[antonreshetov]

Found while testing OpenRouter support (#971). Through OpenRouter, OpenAI models fill every tool parameter, so list_workspace_items received filters nobody asked for (taskType: "task", taskStatus: "todo", isFavorites: false, folderId: null) and returned empty or partial results. A per-folder request also produced 10 parallel tool calls in one response, which exceeded the 8-call limit and failed with invalidResponse.

• isFavorites, taskType and taskStatus filters are nullable: null or omitted means no filter.
• taskType/taskStatus outside Notes return TASK_FILTER_NOTES_ONLY with a hint instead of an empty list.
• folderId accepts a folder ID, "inbox" for items without a folder, or null/omitted for all folders. Other tools keep null as root/Inbox.
• Tool call limit per response raised from 8 to 16 (AI_LIMITS.toolCalls) in the SDK, Responses and chat completions paths and in the message schema.
• invalidResponse in the SDK path now carries a short diagnostic code (finish reason, tool call count, empty response, etc.) in the trace and under the chat error. The trace logs the final error code.
• Tests for nullable filters, folder filtering, the task filter scope, the 16-call limit and diagnostics.