Git Flow

Version control is a record other people (and future you) read to understand why the code is the way it is. Commits should be small, focused, well-described, and safe to share. History should be bisectable for [[fault-recovery]], reviewable for [[review-gate]], and revertible for [[launch-readiness]] — a wall of WIP in one commit defeats all three.

Follow the project's merge and branch conventions when they exist. This skill is the baseline when none are defined.

Pairs with [[incremental-delivery]] for commit-sized slices, [[pipeline-ops]] for CI before merge, [[review-gate]] before opening a PR, [[launch-readiness]] for release tags and rollback, and [[hardening]] for never committing secrets.

When to Use

• Starting work on a feature, fix, or refactor
• Committing — especially before end of day or context switch
• Opening, updating, or merging a pull request
• Rebasing, merging, or resolving conflicts with main
• Cutting a release tag or hotfix branch
• Recovering from a bad commit, lost work, wrong merge, or accidental force-push panic
• Establishing branch and commit conventions for a team

Skip heavy ceremony for trivial one-line fixes — still write a real message and run checks if the repo requires them.

Process

1. Branch from the right base

• Branch off the latest integration branch (main, master, develop — whatever the project uses).
• Pull or fetch before creating the branch — don't branch from stale local main.
• Name for the work, not the person:

feature/add-guest-checkout
fix/null-total-on-empty-cart
chore/upgrade-eslint-9
hotfix/payment-422-regression

• Keep branches short-lived — days, not months. Long drift makes merges painful and review huge.
• One logical initiative per branch. Don't pile unrelated fixes onto a feature branch "while you're here."

2. Commit in small, working slices

Each commit should be one logical change that leaves the tree in a working state ([[incremental-delivery]]):

Commit at green — tests and lint pass for what you changed. Don't commit known-broken intermediate state unless the team explicitly uses stacked PRs / draft WIP on a private branch.

Before each commit:

git status
git diff          # review what you're about to record
# run relevant tests / lint per project norms

Separate refactor from behavior change when possible — reviewers and git bisect depend on it.

3. Write commits that explain why

The diff shows what changed. The message owns why.

Format:

Short imperative subject (~50 chars, no period)

Optional body: context, trade-offs, ticket link, what you rejected.
Wrap at ~72 chars.

Fixes #123

Subject examples:

Add idempotency key to order creation endpoint
Fix race in session refresh when tab wakes from sleep
Revert "cache user permissions globally"

Avoid: fix, wip, stuff, asdf, misc, address comments (say what changed instead).

Body when needed:

• Why this approach vs alternatives ([[decision-docs]] pointer if major)
• Non-obvious constraint or rollback note
• Migration or deploy ordering ("requires migration 0042 first")

Reference tickets (Fixes #, Refs #) so history links to issue context.

4. Never commit noise or secrets

Before every commit, scan the diff for:

• Secrets — API keys, tokens, .env, credentials ([[hardening]]). If leaked: rotate the secret; removing from git isn't enough.
• Generated artifacts — dist/, node_modules/, __pycache__/, build output (use .gitignore).
• Large binaries — use LFS or external storage if truly needed.
• Debug leftovers — console.log, commented code blocks, temporary flags.

Keep .gitignore current when new tooling adds output dirs. Use git check-ignore to verify.

If secrets hit shared history, treat as incident — rotate credentials and consider history rewrite only with team coordination (rare and painful).

5. Prepare a PR for review

Before opening:

1. Self-review the full diff ([[review-gate]]) — you'd catch half the comments yourself.
2. Rebase or merge latest main — resolve conflicts locally, re-run CI.
3. CI green on your branch ([[pipeline-ops]]).
4. PR description — what, why, how to test, screenshots for UI, rollout/rollback notes if risky ([[launch-readiness]], [[migration-path]]).
5. Scope — if the PR grew too large, split ([[incremental-delivery]]).

Update the PR as you push new commits; don't open "empty" and fill in description days later.

6. Merge and integrate — follow project convention

Teams differ; don't improvise on a shared repo:

Updating your branch with latest main:

• Rebase (git rebase origin/main) — linear history; only on your feature branch before merge.
• Merge (git merge origin/main) — preserves branch history; safer if others push to your branch.

Never rebase a branch others are actively committing to without agreement.

After merge: delete the feature branch (remote + local) to reduce clutter.

7. Resolve conflicts deliberately

When merge/rebase stops with conflicts:

1. Understand both sides — don't blindly accept yours or theirs.
2. Open conflicted files; read <<<<<<< markers and integrate intent, not just syntax.
3. Run tests after resolution — conflict markers removed isn't the same as correct merge.
4. git add resolved files; continue merge/rebase.

