Repository Hygiene Sweep

repository-hygiene-sweep — clean a repository without destroying useful work

Maintain a git repository: prune merged branches, compact storage, hunt large/dangling objects, and (only with explicit consent) rewrite history to remove accidental large files or secrets. The discipline is conservation — every step is reversible-first, and anything ambiguous is skipped, not guessed.

⚠️ Critical Constraints

• Never run a destructive command without explicit, scoped confirmation. Branch deletion, history rewrite, reflog expire, and gc --prune=now are the destructive set. Name exactly what will be destroyed and wait for a yes. Why: these are the operations that lose real work, and the user often does not realize what a branch or reflog entry was protecting.
• Skip when in doubt — never guess. If you cannot prove a branch is fully merged, a file is truly junk, or a repo is not shared, do nothing and report the ambiguity. Why: the cost of a wrong delete (lost work) vastly exceeds the cost of leaving a stale branch around.
• Read-only first; mutate only after. Run the diagnosis pass (counts, sizes, candidate lists) and present it before changing anything. Why: the user must see what they are approving.
  • WRONG: git branch --merged | grep -v '\*' | xargs git branch -d (auto-deletes on a single guess)
  • CORRECT: list git branch --merged main, show it, get confirmation, then delete named branches one set at a time.
• Never delete a branch that is not merged with -D. Use -d (refuses to delete unmerged). Why: -D force-deletes branches that contain unique commits — the classic way to silently lose work.
  • WRONG: git branch -D feature/x
  • CORRECT: git branch -d feature/x and, if it refuses, investigate why instead of forcing.
• Never rewrite published/shared history casually. History rewrite (filter-repo, BFG) is acceptable only on a repo the owner confirms is unshared or whose collaborators have agreed to re-clone. Why: rewriting shared history breaks every other clone and can resurrect "deleted" data on the next push.
• Always confirm a clean, backed-up state before any history rewrite. Require a committed/clean working tree and a fresh mirror clone (git clone --mirror) as the backup. Why: history rewrite is the one operation with no in-repo undo once reflogs are expired.
• A leaked secret is not "removed" by rewriting history alone. Always advise rotating/revoking the secret too. Why: the secret may already be cloned, cached, or on a fork — rotation is the only real remediation.

Why This Exists

Repositories accumulate cruft: dozens of merged feature branches, a .git directory bloated by an accidentally-committed binary, loose objects, and .gitignore files that let junk slip in. Cleaning them is routine but genuinely dangerous — the same commands that reclaim space (gc --prune, branch -D, history rewrite) are the ones that permanently destroy work. Agents are biased toward "looks done" and toward force-flags that make errors disappear; that bias is exactly wrong here. This skill exists to make the safe path the default path: diagnose, propose, confirm, act reversibly, and skip anything uncertain.

Quick Start

# 1. Confirm where you are and that nothing is uncommitted (diagnosis is safe, read-only).
git rev-parse --is-inside-work-tree && git status --short

# 2. Run the full read-only diagnosis bundle.
bash {baseDir}/scripts/diagnose.sh        # branches, sizes, large blobs, dangling objects

# 3. Review the report, confirm specific actions, then mutate one category at a time.

{baseDir}/scripts/diagnose.sh is Execute — run it to gather the report. It changes nothing.

Workflow

Phase 1 — Diagnose (read-only, always first)

Establish ground truth before touching anything.

git status --short                                  # uncommitted work? (if yes, STOP — back up first)
git fetch --all --prune --dry-run                   # what would prune; preview only
git branch -vv --merged                             # branches fully merged into HEAD
git branch -vv --no-merged                          # branches with unique commits (DO NOT auto-delete)
git for-each-ref --sort=committerdate refs/heads --format='%(committerdate:short) %(refname:short)'
du -sh .git                                          # current repo storage footprint
git count-objects -vH                               # loose vs packed, size, dangling count

Checkpoint: if the working tree is dirty, or this is a shared repo and the user has not confirmed it, stop and report. Do not proceed to any mutation.

