On 2026‑10‑20 the open‑source project llama.cpp released version b11756. The change makes the server copy the id_slot field from Anthropic /v1/messages requests into its internal OAI‑style body. This lets a proxy keep a session bound to a particular KV cache slot, preventing the model from having to refill its context each time. A new test confirms that requests with id_slot are served by the correct slot.
Why it matters
Developers using llama.cpp with Anthropic‑compatible APIs can keep chat sessions faster by avoiding costly cache reloads.
