The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses the Python exec() function in scripts/rlm_repl.py to run code generated by the LLM. While this is the primary intended functionality of the 'Recursive Language Model' pattern, the skill documentation explicitly notes that the environment is not a hardened sandbox, meaning the LLM-generated code could perform unauthorized actions if the runtime is not separately secured.
[DYNAMIC_EXECUTION]: The ReplSession class in scripts/rlm_repl.py uses the __import__ function to dynamically load libraries (re, math, json, textwrap) into the execution context for the LLM.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data from PDFs and source code repositories, creating a vulnerability to indirect prompt injection.
Ingestion points: scripts/load_codebase.py and scripts/load_pdf.py ingest content from entire directories and PDF files respectively.
Boundary markers: The instructions do not define clear boundaries or 'ignore' rules for content ingested from these external sources.
Capability inventory: The skill possesses powerful capabilities, including arbitrary Python execution (exec) and network-based LLM sub-calls (llm_query).
Sanitization: Document content is loaded directly into the context without sanitization or filtering for malicious instructions.
[DATA_EXFILTRATION]: The DEFAULT_TEXT_EXTS list in scripts/load_codebase.py includes .env files. This allows the skill to load and potentially exfiltrate sensitive environment variables or credentials if they are present in the repository being analyzed.
[COMMAND_EXECUTION]: The skill allows the LLM to generate and execute logic to interact with the file system and external APIs, which poses a command execution risk if used in an environment without restricted permissions.

long-document-evidence-reader