/delegate-agy

Run an agent-mesh in mode agy for the user's request: $ARGUMENTS when the user typed the command, otherwise the request that brought you here.

1. Read ~/.config/agent-mesh/ORCHESTRATE.md and ~/.config/agent-mesh/PROTOCOL.md in full.
2. You are the root. Follow the "Root" section with mesh init --mode agy.
   • Propose before you start ("Root" step 3): each worker's repo, size, model and view, then wait for the user's go, unless they said to run autonomously. A question, a review or a peer ask they asked for needs none.
   • Size every task ("Sizes") and pass --size to every mesh spawn and mesh ask. --kind agy is the default CLI; propose Claude only as "Routing" describes.
   • Pass --model / --effort only if the request names them. A large task plans on the best model of its plan chain with quota and moves to its execution model when you mesh approve its plan (mesh routing shows both). On a quota alert, mesh switch the worker.
   • Your own repo gets a worker like any other; do a part yourself only if the user tells you to. Work they gave you before delegating stays yours.
   • A question about another repo that changes nothing: use mesh ask ("Questions and reviews"). Run it with run_in_background if you're a Claude root. A question for the registered peer projects goes to mesh ask --peers <group|all>; read the one answers file it prints. A PR review the user asks for on its own goes through /delegate-review; the reviews of your own workers' PRs stay here ("Root" step 8).
   • A follow-up to work an earlier worker did ("send the review to the agent that changed it", "ask that agent to make it Y"): resume that worker ("Follow-ups"). It keeps its own CLI and model (a change it must plan first plans on its plan chain), and it's compacted first when its conversation is large.
   • Decide small things yourself and ask the user only where their input adds value; fix review findings and the follow-ups they turn up in the PR they came from, on its branch, by the worker that made it, never in a new PR ("Rules for everyone"); give new small work to the worker that already knows that code, not a fresh agent; put other follow-ups on the repo's list of open work.
   • A list of tasks: follow "Task lists". Triage it into a few themed groups (no size cap: a worker can take a whole batch), each one worker with one PR, sized by its combined work (independent small tasks: small, no plan; one plan for a large group); get the user's one go on the board, then supervise without writing code yourself.
   • If you're a Claude root and mesh prints a send_message, pass it on with the SendMessage tool, exactly as printed.
3. Once the user says go (at once, when they said to run autonomously), start the work and tell the user in one line what you started: run id, and each worker's repo, size and model. Then supervise until every worker reports back. For a question, give the user the answer and the name of the worker that gave it, so they can ask it again.