Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Perform git operations, manage branches, handle multi-remote setups, create PRs/MRs,
cut releases, and maintain audit-grade change history across GitHub, GitLab, and Forgejo.
The goal is clean, signed, traceable history that satisfies both engineering standards and
compliance requirements (PCI-DSS 4.0).
Target versions (July 2026):
git: 2.55.0 (current stable). Major additions include Linux fsmonitor, remote-group push, and parallel compatible hooks. Git 3.0 remains expected later in 2026.
GitHub CLI (gh): 2.96.0
GitLab CLI (glab): 1.109.0
Forgejo CLI (fj): 0.6.0 (verify the current release at codeberg.org/forgejo-contrib/forgejo-cli). Rust-written, official community CLI. Covers PRs (incl. AGit), issues, repos, releases, tags, actions.
Forgejo: v16.0.1 current; v15.0.5 is the current LTS. Critical RCE (CVE-2025-68937) patched in v13.0.2+.
prek: 0.4.10 (Rust, recommended) or pre-commit: 4.6.1 (Python, largest ecosystem)
git-filter-repo: 2.47.x
gitleaks: 8.30.x (secret scanning)
cosign: 3.1.2 (Sigstore, for tag/release signing context)
This skill covers five domains depending on context:
CI/CD pipeline design - use ci-cd; this skill handles git operations within pipelines, not pipeline architecture
PR/MR code review - use code-review; this skill creates PRs, doesn't review code in them
Full application security audit - use security-audit; this skill covers secrets in git history and git-specific security, not application-level vulnerability assessment
Docker image tagging strategy - use docker; this skill handles git tags, not container tags
Post-session or post-merge documentation sweeps (README, changelog, runbooks) - use update-docs
AI Self-Check
AI tools consistently produce the same git mistakes. Before performing any git operation,
verify against this list:
Read before rewrite. Never Edit/Write files without Reading them first. Never git commit without checking git status + git diff.
No destructive ops without confirmation.reset --hard, push --force, branch -D, clean -fd, checkout . - all require explicit user approval. Propose safer alternatives first (--force-with-lease, revert, new branch).
Authorship correct. Check the project's instruction file for author overrides. Many setups have a local git config that's wrong for the remote (e.g., Forgejo identity vs GitHub identity).
No secrets in commits. Check git diff --cached for API keys, tokens, passwords, .env files, private keys before committing. If found, unstage immediately.
No AI attribution in commits. Never add Co-Authored-By lines or any mention of AI tools in commits. No exceptions unless the project explicitly requires it.
Instruction-file policy respected. Follow the repo's policy for AGENTS.md or other instruction files. Never commit local agent state directories like .claude/, and never mention local-only tooling files in commit messages.
No tool-specific paths in commits..claude/, .cursor/, .superpowers/, .worktrees/, docs/local/, PLAN.md, SECURITY-AUDIT.md are local artifacts. Verify covers AI tooling artifacts BEFORE staging (.claude/, .cursor/, .codex/, PLAN.md)
Performance
Use path-limited git log, git diff, and git grep before whole-history scans.
Fetch only needed remotes/branches in large repos when full prune is unnecessary.
Keep commits small enough for review and bisect, but not so small that they split one behavior across many commits.
Best Practices
Prefer revert over history rewrite on shared branches.
Use signed commits/tags where the project or release process requires provenance.
Never run cleanup commands that delete branches, tags, or ignored files without showing the target set first.
Workflow
Step 1: Identify the forge and project context
Signal
Forge
.github/ directory, gh CLI
GitHub
.gitlab-ci.yml, glab CLI, user says "work"
GitLab
.forgejo/ directory, user says "home"
Forgejo
Multiple remotes in git remote -v
Multi-forge
Read the project's instruction file (AGENTS.md or equivalent) for:
Commit conventions (message format, scope list, authorship overrides)
"Pull/update all repos under a directory" -> Bulk repo update workflow (Step 3f)
"Set up signing" / "configure git" -> Configuration (references)
"Clean up history" / "scrub secrets" -> History rewriting (references)
"Set up branch protection" -> Protection rules (references)
Step 3a: Commit workflow
Check state: git status (never -uall on large repos) + git diff (staged and unstaged)
Stage selectively: git add <specific files> - never git add -A or git add . without reviewing what's included. Check for secrets, binaries, generated files, AI tooling artifacts.
Check the diff: git diff --cached - read what you're about to commit. Look for debug code, TODO comments, accidental changes to unrelated files.
