# 30 October: sturdier chat, cheaper Claude calls, own keys in ZeroLog Code 2026-10-30 · Product > Six chat faults were fixed, Claude requests through ZeroLog Code can now read from the provider cache, and an own API key in ZeroLog Code no longer draws on the pool budget. Three releases on 30 October touched the chat, the provider cache and the own-key path of ZeroLog Code. *The three areas of 30 October.* Text version Chat: stop timer, empty placeholder, stream liveness, attachment staging, file hand-over. Cache: volatile facts behind the stable prefix, cache marks through the shim, cache hits booked. Own key: the ZeroLog Code OpenAI path uses the own key. ## Chat The stop timer of an old turn no longer reports on the new turn. An empty placeholder message is closed when a run ends normally. Attachments that arrive while no container is attached wait in a staging area and move to /work/uploads when a container is attached. The staging area is now encrypted at rest. Files the agent created are handed over with present_file at the end of the turn on their own. ## Provider cache Session facts, memory and retrieval used to sit before your message, so each turn changed the request from that point on and the provider re-wrote the whole tool loop instead of reading it. They now ride at the end of the request. Cache marks now also travel through the shim to Claude models on the pool, so Claude Code and ZeroLog Code can read from the provider cache. Cache hits on that path are booked instead of a fixed 0. We have not yet measured the saving in daily use. ## Own key in ZeroLog Code The OpenAI-compatible path of ZeroLog Code did not know own keys. Requests with an own Anthropic key went to the pool, could fail over to the platform key and ran into the pool budget (error 402) although the key was valid. They now use your key, are booked as own-key usage with cost 0 and never fail over. An invalid key returns a clear 403 (byok_key_invalid) instead of silently using the pool. No prices or plans changed.