Setup CI

setup-ci

CI/CD is one of the hardest topics for non-developers because it sits at the intersection of git, servers, authentication, and secrets management. Most tutorials jump straight to YAML. This skill starts earlier.

Your default assumption: the operator has never read a workflow file before. Don't patronize them, but don't skip concepts either.

State assumptions — before acting

Before starting the procedure, write an explicit Assumptions block. Don't pick silently between interpretations; surface the choice. If any assumption is wrong or ambiguous, pause and ask — do not proceed on a guess.

Required block:

Assumptions:
- Stack:             <detected from package.json/Cargo.toml/go.mod/requirements.txt — single stack or monorepo>
- Deploy target:     <none | SSH | Vercel | Fly.io | Cloud Run | Netlify | GitHub Pages>
- Existing CI:       <none | partial (.github/workflows/*.yml with N workflows) | full>
- Secrets state:     <to be configured by operator in GitHub UI — this skill never handles credentials>

Typical items for this skill:

• The detected stack (from package.json / requirements.txt / Cargo.toml / go.mod)
• The deploy target (SSH / Vercel / Fly.io / Cloud Run / Netlify / GitHub Pages / none)
• Whether Secrets are already configured in the GitHub UI — the skill never creates them, only lists what's needed

Stop-and-ask triggers:

• Multiple stacks coexist (e.g., Next.js frontend + Python backend) without clear monorepo structure — ask which one this workflow is for
• No deploy target detectable from config — ask explicitly, never guess

Silent picks are the most common failure mode: the skill runs, produces plausible output, and the operator doesn't notice the wrong interpretation was chosen. The Assumptions block is cheap insurance.

Step 1 — Figure out what the operator actually wants

Before generating any YAML, ask the operator two questions, one at a time. Don't ask all at once.

Question 1: "What happens when you push code today? Do you deploy manually, or is something already automated?"

Their answer tells you:

• "I SSH in and run git pull" → They have a server and a manual deploy. You're replacing the manual step.
• "I use Vercel / Netlify / Fly.io" → They have a managed platform. CI is about tests before deploy, the deploy is already automatic.
• "Nothing, it's only on my laptop" → They need a host choice first. Suggest the simplest option for their stack (Vercel for web frontends, Fly.io for full apps, a VM for anything complicated).
• "What's a deploy?" → Start from the very beginning. Go to the concepts section below.

Question 2: "Which git host are you on? GitHub, GitLab, Bitbucket, self-hosted?"

Their answer tells you which CI system you're writing for:

• GitHub → GitHub Actions
• GitLab → GitLab CI (.gitlab-ci.yml)
• Bitbucket → Bitbucket Pipelines (bitbucket-pipelines.yml)
• Self-hosted / "I don't know" → Default to GitHub Actions since the operator almost certainly knows GitHub

Do not generate YAML until both questions are answered. The #1 failure mode here is producing a GitHub Actions workflow for someone who's actually on GitLab.

If the operator says "just pick whatever's easiest," pick: GitHub + GitHub Actions + their current deploy target. Tell them explicitly that you picked that and why.

CI pipeline components — required, recommended, optional

CI/CD is modular. Not every project needs every component. Use this tree to decide what to scaffold for a given repo.

