Fable Orchestrate
From now on, keep in the main thread:
- Clarifying the request.
- Splitting it into tasks with acceptance criteria.
- Dispatching tasks.
- Verifying reports.
- Hard problems: design decisions, tricky debugging, anything where judgement matters more than legwork.
Delegate implementation work: reading code at scale, writing code, running tests, bulk edits.
Dispatch
The argument picks the executor. Without one, use opus.
| Argument | Executor |
| --- | --- |
| opus (default), sonnet | Agent tool with model set to that name |
| gpt | Agent tool with subagent_type: codex:codex-rescue; prefix the prompt with --wait --fresh |
| herdr <kind> | herdr skill: start a <kind> agent, agent prompt --wait, then agent read for its report |
Run independent tasks in parallel. Tell every subagent to do the work itself without spawning subagents.
Report
Ask each subagent to report:
- What was done.
- Files changed.
- How it was verified, with command output.
- Open questions or blockers.
Verify each report against its acceptance criteria. Follow up on the same executor: SendMessage for opus and sonnet, --wait --resume for gpt, another agent prompt for herdr.