refactor: implement auto-transition from thinking to text content upon detecting </think> tags and remove unused helper functions

internal/adapter/openai/handler_toolcall_test.go

@@ -58,21 +58,6 @@ func parseSSEDataFrames(t *testing.T, body string) ([]map[string]any, bool) {
 	return frames, done
 }
 
-func streamHasRawToolJSONContent(frames []map[string]any) bool {
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			content, _ := delta["content"].(string)
-			if strings.Contains(content, `"tool_calls"`) {
-				return true
-			}
-		}
-	}
-	return false
-}
-
 func streamHasToolCallsDelta(frames []map[string]any) bool {
 	for _, frame := range frames {
 		choices, _ := frame["choices"].([]any)
@@ -100,26 +85,6 @@ func streamFinishReason(frames []map[string]any) string {
 	return ""
 }
 
-func streamToolCallArgumentChunks(frames []map[string]any) []string {
-	out := make([]string, 0, 4)
-	for _, frame := range frames {
-		choices, _ := frame["choices"].([]any)
-		for _, item := range choices {
-			choice, _ := item.(map[string]any)
-			delta, _ := choice["delta"].(map[string]any)
-			toolCalls, _ := delta["tool_calls"].([]any)
-			for _, tc := range toolCalls {
-				tcm, _ := tc.(map[string]any)
-				fn, _ := tcm["function"].(map[string]any)
-				if args, ok := fn["arguments"].(string); ok && args != "" {
-					out = append(out, args)
-				}
-			}
-		}
-	}
-	return out
-}
-
 // Backward-compatible alias for historical test name used in CI logs.
 func TestHandleNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}

internal/adapter/openai/responses_stream_test.go

@@ -325,30 +325,3 @@ func extractSSEEventPayload(body, targetEvent string) (map[string]any, bool) {
 	}
 	return nil, false
 }
-
-func extractAllSSEEventPayloads(body, targetEvent string) []map[string]any {
-	scanner := bufio.NewScanner(strings.NewReader(body))
-	matched := false
-	out := make([]map[string]any, 0, 2)
-	for scanner.Scan() {
-		line := strings.TrimSpace(scanner.Text())
-		if strings.HasPrefix(line, "event: ") {
-			evt := strings.TrimSpace(strings.TrimPrefix(line, "event: "))
-			matched = evt == targetEvent
-			continue
-		}
-		if !matched || !strings.HasPrefix(line, "data: ") {
-			continue
-		}
-		raw := strings.TrimSpace(strings.TrimPrefix(line, "data: "))
-		if raw == "" || raw == "[DONE]" {
-			continue
-		}
-		var payload map[string]any
-		if err := json.Unmarshal([]byte(raw), &payload); err != nil {
-			continue
-		}
-		out = append(out, payload)
-	}
-	return out
-}

internal/adapter/openai/tool_sieve_jsonscan.go

@@ -2,48 +2,6 @@ package openai
 
 import "strings"
 
-func extractJSONObjectFrom(text string, start int) (string, int, bool) {
-	if start < 0 || start >= len(text) || text[start] != '{' {
-		return "", 0, false
-	}
-	depth := 0
-	quote := byte(0)
-	escaped := false
-	for i := start; i < len(text); i++ {
-		ch := text[i]
-		if quote != 0 {
-			if escaped {
-				escaped = false
-				continue
-			}
-			if ch == '\\' {
-				escaped = true
-				continue
-			}
-			if ch == quote {
-				quote = 0
-			}
-			continue
-		}
-		if ch == '"' || ch == '\'' {
-			quote = ch
-			continue
-		}
-		if ch == '{' {
-			depth++
-			continue
-		}
-		if ch == '}' {
-			depth--
-			if depth == 0 {
-				end := i + 1
-				return text[start:end], end, true
-			}
-		}
-	}
-	return "", 0, false
-}
-
 func trimWrappingJSONFence(prefix, suffix string) (string, string) {
 	trimmedPrefix := strings.TrimRight(prefix, " \t\r\n")
 	fenceIdx := strings.LastIndex(trimmedPrefix, "```")
@@ -67,18 +25,3 @@ func trimWrappingJSONFence(prefix, suffix string) (string, string) {
 	consumedLeading := len(suffix) - len(trimmedSuffix)
 	return trimmedPrefix[:fenceIdx], suffix[consumedLeading+3:]
 }
-
-func openFenceStartBefore(s string, pos int) (int, bool) {
-	if pos <= 0 || pos > len(s) {
-		return -1, false
-	}
-	segment := s[:pos]
-	lastFence := strings.LastIndex(segment, "```")
-	if lastFence < 0 {
-		return -1, false
-	}
-	if strings.Count(segment, "```")%2 == 1 {
-		return lastFence, true
-	}
-	return -1, false
-}

internal/sse/parser.go

@@ -3,6 +3,7 @@ package sse
 import (
 	"bytes"
 	"encoding/json"
+	"regexp"
 	"strings"
 
 	"ds2api/internal/deepseek"
@@ -93,6 +94,11 @@ func ParseSSEChunkForContent(chunk map[string]any, thinkingEnabled bool, current
 	if finished {
 		return nil, true, newType
 	}
+	var transitioned bool
+	parts, transitioned = splitThinkingParts(parts)
+	if transitioned {
+		newType = "text"
+	}
 	return parts, false, newType
 }
 
@@ -166,6 +172,9 @@ func updateTypeFromNestedResponse(path string, v any, newType *string) {
 func resolvePartType(path string, thinkingEnabled bool, newType string) string {
 	switch {
 	case path == "response/thinking_content":
+		if newType == "text" {
+			return "text"
+		}
 		return "thinking"
 	case path == "response/content":
 		return "text"
@@ -244,6 +253,59 @@ func appendContentPart(parts *[]ContentPart, content, kind string) {
 	*parts = append(*parts, ContentPart{Text: content, Type: kind})
 }
 
