Struct streamChunk

internal/llm/llm.go:93–109

streamChunk is one OpenAI SSE frame. finish_reason is deliberately not decoded: readSSE dispatches accumulated tool calls at stream end, not on finish_reason=="tool_calls", staying provider agnostic since Ollama's /v1 shim sometimes closes with "stop" even after streaming tool_calls.

// ROUTED on (readSSE dispatches accumulated tool calls at stream end, not on
// finish_reason=="tool_calls", staying provider agnostic since Ollama's /v1 shim
// sometimes closes with "stop" even after streaming tool_calls); it is decoded
// only as one of the end-of-stream completion signals, where any non-empty
// value - right or wrong - means the server finished on purpose.
type streamChunk struct {
	Choices []struct {
		Delta        streamDelta `json:"delta"`
		FinishReason string      `json:"finish_reason"`
	} `json:"choices"`
	Usage *struct {
		CompletionTokens int `json:"completion_tokens"`
		PromptTokens     int `json:"prompt_tokens"`
	} `json:"usage,omitempty"`
	// Error is the mid-stream failure frame OpenAI-compatible backends (and
	// OpenRouter-style proxies) emit when the provider dies after 200 OK:
	// `data: {"error":{...}}`, then the connection closes with no [DONE].
	// Without decoding it, the frame parses to zero choices, the close reads
	// as clean EOF, and a mid-sentence-truncated turn finalizes as a success.
	Error *struct {
		Message string `json:"message"`
	} `json:"error,omitempty"`