Phase 2 — Prune branches (reversible)

Branch deletion is recoverable via reflog for a while, but treat it as consequential.

# Identify the integration branch explicitly (main/master/develop) — do not assume.
git branch --merged main | grep -vE '^\*| main$| master$| develop$'   # safe-to-delete candidates

• Present the candidate list. Delete only confirmed names, using -d (never -D).
• For remote-tracking cleanup of branches already deleted upstream: git fetch --prune (safe; only removes stale remote refs, not your local branches).
• Record each deleted branch's tip SHA in the report so it can be restored: git branch <name> <sha>.

Checkpoint: verify each deletion succeeded with -d (no force). If git refused, the branch had unique commits — leave it and flag it.

Phase 3 — Compact storage (reversible vs destructive split)

git gc                          # safe: repacks, keeps reflog-reachable + recent loose objects
git repack -Ad                  # safe: consolidate packs
# DESTRUCTIVE — only after explicit confirmation:
# git reflog expire --expire=now --all && git gc --prune=now   # drops unreachable objects permanently

Run the safe gc by default. Only offer the prune-now path when the user explicitly wants to reclaim space from already-removed history AND understands it makes recovery impossible.

Phase 4 — Find large / dangling objects (read-only)

See references/large-objects.md for the blob-size ranking and dangling-object triage recipes.

Checkpoint: finding a large blob does not mean removing it. Confirm it is genuinely unwanted (not a needed asset) before considering Phase 5.

Phase 5 — Remove large files / secrets from history (DESTRUCTIVE — gated)

This rewrites history. Do not enter this phase without: clean tree, confirmed unshared-or-agreed repo, and a git clone --mirror backup. Procedure, tool choice (git-filter-repo vs BFG), and the force-push/re-clone protocol are in references/history-rewrite.md.

Phase 6 — .gitignore hygiene (low-risk)

git status --ignored --short                        # what is currently ignored
git ls-files --others --exclude-standard            # untracked, not ignored — should any be ignored?
git check-ignore -v <path>                           # explain why a path is/ignored

• Propose additions for build artifacts, dependency dirs, env/secret files, OS cruft. Do not blanket-ignore source.
• If a file is ALREADY tracked, adding it to .gitignore does nothing — note that it needs git rm --cached (a tracked-content change, confirm separately).

Output Specification

Produce a report named repo-cleanup-report.md (or inline if small) with:

1. Findings — repo size, branch counts (merged/unmerged/stale), large blobs (size + path), dangling object count, .gitignore gaps.
2. Actions taken — each mutation, the exact command, and the result.
3. Skipped (and why) — every ambiguous item deliberately left alone.
4. Recovery — reversible commands (deleted branch SHAs, mirror-backup location).

Quality Rubric

• No destructive command (branch -D, gc --prune=now, history rewrite, reflog expire) ran without a logged, explicit confirmation for that specific target.
• Every branch deletion used -d, and every deleted branch's tip SHA is recorded for restore.
• The report's "Skipped" section is non-empty whenever any candidate was ambiguous — the skill demonstrably chose caution over completeness.

Examples

• Routine pass: Diagnose shows 18 merged branches, 1.2 GB .git, 40 loose objects. → List merged branches, confirm, -d delete; git gc; report 1.2 GB → 1.2 GB (size dominated by a large blob in history, surfaced in Phase 4 for a separate decision).
• Bloat hunt: .git is 4 GB. Phase 4 finds a 3.6 GB dataset.zip committed 200 commits back. → Confirm it is unwanted, confirm repo is solo + clean, mirror-backup, then git filter-repo per references/history-rewrite.md; advise force-push + collaborator re-clone.
• Leaked key: .env with a live API key was committed. → Rewrite history to drop it AND tell the user the key must be rotated immediately — rewriting alone does not secure it.

Troubleshooting

See Also