Fix unhandled JSONDecodeError in AnthropicClient tool call parsing

[hobostay]

Summary

• json.loads(args) in _convert_messages was called without a try/except when parsing tool call arguments
• Malformed JSON in arguments would raise an unhandled json.JSONDecodeError, crashing the inference loop with an unhelpful traceback
• Wrap in try/except and fall back to empty dict on parse failure, consistent with LlamafileClient's handling

Impact

• Without this fix, malformed tool call arguments from upstream systems would crash the entire inference loop instead of producing a recoverable error

Test plan

• Verify malformed JSON in tool call arguments doesn't crash the client
• Verify valid JSON arguments still parse correctly
• Run existing test suite

🤖 Generated with Claude Code

[antoinezambelli]

Thanks for digging into this — the asymmetry you spotted is real. One nuance steered the decision: this json.loads is on the request-construction path (history → backend), where the args were serialized by us, so malformed JSON is an invariant violation we'd rather surface than swallow. Silently coercing to {} would also fire the tool with empty args, which is worse than not firing. (The direct analog, llamafile.py:110, is also unguarded; where llamafile does guard — :521, parsing untrusted model output — it falls back to returning the content as text, not {}.) So rather than soften it, we'd like to keep it loud but friendlier: wrap the decode to raise a clear error with context instead of a bare traceback. Queuing that. Appreciate the report!