The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The MemoryManager class implemented in the skill's code snippets provides a mechanism to record user queries, application usage, and AI responses into a local SQLite database. This data is subsequently retrieved to construct a summary prompt for the AI (generateSummaryPrompt) without sufficient isolation, sanitization, or boundary markers to prevent the AI from following malicious instructions that might be embedded in the stored interaction history.
Ingestion points: User queries and AI responses are captured and stored in the user_intents table via the recordIntent method in SKILL.md.
Boundary markers: The generateSummaryPrompt method interpolates database records directly into a new prompt using simple bullet points, lacking specific delimiters or "ignore previous instructions" directives to separate system instructions from untrusted historical data.
Capability inventory: The skill defines capabilities for network operations (URLSession), local file and database access (FileManager, SQLite), and shell process execution (Process).
Sanitization: While a basic keyword filter (sensitiveKeywords) is present to exclude specific terms like "password" or ".env" from being recorded, it does not validate or sanitize the content for injection patterns.
[EXTERNAL_DOWNLOADS]: The skill's build and configuration instructions direct the user to download the application source code from a GitHub repository (basionwang-bot/HermesPet) and install several third-party CLI tools from external package registries.
Fetches project source code via git clone from a non-vendor repository.
Recommends installation of the @openai/codex-cli and openclaw packages from the NPM registry.
Recommends installation of the anthropic-cli via Homebrew.

hermespet-macos-ai-companion