← MCP server and skill grades

skills

mattpocock/skills/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
5863806dd4e1
Files read
114

This project is a collection of AI coding agent skills and helper scripts for software design, test-driven development, code review, issue triage, and writing workflows. The biggest risk is in the /triage workflow, which instructs the agent to check out external pull requests and run tests or reporter-supplied bug reproduction steps on the user's machine without sandboxing or prior diff review.

What it can reach

Files
Reads and writes repository code, docs, specs, glossaries, ADRs, local session transcripts, global or project CLAUDE.md/AGENTS.md and .claude/settings.json files, and temporary files in $TMPDIR or /tmp.
Shell commands
Directs the coding agent to execute git commands (including worktrees, commits, and branch pushes), test suites, linters, package managers, background claude processes, issue-tracker CLIs (gh and glab), and interactive bash setup/debugging scripts.
Network
Queries and updates GitHub and GitLab issues, pull requests, comments, and labels via gh and glab, fetches external documentation during research, installs npm devDependencies in setup skills, and loads Tailwind and Mermaid scripts from public CDNs in HTML reports.
Secrets
Reads .env and .env.* files when scoping setup wizards, and generated bash wizards prompt the user for API keys and secrets to write into .env and GitHub Actions repository secrets.
Browser
Opens generated HTML architecture reports, interactive lessons, and third-party dashboard URLs in the user's default browser using open, xdg-open, wslview, explorer.exe, or start.

Findings

1 high1 medium3 low
  1. HighTriage workflow checks out external pull requests and runs untrusted code or bug steps locally

    skills/engineering/triage/SKILL.md:74

    When triaging issues or pull requests from external non-collaborators, Step 3 instructs the agent to reproduce bugs from the reporter's steps and, for pull requests, to check out the contributor's branch and run tests or commands locally. An external contributor who opens a pull request (surfaced when PR triage is enabled or triaged directly by number) can modify test files, package.json scripts, or tool configs so running tests executes arbitrary code on the maintainer's machine with access to local credentials and tokens; untrusted issue bodies and comments can also supply harmful reproduction commands or prompt injection to steer the agent and exfiltrate data via gh issue comment.

    Fix: Treat external issue and PR bodies, comments, and diffs as untrusted input: require an explicit diff safety review before checking out external PRs, run untrusted verification only inside an isolated credential-free sandbox, and require explicit user confirmation for every reproduction command.

  2. MediumArchitecture HTML report disables Mermaid XSS protection and loads unpinned CDN scripts

    skills/engineering/improve-codebase-architecture/HTML-REPORT.md:16

    When /improve-codebase-architecture scans a repository, it writes an HTML report and immediately opens it in the user's desktop browser. The HTML scaffold configures Mermaid with securityLevel set to loose (disabling Mermaid's default HTML and script sanitization) and loads floating, unpinned scripts from cdn.tailwindcss.com and cdn.jsdelivr.net without Subresource Integrity hashes. Crafted identifiers, file paths, or commit text in a scanned repo that land unescaped inside a Mermaid diagram—or compromise of either floating CDN URL—can execute arbitrary JavaScript in the browser page holding the codebase's architectural review.

    Fix: Set Mermaid securityLevel to strict, HTML-escape all repository strings inserted into the report, and either inline local styles/rendering or pin exact library versions with Subresource Integrity (integrity) hashes.

  3. LowGit guardrail hook fails open without jq and misses common git flag forms

    skills/misc/git-guardrails-claude-code/scripts/block-dangerous-git.sh:19

    The git-guardrails-claude-code skill installs block-dangerous-git.sh as a Claude Code PreToolUse hook claiming to block all variants of git push, git reset --hard, git clean -f, git branch -D, and git checkout . or git restore .. Because the script lacks error handling around jq, a missing jq binary sets COMMAND to an empty string and exits 0, silently allowing all commands. Even when jq is present, matching fixed substrings misses standard git invocations such as git -C . push, git --no-pager push, git clean -xfd, git clean --force, git reset HEAD --hard, and git checkout -- ..

    Fix: Enable strict error handling and fail closed (exit 2) if jq is missing or input parsing fails, use printf instead of echo, and match git subcommands while accounting for global git flags (-C, -c, --git-dir, --no-pager), combined short flags (-xfd, -df), long flags (--force, --delete), and -- pathspec separators.

  4. LowWizard skill reads plaintext .env and .env.* secret files into model context unprompted

    skills/engineering/wizard/SKILL.md:20

    Although the wizard skill generates a local bash script using hidden terminal input (ask_secret) so credentials stay out of the chat, Step 1 instructs this model-invoked skill to read .env and .env.* directly before asking the user. In repositories that already have .env, .env.local, or .env.production files containing live secrets, reading them in full sends plaintext credentials to the remote model provider and records them in local session transcripts when only variable names are needed.

    Fix: Inspect .env.example or extract only variable names (for example with cut -d= -f1) rather than reading raw .env and .env.* files into the agent's context.

  5. LowPredictable world-readable /tmp files and CLI argument exposure for reports and handoff summaries

    skills/engineering/improve-codebase-architecture/SKILL.md:39

    Architecture reports and session handoff skills (handoff and claude-handoff) fall back to writing directly into /tmp when $TMPDIR is unset (the default on many Linux systems) using predictable filenames and without enforcing 0600 permissions via mktemp, which can expose private codebase summaries to other local users on shared machines. In addition, claude-handoff passes the summary file contents via $(cat <summary file>) on the claude --bg command line, exposing the full handoff prompt in the system process list (ps) for the lifetime of the background agent.

    Fix: Create temporary files or directories using mktemp with 0600/0700 permissions (umask 077), and pass handoff prompts to background processes via standard input or a private file reference instead of command-line arguments.