luisnquin/nixos-config/squad

Squad

Fan a list of items out to subagents and land the approved ones on main. Items can be findings, tickets or tasks. You are the lead: you split, brief and spot-check. You do not write the fix yourself.

Steps

1. Split. Make one work item per finding or ticket. Merge duplicates.
2. Brief the implementer. Each one gets:

• the item, and what you already verified about its cause;
• the files involved, with line anchors;
• the repo's check and test commands;
• "do not commit, do not touch devices".

Use the model the user names for every subagent. Give each implementer its own worktree (isolation: "worktree"), unless items touch disjoint files — then a shared tree is fine, with every commit path-staged.

1. Verdict. The implementer first checks that the item is still real against

current main, then returns one of:

• FIX: the diff, plus proof. Proof is a repro before and after, a test that

fails without the change, or a screenshot pair.

• SKIP: the evidence that the item is invalid, a duplicate or already fixed.
• DECIDE: a product or scope call only the user can make, stated in one line.

1. Review with a reviewer subagent. As each implementer finishes, start a

reviewer for that item. It:

• verifies the implementer's claims independently (contracts, callers,

schema), not from the report;

• rejects missing proof, scope creep, broken conventions, files that belong

to another item;

• sends change requests to the implementer with SendMessage and re-reviews,

or fixes small things itself; loops until it would approve;

• on approval, lands one commit per item on main — path-staged, own files

only, following the repo's convention. Pushes only where the user allows it.

1. Evidence. Shared devices get ONE evidence agent for all items that need

them, so nobody fights over a device. Items from a ticket close with the evidence on the ticket, in the format the project's skill defines (e.g. linear-flow). Never change ticket states.

1. Spot-check. Read every landed commit yourself. Re-run the tests on any

commit a reviewer rebased before pushing.

1. Report. Give a table: item, verdict, commit hash or reason. List the

DECIDE lines and out-of-scope findings last.

Rules

• Do not wait for approval between items. Approval was given by starting the squad.
• A SKIP with evidence is a good outcome. Never fix something to justify a run.
• Kill a subagent that loops or digs past its item. Tokens are the budget.
• Devices: the ones the project names (rose emulators), never nyx or the

personal phone.