During an investigation of Meta’s Muse personal‑agent framework, the author discovered a session that deviated from the default Avocado model and instead invoked a model logged as azure/muse‑special. Searching the Muse source revealed a comment describing a “GPT Responses model client via MAGI native Azure OpenAI lane,” and the model catalogue lists azure/muse‑special alongside azure/gpt‑5.6‑sol. Transcript analysis showed that calls to this model carry a signature tagged gpt_responses_v1 whose encrypted payload begins with the OpenAI‑specific string gAAAAA, and tool‑call identifiers appear as call_ followed by 24 mixed‑case characters, whereas Avocado sessions use call_ plus a 32‑hex‑character suffix. The broader catalogue shipped with Muse’s agent daemon contains roughly fifteen Avocado variants, plus Claude Opus 4.6‑4.8, Sonnet 4.6, Haiku 4.5, GPT‑5.5/5.6 variants via OpenAI, Azure and Codex, and Kimi K3 through Fireworks or Meta‑hosted routes. Anthropic support is implemented via dedicated Rust files (request_flow.rs, convert_prompt.rs, parse_sse_stream.rs). When muse‑special is used, the raw chain‑of‑thought is encrypted and sent back to Azure on the next turn; the binary notes that this encrypted reasoning cannot be overridden by the RL completion server, so Meta only sees the final reply, tool calls and any short reasoning summary. In contrast, Avocado models write thinking text directly into the transcript with an empty signature, making it available for reinforcement‑learning tuning unless the user opts out. No evidence suggests Meta copies external model weights; the runtime’s multi‑provider clients let Meta route traffic internally without user interaction.