Agent Skills: Git Flow Best Practices (2026)

Use when committing, branching, opening PRs, or deciding merge/branch strategy.

UncategorizedID: fusengine/agents/git-flow

Install this agent skill to your local

pnpm dlx add-skill https://github.com/Fusengine/claude-plugins/tree/HEAD/plugins/commit-pro/skills/git-flow

Skill Files

Browse the full folder contents for git-flow.

Download Skill

Loading file tree…

plugins/commit-pro/skills/git-flow/SKILL.md

Skill Metadata

Name
git-flow
Description
Use when committing, branching, opening PRs, or deciding merge/branch strategy.
<objective> Covers Git workflow strategy end to end: the GitHub Flow (default) vs trunk-based vs Git Flow tradeoff table, the branch naming convention (`<type>/<scope-or-summary>`, kebab-case, <50 chars, no personal prefix), protected-branch enforcement (main/master/develop/production never committed to directly, auto-named feature branch proposed instead), the full branch lifecycle (create → work → push → PR → review → merge → delete), merge strategy (real merge commit is fuse-commit-pro's default; rebase vs squash tradeoffs — squash is unused here because it would orphan the release tag's target commit), the CI-gate-before-merge decision tree (three cases by whether required checks are configured, including the `gh pr checks` registration-race workaround), post-merge tag timing, the PR description template, anti-patterns, and solo-dev mode. </objective>

Git Flow Best Practices (2026)

Workflow Choice

| Strategy | When | Verdict | |----------|------|---------| | Trunk-based (direct main) | Solo dev, prototypes, strong CI | OK if you have automated tests | | GitHub Flow (feature branch → PR → merge → delete) | Teams, OSS, code review | ✅ Default | | Git Flow (develop/release/hotfix) | Heavy release cycles | ❌ Outdated for most projects |

fuse-commit-pro default: GitHub Flow.

Branch Naming Convention

Format: <type>/<scope-or-summary> (kebab-case).

| Type | Use | Example | |------|-----|---------| | feat/ | New feature | feat/seo, feat/oauth-google | | fix/ | Bug fix | fix/sniper-loop, fix/csv-parser | | chore/ | Maintenance, deps | chore/bump-deps, chore/rename-files | | docs/ | Documentation | docs/api-reference | | refactor/ | Refactoring (no behavior change) | refactor/extract-utils | | perf/ | Performance | perf/db-indexes | | test/ | Tests only | test/auth-coverage | | ci/ | CI/CD config | ci/github-actions-cache | | build/ | Build system | build/vite-config | | style/ | Formatting | style/prettier-pass |

Rules:

  • kebab-case only (no underscores, no spaces, no caps)
  • < 50 chars total
  • No personal prefix (bruno/...) — collaborators don't know who you are 6 months later
  • No issue number alone (fix/123) — meaningless once issue closed

Protected Branches

main, master, develop, production → never commit directly.

fuse-commit-pro:commit enforces this in Step 0:

  • Detects current branch
  • If protected → blocks + proposes auto-named feature branch from commit type/scope
  • Exceptions: solo prototype (no remote), explicit --no-branch-check, or post-commit version bump

Branch Lifecycle

1. git checkout -b feat/<scope>      # create
2. work + commit                      # multiple commits OK
3. git push -u origin feat/<scope>    # push with upstream
4. gh pr create                       # open PR
5. (review + CI)
6. gh pr merge --merge --delete-branch  # merge + cleanup

Keep branches short-lived (< 3 days ideally). Long-lived branches accumulate conflicts and lose context.

Merge Strategy

| Strategy | When | Result | |----------|------|--------| | Merge commit | fuse-commit-pro default | Branch commits (incl. the bump commit) land on main intact, no rewrite | | Rebase merge | Small atomic commits worth preserving, no merge commit wanted | Linear history, individual commits kept | | Squash merge | not used here: the release tag points at the bump commit, squash would orphan it | 1 commit per feature, but incompatible with fuse-commit-pro's post-merge tagging |

fuse-commit-pro recommendation: real merge commit via gh pr merge --merge --delete-branch (see commands/commit.md Step 7).

CI Gate Before Merge

Cardinal rule: never merge before CI checks are resolved; never assume "zero CI" without verifying.

Known race condition (cli/cli#7401, confirmed by GitHub CLI maintainers): checks take a few seconds to register on a freshly-created PR. Calling gh pr checks --watch immediately after gh pr create can error out — exit code 1, no checks reported on the '<branch>' branch — even though checks are about to appear. The GitHub API gives no way to tell "genuinely zero checks" apart from "not registered yet," so the client must either sidestep the race (native auto-merge) or poll for registration before watching.

