The Agent Skills Directory

[COMMAND_EXECUTION]: The test suite (tests/test_stdlib_parity.py) executes subprocess.run to verify the Python environment and ensure heavy dependencies like numpy or scipy are not loaded. This is a standard environment check for the plugin's performance guarantees.
[DYNAMIC_EXECUTION]: The plugin performs runtime monkey-patching of the agent's internal prompt_builder.build_skills_system_prompt function in __init__.py. This modification is intended to implement the Phase 1 compaction logic described in the documentation, allowing the plugin to override the agent's default system prompt generation for token optimization.
[INDIRECT_PROMPT_INJECTION]: The skill processes user messages and external skill descriptions from ~/.hermes/skills to build a BM25 index and inject relevant content into the LLM context, which constitutes a potential vulnerability surface.
Ingestion points: User messages in _on_pre_llm_call and SKILL.md files located in the agent's skills and plugins directories.
Boundary markers: Injected content is delimited by a "## Retrieved Skills" markdown header to separate it from user input.
Capability inventory: The skill lacks network or sensitive file-write capabilities; it only retrieves text for injection into the agent's context.
Sanitization: The skill implements length truncation for injected descriptions, limiting each to 200 characters.

skill-retrieval