The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES]: The skill installs the graphifyy package from PyPI during the initialization step. The package name differs from the tool name ('graphify'), which is a common indicator of potential typosquatting or shadowing of legitimate libraries.
[DYNAMIC_EXECUTION]: The skill generates and executes Python code blocks using the python -c command within shell scripts. It also records the path to the Python interpreter in a local file (graphify-out/.graphify_python) and dynamically reads this file to execute all subsequent pipeline steps, which makes the execution flow dependent on runtime-generated configuration.
[INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary user-supplied files (code, documents, images, and videos) using LLM subagents. This introduces a risk that malicious instructions embedded within those files could hijack the subagent's logic.
Ingestion points: All files within the user-provided directory are scanned and passed to subagents.
Boundary markers: The subagent prompt in references/pipeline-steps.md enforces a specific JSON output schema but lacks explicit 'ignore previous instructions' or boundary delimiters for the file content itself.
Capability inventory: Subagents are explicitly instructed to use the general-purpose tool type, which provides them with file-writing and bash execution capabilities.
Sanitization: There is no evidence of escaping or filtering of the file content before it is processed by the AI subagents.
[PERSISTENCE]: The hook subcommand installs a git post-commit hook. This configuration causes graph-rebuilding logic to execute automatically every time a user performs a git commit, establishing a persistent execution trigger within the development environment.
[COMMAND_EXECUTION]: The installation step uses the --break-system-packages flag. This allows pip to modify Python packages managed by the system's package manager, which can lead to environment instability or security risks by overriding managed packages.
[EXTERNAL_DOWNLOADS]: The add <url> command allows the agent to fetch arbitrary remote content, which could be used to pull malicious payloads into the local analysis environment.
[DATA_EXFILTRATION]: The skill supports pushing the generated knowledge graph to an external Neo4j instance via the --neo4j-push flag, which involves transmitting processed data over the network to a user-specified URI.

graphify