CI Pipeline
│
├── REQUIRED (always scaffold)
│   ├── Checkout                 # git clone into the runner
│   ├── Runtime install          # the language (Python/Node/Rust/Go/...)
│   ├── Dependency install       # pip/npm/cargo/go mod
│   ├── Test execution           # whatever test command the project has
│   └── Status reporting         # pass/fail to the PR or commit
│
├── RECOMMENDED (scaffold unless operator declines)
│   ├── Lint                     # ruff/eslint/clippy/golangci-lint
│   ├── Typecheck                # mypy/tsc --noEmit/cargo check/go vet
│   ├── Dependency caching       # skip re-downloading node_modules every run
│   ├── Concurrency group        # cancel old runs when a new commit lands
│   └── Branch / PR filter       # only run on main + PRs, not every branch
│
├── OPTIONAL — auto deploy (only if operator asked)
│   ├── Deploy trigger           # workflow_run on success, or direct push
│   ├── Deploy step              # SSH / Vercel / Fly / Cloud Run / Docker push
│   ├── Post-deploy health check # curl -f /health, fail loudly
│   ├── Rollback path            # revert to previous SHA on health fail
│   └── Notification             # Slack/Telegram/email on success or fail
│
├── NICE TO HAVE (suggest, don't auto-add)
│   ├── Coverage report          # pytest-cov, c8, tarpaulin
│   ├── Security scan            # pip-audit, npm audit, cargo audit (see audit-security)
│   ├── Container build + push   # if the deploy is container-based
│   ├── Integration tests        # separate from unit, usually against a real DB
│   ├── Matrix builds            # multiple OSes or language versions
│   ├── Scheduled runs           # nightly cron for dependency updates or flake detection
│   └── Release automation       # semantic-release, changesets, goreleaser
│
└── EXPENSIVE — offer only if operator asks
    ├── E2E browser tests        # Playwright, Cypress (slow, flaky)
    ├── Visual regression        # screenshot diffing
    ├── Load / stress tests      # k6, Locust
    ├── Mutation testing         # Stryker, PIT
    └── Fuzzing                  # per-language fuzzer

How to use this tree with the operator:

1. Start by scaffolding REQUIRED + RECOMMENDED. That's the default.
2. Ask: "Do you also want automatic deployment, or just tests for now?" If yes → add OPTIONAL. If no → stop here.
3. Mention NICE TO HAVE as a one-sentence menu: "You can also add coverage, a security scan, or scheduled runs later. Want any of those now?"
4. Never silently scaffold EXPENSIVE items. They slow every run down and flake on new projects.

Every added component costs something (runner minutes, maintenance, false-positive noise). Bias toward the minimal working pipeline. It's easier to add later than to remove.

Step 2 — Concepts, if the operator wants them

If the operator said "what's a deploy?" or "what is CI?", pause and explain. Short, concrete, concrete, concrete. One analogy per concept.

CI (Continuous Integration) — "Every time you push code, a computer somewhere else runs your tests and tells you if anything broke. If the tests break, you find out in a minute instead of when a user hits the bug."

CD (Continuous Delivery / Deployment) — "If the tests pass, that same computer also deploys your code to production automatically. No more SSH."

Runner — "The computer that does the work. It's not your laptop and it's not your production server. Think of it as a rented robot that runs for 30 seconds and then disappears."

Workflow — "The recipe you give the runner. It's a YAML file that says: install these dependencies, run these tests, if they pass, deploy to this place."

Trigger — "The event that wakes the runner up. Usually 'someone pushed code to the main branch.' Can also be 'someone opened a pull request' or 'every Monday at 9am.'"

Secrets — "Passwords and API keys that the runner needs but that cannot be visible in the YAML (because YAML is in git and git is public). They live in a separate encrypted store on the CI platform and are injected into the runner at runtime."

Artifact — "The thing you built. A zip, a container image, a directory of static HTML. CI produces it; CD ships it."

Environment — "Where the artifact ends up. 'staging' is the pre-production test site; 'production' is what real users see."

After this one-page explanation, ask: "Does that make sense? Should I keep going, or do you want to pause and ask anything?"

Step 3 — Pick the right CI host

The host you generate for depends on:

1. Their git host (from Step 1, question 2)
2. Where they want to deploy (SSH VM? Managed platform?)
3. Whether they already have something working (don't replace it unless asked)

Host-to-file mapping

Skill scaffolds for GitHub Actions by default. For others, state clearly: "I'll write this as GitHub Actions first, then point out the pieces that change for ." The core logic is portable even if the YAML syntax differs.

