wf-run-step

name	wf-run-step
description	Execute the current weft step via the Claude Code Workflow tool, then auto-transition on the structured verdict. Use only when user types /wf-run-step, or when advancing a step whose template declares an executor.
argument-hint	[--resume]
allowed-tools	["Bash","Read","Workflow"]

Run a Weft Step as a Workflow

For steps whose template declares an executor (a Claude Code Workflow script), this runs the workflow, captures its structured verdict, and pipes the result to weft run-result — which records the Workflow runId on the step (for crash-resume) and auto-transitions: complete if the verdict is non-blocking, fail (per on_fail) if blocking.

Steps with no executor are not for this skill — drive them with /wf-step as normal. The executor is for mechanical steps (review fan-out, multi-target verification, regression sweeps); judgment gates stay human-driven.

1. Read the current step's executor

python3 "${CLAUDE_PLUGIN_ROOT}/core/cli.py" status --json

Find the current step (the one with status: running). Read its executor field:

• executor.type — currently only "workflow" is supported.
• executor.script — script name, resolved to ~/.claude/weft/workflows/<script>.js (or the project-local .claude/weft/workflows/<script>.js if present).
• executor.args — JSON args passed to the workflow.

If executor is null: stop. Tell the user this step has no executor and should be advanced with /wf-step. Do not invent one.

2. Invoke the Workflow

If --resume was passed AND the step already has a non-null workflow_run_id, re-invoke with resumeFromRunId so completed sub-agents return cached (this is the crash-resume path):

Workflow({ scriptPath: "<resolved path>", args: <executor.args>, resumeFromRunId: "<workflow_run_id>" })

Otherwise invoke fresh:

Workflow({ scriptPath: "<resolved path>", args: <executor.args> })

The executor script must return an object that includes at least blocking (boolean) and a human-readable summary. Capture the Run ID from the tool result and the returned object.

3. Record the verdict and auto-transition

Build the payload and pipe it to run-result (it transitions the step — complete if not blocking, fail if blocking):

echo '{"run_id":"<RUN_ID>","blocking":<true|false>,"reason":"<one-line>","verdict":<the returned object>}' \
  | python3 "${CLAUDE_PLUGIN_ROOT}/core/cli.py" run-result

Then report the verdict summary and the new workflow state to the user. If blocking, surface the findings so the user can decide whether to fix-and-rerun (/wf-run-step again) or override.

Notes

• weft run-result stores the runId on the step; a fresh session resuming the workflow can re-run this skill with --resume to continue the executor from its last checkpoint.
• The verdict is recorded in the event log (wf.executor_result) and survives weft rebuild.