| name | ralphx-agent-workflow-orchestrator |
| description | Author safe RalphX Agent Workflow scripts that orchestrate durable delegated agents through the high-level workflow API. |
| trigger | RalphX Agent Workflow authoring |
| disable-model-invocation | true |
| user-invocable | false |
RalphX Agent Workflow Orchestrator
Use this skill only in a conversation whose selected capability is Workflow.
Generate a deterministic JavaScript program for the bundled RalphX workflow runner. The program may use only the provided globals: args, meta, phase, log, agent, parallel, pipeline, and checkpoint helpers. It has no filesystem, shell, network, environment, process, import, or module access.
Authoring contract
- Break the request into named phases and call
phase(name) before each phase.
- Give every
agent() call a stable logicalKey. A key must identify the same prompt and output schema across retries.
- Keep prompts self-contained. Do not interpolate untrusted text into executable JavaScript; pass user values through
args.
- Use explicit JSON Schemas for machine-consumed results. A schema makes the delegated agent return JSON-only output; RalphX validates it before caching/completing the invocation and exposes the parsed value as
result.content. Missing, null, extra (when disallowed), or type-invalid required results fail the workflow.
- Keep fanout within the declared
maxConcurrency and maxInvocations metadata. Prefer the smallest useful fanout.
- Use
parallel([{ prompt, logicalKey, agentName, schema, title }, ...]) only for independent agent work; this descriptor form is the backend-enforced concurrent fanout path. Use pipeline() when later work depends on earlier output.
- End with a concise structured result that can be reported truthfully to the user.
Call create_agent_workflow_script with the script, metadata, permission summary, and estimated fanout. This stores the script for user review; it never approves or starts it. Do not claim that a workflow is running until start_agent_workflow_run succeeds after the user-approved hash is present.
Use the run inspection and pause/resume/cancel tools for lifecycle management. Never expose or invent runner attempts, leases, transport frames, delegated job IDs, or recovery bookkeeping.
