Remove vision proxy

package.json

@@ -1,6 +1,6 @@
 {
   "name": "coderelay",
-  "version": "0.1.13",
+  "version": "0.2.0",
   "private": true,
   "type": "module",
   "scripts": {

sidecars/coderelay-proxy/UPSTREAM.md

@@ -4,8 +4,8 @@
 original MIT license and copyright notice are retained in `LICENSE`.
 
 The outer `coderelay-proxy` package contains the CodeBuddy CN relay, request
-policy, account selection, model synchronization, vision routing and the
-newline-delimited JSON lifecycle event protocol consumed by the Tauri host.
+policy, account selection, model synchronization and the newline-delimited JSON
+lifecycle event protocol consumed by the Tauri host.
 
 The sidecar accepts:
 

sidecars/coderelay-proxy/codebuddy.go

@@ -34,19 +34,6 @@ func equalStringSlices(a, b []string) bool {
 	return true
 }
 
-// visionProxyEnabled reports whether the vision-proxy layer is active (mode is
-// "routing" or "preprocess"). Used by /v1/models to report `input_modalities:
-// ["text","image"]` for non-vision models that the proxy will transparently
-// handle — otherwise clients (e.g. Cursor) filter image inputs client-side and
-// the image never reaches the relay.
-func (m *manifest) visionProxyEnabled() bool {
-	if m == nil {
-		return false
-	}
-	mode := strings.ToLower(strings.TrimSpace(m.VisionMode))
-	return mode == "routing" || mode == "preprocess" || mode == "agentic"
-}
-
 // modelIDs returns a snapshot of the current model ID list. It is safe for
 // concurrent use with setModelIDs.
 func (m *manifest) modelIDs() []string {

sidecars/coderelay-proxy/codebuddy_model_cache.go

@@ -19,7 +19,7 @@ const codebuddyModelCacheFilename = "codebuddy_models_cache.json"
 // codebuddyModelCache is the persisted form of the CodeBuddy model catalog.
 // Models are stored in full (not just IDs) so capability fields such as
 // SupportsImages / ContextLength / MaxCompletionTokens survive a reload, which
-// vision-proxy routing and max_tokens clamping depend on.
+// max_tokens clamping and image-capability checks depend on.
 type codebuddyModelCache struct {
 	Version  int                           `json:"version"`
 	SyncedAt string                        `json:"syncedAt,omitempty"`

sidecars/coderelay-proxy/manifest_policy.go

@@ -78,10 +78,6 @@
 	streamIdleTimeout      = 60 * time.Second
 	imageStreamOpenTimeout = 10 * time.Second
 	imageStreamIdleTimeout = 60 * time.Second
-	// visionAgenticStreamIdleTimeout covers the codebuddy vision sub-agent
-	// loop: up to several rounds of text-model + vision-model upstream calls
-	// (each 10-30s), so the relay idle watchdog does not trip mid-loop.
-	visionAgenticStreamIdleTimeout = 300 * time.Second
 )
 
 type accountModelRule struct {
@@ -102,8 +98,6 @@
 	ImmediateSSEResponse       bool                `json:"immediateSseResponse"`
 	MaxConcurrentImageRequests int                 `json:"maxConcurrentImageRequests"`
 	DebugLogs                  *bool               `json:"debugLogs,omitempty"`
-	VisionMode                 string              `json:"visionMode"`
-	VisionModel                string              `json:"visionModel"`
 
 	apiKeyByValue     map[string]*apiKeySpec
 	accountByID       map[string]*accountSpec
@@ -397,7 +391,6 @@
 	Path                    string                     `json:"path,omitempty"`
 	RequestKind             string                     `json:"requestKind,omitempty"`
 	Model                   string                     `json:"model,omitempty"`
-	VisionSubagent          bool                       `json:"visionSubagent,omitempty"`
 	APIKeyID                string                     `json:"apiKeyId,omitempty"`
 	APIKeyLabel             string                     `json:"apiKeyLabel,omitempty"`
 	Transport               string                     `json:"transport,omitempty"`
@@ -749,7 +742,7 @@
 	}
 	e.mu.Lock()
 	defer e.mu.Unlock()
 	fmt.Println(string(data))
 }
 
