Tool-Call Parsing

The problem, from first principles

When a model decides to use a tool it emits a tool call: the tool's name and its arguments, usually as JSON. Models are not parsers, so the JSON arrives wrapped in friendly prose ("Sure! I'll look that up:"), inside a markdown code fence, or with the arguments themselves encoded as a JSON string (the convention of some APIs). A robust agent has to find the call in this mess. A reliable method does not trust a regular expression to match braces, it asks a real JSON decoder to try parsing at every opening brace until one parse succeeds and looks like a tool call.

From theory to code

Implement parse_tool_call.

Constraints

• parse_tool_call(text) returns (name, args). Scan the text for { characters from left to right. At each one try json.JSONDecoder().raw_decode(text, index); skip positions where decoding fails.
• The first decoded value that is a dict containing a string name and an arguments key is the call. If arguments is a string, parse it with json.loads into a dict; otherwise it must already be a dict.
• If no such object is found, or the arguments string is not valid JSON for a dict, raise ValueError.

Hints