Ci Bootstrap

CI Bootstrap

You are going to build a quality and delivery pipeline for THIS repository, inspired by a mature polyglot-monorepo reference, but adapted to the real stack and conventions found here. Do NOT copy technologies that don't exist in this project. Detect, confirm, then implement.

Step 1 — Choose the operating mode

If the user passed a mode in the arguments (autopilot, auto, guided, interactive, step-by-step), use it. Otherwise, ask before doing anything else:

> How do you want to run this? > > 1. Autopilot — I run all phases end to end with sensible defaults. I only stop if I hit a > decision that is genuinely ambiguous and risky (e.g., can't tell which branch is production). > 2. Guided — I pause after each phase, show you what I found / what I'm about to write, and > wait for your confirmation before touching anything.

Mode rules:

• Autopilot: run Phases 0→4 without pausing. Resolve ambiguity with the defaults documented in

references/ (and say which default you picked in the final summary). Only stop for the blocking questions listed at the end of Phase 0 — and even then, first try to resolve them from evidence (remote branches, existing configs, commit history).

• Guided: at the end of each phase, present results and the plan for the next phase as a short

summary plus concrete questions (use structured questions when available). Do not write any file before Phase 0 is confirmed.

In both modes, Phase 0 is strictly read-only.

Step 2 — Execute the phases

Each phase has a detailed reference file. Read it when you reach that phase — not before.

| Phase | What | Reference | |-------|------|-----------| | 0 | Discovery (read-only stack & convention detection) | [references/discovery.md](references/discovery.md) | | 1 | Pre-commit hooks (.pre-commit-config.yaml) | [references/pre-commit.md](references/pre-commit.md) | | 2 | CI (.github/workflows/ci.yml + security workflow) | [references/ci.md](references/ci.md) | | 3 | CD (deploy skeletons, only activated with confirmed targets) | [references/cd.md](references/cd.md) | | 4 | Validation & delivery | [references/validation.md](references/validation.md) |

Non-negotiable principles (the "why" behind the reference)

These apply in every mode, every phase. They are not stylistic preferences — each one prevents a concrete failure mode.

1. Version parity pre-commit ↔ CI. Pin the SAME version of every tool in both places. A hook

that passes locally and fails in CI (or vice versa) destroys trust in the pipeline.

1. Gating by cost.

• On PRs: only fast checks — format check, lint, security scan, light unit tests.
• The heavy suite runs only on push to the integration and production branches:

``yaml if: github.event_name == 'push' && (github.ref == 'refs/heads/' || github.ref == 'refs/heads/') `` Make this EXPLICIT in a YAML comment and tell the user that "a green PR does not guarantee the full suite passed".

1. Concurrency. cancel-in-progress: ${{ github.ref != 'refs/heads/' }} —

production runs are never cancelled.

1. Path scoping. Each linter only runs on the files of its own package/service

(files: regex in pre-commit; paths: filters and working-directory in CI).

1. Cross-platform hooks. For local hooks that call binaries from node_modules, use a Node

wrapper script (node scripts/precommit-runner.mjs ), never bash -c 'cd ... && npx ...' — bash may be missing from PATH on Windows. Only applies if there is a JS frontend.

1. Surgical changes. Do not rewrite CI that works; add the minimum. Do not delete files outside

the task's scope — point at problems, don't erase them.

1. No invented credentials. All secrets via secrets.*. If the deploy target is unclear,

deliver disabled skeletons with documented TODOs instead of guessing.

1. Reuse before replace. If the repo already uses flake8, don't silently switch it to Ruff.

Tool swaps are a user decision, never a default.

Step 3 — Deliver

Finish with a single final summary (see references/validation.md for the checklist): what was created, decisions and defaults taken, secrets the user must register, and next steps. In Guided mode this is the last checkpoint; in Autopilot it is the only long report the user sees — make it count.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: eabs86
• Source: eabs86/ci-bootstrap
• License: MIT

Install and usage instructions live in the source repository linked above.