+var thinkClosePattern = regexp.MustCompile(`(?i)</\s*think\s*>`)
+var thinkOpenPattern = regexp.MustCompile(`(?i)<\s*think\s*>`)
+
+// splitThinkingParts detects </think> inside thinking content and
+// auto-transitions everything after it to text. This handles the
+// DeepSeek API bug where the upstream SSE keeps sending
+// reasoning_content even though the model has finished thinking.
+func splitThinkingParts(parts []ContentPart) ([]ContentPart, bool) {
+	var out []ContentPart
+	thinkingDone := false
+	for _, p := range parts {
+		if thinkingDone && p.Type == "thinking" {
+			// Already transitioned — treat remaining thinking as text.
+			cleaned := stripThinkTags(p.Text)
+			if cleaned != "" {
+				out = append(out, ContentPart{Text: cleaned, Type: "text"})
+			}
+			continue
+		}
+		if p.Type != "thinking" {
+			out = append(out, p)
+			continue
+		}
+		loc := thinkClosePattern.FindStringIndex(p.Text)
+		if loc == nil {
+			out = append(out, p)
+			continue
+		}
+		// Split at </think>: before is still thinking, after becomes text.
+		thinkingDone = true
+		before := p.Text[:loc[0]]
+		after := p.Text[loc[1]:]
+		if before != "" {
+			out = append(out, ContentPart{Text: before, Type: "thinking"})
+		}
+		after = stripThinkTags(after)
+		if after != "" {
+			out = append(out, ContentPart{Text: after, Type: "text"})
+		}
+	}
+	if !thinkingDone {
+		return parts, false
+	}
+	return out, true
+}
+
+// stripThinkTags removes any remaining <think> or </think> tags from text.
+func stripThinkTags(s string) string {
+	s = thinkClosePattern.ReplaceAllString(s, "")
+	s = thinkOpenPattern.ReplaceAllString(s, "")
+	return s
+}
+
 func isStatusPath(path string) bool {
 	return path == "response/status" || path == "status"
 }

internal/sse/parser_test.go

@@ -87,3 +87,65 @@ func TestParseSSEChunkForContentAfterAppendUsesUpdatedType(t *testing.T) {
 		t.Fatalf("unexpected parts: %#v", parts)
 	}
 }
+
+func TestParseSSEChunkForContentAutoTransitionsThinkClose(t *testing.T) {
+	chunk := map[string]any{
+		"p": "response/thinking_content",
+		"v": "deep thoughts</think>actual answer",
+	}
+	parts, _, _ := ParseSSEChunkForContent(chunk, true, "thinking")
+	if len(parts) != 2 {
+		t.Fatalf("expected 2 parts from split, got %d: %#v", len(parts), parts)
+	}
+	if parts[0].Type != "thinking" || parts[0].Text != "deep thoughts" {
+		t.Fatalf("first part should be thinking: %#v", parts[0])
+	}
+	if parts[1].Type != "text" || parts[1].Text != "actual answer" {
+		t.Fatalf("second part should be text: %#v", parts[1])
+	}
+}
+
+func TestParseSSEChunkForContentStripsLeakedThinkTags(t *testing.T) {
+	chunk := map[string]any{
+		"p": "response/thinking_content",
+		"v": "<think>more thoughts</think>  answer",
+	}
+	parts, _, _ := ParseSSEChunkForContent(chunk, true, "thinking")
+	if len(parts) != 2 {
+		t.Fatalf("expected 2 parts, got %d: %#v", len(parts), parts)
+	}
+	if parts[0].Type != "thinking" || parts[0].Text != "<think>more thoughts" {
+		// note: the open tag is before the split, so it remains in the thinking part.
+		// that's fine, the output sanitization handles the final string.
+		t.Fatalf("first part mismatch: %#v", parts[0])
+	}
+	if parts[1].Type != "text" || parts[1].Text != "  answer" {
+		t.Fatalf("second part mismatch: %#v", parts[1])
+	}
+}
+
+func TestParseSSEChunkForContentAutoTransitionsState(t *testing.T) {
+	chunk1 := map[string]any{
+		"p": "response/thinking_content",
+		"v": "end of thought</think>start of text",
+	}
+	parts1, _, nextType1 := ParseSSEChunkForContent(chunk1, true, "thinking")
+	if len(parts1) != 2 || parts1[1].Type != "text" {
+		t.Fatalf("expected split parts, got %#v", parts1)
+	}
+	if nextType1 != "text" {
+		t.Fatalf("expected nextType to transition to text, got %q", nextType1)
+	}
+
+	chunk2 := map[string]any{
+		"p": "response/thinking_content",
+		"v": "more actual text sent to thinking path",
+	}
+	parts2, _, nextType2 := ParseSSEChunkForContent(chunk2, true, nextType1)
+	if len(parts2) != 1 || parts2[0].Type != "text" {
+		t.Fatalf("expected subsequent parts to be text, got %#v", parts2)
+	}
+	if nextType2 != "text" {
+		t.Fatalf("expected nextType2 to remain text, got %q", nextType2)
+	}
+}

internal/toolcall/toolcalls_markup.go

@@ -85,7 +85,7 @@ func parseMarkupSingleToolCall(attrs string, inner string) ParsedToolCall {
 			}
 			if strings.TrimSpace(name) != "" {
 				input := parseToolCallInput(obj["input"])
-				if input == nil || len(input) == 0 {
+				if len(input) == 0 {
 					if args, ok := obj["arguments"]; ok {
 						input = parseToolCallInput(args)
 					}