Merge pull request #3832 from haowang02/feat/disable-image-generation-passthrough

feat(config): add "passthrough" mode for disable-image-generation

config.example.yaml

@@ -116,9 +116,10 @@ max-retry-interval: 30
 # When true, disable auth/model cooldown scheduling globally (prevents blackout windows after failure states).
 disable-cooling: false
 
-# disable-image-generation supports: false (default), true, or "chat".
+# disable-image-generation supports: false (default), true, "chat", or "passthrough".
 # - true: disable image_generation everywhere (also returns 404 for /v1/images/generations and /v1/images/edits).
 # - "chat": disable image_generation injection on non-images endpoints, but keep /v1/images/generations and /v1/images/edits enabled.
+# - "passthrough": never inject or strip image_generation on non-images endpoints (forward the client payload unchanged); behaves like "chat" on /v1/images/* endpoints.
 disable-image-generation: false
 
 # Base model used when proxying gpt-image-2 via the hosted image_generation tool (Responses API).

internal/config/disable_image_generation_mode.go

@@ -9,18 +9,21 @@ import (
 	"gopkg.in/yaml.v3"
 )
 
-// DisableImageGenerationMode is a tri-state config value for disable-image-generation.
+// DisableImageGenerationMode is a four-state config value for disable-image-generation.
 //
 // It supports:
 //   - false: enabled
 //   - true: disabled everywhere (including /v1/images/* endpoints)
 //   - "chat": disabled for all non-images endpoints, but enabled for /v1/images/generations and /v1/images/edits
+//   - "passthrough": never inject and never strip image_generation on non-images endpoints
+//     (the client payload is forwarded unchanged); on /v1/images/* endpoints behave like "chat"
 type DisableImageGenerationMode int
 
 const (
 	DisableImageGenerationOff DisableImageGenerationMode = iota
 	DisableImageGenerationAll
 	DisableImageGenerationChat
+	DisableImageGenerationPassthrough
 )
 
 func (m DisableImageGenerationMode) String() string {
@@ -31,6 +34,8 @@ func (m DisableImageGenerationMode) String() string {
 		return "true"
 	case DisableImageGenerationChat:
 		return "chat"
+	case DisableImageGenerationPassthrough:
+		return "passthrough"
 	default:
 		return "false"
 	}
@@ -42,6 +47,8 @@ func (m DisableImageGenerationMode) MarshalYAML() (any, error) {
 		return true, nil
 	case DisableImageGenerationChat:
 		return "chat", nil
+	case DisableImageGenerationPassthrough:
+		return "passthrough", nil
 	default:
 		return false, nil
 	}
@@ -62,6 +69,8 @@ func (m DisableImageGenerationMode) MarshalJSON() ([]byte, error) {
 		return []byte("true"), nil
 	case DisableImageGenerationChat:
 		return json.Marshal("chat")
+	case DisableImageGenerationPassthrough:
+		return json.Marshal("passthrough")
 	default:
 		return []byte("false"), nil
 	}
@@ -130,7 +139,9 @@ func parseDisableImageGenerationString(s string) (DisableImageGenerationMode, er
 		return DisableImageGenerationAll, nil
 	case "chat":
 		return DisableImageGenerationChat, nil
+	case "passthrough":
+		return DisableImageGenerationPassthrough, nil
 	default:
-		return DisableImageGenerationOff, fmt.Errorf("invalid disable-image-generation value %q (allowed: true, false, chat)", s)
+		return DisableImageGenerationOff, fmt.Errorf("invalid disable-image-generation value %q (allowed: true, false, chat, passthrough)", s)
 	}
 }

internal/config/disable_image_generation_mode_test.go

@@ -41,6 +41,16 @@ func TestDisableImageGenerationMode_UnmarshalYAML(t *testing.T) {
 			t.Fatalf("chat => %v, want %v", w.V, DisableImageGenerationChat)
 		}
 	}
+
+	{
+		var w wrapper
+		if err := yaml.Unmarshal([]byte("disable-image-generation: passthrough\n"), &w); err != nil {
+			t.Fatalf("unmarshal passthrough: %v", err)
+		}
+		if w.V != DisableImageGenerationPassthrough {
+			t.Fatalf("passthrough => %v, want %v", w.V, DisableImageGenerationPassthrough)
+		}
+	}
 }
 
