The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes and monitors LLM interactions, which often involve untrusted user inputs. This creates a surface for indirect prompt injection if malicious instructions in processed data are executed by the agent or evaluators. * Ingestion points: Tracing functions, dataset creation, and evaluation pipelines (SKILL.md). * Boundary markers: Not explicitly enforced in the generic examples. * Capability inventory: The skill interacts with the LangSmith API, reads local files for CSV uploads, and can be configured to read files for trace attachments. * Sanitization: Documentation provides a sanitize_inputs example for redacting sensitive fields.
[DATA_EXFILTRATION]: The skill provides functionality to read and transmit local files and application data to the LangSmith observability platform. * Evidence: Documentation in references/advanced-usage.md shows how to attach local files to runs using the dangerously_allow_filesystem=True flag.
[COMMAND_EXECUTION]: Installation instructions require executing shell commands to install the SDK. * Evidence: SKILL.md and troubleshooting guides contain pip install langsmith commands.
[EXTERNAL_DOWNLOADS]: The skill supports fetching external resources from the LangChain Hub. * Evidence: The code uses client.pull_prompt() to fetch resources from the official langchain-ai organization on the LangSmith Hub.

langsmith-observability