v0.10.0

This release removes the web_search tool the plugin used to add to OpenAI requests. That tool no longer helps prompt caching, and it cost about 4,000 input tokens on every request.

Removed

• The web_search prompt-cache workaround. Earlier backend behaviour sometimes dropped the cached prefix on tool-output continuations unless a native OpenAI tool was present, so the plugin added OpenAI's web_search to every tool-bearing request. A controlled test on the current backend found no difference: with and without the tool, every continuation step reused 98.7–99.4% of the previous request from cache, and no step lost its cache. The tool only added about 4,000 input tokens to each request.
• The webSearch setting and CORTEXKIT_OPENAI_AUTH_NO_WEB_SEARCH. A config file that still sets webSearch loads normally and the key is ignored.
• The web_search tool shown in every session. OpenCode offers a plugin's tools to every session, so sessions on other providers saw a web_search tool that could not search there and returned only its own arguments.

Fixed

• Search tools from other plugins now reach OpenAI models. The plugin used to remove the Exa search tool, and any function tool named web_search, from OpenAI requests, because its own search replaced them. Those tools now pass through unchanged.

Sessions that used OpenAI's hosted search in the past keep working: the backend accepts that history without the tool.