The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains logic for A/B testing that involves executing other skills and skill variants.
Evidence: In SKILL.md, the ab_test_skill function calls run_skill(skill_name, ...) and run_skill_variant(variant_b_path, ...), enabling the dynamic execution of code from file paths determined during analysis.
The "Continuous Improvement Loop" section also describes the automated implementation and testing of skill variants, which implies the agent has the capability to modify its own executable logic.
[DATA_EXFILTRATION]: The skill accesses sensitive local operational data that may contain user-provided information or system secrets.
Evidence: The skill requires access to .omc/metrics.db, .omc/sessions/, and .omc/state/. The database schema explicitly logs a context field (containing session details), error_message, and stack_trace, which are high-value targets for data exposure if the agent is manipulated.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through the analysis of data generated by other skills.
Ingestion points: The skill reads error_message, trigger_keyword, and context from the .omc/metrics.db and .omc/sessions/ files.
Boundary markers: None identified. The skill processes these fields directly without delimiters or instructions to ignore embedded commands.
Capability inventory: The skill can modify the local file system (via Git and optimization logic) and execute local code (run_skill_variant).
Sanitization: None identified. Malicious payloads embedded in the error logs of other skills could influence the analysis logic, potentially leading to the generation of harmful improvement proposals or the execution of malicious variants during A/B testing.

agent-self-improvement