If conflicts are huge, the branch may have drifted too long — consider rebasing earlier next time or splitting the PR.

8. Keep history safe on shared branches

Never force-push (git push --force) to branches others build on — main, shared feature branches, release branches. It rewrites their base and can destroy unpushed work.

Safe to rewrite — your local-only branch, or your feature branch before anyone else pulled it (and team allows force-with-lease on feature branches).

Prefer:

• git revert on main — adds a commit that undoes a bad one; safe for shared history.
• git restore / git checkout -- — discard unstaged or uncommitted local changes.
• git reflog — recover "lost" commits after mistaken reset (local).

If you must undo a pushed commit on a shared branch, revert — don't reset and force-push.

9. Recover from mistakes

Uncommitted work in wrong direction

git stash push -m "describe work"
# or commit on a throwaway branch before experimenting

Wrong files staged

git restore --staged <file>
git restore <file>              # discard working copy if needed

Last commit wrong but not pushed

git commit --amend              # only if HEAD is yours and not pushed shared

Last commit pushed to shared main

git revert <bad-commit-sha>     # new commit that undoes it — don't force-push main

Lost commit after reset

git reflog
git checkout -b recovery <sha-from-reflog>

Cherry-pick one fix onto another branch:

git cherry-pick <sha>

Panic rule: stop before push --force; check git status, git log, and whether others use the branch.

10. Releases, tags, and hotfixes

Releases ([[launch-readiness]]):

• Tag main at the release point — v1.4.0, annotated message with changelog highlights.
• Release branches (release/1.4) if the project maintains patch lines.

Hotfixes:

1. Branch from the tag or release branch that production runs — not from random main tip if they've diverged.
2. Minimal fix + test; fast review.
3. Merge back to main and release branch so the fix isn't lost on next deploy.
4. Tag patch release (v1.4.1).

Document deploy order if migrations or config must land first ([[migration-path]]).

11. Scenario playbooks

Start a feature

Fetch → branch from latest main → small commits at green → push early for backup.

PR feedback

One commit per logical feedback round (or squash locally if team squashes on merge). Reply on threads; don't force-push without re-running CI.

Long-running branch

Merge or rebase main weekly minimum; smaller PRs upstream reduce pain.

CI failed after push

Reproduce locally → fix → one commit → push. Don't pile unrelated fixes.

Accidentally committed secret

Rotate secret immediately → remove from tree → never commit the new secret → team process for history if the old one was pushed.

Need to undo merged PR on main

git revert the merge commit (may need -m 1 for merge commits) → deploy revert → fix forward in new PR.

Experimental spike

Branch spike/name — may be messy locally; don't merge without cleaning history or squash per convention.

Common Rationalizations

• "I'll squash it all into one big commit." — A wall-of-changes commit is unreviewable and unbisectable.
• "The message can just be 'fix'." — Useless later; the why is what the reader needs.
• "Force-push is fine." — On a shared branch it rewrites everyone else's base and loses work.
• "I'll commit everything at the end." — Huge diffs hide mistakes and make rollback all-or-nothing.
• "Rebase is always better." — Only on your branch; rebase shared branches without agreement causes pain.
• "I'll amend the pushed commit." — Amending pushed commits on shared branches requires force-push.
• ".gitignore will catch it later." — Review the diff; ignored files don't help if already tracked.
• "Conflicts resolved — good enough." — Run tests; conflict resolution bugs ship often.

Red Flags

• Commits titled wip, fix, stuff, or asdf
• Single commit mixing unrelated refactor, feature, and dependency bump
• Secrets, build output, or large binaries in history
• Force-pushing main, master, or a branch others use
• Long-lived branches hundreds of commits behind main
• PR opened with no description or test plan
• Merging with failing CI or unresolved conflicts
• Revert needed but team force-pushed instead
• Hotfix only on production branch, never merged back to main
• Amending commits others already pulled

Verification

• Branch named for work, created from latest integration branch
• Each commit is one logical, working change ([[incremental-delivery]])
• Messages: imperative subject + why in body when non-obvious; ticket linked
• Diff reviewed — no secrets, artifacts, debug noise ([[hardening]])
• PR self-reviewed, description complete, CI green ([[review-gate]], [[pipeline-ops]])
• Merge/rebase strategy matches project convention; conflicts resolved and tested
• Shared history not rewritten — revert used on shared branches, not force-push
• Feature branch deleted after merge; hotfix merged back to main if applicable
• Release tagged and documented when shipping ([[launch-readiness]])