The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/conversation_history.py reads private conversation transcripts from the user's home directory and writes them to the repository.\n
Evidence: Lines 27-41 in scripts/conversation_history.py derive paths to ~/.claude/projects/ based on the repository root.\n
Risk: Sensitive chat history containing logic, decisions, or code could be accidentally pushed to public remotes if the generated docs/conversations/ folder is committed.\n- [COMMAND_EXECUTION]: The skill instructions require the execution of a Python script and recommend its integration into automated Git hooks.\n
Evidence: SKILL.md provides commands for chmod +x and python3 scripts/conversation-history.py as part of the installation and pre-push lifecycle.\n- [PROMPT_INJECTION]: The skill introduces a surface for indirect prompt injection by using past session logs as authoritative context for future agent actions.\n
Ingestion points: ~/.claude/projects/*.jsonl (found in scripts/conversation_history.py line 147) and .entire/logs/entire.log (line 52) are parsed for content.\n
Boundary markers: The script uses Obsidian callout syntax (e.g., > [!quote]) in the generate_session_doc function to distinguish roles, providing some structure but not complete protection against adversarial instructions embedded in logs.\n
Capability inventory: The agent consuming these logs typically has access to shell and file system tools via Claude Code, as noted in the SKILL.md metadata.\n
Sanitization: Basic sanitization of XML tags is performed in the _callout_safe function (line 404), but the script does not validate the semantic content for malicious instructions.

knowledge-graph-memory