Step 4 — Scaffold the files

Generate two workflows at minimum:

test.yml — runs on every push and PR

Contents:

1. Checkout the repo
2. Install language runtime (Python, Node, Rust, etc. — auto-detect from repo files)
3. Install dependencies
4. Run lint
5. Run typecheck
6. Run tests

Commented inline for the non-developer reader. Every step has a comment explaining what it does.

deploy.yml — runs on success of test.yml

Triggered by workflow_run (GitHub Actions) or equivalent. Only fires if test.yml passed.

Contents depend on the deploy target (see Step 5). The invariant: deploy the exact revision that passed in test.yml, not whatever main happens to point to when the deploy job starts.

Built-in safety (every workflow this skill scaffolds enforces all of these):

• Top-level permissions: contents: read — start from zero and only grant what the job actually needs. Default-broad token permissions are the #1 GitHub Actions supply-chain risk. Elevate per-job (e.g., id-token: write for OIDC) only where required.
• Third-party actions pinned to a full commit SHA, never a tag. A tag can be moved; a SHA cannot. Every uses: line looks like uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2.
• Carry the tested SHA through to deploy. For workflow_run, use the successful run's head_sha and deploy that exact commit. Never git pull origin main blindly on the server — that is how commit A's passing tests deploy commit B by accident.
• OIDC over long-lived secrets for cloud deploys. Target platforms that support it (AWS, GCP, Azure, Vault, Cloudflare) should use federated tokens instead of storing static keys in GitHub Secrets. SSH to a VM is the one legitimate case for a static DEPLOY_SSH_KEY — there is no OIDC equivalent for arbitrary SSH.
• Protected production environment with required reviewers for any job that deploys to production. environment: name: production in the job turns on the approval gate. Pair with branch protection.
• CODEOWNERS covers .github/workflows/ so a random contributor can't silently change a workflow. Add /.github/workflows/ @maintainer at minimum.
• Self-hosted runners are opt-in and isolated. Never run self-hosted runners on a forked PR without explicit allowlisting. Prefer GitHub-hosted runners for anything that touches secrets. If self-hosted is unavoidable, use ephemeral runners and isolate by repository.
• Log redaction — GitHub masks registered secrets automatically, but echo $SECRET still risks leaking through errors, stack traces, or timing. Never echo a secret. If a command's output might contain one, pipe through a redaction filter.
• concurrency group so two deploys don't race each other.
• Job-level timeout-minutes so a hung deploy eventually fails.
• Secrets cleanup — any temp SSH key or credential file is deleted in an always() step.
• Post-deploy health check — curl -f <url>/health or equivalent. If the health check fails, the deploy job fails loudly.

Working templates in this skill's templates/ directory:

• templates/github-actions-test.yml — language-aware test workflow, permissions-hardened, SHA-pinned
• templates/github-actions-deploy-ssh.yml — deploy to a VM over SSH, with post-deploy health check and cleanup

Templates for Vercel, Fly.io, Cloud Run, and Docker registries follow the same pattern — the skill can emit them on demand, but the canonical reference implementations are the two above.

Repo-aware generation rule

Do not emit generic YAML just because it is the easiest thing to write. Before scaffolding, inspect the repo and tailor the workflow to the actual stack:

• pnpm-lock.yaml → use pnpm, not npm
• yarn.lock → use yarn
• bun.lockb → use bun
• uv.lock → use uv run
• poetry.lock → use poetry run
• Cargo.toml → cargo check / cargo test
• go.mod → go test ./...
• Gemfile → bundle exec ...
• composer.json → composer / phpunit
• mix.exs → mix test

Framework markers matter too:

• Next.js / Vercel / Netlify style app → usually generate test-only CI and let the platform handle deploys
• Django / Rails / Laravel → include migration step explicitly if the app already does it manually
• Multi-service monorepo → ask which service(s) the workflow should cover, or generate one workflow per service path

The operator should feel like the workflow was written for their repo, not copied from a blog post.

