VibeBasic 0.11.2

Changed

• AI reasoning effort defaults to Medium, adds an explicit XHigh choice, and has an adjustable output budget (initially 65536 tokens shared by reasoning and final answer). Fast follow-up effort reduction is opt-in. The server's reported context size is preserved.
• AI Log now contains metadata only: request IDs, model, effort, output cap, usage, HTTP status, finish reason and outcome. Prompts, images, reasoning, answers, tool arguments and raw server errors are excluded.

Fixed

• Reasoning-only, empty, interrupted and tool-call responses cannot be shown or applied as successful final answers. Reasoning and text remain separate, including servers that return JSON instead of SSE.
• Length termination explicitly identifies the output limit and never retries automatically, even if a timeout or stream error follows. Budget errors never trigger an uncapped retry; other recovery paths have a shared four-attempt ceiling.
• HTTP errors, timeouts, cancellation, incomplete streams and refused connections remain distinct. Removed misleading Ollama advice from generic connection failures.
• Added native/browser parity tests for text, reasoning, tools, structured JSON, metadata, usage, completion status and privacy.