Swarm

Plan first, then launch parallel AI agents that build features AND continuously harden the codebase — all running in isolated Git worktrees.

Principle #6 (Find what's hard — that's the value) — the orchestrator's job in Phase 1 is identifying the brittle seam, not splitting work evenly. Principle #10 (Agents drift, gates don't) — the CRAP, Mutation, and Acceptance agents are the gates that keep the Feature agent honest. Principle #7 (Reversibility over correctness) — worktrees make every agent's work cheap to discard. Three principles reinforce each phase. See /principles.

Workflow

Phase 1: PLAN (interactive, human-in-the-loop)
  ├── Activate grill-me: interrogate the plan until every branch is resolved
  ├── Source the backlog: pull tasks from Linear/GitHub/Jira → local file
  └── Pre-flight checklist: detect tooling, verify setup, confirm configuration