-
Run the maestro agent (subagent_type: maestro) via the Agent tool, in the foreground (run_in_background: false), with model: "opus" explicitly set (the environment may default subagents to a cheaper model — override it).
The subagent starts with no memory of this conversation, so the prompt must be self-contained:
- absolute cwd (
pwd)
- the task to plan, verbatim from the user/arguments
- any constraints already discussed in this conversation worth passing along
Do not pre-research the task yourself — that's the agent's job and doing it twice wastes context. Just launch it and wait.
-
When it returns, relay only the file path and its short summary to the user. Don't read the plan file into this conversation unless the user asks to see it.
-
Invoke the crit skill on the plan file so the user can leave inline review comments.
-
After the user finishes commenting, use crit-cli to read the comments and address them yourself with Edit directly on the plan file — no new subagent needed for most feedback, since maestro's research is already captured in the doc. Only re-invoke maestro (step 1) if a comment requires genuinely new codebase research the doc doesn't cover.
-
After the plan is approved, move directly to implementation.
