The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled Python script (scripts/check_saturation.py) to evaluate the progress and depth of research tasks. Technical review of the script confirms it only performs benign text-based heuristic checks, such as counting unique domains and matching keywords against a specification, without any dangerous system calls.
[EXTERNAL_DOWNLOADS]: The skill is designed to perform extensive web exploration via the agent-browser tool. This involves downloading and processing content from external websites, which is the primary intended functionality for a deep research tool.
[PROMPT_INJECTION]: The skill exhibits an inherent surface for indirect prompt injection as it processes untrusted data from the web and stores it in knowledge_fragments.md files for later synthesis.
Ingestion points: External data enters the agent context via web search results and site extractions saved in knowledge_fragments.md files within task-specific directories.
Boundary markers: The instructions do not define specific delimiters or "ignore previous instructions" warnings when the agent reads or synthesizes these fragments.
Capability inventory: The primary risk is limited to the synthesis phase where the agent compiles research; the saturation script used for task management does not execute fragment content.
Sanitization: There is no evidence of content filtering or sanitization of the data retrieved from the web before it is written to the workspace.

deep-research