direct_llm¶
Catalogue-gap fallback: hand a single-shot natural-language task straight to the LLM and return its answer verbatim.
Entry¶
respond
Final output¶
direct_llm_response — the LLM's answer as plain text in response.
How it composes¶
Single-phase skill with no graph branching, no Control IR, and no preprocessor. The respond phase issues one LLM call (max_act_turns: 1) using the user_message artifact as input, then finishes immediately. Routing metadata sets priority: low / tier: fallback so the router always prefers a more specific skill when one matches.
Caveats¶
- No filesystem, shell, or web access — the phase has
allowed_ops: []. - Not appropriate for multi-step tasks or anything requiring side effects.
- When a more specific skill matches, use that instead.
Usage¶
Use when no specialised skill fits and the task can be completed in one LLM call (translate, summarise, answer a knowledge question, format a snippet, reword prose, etc.).
reyn run direct_llm "Translate 'hello' to Japanese."
reyn run direct_llm '{"type":"user_message","data":{"text":"What does idempotent mean?"}}'