The Agent Skills Directory

[COMMAND_EXECUTION]: The 'Researcher OS' component includes automation scripts (researcher/scripts/loop_step.py, researcher/scripts/loop_daily.py) that use subprocess.run to execute other scripts and system utilities such as git and npm. Additionally, the repository provides macOS launchd configuration files and an installation script (researcher/orchestration/launchd/install.sh) designed to run these processes as background daemons, establishing persistence across user sessions.
[EXTERNAL_DOWNLOADS]: The research automation loop implements a mechanism to fetch content from remote URLs using Python's standard library (researcher/scripts/loop_step.py). This allows the ingestion of untrusted data from arbitrary sources into the local research environment for processing.
[PROMPT_INJECTION]: The autonomous research workflow creates a significant surface for indirect prompt injection. Agents are tasked with reading and analyzing untrusted external content to extract patterns and artifacts, which are subsequently used to generate skill updates and Pull Request proposals (researcher/runbooks/autonomous-research-loop.md). Malicious instructions embedded in these retrieved documents could manipulate the agent's reasoning or influence the code it proposes.
[COMMAND_EXECUTION]: An example implementation of a calculator tool provided in examples/interleaved-thinking/examples/03_full_optimization.py utilizes the eval() function to process logic provided in strings. While it attempts to restrict access by clearing __builtins__, this pattern is inherently susceptible to exploitation.

context-engineering-collection