create-workflow

When the user wants a workflow they can run on demand (for example "morning brief"), explain the step recipe clearly before saving.

Use plugin_workflows__create_workflow with kind (never type). Last step must be summarize with prompt (never instruction).

[
  { "id": "news", "kind": "tool", "tool": "web_fetch", "input": { "url": "https://news.ycombinator.com" } },
  { "id": "summary", "kind": "summarize", "prompt": "Write a brief from the receipts only." }
]

• tool — assigned profile or MCP tool that can run locally; input must match that tool's schema. Use web_fetch for URLs. Do not use web_search — it only runs inside a provider chat turn.
• template — substitutes values into text only. Writing “extract stories” in a template returns those instructions; it does not call AI, parse content, or produce structured JSON.
• compare / assert — deterministic checks on prior receipts. compare.op is eq | near | contains. Do not use compare as free-text analysis
• exactly one final summarize — turns the receipt bag into prose (no tools). It cannot be an intermediate extraction step or feed a later database step.

For extraction or analysis before a write, use an available tool that actually performs that work. Verify its input and output schema; do not invent a tool or output field. If no suitable tool is assigned, explain the missing capability instead of saving a recipe that cannot fulfill the request.

References use the tool's actual output shape: {{steps.fetch.content}}, not an assumed .output wrapper. Missing references stop the run. Explicit null values are allowed, so check required data with assert/compare steps before writes when applicable.

For a requested quantity such as five saved stories, check the extraction count and the database result with an assert/compare step. Inspect receipts before claiming success: a completed SQL call with changes: 0 does not prove stories were saved. Do not blindly rerun writes after a failure; inspect which writes already completed.

When the user names a profile to run as, confirm that profile and pass its agentId. Omit agentId to use the current chat profile.

When the user asks to run a saved workflow, use plugin_workflows__list_workflows to find it, then plugin_workflows__run_workflow. Pass input when the recipe uses {{input.*}} bindings.

When the user wants to change an existing workflow, use plugin_workflows__update_workflow.

Do not use workflows for clock-driven jobs — use create_automation for schedules.