The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation focuses purely on technical strategies for context management and compression. No patterns designed to override safety filters, bypass instructions, or extract system prompts were found.
[DATA_EXFILTRATION]: There are no network operations (such as curl, wget, or HTTP libraries) or attempts to access sensitive system files (e.g., credentials or SSH keys). The Python scripts process provided text strings locally.
[REMOTE_CODE_EXECUTION]: No remote code download or execution patterns were identified. The skill does not perform package installations or download external scripts.
[DYNAMIC_EXECUTION]: The test suite (tests/test_compression_evaluator.py) uses importlib.util to dynamically load the local evaluation script for unit testing. This is a standard development practice for testing local modules and does not pose a security risk in this context.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process conversation history, which is untrusted data.
Ingestion points: The ProbeGenerator and StructuredSummarizer classes in scripts/compression_evaluator.py ingest conversation history.
Boundary markers: The skill utilizes structured Markdown headers (e.g., ## Files Modified) to delimit content.
Capability inventory: The provided scripts lack dangerous capabilities such as file system writing, network requests, or subprocess execution.
Sanitization: While regex-based extraction is used without explicit sanitization, the lack of dangerous downstream capabilities makes this a safe implementation surface.

context-compression