The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill utilizes standard shell utilities such as grep, sed, sort, tail, chmod, and zip for internal state management and log processing.
Evidence: These tools are used to maintain the observation logs and package skill updates within the user's workspace.
[DATA_EXPOSURE_AND_EXFILTRATION]: No exfiltration patterns were detected. The skill explicitly states that it does not require external network access for its logic.
Evidence: The skill implements a robust 5-layer confidentiality protocol in references/skill-authoring.md to ensure sensitive client information is stripped from logs and generated skills.
[PROMPT_INJECTION]: The skill provides meta-instructions for behavioral monitoring and workflow optimization. No malicious bypass or safety override patterns were found.
Evidence: Instructions focus on capturing methodology and user corrections for skill improvement rather than overriding system constraints.
[INDIRECT_PROMPT_INJECTION]: The skill processes session data and user-provided handoff documents as sources for observations.
Ingestion points: Current conversation history, skill-observations/log.md, and handoff-doc text (references/environments.md).
Boundary markers: Explicit instructions to attribute derived observations to the handoff document source (references/environments.md).
Capability inventory: Local file system access, shell command execution for text processing, and task scheduling (references/weekly-review.md).
Sanitization: A mandatory 5-layer confidentiality sweep and the requirement that all skill updates must be manually reviewed and installed by the user (references/skill-authoring.md).

task-observer