Module llm 

Optional: let a model phrase the reply.

Off unless serve --llm is asked for and a key is in the environment. The board itself never goes near the model — ids, counts and state transitions are computed in crate::agent and the model is handed the finished sentence to rewrite. That is deliberate: a dogfood app whose assertions depend on a model is a dogfood app that cannot be tested.

Same wire format and the same environment variables as e2e/src/llm.rs, and the same point: an OpenAI-compatible /chat/completions endpoint reached with reqwest and two serde structs. There is no LLM crate here either.

Structs

MissingApiKey

Voice::from_env found no API key, and the endpoint needs one.

Voice

A model that rephrases the agent’s replies.

Constants

API_KEY_ENV

The environment variable holding the API key.

BASE_URL_ENV

The environment variable holding the base URL.

DEFAULT_BASE_URL

Where requests go unless BASE_URL_ENV says otherwise.

DEFAULT_MODEL

The model this talks to by default. Pinned, never a *-latest alias.

FALLBACK_API_KEY_ENV

Read when API_KEY_ENV is unset — the default endpoint is Gemini’s.

MODEL_ENV

The environment variable holding the model id.

QWEN_API_KEY_ENV

The key that goes with QWEN_BASE_URL_ENV.

QWEN_BASE_URL_ENV

Qwen Cloud’s OpenAI-compatible mode, read when BASE_URL_ENV is unset: the base URL, its key, and its model, in that order.

QWEN_DEFAULT_MODEL

The Qwen model used when QWEN_MODEL_ENV is unset. Pinned, like the default.

QWEN_MODEL_ENV

The model that goes with QWEN_BASE_URL_ENV, when MODEL_ENV is unset.