The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes local shell commands and subprocess calls across several scripts (detect-project.sh, instinct-cli.py, observe.sh, start-observer.sh) for project identification, session logging, and process management. These are implemented using safe practices, including argument lists instead of shell strings to prevent command injection.
[EXTERNAL_DOWNLOADS]: The instinct-cli.py utility provides an import command that allows users to fetch instinct definitions from external URLs via urllib.request.urlopen. This is a user-initiated administrative feature where the content is parsed as structured data before being written to local storage.
[DATA_EXFILTRATION]: The skill captures session activity, including tool inputs and outputs, into local JSONL logs. To protect privacy and credentials, a regex-based secret scrubbing mechanism is implemented in observe.sh to redact patterns matching API keys, tokens, and passwords before they are persisted.
[PROMPT_INJECTION]: The architecture presents a surface for indirect prompt injection (Category 8) because it uses an LLM to generate behavioral instructions based on logs of session activity. Malicious tool outputs encountered during a session could potentially influence the generated 'instincts'.
Ingestion points: Session data is logged to observations.jsonl, which is later read by the observer background agent in observer-loop.sh.
Boundary markers: The system uses YAML frontmatter delimiters (---) and structured Markdown headers to separate instruction metadata from the behavioral descriptions.
Capability inventory: The system performs git operations, local file management, and supports user-initiated network imports via the CLI.
Sanitization: Observations are scrubbed for secrets in observe.sh, and the instinct-cli.py tool validates instinct IDs and file paths to prevent traversal attacks.

continuous-learning-v2