Tag Audit — triage accumulated safety tags

Stacked-rebase sessions accumulate safety tags fast — each surgical operation deserves its own rollback anchor, but the cost of regret is high. Once an operation is validated and pushed, those tags decay into redundancy: they either duplicate another tag, point at the same SHA as a current branch ref, or mark an intermediate waypoint no one will ever git reset back to.

This skill audits the safety-tag namespace, surfaces duplicates by sorting on SHA (not name), proposes deletions with reasons, and deletes only after explicit confirmation.

When to Use

• After a successful + pushed + CI-green stacked-rebase or surgical-fix

session, when the operation's safety tags are no longer load-bearing.

• When /stack-cascade Step 2 refuses to proceed because the OP_ID

namespace is non-empty.

• When git tag -l 'safety/*' 'pre-*' 'post-*' returns more entries than

feels healthy and you want to know which ones still earn their keep.

• When inheriting a workspace from a past session and the tag namespace

is unfamiliar.

When NOT to Use

• During an in-flight cascade or surgical fix — tags are still load-bearing

rollback anchors.

• When the operation hasn't been validated end-to-end (CI green, PRs

reviewed, no surprises from teammates) — the rollback window is still open.

• On release tags or other semantically-meaningful tags outside the

safety-tag pattern set — narrow the pattern set first.

The Survivor Heuristic

A safety tag earns its keep only when it points at a unique, semantically-meaningful SHA AND that SHA represents a state you might genuinely need to git reset back to. Anything else is rollback theater — the deletion candidate. Step 3 implements this as a three-question decision tree.

Step 1 — Enumerate tags, sorted by SHA

Sort by SHA (not name). Name-grouped output hides duplicates; SHA-grouped output exposes them — tags pointing at the same commit cluster together, making redundancy visible at a glance.

# Default is conservative: only the canonical `safety/*` namespace.
# Keep in sync with Step 2 below.
PATTERNS=('safety/*')
(
  set -euo pipefail
  count=$(git tag -l "${PATTERNS[@]}" | wc -l | tr -d ' ')
  echo "# Matched $count tag(s) across patterns: ${PATTERNS[*]}" >&2
  [[ "$count" -gt 0 ]] || { echo "# (nothing to audit — check pattern typos or working directory)" >&2; exit 0; }
  git tag -l "${PATTERNS[@]}" | while IFS= read -r t; do
    # Capture-then-print: bare $(...) inside printf's arg list does NOT trigger
    # set -e on substitution failure (the outer printf wins with exit 0).
    # Capturing to a variable lets set -e fire if rev-parse fails.
    # The ^{commit} peel returns the commit SHA for both lightweight AND
    # annotated tags (bare rev-parse on an annotated tag returns the tag-object
    # SHA, which won't cluster with branch SHAs in Step 2).
    sha=$(git rev-parse --short "$t^{commit}") \
      || { echo "rev-parse failed for tag '$t' — aborting audit" >&2; exit 1; }
    printf '%s  TAG    %s\n' "$sha" "$t"
  done | sort
)

The default ('safety/*') is intentionally conservative — it matches only the canonical namespace stack-cascade writes to under the current convention. Widen it explicitly for cleanup that covers other namespaces:

• Legacy stack-*-start/* from cascades before the convention change.
• Hand-named anchors from ad-hoc surgical sessions: pre-mr/*,

pre-port/*, pre-fix-*, post-fix-*, pre-split/*.

• Other safety-tag conventions: backup-*.

PATTERNS=('safety/*' 'pre-*' 'post-*' 'backup-*' 'stack-*')   # broad audit

The audit is read-only; widening the pattern is cheap. But leave the widened pattern in place for Step 5's deletion only after the KEEP/DROP review confirms no release tags or ad-hoc semantic tags slipped in (pre-1.0, pre-release-*, backup-2024-05, safety-audit-2026-q1). See "When NOT to Use" above.

Step 2 — Cross-reference branch tips

Enumerate current branch tips by SHA, then interleave them with Step 1's output so SHA collisions between tags and branch refs are visible adjacently:

# Use the same PATTERNS as Step 1 (default safety/* OR your widened set).
PATTERNS=('safety/*')
(
  set -euo pipefail
  {
    git tag -l "${PATTERNS[@]}" | while IFS= read -r t; do
      # Same capture-then-print + ^{commit} peel as Step 1 — see comments there.
      sha=$(git rev-parse --short "$t^{commit}") \
        || { echo "rev-parse failed for tag '$t' — aborting audit" >&2; exit 1; }
      printf '%s  TAG    %s\n' "$sha" "$t"
    done
    git for-each-ref refs/heads --format='%(objectname:short)  BRANCH %(refname:short)'
  } | sort
)

Any TAG line whose SHA matches an adjacent BRANCH line is a deletion candidate — the branch ref already serves as the anchor.

Step 3 — Decision tree per tag

For each tag in the audit output, apply these three questions in order:

1. Does another tag point at this exact SHA? → drop one; keep the

more recent or more descriptive name.

1. Does a current branch ref point at this exact SHA? → drop the

tag; the branch ref already anchors that SHA.

1. **Is this an intermediate waypoint that was useful during operation

but adds no information now?** → drop unless it marks a uniquely meaningful state worth permanently anchoring (root snapshots, pre-cascade origin SHAs, anchors referenced in narrative docs).

A tag survives only when it has a unique SHA AND no equivalent branch ref AND marks a state genuinely worth git reset-ing back to.

