ollie-kde/prompts/tools-cascade.md

cascade

Delegate subtasks to a cheaper model. Fire-and-forget: returns immediately, worker reports back to your prompt when done.

Trigger condition: a subtask that can be described in a short instruction — the worker can read code, reason about implementation, and make decisions within the scope you define. Examples: implement a function, refactor a module, investigate a bug, extract and summarize information, apply an edit.

Do NOT cascade when: the task requires cross-cutting judgment that depends on context only you have, or coordination across multiple subtasks.

Calling convention:

call_tool: calls=[{tool: "cascade", args: ["<instruction>"]}]
call_tool: calls=[{tool: "cascade", args: ["<instruction>", "<id>"]}]

Parallel — consecutive calls fan out:

call_tool: calls=[
  {tool: "cascade", args: ["Add input validation to handleRequest in /src/server.go"]},
  {tool: "cascade", args: ["Write unit tests for ParseConfig in /src/config.go"]}
]

Returns: session, backend, model as KV lines.

Worker callback: arrives as your next queued turn via PeerSubmit:

[from=<session-id>]
STATUS: done
SUMMARY: <result>

Constraints:

• Instructions must be self-contained — include file paths and the goal, but you don't need to specify how. The worker reads code and reasons about implementation.
• Never wait for results. Continue work or end your turn.
• Results arrive asynchronously — only use cascade for work whose result you don't need in the current turn.
• One task per call.

cascade vs subagent_spawn

	cascade	subagent_spawn
Model	Always cheap (configured globally)	Any model
Identity	Fixed worker	JIT-generated
Use case	Scoped subtasks	Specialized reasoning or custom identity