Merge pull request #194 from PancakeZik/fix/assistant-content-parroting

fix: replace assistant placeholder text to prevent model parroting

config.example.yaml

@@ -236,7 +236,7 @@ nonstream-keepalive-interval: 0
 
 # Global OAuth model name aliases (per channel)
 # These aliases rename model IDs for both model listing and request routing.
-# Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot.
+# Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot, kimi.
 # NOTE: Aliases do not apply to gemini-api-key, codex-api-key, claude-api-key, openai-compatibility, vertex-api-key, or ampcode.
 # You can repeat the same name with different aliases to expose multiple client model names.
 #oauth-model-alias:
@@ -280,6 +280,9 @@ nonstream-keepalive-interval: 0
 #   iflow:
 #     - name: "glm-4.7"
 #       alias: "glm-god"
+#   kimi:
+#     - name: "kimi-k2.5"
+#       alias: "k2.5"
 #   kiro:
 #     - name: "kiro-claude-opus-4-5"
 #       alias: "op45"
@@ -309,6 +312,8 @@ nonstream-keepalive-interval: 0
 #     - "vision-model"
 #   iflow:
 #     - "tstars2.0"
+#   kimi:
+#     - "kimi-k2-thinking"
 #   kiro:
 #     - "kiro-claude-haiku-4-5"
 #   github-copilot:

internal/auth/kiro/aws_auth.go

@@ -238,7 +238,7 @@ func (k *KiroAuth) ListAvailableModels(ctx context.Context, tokenData *KiroToken
 			Description    string  `json:"description"`
 			RateMultiplier float64 `json:"rateMultiplier"`
 			RateUnit       string  `json:"rateUnit"`
-			TokenLimits    struct {
+			TokenLimits    *struct {
 				MaxInputTokens int `json:"maxInputTokens"`
 			} `json:"tokenLimits"`
 		} `json:"models"`
@@ -250,13 +250,17 @@ func (k *KiroAuth) ListAvailableModels(ctx context.Context, tokenData *KiroToken
 
 	models := make([]*KiroModel, 0, len(result.Models))
 	for _, m := range result.Models {
+		maxInputTokens := 0
+		if m.TokenLimits != nil {
+			maxInputTokens = m.TokenLimits.MaxInputTokens
+		}
 		models = append(models, &KiroModel{
 			ModelID:        m.ModelID,
 			ModelName:      m.ModelName,
 			Description:    m.Description,
 			RateMultiplier: m.RateMultiplier,
 			RateUnit:       m.RateUnit,
-			MaxInputTokens: m.TokenLimits.MaxInputTokens,
+			MaxInputTokens: maxInputTokens,
 		})
 	}
 

internal/runtime/executor/kiro_executor.go

@@ -2102,6 +2102,22 @@ func (e *KiroExecutor) parseEventStream(body io.Reader) (string, []kiroclaude.Ki
 				}
 			}
 
+		case "contextUsageEvent":
+			// Handle context usage events from Kiro API
+			// Format: {"contextUsageEvent": {"contextUsagePercentage": 0.53}}
+			if ctxUsage, ok := event["contextUsageEvent"].(map[string]interface{}); ok {
+				if ctxPct, ok := ctxUsage["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: parseEventStream received contextUsageEvent: %.2f%%", ctxPct*100)
+				}
+			} else {
+				// Try direct field (fallback)
+				if ctxPct, ok := event["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: parseEventStream received contextUsagePercentage (direct): %.2f%%", ctxPct*100)
+				}
+			}
+
 		case "error", "exception", "internalServerException", "invalidStateEvent":
 			// Handle error events from Kiro API stream
 			errMsg := ""
@@ -2705,6 +2721,22 @@ func (e *KiroExecutor) streamToChannel(ctx context.Context, body io.Reader, out
 				}
 			}
 
+		case "contextUsageEvent":
+			// Handle context usage events from Kiro API
+			// Format: {"contextUsageEvent": {"contextUsagePercentage": 0.53}}
+			if ctxUsage, ok := event["contextUsageEvent"].(map[string]interface{}); ok {
+				if ctxPct, ok := ctxUsage["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: streamToChannel received contextUsageEvent: %.2f%%", ctxPct*100)
+				}
+			} else {
+				// Try direct field (fallback)
+				if ctxPct, ok := event["contextUsagePercentage"].(float64); ok {
+					upstreamContextPercentage = ctxPct
+					log.Debugf("kiro: streamToChannel received contextUsagePercentage (direct): %.2f%%", ctxPct*100)
+				}
+			}
+
 		case "error", "exception", "internalServerException":
 			// Handle error events from Kiro API stream
 			errMsg := ""

internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_response.go

@@ -14,6 +14,7 @@ import (
 	"time"
 