Write the message: follow the project's convention. Default format:
type(scope): concise description
Optional body explaining *why*, not *what*.
Select type from: feat, fix, docs, refactor, chore, ci, build, revert.
Extended types (use if the project convention allows): test, perf, security.
Always check the project's instruction file and recent git log first - the project may have a
different type list or naming convention. Project instructions override these defaults.
Scopes are strongly recommended but not blindly enforced. A scope adds context that makes
git log --oneline scannable: fix(auth): token expiry vs fix: token expiry. Use the
component, module, or area name. Skip the scope only when the change is truly cross-cutting
or the project doesn't use scopes.
Breaking changes: append ! after the type/scope: feat(api)!: drop v1 endpoints.
For detailed breakage notes, add BREAKING CHANGE: in the commit body footer.
Message guidelines:
Imperative mood ("add feature", not "added feature" or "adds feature")
First line under 72 characters
Body wrapped at 72 characters
Explain why the change was made, not what changed (the diff shows what)
Reference issue/ticket numbers where applicable
Casual, human-sounding tone. Like a competent human wrote it, not an AI.
Good: fix(caddy): self-healing volume permissions, feat(auth): add OIDC login
Bad: "fix stuff", "wip", "update files", "enhance the robustness of the authentication module"
Plain ASCII only. No em-dashes, curly quotes, ligatures, or fancy punctuation in commit messages.
Never mention local instruction files or AI tooling in commit messages or bodies.
If you're updating local-only instruction files alongside code changes, don't include them in the commit.
Never add Co-Authored-By or AI attribution lines. No Signed-off-by: AI. No mentions
of specific AI tools in the commit metadata.
Commit: with any required authorship overrides and signing flags from project config.
Verify: git log -1 --format="%h %s%n Author: %an <%ae>%n Committer: %cn <%ce>" - check authorship is correct before pushing.
Step 3b: PR/MR workflow
First check: does the project use feature branches? Read the project instruction file. Some projects
(especially solo projects) work directly on main by preference. The default here assumes
collaborative development with review. Adapt to the project's actual workflow.
Read references/forge-workflows.md for forge-specific PR/MR creation
patterns (GitHub gh pr create, GitLab glab mr create, Forgejo web UI or API).
Make changes: commit early, commit often. Each commit should be a logical unit.
Rebase onto base: git fetch origin && git rebase origin/main (or whatever the base branch is). Resolve conflicts: look for <<<<<<<, =======, >>>>>>> markers, keep the correct content from both sides (often both additions belong), then git add and git rebase --continue. After resolving conflicts in dependency files (package.json, Cargo.toml, go.mod, etc.), re-run the package manager to regenerate the lock file. Never merge the base into the feature branch. For merge strategy guidance, see references/forge-workflows.md.
Dependency file conflicts (package.json, go.mod, pyproject.toml): keep additions from both branches; when the same dependency has conflicting version pins, take the higher compatible version conservatively and validate. Never hand-edit the lock file - always regenerate it by running the package manager (npm install, go mod tidy, uv sync, etc.) after resolving the manifest conflict. Do not use --ours or --theirs for dependency manifest conflicts - these silently drop one branch's additions.
Push: git push -u origin feat/short-description. Then verify the remote actually
advanced: compare git rev-parse HEAD with the forge PR/MR head SHA (gh pr view ...,
glab mr view ..., or forge API). Do not rely on a terse or aliased push wrapper alone.
If upstream tracking is missing or odd, push the exact refspec explicitly and re-check the
PR/MR head before claiming the update landed.
Create PR/MR: with a clear title (under 70 chars), body with summary + test plan.
Review cycle: address feedback with follow-up commits by default. Rebase and push with --force-with-lease only on a private feature branch after explicit user confirmation and a fresh fetch. Never rewrite a shared branch.
Branch naming convention:
feat/description - new features
fix/description - bug fixes
refactor/description - code restructuring
chore/description - maintenance, deps, config
docs/description - documentation only
ci/description - CI/CD changes
security/description - security fixes (consider if this should be a private advisory instead)
Step 3c: Release workflow
Read references/forge-workflows.md for forge-specific release creation.
Versioning scheme: semver vMAJOR.MINOR.PATCH with optional pre-release suffixes.