Step 5 — Deploy target specifics

SSH to a VM

Needs:

• SSH private key as a GitHub Secret (DEPLOY_SSH_KEY)
• VM host as a Secret or repository variable (DEPLOY_HOST)
• VM user as a Secret or variable (DEPLOY_USER)
• The target already has the repo cloned and a runtime installed
• The sudo commands the deploy needs have NOPASSWD rules set up (sudoers.d)

Common failure: sudo prompts for a password and the CI job hangs. Always verify NOPASSWD is set up or use a user that doesn't need sudo.

Reproduce locally first. If the operator has never done a remote deploy, tell them: "Before we automate this, we should do it manually from your laptop once to make sure it works. Can you SSH in and run the deploy commands by hand?" Automating a broken manual process leads to hours of debugging.

Managed platform (Vercel / Netlify / Railway / Render)

Most managed platforms already auto-deploy on push. CI's job here is only testing before deploy, not deploying itself. The skill should:

1. Confirm the platform is already connected to the repo
2. Generate a test.yml that runs on push and PR
3. Not generate a deploy.yml unless the operator wants a staging/production gate
4. Explain that "deploy" happens automatically because the platform is listening for pushes

Fly.io / Cloud Run / Kubernetes

These typically use an image push + a deploy command:

1. Build the container (docker build or flyctl deploy)
2. Push the image or deploy directly
3. The platform rolls out the new version

Needs the platform's API token as a Secret (FLY_API_TOKEN, GCP_CREDENTIALS, etc.).

"I don't have a target yet"

If the operator doesn't have somewhere to deploy, don't try to set up a full pipeline in one go. Suggest the simplest-possible target for their project:

• Static site / frontend → Vercel or Netlify
• Full-stack web app, want managed → Fly.io or Railway
• Full-stack web app, want control → Hetzner / DO / Oracle Cloud VM + SSH
• Container workload → Fly.io or Cloud Run
• Discord bot / Telegram bot / any long-running process → Fly.io with fly.toml, or a VM

After they pick, come back and run Step 4 / Step 5.

Step 6 — Secrets checklist

Every workflow that uses secrets needs a checklist handed to the operator. The skill produces this as a comment block in the generated YAML and also as a separate message:

## Secrets this workflow needs

Set these in your GitHub repo: Settings → Secrets and variables → Actions → New repository secret

- [ ] `DEPLOY_SSH_KEY` — contents of your SSH private key (e.g., `cat ~/.ssh/my_key`)
- [ ] `DEPLOY_HOST` — IP or hostname of your server (e.g., `192.0.2.1`)
- [ ] `DEPLOY_USER` — username on the server (e.g., `ubuntu`)

After you set all three, re-run the workflow. First run will usually fail on the SSH step if any secret is wrong — the log will tell you which one.

Never print the secret values themselves, and never store them in YAML.

Step 7 — First run = manual dispatch

Always recommend that the first run of a deploy workflow be manual, via workflow_dispatch. Reasons:

1. The operator can watch the log in real time and catch errors at their source
2. Secrets missing / wrong will fail fast instead of on every push
3. If the VM prereq is off, they find out now instead of tomorrow

Tell the operator: "After I generate the files, go to Actions → Deploy → Run workflow. Watch the output. If it passes, you're set up. If it fails, paste me the error and I'll fix it."

Manual dispatch is for a reviewed re-deploy or first-run proof. Make the generated comments explicit about this. In normal operation, workflow_run should be the path that moves tested code to production.

