The Agent Skills Directory

[PERSISTENCE]: The skill provides instructions to append environment variable exports to shell configuration files (~/.bashrc and ~/.zshrc). This allows for persistence across sessions and can be exploited to capture or manipulate environment data in future shell instances.
[METADATA_POISONING]: The skill documentation includes a link to a non-existent repository (https://github.com/langchain-ai/langsmith-fetch) under a trusted organization. This misrepresentation could lead users to trust a third-party tool based on a false association with an established vendor.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of an unverifiable third-party Python package langsmith-fetch. Installing unvetted packages from unknown sources poses a significant supply chain risk.
[INDIRECT_PROMPT_INJECTION]: The skill processes execution traces from LangSmith, which could contain instructions embedded by malicious actors in error logs or tool outputs.
Ingestion points: External traces fetched via langsmith-fetch and stored in recent-traces.json or viewed in the terminal.
Boundary markers: None identified; instructions do not advise the agent to ignore instructions embedded in the traces.
Capability inventory: Shell execution (langsmith-fetch, grep, mkdir, echo), directory creation, and file system interaction.
Sanitization: None identified; data is processed directly as JSON or text.
[CREDENTIALS_UNSAFE]: The skill encourages users to verify sensitive credentials by printing them to the console (echo $LANGSMITH_API_KEY), which can result in secrets being leaked to shell history files or terminal logs.

langsmith-fetch