fix(openai): process reasoning deltas before content in responses stream

- Process `reasoning_content` deltas before message `content` in streaming chat completion translation.
- Ensure reasoning output items receive incremental text before message creation closes reasoning.

Closes: #5659
This commit is contained in:
Luis Pater
2026-09-10 18:59:11 +08:00
parent 09a29bd345
commit c8ecb4f3c9
2 changed files with 181 additions and 31 deletions

View File

@@ -652,6 +652,37 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
idx := int(choice.Get("index").Int())
delta := choice.Get("delta")
if delta.Exists() {
// reasoning_content (OpenAI reasoning incremental text)
rc := delta.Get("reasoning_content")
if !rc.Exists() || rc.String() == "" {
rc = delta.Get("reasoning")
}
if rc.Exists() && rc.String() != "" {
// On first appearance, add reasoning item and part
if st.ReasoningID == "" {
st.ReasoningID = fmt.Sprintf("rs_%s_%d", st.ResponseID, idx)
st.ReasoningIndex = allocOutputIndex()
item := []byte(`{"type":"response.output_item.added","sequence_number":0,"output_index":0,"item":{"id":"","type":"reasoning","status":"in_progress","summary":[]}}`)
item, _ = sjson.SetBytes(item, "sequence_number", nextSeq())
item, _ = sjson.SetBytes(item, "output_index", st.ReasoningIndex)
item, _ = sjson.SetBytes(item, "item.id", st.ReasoningID)
out = append(out, emitRespEvent("response.output_item.added", item))
part := []byte(`{"type":"response.reasoning_summary_part.added","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"part":{"type":"summary_text","text":""}}`)
part, _ = sjson.SetBytes(part, "sequence_number", nextSeq())
part, _ = sjson.SetBytes(part, "item_id", st.ReasoningID)
part, _ = sjson.SetBytes(part, "output_index", st.ReasoningIndex)
out = append(out, emitRespEvent("response.reasoning_summary_part.added", part))
}
// Append incremental text to reasoning buffer
st.ReasoningBuf.WriteString(rc.String())
msg := []byte(`{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"delta":""}`)
msg, _ = sjson.SetBytes(msg, "sequence_number", nextSeq())
msg, _ = sjson.SetBytes(msg, "item_id", st.ReasoningID)
msg, _ = sjson.SetBytes(msg, "output_index", st.ReasoningIndex)
msg, _ = sjson.SetBytes(msg, "delta", rc.String())
out = append(out, emitRespEvent("response.reasoning_summary_text.delta", msg))
}
if c := delta.Get("content"); c.Exists() && c.String() != "" {
// Ensure the message item and its first content part are announced before any text deltas
if st.ReasoningID != "" {
@@ -694,37 +725,6 @@ func ConvertOpenAIChatCompletionsResponseToOpenAIResponses(ctx context.Context,
st.MsgTextBuf[idx].WriteString(c.String())
}
// reasoning_content (OpenAI reasoning incremental text)
rc := delta.Get("reasoning_content")
if !rc.Exists() || rc.String() == "" {
rc = delta.Get("reasoning")
}
if rc.Exists() && rc.String() != "" {
// On first appearance, add reasoning item and part
if st.ReasoningID == "" {
st.ReasoningID = fmt.Sprintf("rs_%s_%d", st.ResponseID, idx)
st.ReasoningIndex = allocOutputIndex()
item := []byte(`{"type":"response.output_item.added","sequence_number":0,"output_index":0,"item":{"id":"","type":"reasoning","status":"in_progress","summary":[]}}`)
item, _ = sjson.SetBytes(item, "sequence_number", nextSeq())
item, _ = sjson.SetBytes(item, "output_index", st.ReasoningIndex)
item, _ = sjson.SetBytes(item, "item.id", st.ReasoningID)
out = append(out, emitRespEvent("response.output_item.added", item))
part := []byte(`{"type":"response.reasoning_summary_part.added","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"part":{"type":"summary_text","text":""}}`)
part, _ = sjson.SetBytes(part, "sequence_number", nextSeq())
part, _ = sjson.SetBytes(part, "item_id", st.ReasoningID)
part, _ = sjson.SetBytes(part, "output_index", st.ReasoningIndex)
out = append(out, emitRespEvent("response.reasoning_summary_part.added", part))
}
// Append incremental text to reasoning buffer
st.ReasoningBuf.WriteString(rc.String())
msg := []byte(`{"type":"response.reasoning_summary_text.delta","sequence_number":0,"item_id":"","output_index":0,"summary_index":0,"delta":""}`)
msg, _ = sjson.SetBytes(msg, "sequence_number", nextSeq())
msg, _ = sjson.SetBytes(msg, "item_id", st.ReasoningID)
msg, _ = sjson.SetBytes(msg, "output_index", st.ReasoningIndex)
msg, _ = sjson.SetBytes(msg, "delta", rc.String())
out = append(out, emitRespEvent("response.reasoning_summary_text.delta", msg))
}
// tool calls
if tcs := delta.Get("tool_calls"); tcs.Exists() && tcs.IsArray() && len(tcs.Array()) > 0 {
if st.ReasoningID != "" {

