The Agent Skills Directory

[DATA_EXFILTRATION]: The skill monitors and logs all tool usage, which may include sensitive data.
Evidence: The hooks/observe.sh script is designed to be used with PreToolUse and PostToolUse hooks, capturing the input and output of every tool call (e.g., Read, Edit, Bash). This data is logged to a local file at ~/.claude/homunculus/observations.jsonl.
Evidence: The background agent defined in agents/start-observer.sh reads this log file and sends its content to an external LLM provider via the claude CLI for pattern analysis. This constitutes a network transfer of all captured session activity.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a feedback loop that processes untrusted data from tool outputs to generate new instructions.
Ingestion points: ~/.claude/homunculus/observations.jsonl via the hooks/observe.sh script, which records data from all tool interactions.
Boundary markers: None. The system does not use delimiters or instructions to differentiate between trusted session context and potentially malicious data returned by tools (e.g., content read from a URL or an external file).
Capability inventory: The system can write new instruction files (instincts) to ~/.claude/homunculus/instincts/personal/ and generate new skills, commands, or agents in ~/.claude/homunculus/evolved/ via the /evolve command.
Sanitization: None. The observe.sh script truncates output to 5000 characters but performs no sanitization or filtering of the content before it is analyzed by the observer agent.
[PERSISTENCE]: The skill configures persistent background processes and hooks that run globally.
Evidence: SKILL.md provides instructions to modify the global ~/.claude/settings.json to include hooks that execute the observe.sh script for every tool invocation.
Evidence: agents/start-observer.sh initiates a long-running background process that monitors the filesystem and performs scheduled analysis.
[EXTERNAL_DOWNLOADS]: The management script allows downloading content from arbitrary remote sources.
Evidence: scripts/instinct-cli.py uses the urllib.request module to fetch instinct definitions from user-supplied URLs.
[DYNAMIC_EXECUTION]: The skill dynamically generates and saves new executable instruction files.
Evidence: The observer agent and the evolve command generate markdown-based skill and agent definitions at runtime based on the analysis of session logs. These files are saved to directories where they are automatically loaded by the agent environment.
[COMMAND_EXECUTION]: The skill relies on shell scripts and Python code to perform management tasks and background analysis, including the invocation of the claude CLI.

continuous-learning-v2