The Agent Skills Directory

[NO_CODE]: The skill consists entirely of instructional Markdown content and documentation. It does not distribute or execute any scripts (Python, Node.js, Shell) or binary files, instead providing a configuration layer for existing literature search MCP tools.
[DATA_EXFILTRATION]: The skill manages search plan persistence by reading and writing files to scoped directories (~/.pubmed-search-mcp/pipelines/ and .pubmed-search/pipelines/). This behavior is consistent with the skill's primary purpose of workflow persistence and is limited to specific user and project data paths.
[COMMAND_EXECUTION]: The skill implements a domain-specific language (DSL) for defining Directed Acyclic Graphs (DAGs) of search steps. Execution is restricted to an internal set of ten actions (e.g., search, pico, metrics, merge) and is protected by a validation engine that enforces schema integrity through auto-fix and semantic rules.
[PROMPT_INJECTION]: The skill processes untrusted pipeline configurations, representing an indirect injection surface. This risk is mitigated by the following controls:
Ingestion points: YAML/JSON configuration strings provided to search tools and local files loaded via specific path triggers (documented in SKILL.md and pipeline-tutorial.md).
Boundary markers: The system requires structured YAML definitions that follow a strict action schema.
Capability inventory: Functionality is confined to reading/writing configurations and executing literature search MCP actions.
Sanitization: The platform applies 21 validation and auto-fix rules to all incoming pipeline definitions before processing.

pipeline-persistence