dsh-engineering-skills

Engineering-discipline skills distilled from a large production agent-system codebase: systematic code review, CI failure triage, shell safety, redundancy/boundary auditing, cross-repo pattern absorption, and release engineering.

Six markdown skills for AI coding agents (DeepSeek Harness, Claude Code, Codex, or any agent that reads SKILL.md). They are methodology-first — no code to install, no runtime, no lock-in: copy the skills/ directory and the agent loads them.

Skills

Install

From PyPI — the pack also ships as a data-only Python package with an installer CLI:

pip install "dsh-engineering-skills"
dsh-engineering-skills list                 # see the 6 skills
dsh-engineering-skills install ~/.dsh/skills          # DeepSeek Harness user skills dir

The wheel contains the same SKILL.md files (packaged as data-files, no code dependencies); the CLI just copies them into any agent's skills directory.

From source / generic agents — copy the skills into your agent's skills directory:

# DeepSeek Harness (user-level skills dir is ~/.dsh/skills, i.e. $DSH_HOME/skills)
cp -r skills/* ~/.dsh/skills/

# Claude Code
cp -r skills/* ~/.claude/skills/

# Codex
cp -r skills/* ~/.codex/skills/

Each skill is a self-contained SKILL.md with YAML frontmatter (name, description, keywords) — no dependencies, no build step.

Why these skills

They come from real incidents and audits, not theory. Notable provenance:

• shell-safety exists because a session froze for 12 minutes on findstr ... | more +0 waiting for a keypress that never arrived.
• redundancy-and-boundary-audit encodes the A/B/C duplication classification that keeps "same name" from being treated as "duplicate" — a domain model / API DTO pair is legitimate layering, and only asymmetric change is drift.
• code-review-methodology accumulates review angles only after each one caught a real bug (single-pass review finds ~30% of bugs in cross-cutting features).
• ci-diagnosis-and-fix documents the two failure modes that turn CI wholesale red: a nonexistent dependency version, and the old $GITHUB_OUTPUT syntax on newer runners.
• release-engineering encodes the flow proven across three same-day releases (a CLI+MCP audit tool, an MCP server, and this skill pack): PyPI name collisions, data-files path flattening, token-scope failures, and requires-python vs verify-venv mismatches all bit once and are now written down.

Validation

python scripts/validate_skills.py   # checks frontmatter + body of every SKILL.md

License

MIT © 2026 Chen (Jarry) Pan