The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen to interface with local authenticated AI CLI tools (claude, codex, kimi, grok, pi) and the herdr visualization tool. This is the intended mechanism for performing multi-engine consensus tasks.
[DYNAMIC_EXECUTION]: The engine_legs.py script utilizes importlib.import_module to load backend-specific logic (e.g., codex_backend, claude_backend) dynamically. These modules are part of the skill's own internal library.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided questions and context files which are interpolated into expert prompts. It mitigates injection risks by using structured XML-like boundary markers (e.g., <вопрос>, <контекст>) and includes a specialized regex-based guard (codex_guard.py) to block actionable offensive requests before they are sent to the AI engines.
Ingestion points: User-provided question arguments and optional --context-file inputs in delphi.py and delphi_viz.py.
Boundary markers: Uses explicit <вопрос>, <контекст>, and <твой-слепой-ответ> tags in prompt construction.
Capability inventory: Can execute shell commands via configured LLM CLIs and write results to a local node_dir for visualization.
Sanitization: Implements codex_guard.py to filter for offensive subjects paired with produce/run verbs.
[EXTERNAL_DOWNLOADS]: The deepseek_router.py utility makes network requests to api.deepseek.com if a DEEPSEEK_API_KEY is provided. This is used for its documented purpose of routing potentially sensitive prompts away from OpenAI to prevent account flags.

delphi