 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/openai/chat-completions"
+	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -85,6 +86,7 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 
 	// Extract and set usage metadata (token counts).
 	if usageResult := gjson.GetBytes(rawJSON, "response.usageMetadata"); usageResult.Exists() {
+		cachedTokenCount := usageResult.Get("cachedContentTokenCount").Int()
 		if candidatesTokenCountResult := usageResult.Get("candidatesTokenCount"); candidatesTokenCountResult.Exists() {
 			template, _ = sjson.Set(template, "usage.completion_tokens", candidatesTokenCountResult.Int())
 		}
@@ -97,6 +99,14 @@ func ConvertCliResponseToOpenAI(_ context.Context, _ string, originalRequestRawJ
 		if thoughtsTokenCount > 0 {
 			template, _ = sjson.Set(template, "usage.completion_tokens_details.reasoning_tokens", thoughtsTokenCount)
 		}
+		// Include cached token count if present (indicates prompt caching is working)
+		if cachedTokenCount > 0 {
+			var err error
+			template, err = sjson.Set(template, "usage.prompt_tokens_details.cached_tokens", cachedTokenCount)
+			if err != nil {
+				log.Warnf("antigravity openai response: failed to set cached_tokens: %v", err)
+			}
+		}
 	}
 
 	// Process the main content part of the response.

internal/translator/kiro/claude/kiro_claude_request.go

@@ -608,18 +608,22 @@ func processMessages(messages gjson.Result, modelID, origin string) ([]KiroHisto
 
 		if role == "user" {
 			userMsg, toolResults := BuildUserMessageStruct(msg, modelID, origin)
+			// CRITICAL: Kiro API requires content to be non-empty for ALL user messages
+			// This includes both history messages and the current message.
+			// When user message contains only tool_result (no text), content will be empty.
+			// This commonly happens in compaction requests from OpenCode.
+			if strings.TrimSpace(userMsg.Content) == "" {
+				if len(toolResults) > 0 {
+					userMsg.Content = kirocommon.DefaultUserContentWithToolResults
+				} else {
+					userMsg.Content = kirocommon.DefaultUserContent
+				}
+				log.Debugf("kiro: user content was empty, using default: %s", userMsg.Content)
+			}
 			if isLastMessage {
 				currentUserMsg = &userMsg
 				currentToolResults = toolResults
 			} else {
-				// CRITICAL: Kiro API requires content to be non-empty for history messages too
-				if strings.TrimSpace(userMsg.Content) == "" {
-					if len(toolResults) > 0 {
-						userMsg.Content = "Tool results provided."
-					} else {
-						userMsg.Content = "Continue"
-					}
-				}
 				// For history messages, embed tool results in context
 				if len(toolResults) > 0 {
 					userMsg.UserInputMessageContext = &KiroUserInputMessageContext{

internal/translator/kiro/common/constants.go

@@ -31,11 +31,23 @@ const (
 
 	// DefaultAssistantContentWithTools is the fallback content for assistant messages
 	// that have tool_use but no text content. Kiro API requires non-empty content.
-	DefaultAssistantContentWithTools = "I'll help you with that."
+	// IMPORTANT: Use a minimal neutral string that the model won't mimic in responses.
+	// Previously "I'll help you with that." which caused the model to parrot it back.
+	DefaultAssistantContentWithTools = "."
 
 	// DefaultAssistantContent is the fallback content for assistant messages
 	// that have no content at all. Kiro API requires non-empty content.
-	DefaultAssistantContent = "I understand."
+	// IMPORTANT: Use a minimal neutral string that the model won't mimic in responses.
+	// Previously "I understand." which could leak into model behavior.
+	DefaultAssistantContent = "."
+
+	// DefaultUserContentWithToolResults is the fallback content for user messages
+	// that have only tool_result (no text). Kiro API requires non-empty content.
+	DefaultUserContentWithToolResults = "Tool results provided."
+
+	// DefaultUserContent is the fallback content for user messages
+	// that have no content at all. Kiro API requires non-empty content.
+	DefaultUserContent = "Continue"
 
 	// KiroAgenticSystemPrompt is injected only for -agentic models to prevent timeouts on large writes.
 	// AWS Kiro API has a 2-3 minute timeout for large file write operations.

