What problem does it solve? Delegating high-volume LLM tasks (diff reviews, log summarization, batch extraction) to a cheaper OpenAI-compatible endpoint saves subscription quota, but doing it safely is hard: API keys leak into command lines, reasoning models silently burn the token budget, and endpoints may forward data to public clouds. This Skill provides the scripts, configuration templates, and verification discipline to delegate work without exposing secrets or trusting unverified output. ## Core Features & Use Cases - Direct chat calls via oa-chat: A Bash script that calls any OpenAI-compatible endpoint (LiteLLM, vLLM, Ollama /v1, LM Studio, OpenRouter) with the API key kept out of argv, token usage reported on stderr, and distinct exit codes for empty or truncated responses. - Agent-mode delegation via agentry: A mergeable settings snippet defining a delegada task-class with an explicit egressClass, plus a wrapper script that resolves the API key from environment, vault command, or file without ever printing it. - Reasoning-model cost control: The --no-think flag disables chain-of-thought on vLLM/SGLang servers, avoiding empty responses caused by finish_reason=length. - Use Case: Pipe a staged git diff into oa-chat with a reviewer system prompt, redirect the full response to a file, and read only a sampled excerpt into the session context. ## Quick Start Ask the agent to delegate the review of the current staged diff to the LiteLLM gateway using the oa-chat script and save the response to a file.
