The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external security reports, SARIF files, and code excerpts to perform analysis. This creates a theoretical surface for indirect prompt injection.
Ingestion points: External findings are ingested in Step 1 of the workflow (described in SKILL.md and references/input-normalization.md), including SARIF results, weAudit annotations, and Markdown finding excerpts.
Boundary markers: The skill documentation includes logical boundaries (warning not to trust scanner severity and requiring manual review targets), but does not explicitly specify technical prompt delimiters or 'ignore' instructions for the LLM when interpolating this data.
Capability inventory: The skill uses Bash, Read, Grep, Glob, and Write tools, and executes Python code snippets to interact with the trailmark query engine.
Sanitization: The skill employs a formal normalization step in references/input-normalization.md that maps various input types into a structured JSON record before graph analysis, which serves as a sanitization and validation layer.
[DYNAMIC_EXECUTION]: The skill utilizes Python code snippets and the trailmark library to perform graph-based queries and pre-analysis on target directories.
Evidence: references/query-recipes.md provides Python templates for initializing the QueryEngine, running engine.preanalysis(), and executing reachability queries. This execution is local to the analysis environment and is the core functionality provided by the vendor to analyze the target codebase.

trailmark-finding-triage