The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes conversation history, which is untrusted external data, to generate summaries and evaluation probes. This ingestion of external content creates a surface for indirect prompt injection. \n
Ingestion points: scripts/compression_evaluator.py (processes the conversation_history input). \n
Boundary markers: The skill employs structured markdown sections for organization but does not implement explicit security boundaries to isolate untrusted content. \n
Capability inventory: The skill logic is limited to text processing and scoring stubs; it does not possess capabilities for network exfiltration or unauthorized file system modification. \n
Sanitization: No sanitization or validation of the ingested conversation history is performed before processing. \n- [DYNAMIC_EXECUTION]: The test suite tests/test_compression_evaluator.py uses dynamic loading to execute the skill's internal evaluation script. \n
Evidence: The script utilizes importlib.util to load and execute scripts/compression_evaluator.py from a relative file path computed at runtime.

context-compression