Cleaning Commit History

Cleaning Commit History

Clean up messy git histories into logical commit sequences that are easy to review and maintain.

VCS Detection

Check which VCS to use first:

jj root 2>/dev/null && echo "USE_JJ=true" || echo "USE_JJ=false"

If jj is available, prefer jj commands -- they're non-interactive and the oplog provides automatic safety (no manual backup branch needed).

Operating Procedure

Phase 0: Safety

Git: Create backup branch before any surgery:

git branch ${CURRENT_BRANCH}-backup

jj: Not needed -- oplog provides safety. Note the current operation ID with jj op log -n 1.

Phase 1: Inventory

1. Determine base branch (main, master, or dev)
2. Use git merge-base (or jj log) to find comparison point
3. Inventory all feature-only changes with git log --oneline $BASE..$FEATURE_BRANCH
4. Note large files, generated paths, vendored code, migrations

Phase 2: Sea of Changes

Compute the net diff from BASE to FEATURE_BRANCH (not commit-by-commit). This represents all changes that need reorganization.

Phase 3: Classify & Cluster

Cluster changes into logical buckets (strict priority):

1. Generated/Vendored/Lockfiles -- isolated to dedicated commits
2. Pure renames/moves -- separated from content changes
3. Formatting-only (whitespace, import order, lint fixes) -- isolated
4. Refactors without behavior change -- separate from logic
5. Feature/Logic changes -- grouped by cohesive unit
6. Tests -- co-located with their corresponding logic changes

Split when a commit mixes mechanical and semantic changes. Squash when multiple tiny edits serve the same concern.

Phase 4: Determine Commit Order

Order for buildability and minimal noise:

1. Pure renames/moves
2. Formatting-only sweep
3. Refactors (non-behavioral)
4. Schema/Migrations
5. Feature/Logic in dependency order
6. Tests (accompany or immediately follow their logic)
7. Docs/Changelog
8. Vendored/lockfile updates

Every intermediate state must build and pass tests.

Phase 5: Rebuild Commits

Git: git reset --mixed $BASE, then stage related hunks per planned commit with git add -p.

jj workflow:

# Squash related changes (always use -m!)
jj squash --from <change1> --into <change2> -m "combined message"

# Selective restore (jj split is interactive, avoid it)
jj new -m "first part"
jj restore --from @- <files-for-first-commit>

# Reorder
jj rebase -r <change> -d <new-parent>

# Clean up messages
jj describe -m "feat(scope): message"

# If anything goes wrong
jj op restore <before-surgery>

Phase 6: Validation

• git diff $BASE..HEAD equals the original sea (no loss of intent)
• Each commit shows clean boundaries with minimal file overlap
• Every commit builds and tests successfully
• No secrets or large binary blobs

Commit Message Style

<type>(<scope>): <short description in present tense, under 72 chars>

- <Bullet point starting with verb>
- <What changed and why>

[Optional: BREAKING CHANGE:, Refs:, Co-authored-by:]

Types: feat, fix, refactor, perf, chore, test, docs, build, ci

Strict Rules

• Never mix formatting/import-order with behavior changes
• Always separate file renames/moves from edits to those files
• Always keep generated and vendored changes isolated
• Always co-locate tests with their logic change
• Never create broken intermediate states

Deliverables

1. Safety Confirmation: Backup branch (git) or oplog snapshot (jj)
2. Commit Plan: Ordered list with title, scope, type, rationale, and files
3. Applied History: Rewritten commits matching the plan
4. Summary Report: Changes vs original, tradeoffs, recovery instructions