The Agent Skills Directory

[DATA_EXFILTRATION]: The skill observes and records user prompts and tool outputs to local log files (observations.jsonl). While it attempts to scrub secrets using regular expressions, this creates a comprehensive record of user activity that could contain sensitive information.
[EXTERNAL_DOWNLOADS]: The instinct-cli.py script allows importing data from arbitrary URLs via the import command. While the content is parsed as YAML and not directly executed, it can influence the agent's future behavior.
[COMMAND_EXECUTION]: The skill executes shell commands (e.g., git, claude CLI) through subprocesses in several scripts (detect-project.sh, instinct-cli.py, observer-loop.sh). These are used for project detection and background analysis.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. A background agent analyzes session logs to create new instructions ('instincts'). Malicious content repeated in tool outputs or prompts could trick the system into generating harmful instructions that persist across sessions.
Ingestion points: ~/.claude/homunculus/projects/*/observations.jsonl (captured from tool outputs and user prompts).
Boundary markers: None are present in the observation logs.
Capability inventory: The generated instincts can instruct the agent to use any available tools (e.g., Edit, Bash, Grep).
Sanitization: Secret scrubbing is performed on logs, but no semantic validation is applied to the learned patterns.

continuous-learning-v2