The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive agent-internal files located at ~/.claude/homunculus/observations.jsonl and ~/.claude/homunculus/buffer.jsonl to collect context about the current session's activities.
[COMMAND_EXECUTION]: Executes external bash and python scripts (scan-inventory.sh and similarity-scorer.py) located in the ~/.claude/skills/skill-factory/ directory to analyze skill patterns.
[INDIRECT_PROMPT_INJECTION]: Ingests potentially untrusted data from git logs, diffs, and session observations which are then processed by subagents to suggest automated actions. 1. Ingestion points: ~/.claude/homunculus/observations.jsonl, ~/.claude/homunculus/buffer.jsonl, and git metadata files. 2. Boundary markers: Prompt templates in references/ specify structured JSON output and scoped tasks for each subagent. 3. Capability inventory: The skill has the ability to modify local files (Edit/Write) and generate new agent behavior files ('instincts'). 4. Sanitization: Uses a structured JSON intermediate format and requires human-in-the-loop confirmation via AskUserQuestion before executing generated actions.
[DYNAMIC_EXECUTION]: The skill generates new behavioral instructions ('instincts') based on session analysis and writes them to ~/.claude/homunculus/instincts/personal/, which can influence the agent's future behavior across sessions.

session-wrap