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/main.go

@@ -385,6 +385,8 @@ func onStreamingResponseBody(ctx wrapper.HttpContext, pluginConfig config.Plugin
 		return chunk
 	}
 
+	promoteThinking := pluginConfig.GetProviderConfig().GetPromoteThinkingOnEmpty()
+
 	log.Debugf("[onStreamingResponseBody] provider=%s", activeProvider.GetProviderType())
 	log.Debugf("[onStreamingResponseBody] isLastChunk=%v chunk: %s", isLastChunk, string(chunk))
 
@@ -392,6 +394,9 @@ func onStreamingResponseBody(ctx wrapper.HttpContext, pluginConfig config.Plugin
 		apiName, _ := ctx.GetContext(provider.CtxKeyApiName).(provider.ApiName)
 		modifiedChunk, err := handler.OnStreamingResponseBody(ctx, apiName, chunk, isLastChunk)
 		if err == nil && modifiedChunk != nil {
+			if promoteThinking {
+				modifiedChunk = promoteThinkingInStreamingChunk(ctx, modifiedChunk, isLastChunk)
+			}
 			// Convert to Claude format if needed
 			claudeChunk, convertErr := convertStreamingResponseToClaude(ctx, modifiedChunk)
 			if convertErr != nil {
@@ -435,6 +440,10 @@ func onStreamingResponseBody(ctx wrapper.HttpContext, pluginConfig config.Plugin
 
 		result := []byte(responseBuilder.String())
 
+		if promoteThinking {
+			result = promoteThinkingInStreamingChunk(ctx, result, isLastChunk)
+		}
+
 		// Convert to Claude format if needed
 		claudeChunk, convertErr := convertStreamingResponseToClaude(ctx, result)
 		if convertErr != nil {
@@ -443,11 +452,12 @@ func onStreamingResponseBody(ctx wrapper.HttpContext, pluginConfig config.Plugin
 		return claudeChunk
 	}
 
-	if !needsClaudeResponseConversion(ctx) {
+	if !needsClaudeResponseConversion(ctx) && !promoteThinking {
 		return chunk
 	}
 
 	// If provider doesn't implement any streaming handlers but we need Claude conversion
+	// or thinking promotion
 	// First extract complete events from the chunk
 	events := provider.ExtractStreamingEvents(ctx, chunk)
 	log.Debugf("[onStreamingResponseBody] %d events received (no handler)", len(events))
@@ -464,6 +474,10 @@ func onStreamingResponseBody(ctx wrapper.HttpContext, pluginConfig config.Plugin
 
 	result := []byte(responseBuilder.String())
 
+	if promoteThinking {
+		result = promoteThinkingInStreamingChunk(ctx, result, isLastChunk)
+	}
+
 	// Convert to Claude format if needed
 	claudeChunk, convertErr := convertStreamingResponseToClaude(ctx, result)
 	if convertErr != nil {
@@ -496,6 +510,16 @@ func onHttpResponseBody(ctx wrapper.HttpContext, pluginConfig config.PluginConfi
 		finalBody = body
 	}
 
+	// Promote thinking/reasoning to content when content is empty
+	if pluginConfig.GetProviderConfig().GetPromoteThinkingOnEmpty() {
+		promoted, err := provider.PromoteThinkingOnEmptyResponse(finalBody)
+		if err != nil {
+			log.Warnf("[promoteThinkingOnEmpty] failed: %v", err)
+		} else {
+			finalBody = promoted
+		}
+	}
+
 	// Convert to Claude format if needed (applies to both branches)
 	convertedBody, err := convertResponseBodyToClaude(ctx, finalBody)
 	if err != nil {
@@ -544,6 +568,49 @@ func convertStreamingResponseToClaude(ctx wrapper.HttpContext, data []byte) ([]b
 	return claudeChunk, nil
 }
 
+// promoteThinkingInStreamingChunk processes SSE-formatted streaming data, buffering
+// reasoning deltas and stripping them from chunks. On the last chunk, if no content
+// was ever seen, it appends a flush chunk that emits buffered reasoning as content.
+func promoteThinkingInStreamingChunk(ctx wrapper.HttpContext, data []byte, isLastChunk bool) []byte {
+	// SSE data contains lines like "data: {...}\n\n"
+	// We need to find and process each data line
+	lines := strings.Split(string(data), "\n")
+	modified := false
+	for i, line := range lines {
+		if !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+		payload := strings.TrimPrefix(line, "data: ")
+		if payload == "[DONE]" || payload == "" {
+			continue
+		}
+		stripped, err := provider.PromoteStreamingThinkingOnEmptyChunk(ctx, []byte(payload))
+		if err != nil {
+			continue
+		}
+		newLine := "data: " + string(stripped)
+		if newLine != line {
+			lines[i] = newLine
+			modified = true
+		}
+	}
+
+	result := data
+	if modified {
+		result = []byte(strings.Join(lines, "\n"))
+	}
+
+	// On last chunk, flush buffered reasoning as content if no content was seen
+	if isLastChunk {
+		flushChunk := provider.PromoteStreamingThinkingFlush(ctx)
+		if flushChunk != nil {
+			result = append(flushChunk, result...)
+		}
+	}
+
+	return result
+}
+
 // Helper function to convert OpenAI response body to Claude format
 func convertResponseBodyToClaude(ctx wrapper.HttpContext, body []byte) ([]byte, error) {
 	if !needsClaudeResponseConversion(ctx) {