v0.7.0

Minor Changes

• 22113ca: sandcastle init now supports fully non-interactive setup. Every interactive prompt has a paired CLI flag (--issue-tracker, --create-label, --build-image, --install-template-deps) on top of the existing --agent / --template / --sandbox / --model / --image-name. When stdin is not a TTY and a flag is missing for a prompt that would otherwise fire, init fails fast with a message naming the missing flag instead of crashing on the prompt library.

Patch Changes

• 0b397a1: Strengthen the simple-loop and sequential-reviewer prompts so an empty pre-expanded LIST_TASKS_COMMAND result is treated as ground truth, not as a stale snapshot. The "do not re-query" hint now frames the filtered list as the sole source of truth, and the # Done completion criterion explicitly equates an empty list with completion. Prevents the agent from running its own unfiltered gh issue list when the filtered list is [].

• c6880a4: Fix createSandbox (and createWorktree) reusing a stale worktree when called twice for the same named branch. A reused worktree holds a local copy of the branch that never moves on its own, so a re-run loop (review → push fixes → re-run) was reading stale code even though origin/<branch> had moved ahead.

  On the clean worktree-reuse path of the branch strategy, sandcastle now runs git fetch origin <branch> followed by git merge --ff-only origin/<branch> so the worktree picks up new upstream commits. The refresh only runs when it is provably safe — clean tree and strictly behind origin. Dirty, diverged (unpushed commits), or fetch fails (offline) → skip the refresh, reuse as-is, log why. Fetch failure is non-fatal and never breaks the run. First creation, the merge-to-head strategy, and the head strategy are untouched. See ADR 0003 for the full rationale.