Version
When
CI behavior
v0.12.0-alpha.1
Start of new minor version, iterating on new features
Tag + release (marked pre-release). Image build optional per project.
v0.12.0-alpha.2
Continued iteration, not stable yet
Same as above.
v0.12.0
Features stable, ready for production
Full release: tag + release + image build + deploy.
v0.12.1
Patch fix on stable release
Full release.
Alpha workflow:
When starting work on a new minor version, tag vX.Y.0-alpha.1
Increment the alpha number (-alpha.2, -alpha.3) as you iterate
When stable, drop the -alpha suffix and release vX.Y.0
Alpha tags still get GitHub/Forgejo releases (marked --prerelease) so CI can optionally build images
The project instruction file defines whether alpha tags trigger image builds or not
General flow:
Version bump: update all version files (check the project instruction file for the list - every project is different).
Commit: chore: bump version to X.Y.Z (or chore: bump version to X.Y.Z-alpha.N)
Tag: git tag -a vX.Y.Z -m "vX.Y.Z" (annotated tags, not lightweight).
Push: push commits AND tags. git push origin main && git push origin vX.Y.Z. If multi-remote, push to all.
Verify: check that CI/CD picked up the tag and started the release pipeline.
Changelog generation: conventional commits feed tools like git-cliff, release-please,
or conventional-changelog to auto-generate changelogs from commit history. GitHub's
gh release create --generate-notes also works (uses PR titles since last tag). The !
breaking change marker and consistent scopes make these tools significantly more useful.
Golden rule: don't panic. Git almost never loses data. The reflog has 90 days of history.
Quick reference:
Undo last commit (keep changes): git reset --soft HEAD~1
Undo last commit (discard changes): git reset --hard HEAD~1 - DESTRUCTIVE, confirm first
Revert a pushed commit: git revert <sha> (creates a new commit, safe for shared branches)
Find lost commits: git reflog - shows every HEAD movement for 90 days
Find which commit broke something: git bisect start && git bisect bad && git bisect good <sha>
Automated bisect with test script: git bisect start HEAD v1.0.0 && git bisect run bun test - src/auth.test.ts - runs the test at each bisect step automatically. Exit codes: 0 = good, 1-127 except 125 = bad, 125 = skip this commit, 128-255 = abort the bisect immediately. Ideal for CI integration: git bisect run ./scripts/ci-check.sh
Squash last N commits (no interactive rebase): git reset --soft HEAD~N && git commit -m "feat: combined change" - resets N commits but keeps all changes staged, then commits them as one. Safer than git rebase -i in automated contexts.
Scrub secrets from history: set GIT_FILTER_REPO_REPLACEMENTS to a mode-0600 temporary replacement file populated from a secret manager or no-echo stdin, then run : "${GIT_FILTER_REPO_REPLACEMENTS:?set a mode-0600 replacement file}" followed by git filter-repo --replace-text "$GIT_FILTER_REPO_REPLACEMENTS". Remove the temporary file normally after use. Never put the leaked value in argv, command text, or shell history. Then force-push ALL branches and tags, coordinated with the team. See references.
Step 3e: Diverged branch resolution
When git pull fails with "divergent branches" or git status shows "have diverged":
Understand the divergence: git log --oneline HEAD..origin/branch (what remote has) and git log --oneline origin/branch..HEAD (what you have locally).
Choose strategy based on context:
Your commits are the ones that matter (common for solo work): git pull --rebase origin branch - replays your local commits on top of remote.
Remote commits are the ones that matter (someone force-pushed, or you want to discard local): git reset --hard origin/branch - DESTRUCTIVE, confirm first.
Both sides have real work (collaboration divergence): git pull --rebase origin branch, resolve conflicts at each step, git rebase --continue.
You don't know yet: git stash && git pull && git stash pop as a safe first attempt (only works if the divergence is minor).
After resolving: verify with git log --oneline -10 that history looks correct.
Prevent recurrence: if this was caused by force-push on a shared branch, set up branch protection. If caused by forgetting to pull before committing, consider git config pull.rebase true for rebase-on-pull default.
Step 3f: Bulk repo update workflow
When asked to pull or update many repositories under a directory, keep the run sequential,
auditable, and conservative:
Discover only real repo roots first, usually direct children of the requested directory
unless the user explicitly asks for a recursive scan. Avoid vendored repos, worktrees,
dependency caches, and generated checkouts.
Process repos sequentially, not in parallel. This avoids hammering SSH remotes, rate
limits, or firewall protections.
For each repo, record the branch, upstream, and dirty state before changing anything.
