martech-change-guardlisted

# MarTech Change Guard Use the deterministic guard before a bulk data mutation. Do not substitute an LLM review for the script's changeset, invariant checks, rollback operations, or verification result. Resolve `scripts/guard.py` relative to this `SKILL.md` and invoke that resolved absolute path. Installed plugins run from the user's project, so do not assume `scripts/guard.py` exists under the current working directory. Keep input and output paths in the user's workspace; never write plan artifacts into the installed skill or plugin cache. ## Plan a change Obtain a current-state export and a proposed-state export with the same stable record key. Export the same columns in both files; schema drift fails closed. Read [references/input-guide.md](references/input-guide.md) when the files or export scope are unclear. If the user has only described a desired transformation, create the proposed export without touching the live system. Then run: ```bash python <absolute-skill-directory>/scripts/guard.py plan --before current.csv --proposed proposed.csv \ --key record_id --policy policy.json --out guard-plan ``` Read `guard-plan/risk-report.json` and lead with its decision: - `allow`: the plan passed configured controls, but a live write still requires the user's authorization and the destination tool's normal safeguards. - `review`: show the blast radius, warnings, and canary. Ask for approval immediately before any live write. - `block`: do not execute. Explain each in