The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard shell utilities (find, grep, awk, sed, mv, cp, chmod) to manage a directory-based observation log and perform skill updates.
Evidence: Bash snippets in SKILL.md and references/weekly-review.md perform file operations within a pinned absolute workspace path.
Safety Measures: The skill enforces strict quoting of paths to prevent word-splitting attacks, uses the noclobber (set -C) redirect to prevent accidental file overwrites, and employs find -print0 with read -d '' to safely handle filenames containing spaces.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface by ingesting and acting upon instructions found within its own observation log files during review tasks.
Ingestion Points: The agent reads Markdown files from the observation-log/ directory during the Session Start Protocol and the Weekly Review (weekly-review.md).
Boundary Markers: Observations are encapsulated in structured Markdown files with YAML frontmatter delimiters (---) and predefined headers (Issue:, Principle:).
Capability Inventory: The skill executes shell commands for file management (mv, cp, mkdir) and can stage new skill versions.
Sanitization: The migrate-log.py script uses json.dumps to sanitize values emitted to YAML frontmatter. The skill's documentation explicitly requires a "Pre-Flight" verification step and adversarial review for any principle being promoted to a core skill.

task-observer