 func (e *eventEmitter) emitStartupStage(stage string) {
@@ -979,7 +972,7 @@
 			if isCodexClientModelsRequest(c.Request) {
 				c.JSON(http.StatusOK, buildCodexClientModelsResponse(models, spec, contextWindowsForAPIKey(p.manifest, spec)))
 			} else {
-				c.JSON(http.StatusOK, buildModelsResponse(models, p.manifest.visionProxyEnabled()))
+				c.JSON(http.StatusOK, buildModelsResponse(models))
 			}
 			c.Abort()
 			return
@@ -1138,13 +1131,12 @@
 		RequestID:     requestID,
 		Method:        c.Request.Method,
 		Path:          requestPath(c.Request),
-		RequestKind:    requestKind,
-		Model:          model,
-		VisionSubagent: internallogging.GetVisionSubagent(c.Request.Context()),
-		APIKeyID:       stringFromAPIKey(spec, "id"),
-		APIKeyLabel:    stringFromAPIKey(spec, "label"),
-		Transport:      diagnosticTransport(c.Request),
-		Status:         status,
+		RequestKind:   requestKind,
+		Model:         model,
+		APIKeyID:      stringFromAPIKey(spec, "id"),
+		APIKeyLabel:   stringFromAPIKey(spec, "label"),
+		Transport:     diagnosticTransport(c.Request),
+		Status:        status,
 		LatencyMS:     latencyMS,
 		CompletedAtMS: completedAtMS,
 		Aborted:       c.IsAborted(),
@@ -1248,15 +1240,16 @@
 // carries `input_modalities` so clients (e.g. Cursor) can detect vision-capable
 // models instead of defaulting to text-only and filtering images client-side.
 //
-// A model reports image capability when either:
-//  1. the backend natively supports images for it (registry.CodebuddyModelSupportsImages), or
-//  2. the vision-proxy layer is active and will transparently describe/handle
-//     images for it (preprocess/routing via hy3-preview).
-func buildModelsResponse(models []string, visionProxyEnabled bool) gin.H {
+// A model reports image capability when the backend natively supports images
+// for it (registry.CodebuddyModelSupportsImages, backed by the online catalog
+// plus the measured capability overrides in the registry). Models without
+// native image support report text-only so clients do not send images the
+// upstream would reject.
+func buildModelsResponse(models []string) gin.H {
 	data := make([]gin.H, 0, len(models))
 	for _, model := range models {
 		modalities := []any{"text"}
-		if internalregistry.CodebuddyModelSupportsImages(model) || visionProxyEnabled {
+		if internalregistry.CodebuddyModelSupportsImages(model) {
 			modalities = []any{"text", "image"}
 		}
 		data = append(data, gin.H{
@@ -1799,41 +1792,6 @@
 	return resolveSupportedModelAlias(m, withoutPrefix)
 }
 
-// requestHasVisionInput reports whether a request body carries image input
-// (OpenAI image_url parts or Responses input_image parts). It is used by the
-// stream watchdog to extend the idle timeout for vision sub-agent loops.
-func requestHasVisionInput(body []byte) bool {
-	if len(body) == 0 || !json.Valid(body) {
-		return false
-	}
-	var payload any
-	if err := json.Unmarshal(body, &payload); err != nil {
-		return false
-	}
-	return valueHasVisionInput(payload)
-}
-
-func valueHasVisionInput(value any) bool {
-	switch typed := value.(type) {
-	case map[string]any:
-		if typ, _ := typed["type"].(string); strings.EqualFold(strings.TrimSpace(typ), "input_image") || strings.EqualFold(strings.TrimSpace(typ), "image_url") {
-			return true
-		}
-		for _, child := range typed {
-			if valueHasVisionInput(child) {
-				return true
-			}
-		}
-	case []any:
-		for _, child := range typed {
-			if valueHasVisionInput(child) {
-				return true
-			}
-		}
-	}
-	return false
-}
-
 func stripModelPrefix(model string, spec *apiKeySpec) string {
 	trimmed := strings.TrimSpace(model)
 	if spec == nil || strings.TrimSpace(spec.ModelPrefix) == "" {