Run git fetch --all --prune, then git pull --ff-only --recurse-submodules only
when an upstream exists. Do not create merge commits during bulk maintenance.
When submodules are present, also run git submodule update --init --recursive after
the pull - --recurse-submodules updates existing submodule checkouts but does not
initialize new submodules added since the last pull.
If a pull is blocked by local changes, use a named stash, fast-forward the branch,
then pop the stash. If conflicts appear, preserve both upstream updates and local
user additions when possible.
After resolving stash conflicts, run git restore --staged . so previously unstaged
user changes do not stay staged accidentally. Keep the stash entry if conflict
resolution required judgment, and mention it in the final response.
Verify every repo with git rev-list --left-right --count HEAD...@{u} and
git status --porcelain. Report repos that remain dirty or have no upstream.
Multi-Forge Patterns
Many projects use multiple git remotes (e.g., Forgejo for self-hosted CI + GitHub for public releases,
or GitLab at work + GitHub for open source mirrors).
Keep origin as the primary development remote and name mirrors explicitly.
Push to multiple remotes explicitly; do not make "push everywhere" the default.
Use separate SSH keys per forge.
Remember that CLI auth can be overridden by environment variables, especially GITHUB_TOKEN.
Detailed multi-forge workflows stay in references/forge-workflows.md.
Commit Signing
Sign commits because unsigned authorship is trivial to spoof.
SSH signing is the default recommendation for most teams.
GPG still matters where long-lived key lifecycles or policy require it.
gitsign is attractive for low-management open-source flows, but it changes the trust model.
Detailed signing setup stays in references/security-and-signing.md.
Security
Keep credentials out of the repo, including .env files, keys, and embedded connection strings.
Use secure credential helpers, not plaintext storage.
Use hooks for secrets, lint, and branch protection, but treat untrusted hook configs as code execution surfaces.
If a secret hit history, scrub it and rotate it. History rewriting does not undo the exposure.
Keep .gitignore and .gitattributes intentional; AI tooling artifacts belong out of the repo by default.
Detailed security and signing guidance stays in references/security-and-signing.md.
PCI-DSS 4.0: Change Management Mapping
Source code management requirements that map to git practices.
PCI-DSS Req
What it means for git
Implementation
6.2.2
Annual security training for developers
Not a git control, but signed commits prove who was trained
6.2.4
Access control + change tracking
Branch protection, required reviewers, signed commits, audit trail
6.4.1
Separate dev/test/prod environments
Branch-per-environment or tag-based promotion
6.4.2
Changes approved, documented, tested
PR/MR with required approvals, linked CI checks, merge audit log
6.5.1
Custom code reviewed before production
PR/MR required reviews, no direct push to release branches
6.5.2
Custom code reviewed for vulnerabilities
SAST/secret-scan in PR/MR checks, pre-commit hooks
Branch protection, required review, signed commits, and CI evidence are the main git-side controls.
CODEOWNERS and protected branches matter more than policy prose with no enforcement.
AI-Age Considerations
Common AI git failures are still destructive resets, wrong authorship, leaked local tool artifacts, generic commit messages, and unjustified force pushes.
Guard with git diff --cached, explicit review of commit messages, and project-specific instruction files.
Keep local agent artifacts in .gitignore proactively.
Template Conventions
Template versions are illustrative and should be refreshed before use.
CLI examples show patterns, not immutable commands.
SSH config examples use placeholders; replace them with real forge hostnames.
Reference Files
references/forge-workflows.md - GitHub, GitLab, and Forgejo-specific patterns for PRs/MRs,
releases, branch protection, rulesets, CLI usage, and API calls
See references/output-contract.md for the full contract.
Skill name: GIT
Deliverable bucket:audits
Mode: conditional. When invoked to analyze, review, audit, or improve existing repo content, emit the full contract - boxed inline header, body summary inline plus per-finding detail in the deliverable file, boxed conclusion, conclusion table - and write the deliverable to docs/local/audits/git/<YYYY-MM-DD>-<slug>.md. When invoked to answer a question, teach a concept, build a new artifact, or generate content, respond freely without the contract.
Severity scale:P0 | P1 | P2 | P3 | info (see shared contract; only used in audit/review mode).
