The Agent Skills Directory

[DYNAMIC_EXECUTION]: The framework features a 'Recursive Reflector' component described as writing and executing Python code at runtime to analyze agent execution traces and extract actionable strategies. This dynamic code generation is a high-risk capability that could be subverted if the input traces are compromised.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data which serves as the input for its learning engine.\n
Ingestion points: The agent ingests execution traces (conversation histories) and user feedback via the 'learn_from_traces' and 'learn_from_feedback' methods, as well as external content via the 'browser-use' integration.\n
Boundary markers: There are no explicit delimiters or security warnings shown to prevent the agent or the reflection engine from being influenced by instructions embedded within the processed traces.\n
Capability inventory: The skill has the ability to execute generated Python code, perform network operations to LLM providers and the kayba.ai API, and modify local JSON files.\n
Sanitization: The documentation does not indicate that trace content is sanitized or validated before being used by the Reflector to generate code.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of the 'ace-framework' and several sub-packages from the Python Package Index (PyPI). It also references a repository and documentation hosted under the 'kayba-ai' organization on GitHub, which is not a source on the trusted organizations list.\n- [COMMAND_EXECUTION]: The skill documentation includes instructions for executing shell commands, including environment variable exports for API keys and the use of the 'ace' and 'kayba' command-line interfaces for setup, configuration, and hosted service interaction.\n- [DATA_EXFILTRATION]: The skill provides commands to upload execution traces, which may contain sensitive conversation data, to the 'kayba.ai' hosted service via the 'kayba upload-traces' command.

agentic-context-engine