feat(ai-proxy): add promoteThinkingOnEmpty and hiclawMode config options (#3625)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

plugins/wasm-go/extensions/ai-proxy/provider/model.go

@@ -255,6 +255,70 @@ func (m *chatMessage) handleStreamingReasoningContent(ctx wrapper.HttpContext, r
 	}
 }
 
+// promoteThinkingOnEmpty promotes reasoning_content to content when content is empty.
+// This handles models that put user-facing replies into thinking blocks instead of text blocks.
+func (r *chatCompletionResponse) promoteThinkingOnEmpty() {
+	for i := range r.Choices {
+		msg := r.Choices[i].Message
+		if msg == nil {
+			continue
+		}
+		if !isContentEmpty(msg.Content) {
+			continue
+		}
+		if msg.ReasoningContent != "" {
+			msg.Content = msg.ReasoningContent
+			msg.ReasoningContent = ""
+		}
+	}
+}
+
+// promoteStreamingThinkingOnEmpty accumulates reasoning content during streaming.
+// It strips reasoning from chunks and buffers it. When content is seen, it marks
+// the stream as having content so no promotion will happen.
+// Call PromoteStreamingThinkingFlush at the end of the stream to emit buffered
+// reasoning as content if no content was ever seen.
+// Returns true if the chunk was modified (reasoning stripped).
+func promoteStreamingThinkingOnEmpty(ctx wrapper.HttpContext, msg *chatMessage) bool {
+	if msg == nil {
+		return false
+	}
+	hasContentDelta, _ := ctx.GetContext(ctxKeyHasContentDelta).(bool)
+	if hasContentDelta {
+		return false
+	}
+
+	if !isContentEmpty(msg.Content) {
+		ctx.SetContext(ctxKeyHasContentDelta, true)
+		return false
+	}
+
+	// Buffer reasoning content and strip it from the chunk
+	reasoning := msg.ReasoningContent
+	if reasoning == "" {
+		reasoning = msg.Reasoning
+	}
+	if reasoning != "" {
+		buffered, _ := ctx.GetContext(ctxKeyBufferedReasoning).(string)
+		ctx.SetContext(ctxKeyBufferedReasoning, buffered+reasoning)
+		msg.ReasoningContent = ""
+		msg.Reasoning = ""
+		return true
+	}
+	return false
+}
+
+func isContentEmpty(content any) bool {
+	switch v := content.(type) {
+	case nil:
+		return true
+	case string:
+		return strings.TrimSpace(v) == ""
+	default:
+		return false
+	}
+}
+
 type chatMessageContent struct {
 	CacheControl map[string]interface{}      `json:"cache_control,omitempty"`
 	Type         string                      `json:"type,omitempty"`
@@ -648,3 +712,87 @@ func (r embeddingsRequest) ParseInput() []string {
 	}
 	return input
 }
+
+// PromoteThinkingOnEmptyResponse promotes reasoning_content to content in a non-streaming
+// response body when content is empty. Returns the original body if no promotion is needed.
+func PromoteThinkingOnEmptyResponse(body []byte) ([]byte, error) {
+	var resp chatCompletionResponse
+	if err := json.Unmarshal(body, &resp); err != nil {
+		return body, fmt.Errorf("unable to unmarshal response for thinking promotion: %v", err)
+	}
+	promoted := false
+	for i := range resp.Choices {
+		msg := resp.Choices[i].Message
+		if msg == nil {
+			continue
+		}
+		if !isContentEmpty(msg.Content) {
+			continue
+		}
+		if msg.ReasoningContent != "" {
+			msg.Content = msg.ReasoningContent
+			msg.ReasoningContent = ""
+			promoted = true
+		}
+	}
+	if !promoted {
+		return body, nil
+	}
+	return json.Marshal(resp)
+}
+
+// PromoteStreamingThinkingOnEmptyChunk buffers reasoning deltas and strips them from
+// the chunk during streaming. Call PromoteStreamingThinkingFlush on the last chunk
+// to emit buffered reasoning as content if no real content was ever seen.
+func PromoteStreamingThinkingOnEmptyChunk(ctx wrapper.HttpContext, data []byte) ([]byte, error) {
+	var resp chatCompletionResponse
+	if err := json.Unmarshal(data, &resp); err != nil {
+		return data, nil // not a valid chat completion chunk, skip
+	}
+	modified := false
+	for i := range resp.Choices {
+		msg := resp.Choices[i].Delta
+		if msg == nil {
+			continue
+		}
+		if promoteStreamingThinkingOnEmpty(ctx, msg) {
+			modified = true
+		}
+	}
+	if !modified {
+		return data, nil
+	}
+	return json.Marshal(resp)
+}
+
+// PromoteStreamingThinkingFlush checks if the stream had no content and returns
+// an SSE chunk that emits the buffered reasoning as content. Returns nil if
+// content was already seen or no reasoning was buffered.
+func PromoteStreamingThinkingFlush(ctx wrapper.HttpContext) []byte {
+	hasContentDelta, _ := ctx.GetContext(ctxKeyHasContentDelta).(bool)
+	if hasContentDelta {
+		return nil
+	}
+	buffered, _ := ctx.GetContext(ctxKeyBufferedReasoning).(string)
+	if buffered == "" {
+		return nil
+	}
+	// Build a minimal chat.completion.chunk with the buffered reasoning as content
+	resp := chatCompletionResponse{
+		Object: objectChatCompletionChunk,
+		Choices: []chatCompletionChoice{
+			{
+				Index: 0,
+				Delta: &chatMessage{
+					Content: buffered,
+				},
+			},
+		},
+	}
+	data, err := json.Marshal(resp)
+	if err != nil {
+		return nil
+	}
+	// Format as SSE
+	return []byte("data: " + string(data) + "\n\n")
+}

