The Agent Skills Directory

EXTERNAL_DOWNLOADS (SAFE): The skill installs the langsmith package via pip.
Evidence: Found langsmith>=0.2.0 in the dependencies field.
Trust: LangSmith is maintained by the langchain-ai organization, which is a recognized trusted source in the AI ecosystem.
DATA_EXFILTRATION (SAFE): The skill transmits application execution traces, including prompt inputs and model outputs, to the external domain smith.langchain.com.
Evidence: The skill utilizes @traceable decorators and the LangSmith Client to send data to the cloud service.
Mitigation: This activity is the primary intended function of the skill. The documentation includes a specific remediation example using the process_inputs hook to sanitize sensitive information (e.g., passwords) before transmission.
CREDENTIALS_UNSAFE (SAFE): The skill utilizes environment variables for authentication and does not contain hardcoded secrets.
Evidence: The documentation correctly instructs users to set LANGSMITH_API_KEY as an environment variable and uses placeholders in examples.
INDIRECT_PROMPT_INJECTION (SAFE): The skill presents an attack surface by processing untrusted prompt data.
Ingestion points: Prompts and queries passed to traced functions in SKILL.md.
Boundary markers: Not explicitly defined in the trace wrapper code.
Capability inventory: Data transmission to external observability API.
Sanitization: Explicit guidance provided in the 'Process inputs/outputs' section for filtering sensitive content.

langsmith-observability