git-best-practices

name	git-best-practices
description	Git workflow patterns for commits, branching, PRs, and history management across heterogeneous repositories. Use when creating commits, managing branches, opening pull requests, or rewriting history. Do not use for non-git implementation tasks or repo-specific release policy decisions without repository documentation.

Git Best Practices

Always Active Principles

When this skill is loaded, follow these directives for all git operations:

1. Discover before acting — run branch discovery to determine the repo's default and production branches before branching, merging, or opening PRs
2. Conventional commits — every commit uses type(scope): description format
3. Stage explicitly — add files by name so only intended changes are committed
4. Protect shared history — use --force-with-lease for force pushes; confirm with the user before any force push

Agent Git Workflow

Follow this sequence when performing git operations:

1. Check state — run git status and git diff HEAD; output: working tree and unstaged/staged delta
2. Discover branches — identify and store default/current/(optional) production branch names (see Branch Discovery)
3. Stage by name — git add path/to/file for each file; verify with git status
4. Write a conventional commit — type(scope): description with optional body
5. Push safely — use regular push by default; use git push --force-with-lease origin {branch} only for rewritten history and only after user confirmation

Checkpoint Commits

Agents may create WIP checkpoint commits during long-running tasks. These are development artifacts, cleaned up before PR.

• Prefix with wip: or use standard conventional commit format
• Keep changes logically grouped even in WIP state
• Run /rewrite-history before opening a PR to craft a clean narrative

Commit Discipline

• Stage files explicitly by name: git add src/auth.ts src/auth.test.ts
• Verify staged content with git status before committing
• Keep secrets, .env files, credentials, and large binaries out of commits — warn the user if staged files look sensitive
• Target one logical change per commit in final PR-ready state

Force Push

Use --force-with-lease exclusively to protect against overwriting upstream changes:

git push --force-with-lease origin feat/my-branch

Always confirm with the user before any force push, regardless of branch.

Conventional Commits

Format: type(scope): description

Subject line rules:

• Lowercase, imperative mood, no trailing period
• Under 72 characters
• Scope is optional but preferred when a clear subsystem exists

Common types:

Type	Use for
feat	New functionality
fix	Bug fix
docs	Documentation only
refactor	Restructuring without behavior change
perf	Performance improvement
chore	Maintenance, dependencies, tooling
test	Adding or updating tests
ci	CI/CD pipeline changes
build	Build system changes
style	Formatting, whitespace (no logic change)

Commit Bodies

Body is optional — only add one when the change is genuinely non-obvious. The subject line carries the "what"; the body explains "why."

Add a body when:

• The motivation or tradeoff is non-obvious
• Multi-part changes benefit from a bullet list
• External context is needed (links, issue references, root cause)

Examples

Single-line fix, no body needed:

fix(shell): restore Alt+F terminal navigation

Non-obvious fix with body explaining root cause:

fix(shell): use HOMEBREW_PREFIX to avoid path_helper breaking plugins in login shells

macOS path_helper reorders PATH in login shells, putting /usr/local/bin
before /opt/homebrew/bin. This caused `brew --prefix` to resolve the stale
Intel Homebrew, so fzf, zsh-autosuggestions, and zsh-syntax-highlighting
all silently failed to load in Ghostty (which spawns login shells).

Use the HOMEBREW_PREFIX env var (set by brew shellenv in .zshenv) instead
of calling `brew --prefix` — it survives path_helper and is faster.

Feature with bullet-list body for multi-part changes:

feat(install): add claude bootstrap runtime management

- migrate Claude defaults to declarative files under claude/defaults
- add claude-bootstrap check/fix/uninstall with backup-first migration
- stop stowing full claude/codex runtime trees and tighten drift checks

Monorepo commit with ticket reference in branch and scope:

fix(pool-party): handle stale settlement state on reconnect

PoolSettlement contract stays in pending state when the participant
disconnects mid-settlement. Check settlement timestamp and expire
stale entries on reconnect.

Fixes SEND-718

Submodule update with downstream commit info:

chore(submodule): update claude-code

Bump claude-code to 88d0c75 (feat(skills): add tiltup, specalign, and e2e skills).

For trivial bumps, bump or bump claude-code submodule is acceptable.

Breaking change using `!` suffix:

refactor(api)!: change auth endpoint response format

The /auth/token endpoint now returns { access_token, expires_in }
instead of { token, expiry }. All clients must update their parsers.

Branch Discovery

Before branching or opening a PR, discover the repo's branch topology. Run these commands and store the results:

# Default branch (PR target for most repos)
gh repo view --json defaultBranchRef --jq '.defaultBranchRef.name'

# Current branch
git branch --show-current

# Production branch (if different from default)
git branch -r --list 'origin/main' 'origin/master' 'origin/production'

Fallback when gh is unavailable or the repo has no remote:

# Infer default branch from local refs
git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@'

# Last resort: check local branches and fail loudly if unknown
if git rev-parse --verify main >/dev/null 2>&1; then
  echo main
elif git rev-parse --verify master >/dev/null 2>&1; then
  echo master
else
  echo "ERROR: unable to determine default branch (main/master not found)." >&2
  exit 1
fi

Store the discovered branch name and reference it throughout. Use the actual branch name in all subsequent commands.

Branch Naming

Use repository branch naming conventions first. If no convention is documented, use:

Format: type/description-TICKET-ID

Examples:

• feat/add-login-SEND-77
• fix/pool-party-stall-SEN-68
• chore/update-deps
• hotfix/auth-bypass

Include the ticket ID when an issue exists. Omit when there is no ticket.

Branch Flow

Use repository branch flow policy first. If policy is undocumented, a common baseline is:

{production-branch} (production deploys)
 └── {default-branch} (staging/testnet deploys, PR target)
      ├── feat/add-feature-TICKET
      ├── fix/bug-description-TICKET
      └── hotfix/* (branches off production branch for hotfixes)

• Feature and fix branches start from the default branch
• Hotfix branches start from the production branch
• PRs target the default branch unless the repo uses a single-branch flow
• When default branch and production branch are the same, all PRs target that branch directly

Merge Strategy

Use repository merge policy first (required in many organizations).

If no policy exists, these defaults are reasonable:

PR target	Strategy	Rationale
Feature → default branch	Squash merge	Clean history, one commit per feature
Default → production	Merge commit	Preserves the release boundary; visible deploy points
Hotfix → production	Squash merge	Single atomic fix on production

PR Workflow

Sizing

Pragmatic sizing over arbitrary limits. Each commit tells a clear story regardless of PR size. A PR should be reviewable as a coherent unit — if a reviewer cannot hold the full change in their head, consider splitting.

PR Creation

Use repo-native PR tooling (gh pr create, GitLab CLI, or web UI) with:

• Short title under 70 characters
• Summary section with 1-3 bullet points
• Test plan as a bulleted checklist

History Rewriting Before PR

For branches with messy WIP history, use /rewrite-history to:

1. Backup the branch
2. Reset to the base branch tip
3. Recommit changes as a clean narrative sequence
4. Verify byte-for-byte match with backup
5. Confirm with the user before force-pushing rewritten history
6. Open PR with link to backup branch

Each rewritten commit introduces one coherent idea, building on the previous — like a tutorial teaching the reader how the feature was built.