The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs dynamic module loading in scripts/search.py to import sibling scripts using computed file paths at runtime.
[CREDENTIALS_UNSAFE]: The skill reads sensitive information from local files such as ~/.git-credentials and ~/.openclaw/credentials/search.json to obtain authentication tokens for GitHub and various search APIs.
[EXTERNAL_DOWNLOADS]: The skill retrieves data from numerous external services, including the Tavily, Exa, and Grok APIs, as well as platforms like GitHub, Reddit, and Hacker News. These are well-known services and the downloads are essential for the skill's functionality.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from the web and includes it in prompts sent to an LLM for relevance evaluation.
Ingestion points: scripts/fetch_thread.py and scripts/search.py fetch data from arbitrary external URLs.
Boundary markers: Instructions to ignore embedded commands are present in some components but absent in the relevance_gate.py logic.
Capability inventory: The skill has full network access via Python's requests and urllib modules.
Sanitization: The skill performs basic HTML tag removal but does not sanitize content against adversarial prompt instructions.

search-layer