View File

@@ -2,6 +2,7 @@ package responses
import (
"context"
"fmt"
"strings"
"testing"
@@ -1425,3 +1426,152 @@ func TestConvertOpenAIChatCompletionsResponseToOpenAIResponses_EmptyToolCallsArr
t.Fatalf("unexpected completed response message text: got %q, want %q", got, "Hello world!")
}
}
func TestConvertOpenAIChatCompletionsResponseToOpenAIResponses_ChunkWithContentAndReasoningContent(t *testing.T) {
request := []byte(`{"model":"deepseek-v4-flash"}`)
chunks := []string{
`data: {"id":"chatcmpl_ds","object":"chat.completion.chunk","created":1773896263,"model":"deepseek-v4-flash","choices":[{"index":0,"delta":{"role":"assistant","reasoning_content":"Thinking part 1,"},"finish_reason":null}]}`,
`data: {"id":"chatcmpl_ds","object":"chat.completion.chunk","created":1773896263,"model":"deepseek-v4-flash","choices":[{"index":0,"delta":{"content":"Bien","reasoning_content":" Just professional."},"finish_reason":null}]}`,
`data: {"id":"chatcmpl_ds","object":"chat.completion.chunk","created":1773896263,"model":"deepseek-v4-flash","choices":[{"index":0,"delta":{"content":" continues here."},"finish_reason":"stop"}]}`,
`data: [DONE]`,
}
var param any
var reasoningAddedCount, reasoningDoneCount int
var messageAddedCount, messageDoneCount int
var completedCount int
var lastReasoningSummaryText string
var lastMessageContentText string
var eventOrder []string
var completedData gjson.Result
var addedItemIDs []string
for _, line := range chunks {
for _, chunk := range ConvertOpenAIChatCompletionsResponseToOpenAIResponses(context.Background(), "deepseek-v4-flash", request, request, []byte(line), &param) {
event, data := parseOpenAIResponsesSSEEvent(t, chunk)
switch event {
case "response.output_item.added":
itemType := data.Get("item.type").String()
itemID := data.Get("item.id").String()
addedItemIDs = append(addedItemIDs, itemID)
eventOrder = append(eventOrder, fmt.Sprintf("added:%s:%d", itemType, data.Get("output_index").Int()))
if itemType == "reasoning" {
reasoningAddedCount++
} else if itemType == "message" {
messageAddedCount++
}
case "response.output_item.done":
itemType := data.Get("item.type").String()
eventOrder = append(eventOrder, fmt.Sprintf("done:%s:%d", itemType, data.Get("output_index").Int()))
if itemType == "reasoning" {
reasoningDoneCount++
lastReasoningSummaryText = data.Get("item.summary.0.text").String()
} else if itemType == "message" {
messageDoneCount++
lastMessageContentText = data.Get("item.content.0.text").String()
}
case "response.completed":
completedCount++
completedData = data
}
}
}
if reasoningAddedCount != 1 {
t.Fatalf("expected exactly 1 reasoning output_item.added, got %d (events: %v)", reasoningAddedCount, eventOrder)
}
if reasoningDoneCount != 1 {
t.Fatalf("expected exactly 1 reasoning output_item.done, got %d (events: %v)", reasoningDoneCount, eventOrder)
}
if lastReasoningSummaryText != "Thinking part 1, Just professional." {
t.Fatalf("unexpected reasoning summary text: got %q, want %q", lastReasoningSummaryText, "Thinking part 1, Just professional.")
}
if messageAddedCount != 1 {
t.Fatalf("expected exactly 1 message output_item.added, got %d (events: %v)", messageAddedCount, eventOrder)
}
if messageDoneCount != 1 {
