sdlc-parallel-worklisted

# Several streams at once ## Two different tools **A session** is a full instance working one task in its own checkout. Sessions share nothing and know nothing about each other — the person steering them is the only connection. **A subagent** runs inside a session as a scoped helper with its own context window and its own tool access. Sessions increase how much is in flight. Subagents keep one session from drowning in its own context. They compose; neither substitutes for the other. ## What has to be true first **A real `CLAUDE.md`** (`sdlc-claude-md`), because it's the only thing keeping independent sessions consistent with each other. **A working feedback loop** (`sdlc-feedback-loop`) — this is the actual prerequisite. A session that verifies itself needs supervision at the end; one that can't needs supervision throughout, and you cannot supervise three of those at once. Everything here depends on that. **Permissions tuned** so sessions aren't parked on prompts for commands you consider routine. Three blocked sessions are worse than one unblocked one. ## Splitting the work Split along **files**, not features. Use the plan (`sdlc-plan`) to see where the work is genuinely independent. Tasks that touch the same files go in one session, sequentially — parallelizing them just manufactures merge conflicts you then resolve by hand. Give each task its own worktree, so sessions can't collide: ```bash git worktree add ../app-draft-persistence -b draft-persistence git work