glebis/claude-skills/repo-publish

Repo Publish

End-to-end pipeline for taking a project from "I made changes" to "live on a remote". Runs quality gates, generates an intelligent multi-commit history, pushes, and verifies the deploy.

When to Use

Use this skill when:

• The user says "publish", "release", "ship it", "push to remote", "deploy this"
• A project has accumulated local changes that need to land cleanly
• A version bump or release needs to be cut
• The user explicitly asks to run the full landing-the-plane workflow

Do not use this skill for:

• Single quick commits the user is doing themselves
• Documentation-only edits (use the project's own docs workflow)
• Anything involving secrets/credentials (the scripts will skip those files)

Workflow Decision Tree

Start here. Pick the lane based on the user's intent:

User wants to publish
  ├─ Has local changes? ──── no ──► Just push / bump version
  ├─ Want a release? ──── yes ──► [Release mode](#release-mode)
  └─ Just routine commit? ──────► [Push mode](#push-mode)

Push Mode

For ordinary commits that need to land.

Step 1 — File remaining work

Run the issue tracker step:

bd ready              # What's pending
bd show <id>          # Inspect any blockers
bd update <id> --status in_progress  # Claim work

If you can't bd, ask the user how they track issues for this project. Don't silently skip it — issue hygiene is part of the gate.

Step 2 — Run quality gates

Inspect package.json, pyproject.toml, Makefile, or whatever exists and run whatever the project defines as "test + lint + build":

# Node
npm test && npm run lint && npm run build
 
# Python
pytest && ruff check . && mypy .
 
# Go
go test ./... && go vet ./...
 
# Rust
cargo test && cargo clippy -- -D warnings
 
# Makefile-driven
make test lint build

If a gate fails, stop. Report which gate failed and the relevant error. Don't try to fix the code in the same publish flow — open a fix-PR first.

For projects without an obvious test command, run ls to find the closest script and ask the user.

Step 3 — Stage and commit intelligently

Use the scripts/intelligent_commit.sh helper. It groups files by change type (features, fixes, docs, chore) and produces one commit per group:

~/.agents/skills/repo-publish/scripts/intelligent_commit.sh

Read its docs at references/commit-grouping.md for the heuristics.

If the user prefers a single commit, ask explicitly before combining.

Step 4 — Pull rebase, push, verify

git pull --rebase        # Resolve any rebase conflicts before pushing
git push                 # Pushes the working branch
git status               # MUST show "Your branch is up to date"

If push is rejected: there's a remote-side change. Pull rebase again, resolve, retry. Never --force without explicit user consent.

Release Mode

For cutting versions. Adds three extra steps before push.

Step R1 — Choose bump type

Ask the user: patch (x.y.Z → x.y.(Z+1)), minor (X.Y → X.(Y+1)), or major. Default to patch unless they say otherwise. Read the change log to confirm.

Step R2 — Bump version

# Node
npm version <patch|minor|major>  # Updates package.json + creates tag
 
# Python (with bumpversion or similar)
bumpversion <patch|minor|major>
 
# Manual: edit pyproject.toml / VERSION / package.json by hand, then:
git tag -a v<VERSION> -m "Release v<VERSION>"

Step R3 — Generate / update CHANGELOG

If the project has a CHANGELOG.md, append a new section for this version with bullet points summarizing the commits in git log --oneline v<PREV>..HEAD. If it doesn't, skip — don't create empty changelogs.

Step R4 — Commit the release

git add package.json CHANGELOG.md  # or equivalents
git commit -m "release: v<VERSION>"
git push --follow-tags

The --follow-tags pushes the new tag along with the commit.

Step R5 — Post-release verification

git tag --list "v<VERSION>"   # Tag should appear
git ls-remote --tags origin    # Verify remote has the tag

For projects with deployment automation (Vercel, npm publish, etc.), the user usually has a manual post-push step. Ask, don't assume.

Common Pitfalls

• Never `--force` push without explicit user consent and a clear reason. It rewrites shared history.
• Don't run tests in a publish flow that you can't read the output of. If the runner is opaque, surface the URL or command to the user.
• Don't commit secrets. .env, *.pem, id_rsa, etc. Check git status for these before staging.
• If `bd` (or whatever tracker) has open issues, file the remaining work as a new issue and reference the ID in the commit before pushing.
• If multiple branches diverge, the right move is bd sync followed by a fresh git pull --rebase. Don't try to force-push your way out of sync.

Environment Variables

None required. The skill uses whatever git is configured to use for push. If the user needs a specific remote credential, they'll have it set in their git config.

Related Workflows

• Single-commit hotfix → skip intelligent_commit.sh, use git add -p + a single git commit
• Bisecting a regression → not the publish skill's job; use git bisect directly
• Pre-release testing → use the project's own staging environment before invoking this skill