sidecars/coderelay-proxy/manifest_policy_test.go

@@ -255,7 +255,7 @@ func TestCodexClientModelsResponseShape(t *testing.T) {
 }
 
 func TestCodebuddyModelsResponseReportsInputModalities(t *testing.T) {
-	models := []string{"hy3", "hy3-preview", "deepseek-v4-pro", "deepseek-v4-flash", "hunyuan-2.0-instruct"}
+	models := []string{"hy3", "hy3-preview", "deepseek-v4-pro", "deepseek-v4-flash", "hunyuan-2.0-instruct", "glm-5v-turbo"}
 
 	modalitiesFor := func(response gin.H, modelID string) []any {
 		data, ok := response["data"].([]gin.H)
@@ -281,22 +281,20 @@ func TestCodebuddyModelsResponseReportsInputModalities(t *testing.T) {
 		}
 	}
 
-	// 情况1：vision-proxy 启用（默认 preprocess）——所有模型都报 image，
-	// 因为反代能透明处理纯文本模型的图片（描述后回填）。
-	enabled := buildModelsResponse(models, true)
-	for _, m := range models {
-		assertModalities(enabled, m, []any{"text", "image"})
-	}
-
-	// 情况2：vision-proxy 关闭——仅后端原生支持视觉的模型报 image。
-	// hy3/hy3-preview（app.asar supportsImages）报 image；
-	// deepseek（已移出白名单，后端返回拒绝）与 hunyuan（假视觉）报 text。
-	disabled := buildModelsResponse(models, false)
-	assertModalities(disabled, "hy3", []any{"text", "image"})
-	assertModalities(disabled, "hy3-preview", []any{"text", "image"})
-	assertModalities(disabled, "deepseek-v4-pro", []any{"text"})
-	assertModalities(disabled, "deepseek-v4-flash", []any{"text"})
-	assertModalities(disabled, "hunyuan-2.0-instruct", []any{"text"})
+	// 图片能力判定 = 在线清单 supportsImages + registry 里的实测校正表
+	// （排除 glm-5v-turbo，补入 glm-5.1 / deepseek-v3-2-volc）。
+	// 单测环境下 codebuddySynced 为空、回退静态 models.json：
+	// hy3/hy3-preview 在静态目录中含 supportsImages，报 image；
+	// deepseek（静态目录未标记图片能力）与 hunyuan（无图片能力）报 text；
+	// glm-5v-turbo 由校正表强制排除，即使目录误标也报 text。
+	// 注意：运行态同步到在线清单后，deepseek-v4.x 会被标记为图片能力 → 报 image。
+	response := buildModelsResponse(models)
+	assertModalities(response, "hy3", []any{"text", "image"})
+	assertModalities(response, "hy3-preview", []any{"text", "image"})
+	assertModalities(response, "deepseek-v4-pro", []any{"text"})
+	assertModalities(response, "deepseek-v4-flash", []any{"text"})
+	assertModalities(response, "hunyuan-2.0-instruct", []any{"text"})
+	assertModalities(response, "glm-5v-turbo", []any{"text"})
 }
 
 func TestCodexClientModelsResponsePreserves56Template(t *testing.T) {
@@ -446,23 +444,6 @@ func TestCodexClientModelsResponseDoesNotInjectFastMode(t *testing.T) {
 	}
 }
 