sdk/cliproxy/auth/oauth_model_alias.go

@@ -221,7 +221,7 @@ func modelAliasChannel(auth *Auth) string {
 // and auth kind. Returns empty string if the provider/authKind combination doesn't support
 // OAuth model alias (e.g., API key authentication).
 //
-// Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot.
+// Supported channels: gemini-cli, vertex, aistudio, antigravity, claude, codex, qwen, iflow, kiro, github-copilot, kimi.
 func OAuthModelAliasChannel(provider, authKind string) string {
 	provider = strings.ToLower(strings.TrimSpace(provider))
 	authKind = strings.ToLower(strings.TrimSpace(authKind))
@@ -245,7 +245,7 @@ func OAuthModelAliasChannel(provider, authKind string) string {
 			return ""
 		}
 		return "codex"
-	case "gemini-cli", "aistudio", "antigravity", "qwen", "iflow", "kiro", "github-copilot":
+	case "gemini-cli", "aistudio", "antigravity", "qwen", "iflow", "kiro", "github-copilot", "kimi":
 		return provider
 	default:
 		return ""

sdk/cliproxy/auth/oauth_model_alias_test.go

@@ -79,6 +79,15 @@ func TestResolveOAuthUpstreamModel_SuffixPreservation(t *testing.T) {
 			input:   "gemini-2.5-pro(none)",
 			want:    "gemini-2.5-pro-exp-03-25(none)",
 		},
+		{
+			name: "kimi suffix preserved",
+			aliases: map[string][]internalconfig.OAuthModelAlias{
+				"kimi": {{Name: "kimi-k2.5", Alias: "k2.5"}},
+			},
+			channel: "kimi",
+			input:   "k2.5(high)",
+			want:    "kimi-k2.5(high)",
+		},
 		{
 			name: "case insensitive alias lookup with suffix",
 			aliases: map[string][]internalconfig.OAuthModelAlias{
@@ -161,6 +170,8 @@ func createAuthForChannel(channel string) *Auth {
 		return &Auth{Provider: "qwen"}
 	case "iflow":
 		return &Auth{Provider: "iflow"}
+	case "kimi":
+		return &Auth{Provider: "kimi"}
 	case "kiro":
 		return &Auth{Provider: "kiro"}
 	default:
@@ -168,6 +179,14 @@ func createAuthForChannel(channel string) *Auth {
 	}
 }
 
+func TestOAuthModelAliasChannel_Kimi(t *testing.T) {
+	t.Parallel()
+
+	if got := OAuthModelAliasChannel("kimi", "oauth"); got != "kimi" {
+		t.Fatalf("OAuthModelAliasChannel() = %q, want %q", got, "kimi")
+	}
+}
+
 func TestApplyOAuthModelAlias_SuffixPreservation(t *testing.T) {
 	t.Parallel()
 

sdk/cliproxy/service_oauth_excluded_models_test.go

@@ -0,0 +1,45 @@
+package cliproxy
+
+import (
+	"testing"
+
+	"github.com/router-for-me/CLIProxyAPI/v6/sdk/config"
+)
+
+func TestOAuthExcludedModels_KimiOAuth(t *testing.T) {
+	t.Parallel()
+
+	svc := &Service{
+		cfg: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"kimi": {"kimi-k2-thinking", "kimi-k2.5"},
+			},
+		},
+	}
+
+	got := svc.oauthExcludedModels("kimi", "oauth")
+	if len(got) != 2 {
+		t.Fatalf("expected 2 excluded models, got %d", len(got))
+	}
+	if got[0] != "kimi-k2-thinking" || got[1] != "kimi-k2.5" {
+		t.Fatalf("unexpected excluded models: %#v", got)
+	}
+}
+
+func TestOAuthExcludedModels_KimiAPIKeyReturnsNil(t *testing.T) {
+	t.Parallel()
+
+	svc := &Service{
+		cfg: &config.Config{
+			OAuthExcludedModels: map[string][]string{
+				"kimi": {"kimi-k2-thinking"},
+			},
+		},
+	}
+
+	got := svc.oauthExcludedModels("kimi", "apikey")
+	if got != nil {
+		t.Fatalf("expected nil for apikey auth kind, got %#v", got)
+	}
+}
+

sdk/cliproxy/service_oauth_model_alias_test.go

@@ -90,3 +90,27 @@ func TestApplyOAuthModelAlias_ForkAddsMultipleAliases(t *testing.T) {
 		t.Fatalf("expected forked model name %q, got %q", "models/g5-2", out[2].Name)
 	}
 }
+
+func TestApplyOAuthModelAlias_KimiRename(t *testing.T) {
+	cfg := &config.Config{
+		OAuthModelAlias: map[string][]config.OAuthModelAlias{
+			"kimi": {
+				{Name: "kimi-k2.5", Alias: "k2.5"},
+			},
+		},
+	}
+	models := []*ModelInfo{
+		{ID: "kimi-k2.5", Name: "models/kimi-k2.5"},
+	}
+
+	out := applyOAuthModelAlias(cfg, "kimi", "oauth", models)
+	if len(out) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(out))
+	}
+	if out[0].ID != "k2.5" {
+		t.Fatalf("expected model id %q, got %q", "k2.5", out[0].ID)
+	}
+	if out[0].Name != "models/k2.5" {
+		t.Fatalf("expected model name %q, got %q", "models/k2.5", out[0].Name)
+	}
+}