The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves ingesting user queries to match them against a registry of available agents, which creates a surface for indirect prompt injection.
Ingestion points: User input is passed as a command-line argument to match_skills.py and orchestrate.py in the skill's mandatory workflow.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands within the user query interpolation.
Capability inventory: The skill maintains a local registry (registry.json) and performs file system operations to discover other SKILL.md files. It executes local scripts via subprocess.run to update its state.
Sanitization: The Python scripts use tokenization and keyword matching via regular expressions to process the query, which naturally limits the potential for control-flow hijacking within the scoring logic.
[COMMAND_EXECUTION]: The skill requires the execution of several Python scripts to perform its discovery and orchestration tasks.
Evidence: The workflow utilizes python agent-orchestrator/scripts/scan_registry.py, match_skills.py, and orchestrate.py to manage the ecosystem.
The match_skills.py script uses subprocess.run([sys.executable, ...]) to invoke the registry scanner. This implementation is safe as it targets an internal script within the same package using the current interpreter, avoiding path injection or remote execution risks.

agent-orchestrator