Things not to do

• Don't write YAML blind. Always ask the two Step 1 questions first
• Don't invent secret names. If you reference DEPLOY_HOST in the YAML, the same name has to be in the checklist
• Don't use third-party actions unnecessarily. actions/checkout and actions/setup-* are fine; random actions from marketplace are a supply-chain risk. If you reference one, pin to a commit SHA
• Don't copy an outdated template. CI/CD conventions move fast. Check that the syntax matches the current docs before emitting
• Don't auto-deploy to production from the first run. Require workflow_dispatch manually for the first run, then switch to workflow_run for subsequent automated runs
• Don't leave framework mismatch in place. A Python workflow in a pnpm repo is not a starting point, it's a bug. Detect the repo and emit the right commands the first time.
• Don't scaffold workflows the operator didn't ask for. If they asked for test.yml, do not also drop in deploy.yml, release.yml, or codeql.yml because they "looked useful". Adjacent needs (Dependabot, Renovate, CodeQL) go in the output as hand-off suggestions — a single extra workflow can leak credentials or trigger unwanted billing on the first push.
• Don't scaffold gate-bypass mechanisms into generated workflows. Forbidden in emitted YAML: continue-on-error: true on a quality gate (lint / typecheck / test / audit), if: false-skipped jobs that look real, allowed_failures lists for required checks, conditional skips on commit-message tokens ([skip-ci] exceptions for the full quality gate), --no-verify flags in scripted steps. The pipeline's value is in failing loudly on regressions; bypass mechanisms are a regression dressed as flexibility. If the operator wants a soft check, scaffold it as a separate non-required job — never weaken a required one.

Common failure modes

Things to watch for in your own output:

• Deploy fix loops — iterating on the YAML remotely when the problem is actually that the manual deploy doesn't work yet. Always have the operator prove the manual path works before automating
• Platform env var confusion — Vercel, Netlify, and similar have their own secrets UI, separate from git host secrets. If the deploy uses env vars on the platform, tell the operator to set them there, not in GitHub Secrets
• Assumption that git pull is enough — if the project has migrations, native dependencies, cache warming, or a restart step, git pull alone will silently produce a broken deployment. Enumerate every step
• Secrets in logs — echo $DEPLOY_KEY in a step leaks the key to the public log. Never echo secrets
• Long-running runners forgotten — if the workflow starts a server and doesn't stop it, the runner eventually times out. Scope runners to single commands

Sweep mode — read-only audit

When invoked from /vibesubin (the umbrella skill's parallel sweep), this skill runs in read-only audit mode. Do not scaffold workflows, do not write .github/workflows/*.yml, do not touch any files.

Instead, produce a findings-only report:

• What CI/CD configuration currently exists (.github/workflows/, .gitlab-ci.yml, etc.) — list files with one-line purposes.
• What's missing (no test workflow, no deploy workflow, no secret management pattern, no OIDC setup where it would fit).
• What looks broken or outdated (SHA pinning drift, deprecated actions, missing permissions: blocks, overly-broad GITHUB_TOKEN scopes).
• Stoplight verdict: 🟢 CI is in good shape / 🟡 gaps exist but nothing critical / 🔴 critical issues (secrets in logs, write-all token, no concurrency guards on deploy).
• A one-line "fix with" pointer indicating that running /setup-ci directly will scaffold the missing pieces.

The operator reviews the sweep report and, if they want the fixes applied, invokes /setup-ci directly — which then runs the full scaffolding procedure.

How to tell: the task context from the umbrella will include a sweep=read-only marker or an explicit instruction like "produce findings only, do not edit". If that's present, never write. If the operator invokes this skill by name (/setup-ci), the full procedure applies and editing is expected.

Harsh mode — no hedging

When the task context contains the tone=harsh marker (usually set by the /vibesubin harsh umbrella invocation, but can also come from direct requests like "don't sugarcoat" / "brutal review" / "매운 맛" / "厳しめ"), switch output rules on the CI audit report:

