workflow-kit

name	workflow-kit
description	Use Agent Workflow Kit workflow commands in Grok Build.

Workflow Kit

Use plugin commands, CLI tool calls, or script calls to run workflows. Keep intermediate transcripts in artifacts.

CLI fallback:

agent-workflow-kit workflow "<task>" --json
agent-workflow-kit workflow-run no-write-probe --json
agent-workflow-kit workflow-run <workflow> --args-json '{"key":"value"}' --json
agent-workflow-kit workflow-run <workflow> --stream --json
agent-workflow-kit workflow-status <run-id> --json
agent-workflow-kit workflow-status <run-id> --tree
agent-workflow-kit workflow-events <run-id> --json
agent-workflow-kit workflow-events <run-id> --follow
agent-workflow-kit workflow-resume <run-id> --json
agent-workflow-kit workflow-stop <run-id> --json
agent-workflow-kit workflows --json
agent-workflow-kit workflows --watch
agent-workflow-kit deep-research "<question>" --json
agent-workflow-kit ultracode on --json
agent-workflow-kit ultracode off --json
agent-workflow-kit ultracode status --json

Saved workflow scripts read structured args from context.args. Generated workflow runs return args.workflow.name and args.workflow.path for later workflow-run. Use workflow-run --stream --json for long launches: event lines go to stderr and the final run JSON stays on stdout. Use workflow-events --follow for an existing run id instead of repeatedly polling.

Executable smoke:

generated_json="$(agent-workflow-kit workflow "file command generated" --json)"
generated_name="$(printf '%s' "$generated_json" | bun -e 'const fs = require("fs"); const data = JSON.parse(fs.readFileSync(0, "utf8")); console.log(data.args.workflow.name);')"
printf '%s\n' "$generated_json"
agent-workflow-kit workflow-run "$generated_name" --json
run_json="$(agent-workflow-kit workflow-run no-write-probe --json)"
run_id="$(printf '%s' "$run_json" | bun -e 'const fs = require("fs"); const data = JSON.parse(fs.readFileSync(0, "utf8")); console.log(data.runId);')"
printf '%s\n' "$run_json"
agent-workflow-kit workflow-status "$run_id" --json
agent-workflow-kit workflow-status "$run_id" --tree
agent-workflow-kit workflow-events "$run_id" --json
agent-workflow-kit workflow-resume "$run_id" --json
agent-workflow-kit workflow-stop "$run_id" --json
agent-workflow-kit workflows --json
AGENT_WORKFLOW_KIT_WATCH_ITERATIONS=1 agent-workflow-kit workflows --watch
agent-workflow-kit deep-research "file-command-smoke" --json

Ultracode & multi-phase orchestration

Authoring and running a workflow spins up many subagents and spends real tokens. Author a workflow only when the user typed "workflow"/"workflows", asked for multi-agent orchestration, when the ultracode standing opt-in or keyword trigger applies, when a skill or command instructed it, or when the user named a saved workflow. Otherwise act directly or offer a workflow and let the user opt in.

Ultracode has three separate meanings in this kit. The standing opt-in is the persisted project behavior (agent-workflow-kit ultracode on|off|status, stored in .agent-workflow-kit/config.json) that author-runs a workflow for substantive tasks by default. The keyword trigger is recorded separately as ultracodeKeywordTriggerEnabled and is disabled when workflows are disabled. Model effort is host-owned: this standalone CLI reports model effort as unsupported and records ultracodeEffortMode: "orchestration-only" when orchestration is enabled. Turn ultracode on only when the user asks; when off, revert to the opt-in gate.

For larger work, decompose into a sequence of workflows (understand -> design -> implement -> review), inspecting each run with workflow-run --stream, workflow-events --follow, or workflow-status --tree between phases rather than one giant run.