-func TestRequestVisionDetectionIgnoresToolSchemaFieldNames(t *testing.T) {
-	body := []byte(`{
-		"model":"deepseek-v4-pro",
-		"tools":[{
-			"type":"function",
-			"name":"inspect_url",
-			"parameters":{
-				"type":"object",
-				"properties":{"image_url":{"type":"string"}}
-			}
-		}]
-	}`)
-	if requestHasVisionInput(body) {
-		t.Fatal("tool schema field names must not be treated as image input")
-	}
-}
-
 func TestCodexClientModelsResponseEnablesWebsocketsWhenConfigured(t *testing.T) {
 	response := buildCodexClientModelsResponse([]string{"gpt-5.6-sol"}, &apiKeySpec{
 		ResponsesWebsockets: true,

sidecars/coderelay-proxy/relay_server.go

@@ -574,7 +574,7 @@ func (s *relayServer) handleModels(c *gin.Context) {
 		c.JSON(http.StatusOK, buildCodexClientModelsResponse(models, spec, contextWindowsForAPIKey(s.manifest, spec)))
 		return
 	}
-	c.JSON(http.StatusOK, buildModelsResponse(models, s.manifest.visionProxyEnabled()))
+	c.JSON(http.StatusOK, buildModelsResponse(models))
 }
 
 func (s *relayServer) handleResponses(c *gin.Context) {

sidecars/coderelay-proxy/stream_protocol.go

@@ -62,15 +62,6 @@ func (s *relayServer) streamTimeoutsForRequest(r *http.Request, body []byte, mod
 		profile.open = durationFromConfigMillis(s.cfg.Streaming.StreamOpenTimeoutMS, profile.open)
 		profile.idle = durationFromConfigMillis(s.cfg.Streaming.StreamIdleTimeoutMS, profile.idle)
 	}
-	// Requests carrying vision input are handled by the codebuddy vision
-	// sub-agent loop, which performs multiple rounds of upstream calls before
-	// emitting content. Give them a much longer idle timeout so the relay
-	// watchdog does not cancel mid-loop.
-	if requestHasVisionInput(body) {
-		if profile.idle < visionAgenticStreamIdleTimeout {
-			profile.idle = visionAgenticStreamIdleTimeout
-		}
-	}
 	if !isImageGenerationRequest(r, body, model) {
 		return profile
 	}
@@ -158,12 +149,6 @@ func relayContext(c *gin.Context) context.Context {
 	if c == nil || c.Request == nil {
 		return context.Background()
 	}
-	// Attach the vision sub-agent holder to the request context so the
-	// downstream executor's SetVisionSubagent and the request-completed
-	// diagnostic's GetVisionSubagent observe the same holder instance. Without
-	// this, both sides would hold distinct (or missing) holders and the
-	// request_completed event's visionSubagent flag would always be false.
-	c.Request = c.Request.WithContext(internallogging.WithVisionSubagentHolder(c.Request.Context()))
 	endpoint := c.Request.Method
 	if c.Request.URL != nil {
 		endpoint += " " + c.Request.URL.Path

sidecars/coderelay-proxy/third_party/CLIProxyAPI/internal/config/sdk_config.go

@@ -4,8 +4,6 @@
 // debug settings, proxy configuration, and API keys.
 package config
 
-import "strings"
-
 // SDKConfig represents the application's configuration, loaded from a YAML file.
 type SDKConfig struct {
 	EnableGeminiCLIEndpoint bool `yaml:"enable-gemini-cli-endpoint,omitempty" json:"enable-gemini-cli-endpoint,omitempty"`
@@ -64,12 +62,6 @@ type SDKConfig struct {
 	// NonStreamKeepAliveInterval controls how often blank lines are emitted for non-streaming responses.
 	// <= 0 disables keep-alives. Value is in seconds.
 	NonStreamKeepAliveInterval int `yaml:"nonstream-keepalive-interval,omitempty" json:"nonstream-keepalive-interval,omitempty"`
-
-	// CodebuddyVision configures the CodeBuddy vision-proxy layer. When a chat
-	// request carries image input for a model that does not natively support
-	// images, the proxy either swaps the model to a vision model (routing) or
-	// converts the images to text descriptions first (preprocess).
-	CodebuddyVision CodebuddyVisionConfig `yaml:"codebuddy-vision" json:"codebuddy-vision"`
 }
 
