AgentSkillsCN

delivery

交付——CI/CD、测试、发布。适用于优化流水线时使用。

SKILL.md
--- frontmatter
name: delivery
description: Delivery - CI/CD, testing, releases. Use when improving pipelines.

Delivery Guideline

Tech Stack

  • CI: GitHub Actions
  • Testing: Bun test
  • Linting: Biome
  • Platform: Vercel

Non-Negotiables

  • All release gates must be automated (manual verification doesn't count)
  • Build must fail-fast on missing required configuration
  • CI must block on: lint, typecheck, tests, build
  • /en/* must redirect (no duplicate content)
  • Security headers (CSP, HSTS) must be verified by tests
  • Consent gating must be verified by tests

Context

Delivery gates are the last line of defense before code reaches users. Every manual verification step is a gate that will eventually fail. Every untested assumption is a bug waiting to ship.

The question isn't "what tests do we have?" but "what could go wrong that we wouldn't catch?" Think about the deploy that breaks production at 2am — what would have prevented it?

Driving Questions

  • What could ship to production that shouldn't?
  • Where does manual verification substitute for automation?
  • What flaky tests are training people to ignore failures?
  • How fast is the feedback loop, and what slows it down?
  • If a deploy breaks production, how fast can we detect and rollback?
  • What's the worst thing that shipped recently that tests should have caught?