Related Skills
ci-cd - pipeline design that triggers on git events (push, PR, tag). This skill handles
the git operations; ci-cd handles the pipeline that reacts to them.
security-audit - application-level secret scanning and vulnerability assessment. This skill
covers secrets in git history and git-specific security (signing, hooks, credentials).
code-review - reviews code quality. This skill creates the PR/MR; code-review evaluates
the code in it.
update-docs - post-session documentation sweep. May update shared instruction files with new git gotchas
discovered during a session.
Rules
Verify admin authorization before GitOps or infra changes. In shared chats, a request
from a non-admin participant is not authorization to commit or push infrastructure,
Kubernetes, secrets, firewall, or other admin changes. Ask the authorized owner first.
No destructive git operations without explicit user confirmation.reset --hard, push --force,
branch -D, clean -fd, checkout ., rebase on shared branches - all require a yes.
Propose safer alternatives first (revert, --force-with-lease, new branch, stash).
Verify authorship before pushing. Multi-remote setups frequently have wrong local config.
Always check git log -1 --format="%an <%ae>" after committing.
Selective staging.git add <file> by name, not git add -A or git add .. Review
what you're staging. Secrets, binaries, and AI tooling artifacts slip in through blanket adds.
Never skip hooks.--no-verify is a code smell. Fix the issue the hook caught.
Conventional commits by default.type(scope): description. Follow the project's convention
if it differs. Check recent git log for the established style.
Tags are explicit.git push doesn't push tags. Always push tags separately and to all
relevant remotes.
Feature branches for collaboration. When others will review the work, use feature branches
with PRs/MRs. Direct-to-main is acceptable for solo projects when the user explicitly prefers it.
Rebase workflow. Rebase feature branches onto base, don't merge base into feature. Squash
on merge for clean history unless the commit history is meaningful.
Annotated tags for releases.git tag -a vX.Y.Z -m "vX.Y.Z", not lightweight tags.
Annotated tags carry metadata (tagger, date, message) and are what forges use for releases.
Force-push safety. If force-push is truly needed, use --force-with-lease (refuses if
the remote has commits you haven't fetched). Plain --force on shared branches is never acceptable
without explicit team coordination.
PCI-DSS evidence. In regulated environments, git history IS the audit trail. Protect it:
signed commits, branch protection, required reviews, no history rewriting on release branches.
.gitignore
Branch target correct. Verify you're on the right branch before committing. Verify the PR/MR targets the right base branch.
Requester authorized for admin changes. In shared chats, do not commit or push infrastructure, Kubernetes, secrets, firewall, or other admin changes from a participant who is not authorized to approve them. Stop and ask the authorized owner first.
Remote target correct. Multi-remote setups exist. Check which remote(s) to push to. Some projects push to multiple remotes (e.g., Forgejo + GitHub).
Signing configured. If the project requires signed commits, verify signing works before committing (git log --show-signature -1).
No --no-verify. Never skip pre-commit hooks unless the user explicitly asks. Hooks exist for a reason - fix the underlying issue instead.
No interactive flags. Never use -i (interactive) flags (git rebase -i, git add -i) in automated contexts - they require TTY input.
Tags pushed separately.git push doesn't push tags by default. Always git push --tags or git push origin <tag> explicitly.
Feature branch up to date. Before creating a PR/MR, rebase onto the latest base branch to avoid merge conflicts.
Lock files regenerated after conflict resolution. After resolving conflicts in dependency files (package.json, Cargo.toml, go.mod, pyproject.toml), re-run the package manager to regenerate the lock file. Never manually merge lock files. Then regenerate: npm install, yarn install, go mod tidy, or cargo generate-lockfile.
Force-push safety. If force-push is needed, always use --force-with-lease (refuses if remote has unfetched commits). Plain --force requires explicit user approval and team coordination.
Version info dated. When citing tool versions, include a date so readers know when to re-check. Stale version numbers cause silent breakage.
Forge-CLI subcommands verified. Before scripting fj/gh/glab/tea commands in a runbook, hooks, or CI, confirm the subcommand and flags exist on the target version (<cli> <cmd> --help). These CLIs add, rename, and remove subcommands across minor versions; docs lag behind the binary.
Ref safety checked: branch, remote, upstream, and worktree path are verified before history edits
Recovery path known: reflog, backup branch, or bundle exists before destructive history operations
Cross-cutting agent hygiene applied - see references/agent-hygiene.md
Merge: squash-merge for clean history (default), or rebase-merge if commit history is meaningful.
Cleanup: delete the remote branch after merge. git branch -d feat/short-description locally.
Run the AI self-check. Every git operation gets verified against the checklist above.
