The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill correctly identifies that content fetched from the web is untrusted. Instructions in SKILL.md and evals/evals.json (specifically Eval #20) mandate that the agent must ignore embedded instructions in fetched Markdown and flag injection attempts to the user.
Ingestion points: scripts/deepseek_search.js ingests search snippets and full page content from the web.
Boundary markers: Explicit instructional warnings in SKILL.md serve as conceptual boundaries.
Capability inventory: The skill allows file writing (--save flag) and shell execution via the Node.js helper.
Sanitization: The helper script unescapes Markdown backslashes to ensure data integrity but relies on the LLM's instructions for behavioral sanitization.
[COMMAND_EXECUTION]: The regression test suite in scripts/test_helper.py utilizes subprocess.run() to execute the Node.js helper script. These calls are securely constrained to local paths and hardcoded test arguments, posing no risk to the host system during normal operation.
[EXTERNAL_DOWNLOADS]: The skill communicates exclusively with the official DeepSeek API endpoints (api.deepseek.com). This communication is necessary for the skill's primary function and follows standard security practices, including the use of environment variables for API key management and discouraging the printing of secrets.

deepseek-search