The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the local repository (code, configuration, measurements) to provide context for the Jev model. • Ingestion points: SKILL.md (Step 3) instructs the agent to read various repository files. • Boundary markers: SKILL.md (Step 5) requires the agent to include instructions for the model to treat quoted content as data and not instructions. • Capability inventory: The skill performs network calls to api.typesafe.ai (references/hosted-api.md) and executes local syntax checks using python3 (references/hosted-api.md, tests/test_procedure.py). • Sanitization: SKILL.md (Step 4) specifies a sanitization step to exclude credentials, personal data, and unnecessary private paths from the payload.
[COMMAND_EXECUTION]: The skill documentation and test suite use shell commands for local validation. • Evidence: tests/test_procedure.py uses subprocess.run to execute a Python one-liner for JSON syntax validation. references/hosted-api.md suggests using a similar python3 -c command for local payload verification.
[DATA_EXFILTRATION]: The skill sends gathered evidence to an external API endpoint. • Evidence: The skill communicates with https://api.typesafe.ai/v1/systemone as part of its primary workflow. Risk is mitigated by explicit instructions to sanitize payloads and exclude sensitive credentials or personal data before transmission.
[DYNAMIC_EXECUTION]: The skill utilizes runtime command execution for configuration validation. • Evidence: The test suite in tests/test_procedure.py dynamically constructs and executes a Python command to verify JSON integrity.

ask-jev