agent-skills
A Claude Code plugin marketplace that ships engineering skills and LSP server configurations. Install it once and get a suite of AI-driven PR workflow automations available in any project.
What this is
agent-skills is a declarative, no-build repository that acts as a Claude Code plugin marketplace. It contains no runtime process and no server — it is pure configuration (JSON + Markdown). Claude Code reads the manifests and makes the skills available as slash commands and automated workflows.
The marketplace currently ships one plugin, engineering, which provides:
- Three skills that automate common PR tasks end-to-end.
- Three LSP server configurations (Ruby, Python, TypeScript/JavaScript) that activate language intelligence when the plugin is installed.
Requirements
| Tool | Purpose |
|---|
| Claude Code | Reads manifests and executes skills |
| gh CLI | Used by all three skills to interact with GitHub |
| jq | Used by track-pr-review-request for JSON processing |
ruby-lsp | LSP server for Ruby (developer environment) |
pyright-langserver | LSP server for Python (developer environment) |
typescript-language-server | LSP server for TypeScript/JavaScript (developer environment) |
LSP servers are expected to be installed in your development environment; they are not bundled in this repo.
For track-pr-review-request on macOS, curl, security (Keychain), and launchctl are also required — all are built-in on macOS.
Installation
1. Add the marketplace
Run once per machine to register this repo as a Claude Code marketplace source:
claude plugins marketplace add /path/to/agent-skills
Verify it was added:
claude plugins marketplace list
# st-marketplace Source: Directory (/path/to/agent-skills)
2. Install the engineering plugin
claude plugins install engineering@st-marketplace
Verify the installation:
claude plugins details engineering
# Shows: 3 skills, 3 LSP servers
The skills are now available in any Claude Code session.
Available plugins
engineering
Version: 0.0.1 — Author: Samir Tapiero
Skills
| Skill | Trigger | What it does |
|---|
fix-pr-feedback | "address PR comments", "fix review feedback", "respond to bugbot" | Fetches all unresolved PR review comments (inline threads and top-level), triages each with you, applies fixes, lints, commits, pushes, and posts threaded replies. |
fix-pr-ci-checks | "fix CI", "fix the failing checks", "make the build pass" | Fetches failing GitHub Actions checks and their logs, classifies each failure, triages with you, applies fixes, verifies locally, and pushes to re-trigger CI. |
track-pr-review-request | "track my PR reviews", "set up daily PR review reminders" | Installs a macOS launchd agent that fires daily, checks for pending PR review requests, and sends a Slack notification. Also supports cancel and status sub-commands. |
LSP servers
| Server | Language(s) | Command |
|---|
| ruby-lsp | .rb, .rake, .gemspec | ruby-lsp |
| pyright-langserver | .py, .pyi | pyright-langserver --stdio |
| typescript-language-server | .ts, .tsx, .js, .jsx, .mts, .mjs, .cjs | typescript-language-server --stdio |
Skill reference
fix-pr-feedback
Address all unresolved review feedback on a PR end-to-end.
fix-pr-feedback [PR-number-or-URL]
- Accepts a PR number or full GitHub URL. Auto-detects from the current branch if omitted.
- Fetches both inline review threads and top-level conversation comments from any author (humans, Copilot, Cursor Bugbot, other bots).
- Reads the relevant file lines before asking you anything — pre-validates whether each concern still applies in the current code.
- Offers four modes: post your own replies, draft replies for review, analyze and fix + reply, or skip.
- Smart commit strategy: amends history before approval; creates a follow-up commit after approval to preserve the review record.
- Posts threaded replies per inline comment. Never collapses inline feedback into a single top-level comment.
- Gates: linter must pass before pushing; explicit user confirmation required before any push or reply.
fix-pr-ci-checks
Fix failing CI checks on a PR end-to-end.
fix-pr-ci-checks [PR-number-or-URL]
- Fetches failing GitHub Actions checks and the relevant log excerpts.
- Classifies each failure: fixable code failure, human-gate (waiting-on-review, merge-queue — never touches these with code), or flaky/infra (surfaces to you, never blind-retries).
- Delegates test failure diagnosis to
test:triage-tests before proposing a fix.
- Runs local linters and tests to verify fixes before pushing.
- Single-pass: it does not watch CI go green. Hands off to
release:read-ci-status --watch after pushing.
