The Agent Skills Directory

[SAFE]: The skill implements a persistent world model using a local SQLite database for project-specific knowledge retention.
[SAFE]: It installs lifecycle hooks by additively merging entries into the agent's configuration, which is the standard mechanism for integrating such functionality.
[SAFE]: Fact-harvesting logic in assets/harvest.py maintains a strict trust boundary by only processing user and assistant messages, explicitly ignoring tool outputs to prevent indirect prompt injection.
[SAFE]: The skill relies solely on the Python standard library and performs no network operations or exfiltration of sensitive data.
[SAFE]: Execution observation is performed through structural parsing of shell commands without actually running them, ensuring a safe observation loop.
[SAFE]: Subprocess calls in the evaluation harness are used for testing purpose and are not part of the runtime skill hooks.

world-model-ledger