The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements an autonomous 'while(true)' loop designed to have the agent execute system commands for testing, building, and environment scanning. While intended for self-improvement, this provides a wide capability for command execution within the agent's runtime environment.
[DYNAMIC_EXECUTION]: The core functionality of the skill is to enable the agent to dynamically generate, modify, and execute its own code across multiple iterations. The rules/autonomous-loop.md and SKILL.md files detail logic for writing to files, running scripts, and updating the agent's own ruleset.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection during its 'Perceive' phase.
Ingestion points: The agent is instructed to scan the workspace environment (scan_environment), detect issues in code, and potentially perform web searches (web_search mentioned in rules/safety.md), all of which may contain untrusted data.
Boundary markers: The provided pseudocode and instructions do not specify explicit boundary markers or delimiters to separate untrusted data from the agent's core instructions.
Capability inventory: The skill has capabilities for file system writes (safe_write, write_file), command execution (running tests and builds), and network access (search).
Sanitization: While the skill includes a 'Security review' phase to detect dangerous patterns (e.g., sudo, rm -rf), there is no explicit sanitization or escaping of ingested external content before it is processed by the LLM in the decision-making phase.

zi-kong