 // ClaudeCodeConfig configures Claude Code compatibility behavior.
@@ -97,72 +89,4 @@ type StreamingConfig struct {
 	BootstrapRetryMaxDelayMS  int `yaml:"bootstrap-retry-max-delay-ms,omitempty" json:"bootstrap-retry-max-delay-ms,omitempty"`
 }
 
-// CodebuddyVisionConfig controls the CodeBuddy vision-proxy layer.
-//
-// The Tencent CodeBuddy backend accepts image input on a per-model basis. Some
-// text-only models (e.g. hunyuan-2.0-instruct) silently ignore images and reply
-// with "this model does not support image input" instead of an error. When
-// enabled, the proxy detects image input and handles it for non-vision models.
-type CodebuddyVisionConfig struct {
-	// Mode selects the strategy:
-	//   - "off" (default): disabled; images pass through unchanged.
-	//   - "routing": swap the request model to Model for non-vision models.
-	//   - "preprocess": describe images with Model first, then continue with the
-	//     original model.
-	//   - "agentic": inject an inspect_image tool and run a server-side tool-calling
-	//     loop so the text-only model can autonomously query the vision model
-	//     multiple times during reasoning.
-	// Any other value falls back to "off".
-	Mode string `yaml:"mode" json:"mode"`
-
-	// Model is the vision model used as the routing target / preprocess engine.
-	// Default "hy4-preview".
-	Model string `yaml:"model" json:"model"`
-
-	// PreprocessPrompt overrides the user-visible prompt sent to the vision model
-	// in preprocess mode. Empty uses a built-in default.
-	PreprocessPrompt string `yaml:"preprocess-prompt" json:"preprocess-prompt"`
-
-	// MaxToolRounds caps the number of inspect_image tool-call iterations in
-	// agentic mode. Non-positive falls back to a default of 3.
-	MaxToolRounds int `yaml:"max-tool-rounds" json:"max-tool-rounds"`
-}
-
-// VisionMode constants for CodebuddyVisionConfig.Mode.
-const (
-	CodebuddyVisionModeOff         = "off"
-	CodebuddyVisionModeRouting     = "routing"
-	CodebuddyVisionModePreprocess  = "preprocess"
-	CodebuddyVisionModeAgentic     = "agentic"
-)
-
-// NormalizedVisionMode returns the effective mode, mapping unknown values to "off".
-func (c CodebuddyVisionConfig) NormalizedVisionMode() string {
-	switch strings.ToLower(strings.TrimSpace(c.Mode)) {
-	case CodebuddyVisionModeRouting:
-		return CodebuddyVisionModeRouting
-	case CodebuddyVisionModePreprocess:
-		return CodebuddyVisionModePreprocess
-	case CodebuddyVisionModeAgentic:
-		return CodebuddyVisionModeAgentic
-	default:
-		return CodebuddyVisionModeOff
-	}
-}
-
-// MaxVisionToolRounds returns the effective agentic iteration cap (default 3).
-func (c CodebuddyVisionConfig) MaxVisionToolRounds() int {
-	if c.MaxToolRounds > 0 {
-		return c.MaxToolRounds
-	}
-	return 3
-}
 
-// VisionModel returns the configured vision model, defaulting to "hy4-preview".
-func (c CodebuddyVisionConfig) VisionModel() string {
-	model := strings.TrimSpace(c.Model)
-	if model == "" {
-		return "hy4-preview"
-	}
-	return model
-}