Step 4 — Propose KEEP/DROP with reasons before deleting

Always show the proposal before any deletion. Example output (with widened patterns covering an ad-hoc pre-* / post-* session):

Found 21 safety tags. After audit:

KEEP (5 — unique semantically-meaningful SHAs):
  3be7b11a  pre-split/improved-disabled        [root anchor]
  4a9fac77  pre-split/pr10                     [pre-cascade origin]
  e5f1dd7d  pre-split/pr11                     [pre-cascade origin]
  aa71726b  pre-split/pr12                     [pre-cascade origin]
  f675fe31  pre-split/pr13                     [pre-cascade origin]

DROP (16):
  pre-mr/pr10-cascade        [duplicate: same SHA as pre-fix-pr11ci/pr10]
  pre-mr/pr11-cascade        [duplicate: same SHA as pre-fix-pr11ci/pr11]
  pre-fix-pr11ci/pr10        [branch-ref: SHA matches branch pr10]
  post-fix-pr11ci/pr11       [branch-ref: SHA matches branch pr11]
  post-fix-pr11ci/pr12       [branch-ref: SHA matches branch pr12]
  post-fix-pr11ci/pr13       [branch-ref: SHA matches branch pr13]
  pre-mr/pr9.6-post-triage   [intermediate-waypoint]
  pre-mr/pr9.6-post-c5       [intermediate-waypoint]
  ... (etc)

Proceed with deletion? [y/N]

Reason codes: [duplicate], [branch-ref], [intermediate-waypoint], or a one-line free-text reason for KEEP entries ([root anchor], [pre-cascade origin], [narrative anchor]).

Step 5 — Confirm and delete (local)

Two-pass verify-then-mutate under set -euo pipefail. Use quoted array iteration — bash unquoted-var word-splitting fails silently under zsh (see references/shell-portability.md).

TAGS_TO_DELETE=(
  "pre-mr/pr10-cascade"
  "pre-mr/pr11-cascade"
  "pre-fix-pr11ci/pr10"
  # ... full list from Step 4 DROP
)
(
  set -euo pipefail
  # Verify pass: confirm every tag exists before deleting any
  for t in "${TAGS_TO_DELETE[@]}"; do
    git rev-parse --verify "refs/tags/$t" >/dev/null 2>&1 \
      || { echo "Tag '$t' not found — aborting (no tags deleted)" >&2; exit 1; }
  done
  # Mutation pass: delete each
  for t in "${TAGS_TO_DELETE[@]}"; do
    git tag -d "$t"
  done
)

The verify pass is a pre-flight check: a missing tag name aborts before any deletion, so the namespace can't be left half-cleaned by a typo or a stale paste from Step 4. (Concurrent tag mutation by another process between the verify and mutate passes is still possible — this is shell, not a transaction. The window is narrow if you're the only writer.)

Step 6 — Remote tag cleanup (opt-in only)

Most safety tags are local-only: git push does not push tags by default. Skip this step unless the tags were explicitly pushed (e.g., via git push --tags or a CI workflow that forwards tags).

When remote cleanup is needed, push deletion refspecs in a single atomic operation so a partial failure doesn't leave the remote half-cleaned:

TAGS_TO_DELETE=(
  "pre-mr/pr10-cascade"
  "pre-mr/pr11-cascade"
  # ... same list as Step 5
)
(
  set -euo pipefail
  # Guard against empty array: `git push --atomic origin` with NO refspecs
  # falls back to push.default and pushes the current branch — exactly the
  # opposite of "delete zero remote tags".
  [[ ${#TAGS_TO_DELETE[@]} -gt 0 ]] \
    || { echo "TAGS_TO_DELETE is empty — refusing to push" >&2; exit 1; }
  refspecs=()
  for t in "${TAGS_TO_DELETE[@]}"; do
    refspecs+=(":refs/tags/$t")
  done
  git push --atomic origin "${refspecs[@]}"
)

:refs/tags/<name> is the deletion refspec; --atomic makes the multi-ref push all-or-nothing.

Common Failure Modes

• Bash idiom failing silently under zsh — `for t in $unquoted_var;

do git tag -d $t; done passes the whole string as one arg to git tag -d, which complains about a single-revision-not-found error. The loop "completes" without deleting anything. Always use quoted array iteration (for t in "${arr[@]}") or xargs. See references/shell-portability.md` for the full mitigation set.

• Too-broad pattern matches a release tag — git tag -l '*' is

obviously dangerous, but safety-* accidentally matches a hand-named safety-audit-2026-q1 tag, or backup-* matches a release naming scheme. The proposal-before-delete gate at Step 4 is the safety net — read the KEEP/DROP list before confirming.

• **Deleting a tag whose SHA equals a branch tip — but the branch

itself is the one you wanted to restore** — if the cascade's pushed state turns out to be wrong and you want to roll back, the safety tag was the rollback anchor. Branch tips can move; safety tags do not. This is why "rollback window closed" is a precondition: don't audit until you're committed to the current state.

• Audit run mid-cascade — running this skill while another git

operation is in progress in the same repo (a rebase paused on conflict, a cherry-pick mid-application) can delete tags the paused operation needs for --abort. Check git status first.

Pairing with /stack-cascade

/stack-cascade creates safety tags under safety/${OP_ID}/before/<branch> (and historically under stack-${PURPOSE}-start/<branch> before the convention change). Step 2 of /stack-cascade blocks if its OP_ID namespace is non-empty — that collision-check error names this skill as the remediation. The error shows a literal git tag -l 'safety/${OP_ID}/*' preview command; running this skill against that pattern set produces the targeted DROP list. For legacy stack-*-start/* tags, widen PATTERNS to include 'stack-*' in Step 1.

/stack-cascade and /rebase-validate compose tightly along the cascade flow: cascade executes, validate verifies meaning preservation. /tag-audit is a separate utility that pairs with cascade's tag namespace as post-success aftercare, but operates independently — it's useful for any safety-tag accumulation, not just cascades, and its preconditions (rollback window closed) decouple it from the cascade's session.