The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the 'sentence-transformers' package via pip and fetches pre-trained embedding models (e.g., 'BAAI/bge-small-zh-v1.5') from HuggingFace. These are standard operations for a skill implementing semantic search using established libraries and well-known services.
[COMMAND_EXECUTION]: Several Python scripts use subprocess.run or subprocess.call to manage internal operations. These include:
Forwarding installation commands to the initialization script in main.py.
Installing necessary Python dependencies (sentence-transformers) in scripts/service/init/helpers.py.
Triggering the index synchronization script (sync_index.py) in scripts/service/manage/commands/_helpers.py and scripts/service/hooks/sync_and_cleanup.py.
Launching 'datasette' for database visualization in scripts/service/manage/commands/cmd_db.py. All detected executions target the local Python interpreter or its installed modules for legitimate system functionality.
[DATA_EXFILTRATION]: No unauthorized data exfiltration patterns were detected. Network activity is limited to downloading models from HuggingFace. Local memory data (facts, summaries, logs) is stored within the project's .cursor/skills/memory-data/ directory or a global ~/.memory/ folder.
[PROMPT_INJECTION]: The instructions in SKILL.md and memory-rules.mdc utilize strong imperative language (e.g., "Mandatory", "Must") to define the agent's memory-handling logic. This is standard for technical guidelines and does not represent an attempt to bypass AI safety filters.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (user messages and prior conversation history) and stores it in memory files. While this creates a surface for indirect injection, the skill includes basic content sanitization in save_fact.py and uses structured data formats to minimize risk.

memory