• Lead with the worst safety gap. First line is the single most dangerous issue — "GITHUB_TOKEN has write-all on every workflow in this repo", "no pinned action SHAs, every uses: line is a moving target", ".github/workflows/deploy.yml echoes $DEPLOY_KEY in a run step — rotate that key and fix the workflow". No preamble.
• No softening on supply-chain issues. Drop "consider pinning", "might be a good idea", "you may want to add". Replace with "pin every third-party action to a SHA before your next commit — tags can be moved silently".
• Missing permissions blocks are an immediate finding, not a nice-to-have. Balanced mode says "no top-level permissions: block". Harsh mode says "every job in this workflow gets a write-scoped GITHUB_TOKEN by default — this is the #1 supply-chain risk on GitHub Actions. Add permissions: contents: read now."
• Manual-deploy suggestions become directives. "Do not automate the deploy until the manual path works" — not "it's usually a good idea to confirm the manual path first."
• First-run warnings are absolute. Harsh mode refuses to generate a deploy workflow without workflow_dispatch for the first run, and says so in one sentence.
• No "you're mostly set up" closures when real safety gaps exist. If the repo has any unpinned third-party action, any write-scoped token, or any echoed secret, the verdict line does not end with reassurance.

Harsh mode does not invent findings, exaggerate CVE severity, or become rude. Every harsh statement must cite the same workflow file, job name, or secret reference the balanced version would cite. The change is framing, not substance.

Layperson mode — plain-language translation

When the task context contains explain=layperson (from /vibesubin explain, /vibesubin easy, "쉽게 설명해줘", "일반인도 이해되게", "explain like I'm non-technical", "非開発者でも分かるように", "用通俗的话解释"), add a plain-language layer to every finding this skill emits. Combines freely with tone=harsh. Full rules at /plugins/vibesubin/skills/vibesubin/references/layperson-translation.md.

Three dimensions per finding

Every finding gets three questions answered in plain language, in the operator's language (Korean / English / Japanese / Chinese):

• 왜 이것을 해야 하나요? / Why should you do this? — "CI가 없으면 매 배포가 수동입니다. 테스트를 까먹고 배포, lint 안 돌리고 배포, 시크릿 까먹고 배포 — 한 번의 실수가 프로덕션을 내립니다."
• 왜 중요한 작업인가요? / Why is it an important task? — "CI는 비개발자 pack에서 가장 큰 생산성 레버예요. 한 번 세팅하면 git push가 모든 걸 다 해요. 안 세팅하면 매주 같은 실수가 반복됩니다."
• 그래서 무엇을 하나요? / So what gets done? — "package.json·requirements.txt·Cargo.toml 같은 걸로 스택을 감지하고, 테스트·lint·배포 workflow를 .github/workflows/에 올리고, 배포 후 health check까지 포함한 yaml을 scaffold합니다. 시크릿 값은 만지지 않고 — 어떤 이름·어디에 넣어야 하는지만 알려줍니다."

Severity translation

• 🔴 none → "CI 자체가 없음 — 지금 모든 배포가 수동 눈대중"
• 🟡 partial → "테스트는 돌아가는데 배포가 수동 / 배포는 되는데 health check 없음"
• 🟢 healthy → "테스트 + 배포 + health check까지 자동"

Box format

Wrap each finding in the box format from the shared reference. Header uses urgency phrase and the finding number. Footer names the hand-off skill.

What does NOT change

Findings, counts, file:line references, evidence, and severity are identical to balanced/harsh output. Only the wrapping and dimension annotations are added.

Hand-offs

• Deploy touches .env or secrets → manage-secrets-env for the .env pattern and lifecycle
• Deploy touches branch strategy, dep pinning, or directory layout → project-conventions
• Deploy leaks secrets → audit-security immediately
• Deploy breaks and touches production code → refactor-verify for the fix

Details

Canonical working templates live in templates/:

• templates/github-actions-test.yml — test workflow with permissions hardening and SHA-pinned actions
• templates/github-actions-deploy-ssh.yml — deploy to a VM over SSH with post-deploy health check and secret cleanup

The CI/CD concept explainer (runners, triggers, environments, OIDC, secrets, concurrency) is embedded in the "Step 2 — Concepts" section of this SKILL.md rather than split into a reference file. Per-platform deploy gotchas live in the "Step 5 — Deploy target specifics" section.