The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied query text directly via command-line arguments or agent context and interpolates it into prompts sent to external LLMs.
Ingestion points: User text is read via sys.argv[1:] in lookup.py and research_lookup.py and passed into structural query objects.
Boundary markers: No strict boundary markers or delimiters are automatically added to separate user input from internal system instructions within the custom academic prompt format constructed inside _format_academic_prompt.
Capability inventory: The ingested queries trigger outbound HTTP requests using requests.post to https://openrouter.ai/api/v1/chat/completions or invoke the openai SDK to talk to https://api.parallel.ai.
Sanitization: Input text is converted to lowercase for simple keyword routing checks (any(kw in query_lower for kw in self.ACADEMIC_KEYWORDS)), but there is no explicit sanitization to neutralize malicious embedded markdown or LLM instructions.
[COMMAND_EXECUTION]: The documentation showcases how users can manually execute the python scripts using standard terminal command lines (python scripts/research_lookup.py --model-info), which is standard utility orchestration rather than a vulnerability.
[EXTERNAL_DOWNLOADS]: The skill uses Python requests and the openai client to fetch analytical and academic text over the network from remote services (api.parallel.ai and openrouter.ai), which represents normal vendor infrastructure connectivity.

research-lookup