tp-revert

tp-revert — Safe Design Revert

Reverts a previously-landed design merge commit. The skill is depth-gated (newest landing, probe depth 0 only), runs a dry-run forecast, applies git revert -m 1, preserves living docs via carve-out, regenerates amendments, and lands the result through the standard gate.

Not lock-enforcing. The target design's lock is already archived; nothing to claim.

---

Steps

1. Run first-run preflight per skills/_shared/first-run.md. Resolve $TP_ROOT with the canonical form:

TP_ROOT="$(bash <skill-dir>/../../skills/_shared/resolve_root.sh --skill-dir <skill-dir>)"

The {pr#} argument requires no design-name validation (it is a PR number or merge sha, not a design slug). Any slug extracted from the PR's headRefName is expected to match [a-z0-9-]+ per skills/_shared/validate-name.md — print a warning if it does not but continue (the slug is used for labeling, not enforcement here).

Step 1 — Probe

python3 "$TP_ROOT"/skills/tp-revert/scripts/revert_probe.py \
    --repo . (--pr {pr#} | --sha {merge-sha}) [--base master] --json

On error field non-null: print the error and stop.

Step 2 — Depth gate (hard line)

• depth > 0 → REFUSE with reality. Print the forecast conflicted[] set, the documented recipe below, and stop. Proceeding past newest landing (probe depth 0) is manual-only by the operator using the recipe. The skill never automates past newest landing (probe depth 0).
• depth == 0 but clean == false (unexpected) → same refusal path. Never pretend it is safe.

The depth caveat: a revert is clean only while newest (probe depth 0). Once any later merge lands, the probability of a clean revert drops sharply (probe result: 1/12 merged PRs reverted clean overall — the one clean WAS the newest landing (probe depth 0); past it, zero clean).

Step 3 — Provision workspace

git worktree add .claude/worktrees/revert-{slug} -b revert/{slug} origin/{base}