fix: ensure input_schema is always present for Vertex AI proxy compatibility

[hanzheng1954]

Problem

When using Shelley with a custom Anthropic-compatible endpoint backed by Vertex AI (e.g. Langdock), requests fail with:

LLM request failed: status 400 Bad Request: {"message":[{"expected":"object","code":"invalid_type","path":["tools",9,"input_schema"],"message":"Invalid input: expected object, received undefined"}]}

Root Cause

Server-side tools (e.g. web_search_20250305) are created without an InputSchema in claudetool/toolset.go. The wire struct in ant.go uses json:"input_schema,omitempty", so a nil schema causes the input_schema key to be completely absent from the JSON.

Anthropic's own API is lenient and accepts this. However, Vertex AI-backed proxies strictly require input_schema to be present as an object on every tool.

Fix

In fromLLMTool(), default to an empty object schema {"type":"object","properties":{}} when InputSchema is nil. This is the defensive serialization-layer fix that covers any current or future tool that omits InputSchema.

Testing

Verified the fix resolves 400 errors against Langdock's Vertex AI proxy (https://api.langdock.com/anthropic/eu/v1/messages) while remaining fully compatible with the official Anthropic API.

[cla-bot]

We require contributors to sign our Contributor License Agreement, and we don't have you on file. In order for us to review and merge your code, please contact @crawshaw at david@bold.dev to get yourself added.