The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The skill executes uv run python -m runtime.harness scripts/braintrust_analyze.py. Because the content of scripts/braintrust_analyze.py is not included in the skill definition, its specific actions cannot be verified, representing a risk of arbitrary local code execution.
CREDENTIALS_UNSAFE (LOW): The documentation directs users to store BRAINTRUST_API_KEY in sensitive environment files like ~/.claude/.env or project .env. This highlights an attack surface for credential access.
DATA_EXFILTRATION (MEDIUM): The skill is designed to read and process Braintrust tracing data, which contains full session histories, including prompts and tool outputs. Without visibility into the script, there is a risk this sensitive data could be sent to external endpoints.
Indirect Prompt Injection (MEDIUM): This skill has an inherent vulnerability surface due to its primary function.
Ingestion points: Reads historical session logs via the Braintrust API.
Boundary markers: None identified in the command descriptions or provided documentation.
Capability inventory: Executes local Python code via uv run, which can access the file system and network.
Sanitization: No evidence of sanitization or escaping of the historical session content before it is processed and presented back to the agent.

braintrust-analyze