Cline patched a bug that silently stripped every tool

Cline fixed a bug that misread an inferred capability list as a denial list. The mistake stripped every tool from requests to custom OpenAI-compatible models, per the v4.1.12 release notes. Anyone running a custom backend in Cline lost every tool call with no crash and no warning explaining why.

That capability list came from convenience flags like supportsReasoning, meant as a hint about what a model can do. Cline's code read the hint as authoritative instead, treating an inference default as a hard rule. The fix took two releases to land.

The same release window added enterprise MCP allowlists. Remote config can now disable the MCP marketplace outright, or restrict servers to an allowedMCPServers list that hides marketplace entries falling outside it. It's a straightforward permissions feature for anyone running Cline in a managed environment.

Hook timing got fixed too. PreToolUse contextModification now actually reaches the model instead of getting dropped. PostToolUse hooks are awaited instead of firing and forgetting.