branch-lifecyclelisted

# Branch & Worktree Lifecycle The roster's core safety rule is **serialized writing, parallel reading** — only the `executor` writes, one scope at a time (`references/core/agents.md`). Prose cannot enforce that. A worktree can: two agents in two trees cannot corrupt each other's files no matter what they do. ## Start — a branch per scope, and the scope is on disk ```bash git checkout -b <scope> # sequential work git worktree add ../wt-<scope> -b <scope> # parallel work: an isolated tree per scope ``` Before any agent starts, **write the scope down**: which pins it closes, and which file globs it may touch. That file is what makes "one scope at a time" checkable instead of promised — an agent's scope is a declared fact, not a shared assumption, and two scopes whose globs intersect must not run concurrently. Distinguish two relations, because conflating them is how a schedule deadlocks or corrupts: - **`depends_on`** — B needs A's *result*. Ordering. - **`conflicts_with`** — B and A touch the same files. Mutual exclusion, no ordering implied. Independent scopes go in parallel worktrees. Conflicting scopes serialize, whatever the DAG says. ## During — commit against pins - Commit at each green step, referencing the pin: the ledger says *why*, the commit says *what*. - **Never commit on the default branch.** Branch first, always. - A shared file (types, config, schema) belongs to **one** scope. Assign it explicitly rather than hoping two agents edit