The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell scripts and Python subprocesses to automate background observation and analysis tasks. It executes standard tools like git for project identification and the claude CLI for pattern extraction. The implementation uses secure patterns, such as passing data via environment variables and stdin, to prevent shell injection vulnerabilities.
[DATA_EXFILTRATION]: The observe.sh hook captures session interaction data. To mitigate the risk of credential exposure, the skill includes logic to automatically scrub sensitive patterns (API keys, tokens, passwords) using Python-based regex redaction before the data is stored locally or processed by the background agent.
[EXTERNAL_DOWNLOADS]: The CLI tool provides an import command that can fetch instinct definitions from remote URLs. This is a primary feature for sharing learned behaviors, and the tool validates target file paths to prevent directory traversal during the import process.
[PROMPT_INJECTION]: The background observer agent (Haiku) processes session logs that include untrusted user input and tool outputs. This presents an indirect prompt injection surface where malicious log content could theoretically influence the creation of new instincts, although the impact is confined to the skill's internal knowledge base.

continuous-learning-v2