StreamingAccumulator

Add a streaming chunk to the accumulator

Add thinking content delta directly

Clear the accumulator state

Get the current accumulated content

Get the current accumulated thinking content

Get the current tool calls, in the order each was first seen in the stream

Check if there is any thinking content

Check if accumulation is complete

Convert accumulated data to a Completion, stamped with the current time

Convert accumulated data to a Completion, stamped with the given creation time. Pure given the accumulator's current state - split out from the no-arg overload so the conversion itself is testable without a real clock.

Update the prompt and completion token counts.

The completion's total becomes their sum plus any thinking tokens; cache counts set by updateUsage are kept. Use updateUsage when the provider reports more than these two.

Update token counts including thinking tokens.

The completion's total becomes their sum; cache counts set by updateUsage are kept.

Merge a provider's usage report into the accumulated usage, every org.llm4s.llmconnect.model.TokenUsage field included, so a streamed org.llm4s.llmconnect.model.Completion reports what a non-streaming one does.

Call it once per usage report the stream carries. Each field follows the latest report that states it: promptTokens and completionTokens when positive, the Option fields (thinkingTokens, cachedTokens, cacheCreationTokens) when defined, and totalTokens when positive. A field a report leaves at zero or None keeps its earlier value, so a provider that splits its usage across events (input and cache counts first, output last) can pass each part as it arrives. When no report states a total, the completion's total is prompt plus completion plus thinking tokens.