t.Fatalf("expected exactly 1 message output_item.done, got %d", messageDoneCount)
}
if lastMessageContentText != "Bien continues here." {
t.Fatalf("unexpected message content text: got %q, want %q", lastMessageContentText, "Bien continues here.")
}
if completedCount != 1 {
t.Fatalf("expected exactly 1 response.completed, got %d", completedCount)
}
// Verify strict event ordering: reasoning completes before message starts
wantOrder := []string{
"added:reasoning:0",
"done:reasoning:0",
"added:message:1",
"done:message:1",
}
if len(eventOrder) != len(wantOrder) {
t.Fatalf("unexpected event order length: got %v, want %v", eventOrder, wantOrder)
}
for i, want := range wantOrder {
if eventOrder[i] != want {
t.Fatalf("eventOrder[%d] = %q, want %q; full sequence: %v", i, eventOrder[i], want, eventOrder)
}
}
// Verify item IDs are unique
if len(addedItemIDs) != 2 || addedItemIDs[0] == addedItemIDs[1] {
t.Fatalf("expected 2 distinct item IDs, got %v", addedItemIDs)
}
// Verify completed response contains both outputs in ascending order
if got := completedData.Get("response.output.0.summary.0.text").String(); got != "Thinking part 1, Just professional." {
t.Fatalf("unexpected completed reasoning: got %q", got)
}
if got := completedData.Get("response.output.1.content.0.text").String(); got != "Bien continues here." {
t.Fatalf("unexpected completed message: got %q", got)
}
}
func TestConvertOpenAIChatCompletionsResponseToOpenAIResponses_SingleChunkWithBothContentAndReasoningContent(t *testing.T) {
request := []byte(`{"model":"deepseek-v4-flash"}`)
chunks := []string{
`data: {"id":"chatcmpl_single","object":"chat.completion.chunk","created":1773896263,"model":"deepseek-v4-flash","choices":[{"index":0,"delta":{"content":"Answer","reasoning_content":"Thought"},"finish_reason":"stop"}]}`,
`data: [DONE]`,
}
var param any
var eventOrder []string
var completedData gjson.Result
for _, line := range chunks {
for _, chunk := range ConvertOpenAIChatCompletionsResponseToOpenAIResponses(context.Background(), "deepseek-v4-flash", request, request, []byte(line), &param) {
event, data := parseOpenAIResponsesSSEEvent(t, chunk)
switch event {
case "response.output_item.added":
eventOrder = append(eventOrder, fmt.Sprintf("added:%s:%d", data.Get("item.type").String(), data.Get("output_index").Int()))
case "response.output_item.done":
eventOrder = append(eventOrder, fmt.Sprintf("done:%s:%d", data.Get("item.type").String(), data.Get("output_index").Int()))
case "response.completed":
completedData = data
}
}
}
wantOrder := []string{
"added:reasoning:0",
"done:reasoning:0",
"added:message:1",
"done:message:1",
}
if len(eventOrder) != len(wantOrder) {
t.Fatalf("unexpected event order length: got %v, want %v", eventOrder, wantOrder)
}
for i, want := range wantOrder {
if eventOrder[i] != want {
t.Fatalf("eventOrder[%d] = %q, want %q; full sequence: %v", i, eventOrder[i], want, eventOrder)
}
}
if got := completedData.Get("response.output.0.summary.0.text").String(); got != "Thought" {
t.Fatalf("unexpected completed reasoning: got %q", got)
}
if got := completedData.Get("response.output.1.content.0.text").String(); got != "Answer" {
t.Fatalf("unexpected completed message: got %q", got)
}
}