plugins/wasm-go/extensions/ai-proxy/provider/provider.go

@@ -178,6 +178,8 @@ const (
 	ctxKeyPushedMessage          = "pushedMessage"
 	ctxKeyContentPushed          = "contentPushed"
 	ctxKeyReasoningContentPushed = "reasoningContentPushed"
+	ctxKeyHasContentDelta        = "hasContentDelta"
+	ctxKeyBufferedReasoning      = "bufferedReasoning"
 
 	objectChatCompletion      = "chat.completion"
 	objectChatCompletionChunk = "chat.completion.chunk"
@@ -474,6 +476,12 @@ type ProviderConfig struct {
 	// @Title zh-CN 合并连续同角色消息
 	// @Description zh-CN 开启后，若请求的 messages 中存在连续的同角色消息（如连续两条 user 消息），将其内容合并为一条，以满足要求严格轮流交替（user→assistant→user→...）的模型服务商的要求。
 	mergeConsecutiveMessages bool `required:"false" yaml:"mergeConsecutiveMessages" json:"mergeConsecutiveMessages"`
+	// @Title zh-CN 空内容时提升思考为正文
+	// @Description zh-CN 开启后，若模型响应只包含 reasoning_content/thinking 而没有正文内容，将 reasoning 内容提升为正文内容返回，避免客户端收到空回复。
+	promoteThinkingOnEmpty bool `required:"false" yaml:"promoteThinkingOnEmpty" json:"promoteThinkingOnEmpty"`
+	// @Title zh-CN HiClaw 模式
+	// @Description zh-CN 开启后同时启用 mergeConsecutiveMessages 和 promoteThinkingOnEmpty，适用于 HiClaw 多 Agent 协作场景。
+	hiclawMode bool `required:"false" yaml:"hiclawMode" json:"hiclawMode"`
 }
 
 func (c *ProviderConfig) GetId() string {
@@ -699,6 +707,12 @@ func (c *ProviderConfig) FromJson(json gjson.Result) {
 		}
 	}
 	c.mergeConsecutiveMessages = json.Get("mergeConsecutiveMessages").Bool()
+	c.promoteThinkingOnEmpty = json.Get("promoteThinkingOnEmpty").Bool()
+	c.hiclawMode = json.Get("hiclawMode").Bool()
+	if c.hiclawMode {
+		c.mergeConsecutiveMessages = true
+		c.promoteThinkingOnEmpty = true
+	}
 }
 
 func (c *ProviderConfig) Validate() error {
@@ -833,6 +847,10 @@ func (c *ProviderConfig) IsOriginal() bool {
 	return c.protocol == protocolOriginal
 }
 
+func (c *ProviderConfig) GetPromoteThinkingOnEmpty() bool {
+	return c.promoteThinkingOnEmpty
+}
+
 func (c *ProviderConfig) ReplaceByCustomSettings(body []byte) ([]byte, error) {
 	return ReplaceByCustomSettings(body, c.customSettings)
 }