirty worktree with a snapshot. The helper default is .orca/baselines/; commit/sweep templates may use the repo’s git dir so snapshots stay out of commits.

bunx -p @twelvehart/orcats orcats --backend codex .orca/workflows/my-workflow.ts -- --baseline=strict
bunx -p @twelvehart/orcats orcats --backend codex .orca/workflows/my-workflow.ts -- --baseline=accept-dirty
ORCA_BASELINE_POLICY=strict bunx -p @twelvehart/orcats orcats --backend codex .orca/workflows/my-workflow.ts

The CLI arg wins over ORCA_BASELINE_POLICY. accept-dirty records git status --porcelain=v1, staged and unstaged diffs, untracked files, and initial gate output before any backend repair turn.

Use a loop module instead when the artifact should be discoverable with orcats loops, run with orcats run, or served as a long-lived trigger.