ControlBird AI: Prompt caching now applies on every model provider, not only Anthropic. Models reached through OpenRouter - including Claude and Qwen, which only cache when asked to explicitly - previously got no caching at all, so repeated context was paid for in full on every turn.
ControlBird AI: Each assistant reply now shows how much of its input was served from cache next to the existing token counts, so the hit rate is visible in the same place as the cost.
ControlBird AI: The assistant can now file a problem report when you ask it to, through the same feedback channel the rest of the product uses. Asked to report a problem before, it would name a channel that does not exist.
Bug Fixes
ControlBird AI: Asking the assistant to set up an integration could be answered with a flat statement that no such integration exists, when it does. The assistant could only ever see 30 of the 81 integrations in the marketplace, and because the list is alphabetical, everything sorting after roughly the letter G - Modbus, MQTT, OPC UA, Zigbee2MQTT and the rest - was permanently invisible to it. It now searches and reads the full catalog.
ControlBird AI: The assistant now recognises an integration that is already installed on the node and offers to configure it, instead of proposing to install it again.
ControlBird AI: Reported input-token counts are now consistent between providers. Anthropic counts only the portion of a request it did not serve from cache, so the more caching worked, the lower the reported input looked; the figure shown is now the full input processed regardless of provider.
Performance
ControlBird AI: Live system readings - entity and alarm counts, service status - are no longer mixed into the cached part of the assistant's instructions. Any one of them changing used to invalidate the entire cached prefix, so on a busy node the cache almost never paid off.
ControlBird AI: Reading the integration marketplace now uses roughly a tenth of the conversation's available context while returning all 81 integrations rather than 30, leaving far more room for the actual conversation.