gsong-marketplace

George Song's Claude Code plugin marketplace.

Add the marketplace, then install plugins individually:

/plugin marketplace add gsong/claude-marketplace

Full lifecycle AI documentation for Claude Code projects: bootstrap, lookup, update, check, and audit.

/plugin install ai-docs@gsong-marketplace

Skills: /ai-docs:init, /ai-docs:lookup, /ai-docs:update, /ai-docs:check, /ai-docs:audit

Hooks: UserPromptSubmit reminds Claude to consult ai-docs:lookup before code changes

Session memory and project instruction management for Claude Code.

/plugin install ai-memory@gsong-marketplace

Skills: /ai-memory:save, /ai-memory:review

OpenAI Codex CLI integration for parallel PR reviews, task delegation, and multi-round consensus discussions via the codex plugin.

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex-tools@gsong-marketplace

Skills: /codex-tools:review, /codex-tools:run, /codex-tools:discuss

Requires: Codex CLI and the codex plugin (installs with codex-tools once the openai-codex marketplace is added); review also needs gh, uv, and the gh-tools directory. See the plugin README.

GitHub CLI PR review, triage, comment posting, review replies, and project management skills.

/plugin install gh-tools@gsong-marketplace

Skills: /gh-tools:review, /gh-tools:triage, /gh-tools:post-comments, /gh-tools:address-review, /gh-tools:project-manager

Requires: gh CLI; review also needs uv and the mattpocock-skills and feature-dev plugins (superpowers optional). See the plugin README.

Git commit, worktree, and auto-squash skills.

/plugin install git-tools@gsong-marketplace

Skills: /git-tools:commit, /git-tools:worktree, /git-tools:auto-squash

Requires: pnpm, as the install fallback in worktrees. See the plugin README.

Dependency upgrades and CI/CD security auditing for project repositories.

/plugin install repo-maintenance@gsong-marketplace

Skills: /repo-maintenance:pnpm-deps, /repo-maintenance:pnpm, /repo-maintenance:mise, /repo-maintenance:gha, /repo-maintenance:zizmor

Requires: pnpm, mise, zizmor, gh CLI, and Node.js with npx (runs actions-up), each for the skills that use it. See the plugin README.

General-purpose utilities for Claude Code.

/plugin install utilities@gsong-marketplace

Skills: /utilities:date

Requires: macOS/BSD date.

Keep text an audience reads in your voice: a drafting pipeline, a send-time lint, a draft gate, and correction capture.

/plugin install writing@gsong-marketplace

Skills: /writing:draft

Hooks: PreToolUse send-time lint and smart-quote check, PostToolUse draft gate and correction capture, Stop correction promotion

Requires: jq, perl, and uv. See the plugin README.

CI scans every skill in plugins/ and .claude/skills/ with Cisco skill-scanner. It fails on a high or critical finding. The scan runs offline with rules only, so it needs no account and sends nothing out. It runs on every pull request and push to main, and a failed scan blocks the merge. The report goes to GitHub code scanning, where medium and low findings show as alerts.

Run the same scan locally:

mise run scan:skills

Pass a path to also write a SARIF report. The scan reads a copy of the files that git tracks or does not ignore, so ignored files such as __pycache__ stay out. A local run also reads untracked files, which CI does not have, so it can report findings that CI does not.

To triage a finding:

• A real risk: fix the skill.
• A false positive: add a suppression to skill-scanner-policy.yaml. Scope it by rule_id, skills, and paths, and give a reason.
• A rule that can never apply here: add it to disabled_rules.

The scan has limits:

• It reads skill directories only. The hook scan below covers plugin hooks, but a script outside skills/ that no hook runs goes unscanned.
• Rules alone catch 7.7% of malicious skills by Cisco's own count. The scan finds known patterns only, and it does not replace a review.

CI also scans the plugin hooks with scripts/scan-hooks.py. The hook scan reads hooks/hooks.json and any hooks that plugin.json declares. It reads each hook command, and every code file that a command or a scanned file names. It skips files under skills/, because the skill scan reads them. It runs in the same CI job as the skill scan, so a failed hook scan also blocks the merge. The CI log lists each file it scanned.

The hook scan fails when a line does one of these things:

• It calls a network tool, such as curl, wget or nc.
• It pipes text into a shell, or into an interpreter that reads its program from the pipe.
• It runs a stream or command output as code, such as source <(…) or eval "$(…)".
• It opens a connection through /dev/tcp/ or /dev/udp/.
• It loads a network module, such as Python's urllib, Node's https or Perl's LWP, or calls fetch.
• It fetches code or packages, such as uvx, npx, npm install, pip install, git clone, or a script header that lists dependencies for uv.

It also fails on an http hook. It fails on a command that runs a script by a path outside ${CLAUDE_PLUGIN_ROOT}, or whose script is missing. It fails on a file it cannot read or parse. A whole-line comment does not count. Run the same scan locally:

mise run scan:hooks

The hook scan has no suppression file. If a hook needs the network, change the rule in scripts/scan-hooks.py and say why in the pull request. The hook scan has these limits:

• It matches text only. It does not detect a command name that a script builds at run time.
• It does not detect a hook that runs a script under skills/, unless the command names that script through ${CLAUDE_PLUGIN_ROOT}.
• It checks network use and code fetches only. It does not detect a hook that reads secrets or deletes files.

MIT