The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes background processes (nohup) and platform hooks to maintain a continuous observation and analysis loop. agents/start-observer.sh spawns agents/observer-loop.sh as a background daemon to process session logs.- [EXTERNAL_DOWNLOADS]: The scripts/instinct-cli.py utility includes an import feature that can fetch 'instinct' definitions from arbitrary HTTP/HTTPS URLs. These definitions directly influence the agent's learned behaviors and generated skills.- [DATA_EXFILTRATION]: The background observer agent sends session observations to the LLM provider for pattern analysis via the platform's CLI. While the skill attempts to scrub common secrets (API keys, tokens) from these logs using regex, the underlying session activity is transmitted externally for processing.- [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface (Category 8) because the observer agent analyzes untrusted data from tool inputs and outputs. An attacker who can influence tool output (e.g., via a malicious web page or file) could inject instructions that the observer agent might misinterpret as legitimate patterns to be learned.
Ingestion points: agents/observer-loop.sh reads and processes the observations.jsonl file containing recorded tool calls.
Boundary markers: The prompt used to invoke the observer agent lacks explicit boundary markers or instructions to disregard potential commands embedded within the observation data.
Capability inventory: The system can write local files, generate new skill and agent definitions, and execute shell commands through its various scripts.
Sanitization: While the skill includes regex-based scrubbing for credentials in hooks/observe.sh, it does not perform semantic sanitization or validation of the natural language content processed by the background agent.

continuous-learning-v2