# `Rho.LLM.ReqLLM`
[🔗](https://github.com/GenericJam/rho/blob/v0.1.0/lib/rho/llm/req_llm.ex#L1)

The model calls through req_llm's providers, streamed, with the
provider's credential from `Rho.Auth` (fetched, and for a subscription
refreshed when due, per call):

  * `anthropic:` signed in to a subscription (Claude Pro/Max) →
    `auth_mode: :oauth`, `access_token`, `with_claude_subscription: true`
    (Claude Code's betas and identity); with an API key → `api_key`.
    Either way prompt caching (`anthropic_prompt_cache: true`,
    `anthropic_cache_messages: -1`): req_llm puts a `cache_control`
    breakpoint on the last tool, the system prompt and the last message,
    so each call reads the previous one's prefix from the cache instead
    of paying (and spending rate limit on) the whole prompt again
  * `openai_codex:` (ChatGPT Plus/Pro) → `auth_mode: :oauth`,
    `access_token`, `chatgpt_account_id`, and the session id as
    `session_id` (the prompt cache key)
  * `openai:` (an API key) → `api_key`, and the session id as
    `prompt_cache_key`
  * `google:` (a Gemini API key) → `api_key`, sent in the
    `x-goog-api-key` header (`google_auth_header: true`) rather than the
    URL, so it stays out of logged URLs

A provider without a credential fails with `{:signed_out, provider}`;
any other model prefix with `{:other, _}`. `max_tokens` is always
explicit (req_llm's default is the model's whole output limit).

Pictures the model makes come back as the reply's `:images`: the
`:image` content parts (bytes and `media_type`) of the response's
message. That message already holds the streamed `:content_part`
chunks (req_llm's response builder materializes them; preview frames,
`stream_only?`, it leaves out), so each picture counts once. An
`:image_url` part (a link, not bytes) is skipped: rho doesn't fetch it.

Every call that reached the provider is recorded with
`Rho.Usage.record/3`: its tokens and cost, the subscription
windows its response headers report, and a 429's reset.

A 429 (or a 529 overloaded) comes back as `{:rate_limited, limit}`
(`t:Rho.LLM.limit/0`): its `retry-after`, and when
a usage window is used up, which one and until when: from the response's
headers, a Codex usage-limit body, else the windows
`Rho.Usage` recorded from earlier calls (an Anthropic 5-hour
window at 100 % answers with the same message as a short throttle).

---

*Consult [api-reference.md](api-reference.md) for complete listing*