Imported from yama-otoko/taskwish (packages/create-project/templates/revenue-monitor/.agents/skills/taskwish/SKILL.md). Install upstream with npx skills add yama-otoko/taskwish --skill taskwish. Copyright stays with the author.

TaskWish

Keep taskwish.ts as the runnable project entrypoint. Organize each actor under src/<service-name>/ with its actor definition in <service-name>.ts, one file per action, an index.ts service export, and a colocated <service-name>.test.ts. Do not use actor.ts, <service-name>.actor.ts, or actor.test.ts filenames.

An actor owns its domain state and commands. Coordinate workflows across actors through a dedicated actor that declares dependencies with .use(...) and calls them through this.actions.<service>.<action>.

Declare each actor-owned event directly inside that actor's .scope(...), for example Actor("IncidentOrchestrator").scope(Event("InvestigationRequested", { runId: "string" })). Use the local event name there; TaskWish qualifies it as IncidentOrchestrator::InvestigationRequested. Do not pre-qualify event names or collect event declarations in a separate events object before passing them to the actor.

Define public actions with .on("Command", "actionName"). Give command names enough domain context to stand on their own: prefer verb-and-noun names such as addTodo, scheduleTodoReminder, or createCustomer over vague names such as add, record, or process. Add .input(...) when the action accepts input. Define the input object inline in .input({ ... }); do not export, import, or spread standalone input definitions such as intakeDefinition. Repeat shared fields at each public action boundary so the complete contract remains visible there. Add .meta(...) to public commands so Console users can understand them: provide a concise description, and add input descriptions and examples when fields need guidance. Use lower camelCase names for every Step(...), such as validateInput or createCustomer. Prefer multiple focused steps when an action has distinct phases or coordinates a workflow, and access earlier results through this.<stepName>. Keep truly atomic actions as one step instead of adding artificial stages. Let State.List generate primary IDs instead of supplying them when inserting records.

Format fluent actor and action definitions with a blank line between .use(...) and .scope(...) when both are present, and between the .on(...), .input(...), .run(...), and .meta(...) action sections. Inside .run(...), put a blank line between adjacent Agent(...) declarations.

Add Bun unit tests for public service behavior and multi-actor workflows. Exercise public actions with Service.action.ctx({ ...mocks }).run(input). Mock default agents with agent, named agents with their configured names, and state with state; keep provider calls and persistent state out of unit tests. When importing a stateful actor still initializes its store, point TW_DEFAULT_STORE_PATH to a temporary directory before dynamically importing it, then restore the environment and remove the directory after the test so state cannot leak between runs.

Update taskwish.ts when a runnable scenario helps demonstrate a changed workflow. Run bun run check and bun test after changing actor contracts. Run bun start by default, or npm run start:node when specifically verifying Node.js compatibility.