org.llm4s.llmconnect.streaming

Server-Sent Events (SSE) parser for streaming LLM responses.

Handles the SSE format used by both OpenAI and Anthropic streaming endpoints. SSE format consists of fields like:

• data:
• event:
• id:
• retry:
• :

Accumulates streaming chunks into a complete response. Handles content accumulation, tool call accumulation, thinking content, and token tracking.

Mutable and not thread-safe: use one accumulator per stream, fed from the thread that reads it. Tool calls are keyed by id, so every chunk of a streamed tool call - continuations included - must carry its call's id; a tool-call chunk with an empty id is ignored. Clients whose wire format identifies continuations only by index must map them back to the id first.

Factory for creating accumulators

Shared boundary parser for streaming tool-call arguments.

• Empty input yields an empty object sentinel for accumulators.
• Valid JSON is parsed as-is.
• Invalid/partial JSON is preserved as a raw string for later assembly.
• JSON nested more than 512 levels deep is preserved as a raw string too: the arguments are model output, and a value that deep overflows the stack of whatever renders it next (#1562).