The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions direct users and agents to download and immediately execute shell scripts from a remote, unverified personal GitHub repository using high-risk piping methods.
Evidence (macOS/Linux): curl -fsSL https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.sh | sh in SKILL.md and references/installation.md.
Evidence (Windows): irm https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.ps1 | iex in references/installation.md and references/troubleshooting.md.
[EXTERNAL_DOWNLOADS]: The skill promotes the installation of a global NPM package and other scripts from an external source that is not recognized as a trusted vendor.
Evidence: npm i -g @colbymchenry/codegraph in references/installation.md.
[COMMAND_EXECUTION]: The skill requires the execution of various CLI commands that modify the local filesystem, install software into agent configurations, and host a local MCP server.
Evidence: codegraph install, codegraph init -i, and codegraph serve --mcp documented throughout references/cli-reference.md.
[PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection by ingesting untrusted data (project source code) and returning it to the agent's context where malicious instructions could be embedded.
Ingestion points: Local codebase indexing via codegraph init -i described in SKILL.md.
Boundary markers: Lacks explicit delimiters to isolate indexed code from the agent's instruction set, although it uses a ⚠️ staleness banner for edited files.
Capability inventory: Extensive access to project source code via MCP tools like codegraph_explore and codegraph_node which return code snippets to the agent.
Sanitization: No evidence of sanitization or escaping of external code content before it is processed by the AI agent.

codegraph