How to handle streaming tool calls in LLM APIs without breaking JSON parsers

Streaming tool calls in large language model APIs like Anthropic's present a unique challenge because the data being streamed is structured JSON, which remains invalid until fully received. Unlike plain text streaming, partial JSON fragments cannot be parsed mid-stream without triggering errors on every standard parser. Developers commonly attempt three workarounds: catching parse exceptions on every chunk, buffering until the stream ends, or guessing structure via string matching — each with significant drawbacks. Exception-driven parsing on every delta pollutes logs and masks genuine errors, while full buffering effectively cancels the user-experience benefits of streaming. The recommended approach is to separate display-side parsing from execution-side parsing, since the two operations have different tolerances for incomplete or malformed data.