Determine which of the three cases applies from what actually exists on the PR — never from assumption. The key branch point is whether the repo enforces required status checks in branch protection, not merely whether auto-merge is available: gh pr merge --auto only ever waits for required checks. On a repo where checks run but aren't marked required (e.g. this repo — CodeQL + Analyze execute but aren't required), --auto merges immediately without waiting, silently reproducing the exact race this section exists to close. Check first with gh pr checks <pr> --required (empty output / no required checks reported on the '<branch>' branch → no required checks configured) before picking a branch:

  1. Repo has required status checks in branch protection → the only case where --auto actually gates. GitHub waits server-side for the required checks to appear and pass — no race, regardless of their state at call time:
    gh pr merge <pr> --auto --merge --delete-branch
    
    Also requires "Allow auto-merge" enabled in repo Settings → General (otherwise GraphQL: Auto merge is not allowed for this repository).
  2. Repo has checks but none are required (this repo's current state — CodeQL/Analyze run but aren't required) — --auto would merge immediately without waiting, so don't use it here. Poll client-side until checks register, then watch: don't --watch right after gh pr create/a push. Poll gh pr checks until it stops reporting "no checks reported" (bounded retries), THEN watch. The poll's pipe to grep only detects check registration — it never gates the merge, which still runs un-piped via &&:
    pr=<pr>
    max_attempts=18   # 18 × 5s ≈ 90s — raise if this repo's CI is known to queue slower
    attempt=0
    until gh pr checks "$pr" 2>&1 | grep -qv "no checks reported"; do
      attempt=$((attempt + 1))
      if [ "$attempt" -ge "$max_attempts" ]; then
        echo "Timeout: no checks registered after $((max_attempts * 5))s — verify .github/workflows/ before treating this as zero-CI." >&2
        break
      fi
      sleep 5
    done
    gh pr checks "$pr" --watch && gh pr merge "$pr" --merge --delete-branch
    
    Never pipe the final gh pr checks <pr> --watch call itself (e.g. | tail): a pipe swallows the exit code and lets a merge proceed after failing CI. Chain with && so the merge only runs if checks passed.
  3. Repo has zero CI checks (verified, not assumed — e.g. no .github/workflows/* and the poll above timed out) → immediate merge is allowed:
    gh pr merge <pr> --merge --delete-branch
    

A non-required check never blocks the merge — only required checks actually gate. Mark CI checks as required in branch protection so case 1's --auto has teeth; without that, checks stay in case 2 forever.

Note on --required: the poll loop in case 2 intentionally queries gh pr checks without --required — it only needs to know "has any check registered yet." If a future revision adds --required to that same loop, the zero-checks message becomes no required checks reported on the '<branch>' branch, which does not contain the substring no checks reported (the inserted word "required" breaks the contiguous match) — the grep would need to change to checks reported on the (matches both variants) or handle both strings explicitly.

Merge is always --merge (real merge commit) — never --squash, it would orphan the release tag's target (see Tagging timing below).

Tagging timing: never push the tag before the merge is validated — CI could still fail or branch protection could still block the merge, and a tag pushed early would point at a commit that never lands on main. Tag vX.Y.Z on main AFTER the merge completes, then push the tag (fuse-commit-pro:commit does this automatically in Step 8 — see also commands/commit.md Step 8 and post-commit/references/tag-timing.md).

After the Tag — Release Workflow

Pushing the tag is not the end of a release: a tag-triggered workflow (e.g. Publish to npm) must be watched to the end, reruns included. Full procedure: commands/commit.md Step 9. In short:

  • Find the runs: gh run list --commit "$(git rev-list -n1 vX.Y.Z)" --event push (poll ~90s, they register late).
  • Watch each: gh run watch <id> --exit-status — never piped.
  • Failure → read gh run view <id> --log-failed at once and classify. Flaky (timing/perf threshold just over, network/registry timeout, runner crash) → gh run rerun <id> --failed, watch again, max 2 reruns then escalate. Real defect (compile/type/lint, deterministic test, auth, same failure twice) → stop and report the cause.
  • Say "published" only when the final attempt is fully green (post-publish verification jobs included) and the registry serves the version (npm view <pkg> version == X.Y.Z).
  • The release runs through the fuse-ai-pilot:commit agent launched by the lead — never through a code delegate (Kimi, Codex, Cursor): they write code, not releases.

After Merge

  • Delete branch automatically (--delete-branch flag or GitHub auto-delete setting)
  • Pull main: git checkout main && git pull --ff-only origin main
  • Delete local: git branch -d feat/<scope> (automatic if --delete-branch used remotely)

Pull Request Template

## Summary
<1-3 bullet points of what changed>

## Changes
<list of major files/components touched>

## Test plan
- [ ] Manual test on X
- [ ] CI green
- [ ] Sniper validation clean

## Breaking changes
None / <description with migration path>

Anti-Patterns

  • ❌ Long-lived feature branches (> 1 week) — rebase early or split
  • ❌ Commits on main "to save time" — bypasses review, breaks CI gates
  • ❌ Force push to main — never. Forbidden in fuse-commit-pro.
  • ❌ Branch named wip, temp, test123 — meaningless, can't be found later
  • ❌ PR without description — reviewers can't context-switch
  • ❌ Merging your own PR without review (when working in a team)
  • ❌ Stale branches (no commits > 30 days) — delete or close

Solo Dev Mode

If you're alone on a repo with no PR review:

  • Still use feature branches (rollback safety net)
  • Self-PR is fine for sanity check (you'll see the diff fresh)
  • Or commit-then-push on main with strong CI as safety
  • fuse-commit-pro:commit detects no-remote case and skips Step 7