flow-reviewlisted

## Procedure 0. Resolve external context (`rules/workflow.md` section 4). Then dispatch the reviewer agent (subagent_type: reviewer, foreground) with steps 1 onward and the resolved arguments; it writes the report, you relay its summary. 1. Get the scratch directory via `scratch-dir.sh`. Stack is in the repo context your startup produced (`agent-context.sh`). 2. Mechanical-skip check. Opt-in by signal - only fires when the plan marks it, not by default: - Fires: the most recent plan for this project is marked `plan-shape: mechanical` and the implement step's verification passed cleanly (no failures recorded in the implement output). Emit a one-line review ("verification passed, no findings, mechanical change per the plan report"), state that it fired and why, then stop. - Does not fire: state that it did not fire and fall through to the full review (step 3). - `--full` in $ARGUMENTS skips this check entirely and always forces a full review. 3. If the diff exceeds 500 changed lines, propose splitting the review into logical chunks before starting. 4. Determine the review scope: - If $ARGUMENTS looks like a commit range (`main..HEAD`, SHA range): review that range - If $ARGUMENTS is a path: review that path's current state - Otherwise: review `git diff HEAD` (working copy) 5. Run `run-checks.sh` to establish baseline check state. Note any pre-existing failures before reviewing. 6. Load patterns skills matching the detected stack (react-patterns, django-patte