The Agent Skills Directory

[DATA_EXFILTRATION]: The script export_owner_prompts.py reads sensitive interaction history from ~/.claude/projects and ~/.codex/sessions.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical chat data through a subagent. Malicious instructions in past prompts could influence the goal-mapping process.
Ingestion points: User interaction logs in JSONL format from Claude Code and Codex.
Boundary markers: No specific delimiters or safety instructions are added to the aggregated prompts file.
Capability inventory: The skill writes results to disk and updates the claude-mem work state through a local worker API.
Sanitization: The tool strips specific memory tags but does not perform content-based filtering of the prompt text.
[COMMAND_EXECUTION]: The export_owner_prompts.py script executes local git commands via subprocess.run to manage output directory visibility.

owner-goals