 func TestDisableImageGenerationMode_UnmarshalJSON(t *testing.T) {
@@ -73,4 +83,14 @@ func TestDisableImageGenerationMode_UnmarshalJSON(t *testing.T) {
 			t.Fatalf("chat => %v, want %v", v, DisableImageGenerationChat)
 		}
 	}
+
+	{
+		var v DisableImageGenerationMode
+		if err := json.Unmarshal([]byte(`"passthrough"`), &v); err != nil {
+			t.Fatalf("unmarshal passthrough: %v", err)
+		}
+		if v != DisableImageGenerationPassthrough {
+			t.Fatalf("passthrough => %v, want %v", v, DisableImageGenerationPassthrough)
+		}
+	}
 }

internal/config/sdk_config.go

@@ -17,6 +17,8 @@ type SDKConfig struct {
 	//     and returns 404 for /v1/images/generations and /v1/images/edits.
 	//   - "chat": disable image_generation injection for all non-images endpoints (e.g. /v1/responses, /v1/chat/completions),
 	//     while keeping /v1/images/generations and /v1/images/edits enabled and preserving image_generation there.
+	//   - "passthrough": do not modify the tool list on non-images endpoints — keep image_generation if the client
+	//     sent it and do not inject it otherwise; on /v1/images/generations and /v1/images/edits behave like "chat".
 	DisableImageGeneration DisableImageGenerationMode `yaml:"disable-image-generation" json:"disable-image-generation"`
 
 	// GPTImage2BaseModel sets the base (mainline) model used when proxying GPT Image 2

internal/runtime/executor/helps/payload_helpers.go

@@ -33,11 +33,9 @@ func ApplyPayloadConfigWithRequest(cfg *config.Config, model, protocol, fromProt
 
 	// Apply disable-image-generation filtering before payload rules so config payload
 	// overrides can explicitly re-enable image_generation when desired.
-	if cfg.DisableImageGeneration != config.DisableImageGenerationOff {
-		if cfg.DisableImageGeneration != config.DisableImageGenerationChat || !isImagesEndpointRequestPath(requestPath) {
-			out = removeToolTypeFromPayloadWithRoot(out, root, "image_generation")
-			out = removeToolChoiceFromPayloadWithRoot(out, root, "image_generation")
-		}
+	if shouldStripImageGeneration(cfg.DisableImageGeneration, requestPath) {
+		out = removeToolTypeFromPayloadWithRoot(out, root, "image_generation")
+		out = removeToolChoiceFromPayloadWithRoot(out, root, "image_generation")
 	}
 
 	rules := cfg.Payload
@@ -199,6 +197,23 @@ func isImagesEndpointRequestPath(path string) bool {
 	return false
 }
 
+// shouldStripImageGeneration reports whether the built-in image_generation tool must be
+// removed from the outbound payload for the given mode and request path.
+//   - All: strip on every endpoint.
+//   - Chat: strip only on non-images endpoints; keep it on /v1/images/* endpoints.
+//   - Off / Passthrough: never strip. Off injects the tool elsewhere; Passthrough forwards
+//     the client payload untouched.
+func shouldStripImageGeneration(mode config.DisableImageGenerationMode, requestPath string) bool {
+	switch mode {
+	case config.DisableImageGenerationAll:
+		return true
+	case config.DisableImageGenerationChat:
+		return !isImagesEndpointRequestPath(requestPath)
+	default:
+		return false
+	}
+}
+
 func payloadModelRulesMatch(rules []config.PayloadModelRule, protocol string, fromProtocol string, headers http.Header, payload []byte, root string, models []string) bool {
 	if len(rules) == 0 || len(models) == 0 {
 		return false

internal/runtime/executor/helps/payload_helpers_disable_image_generation_test.go

@@ -97,6 +97,33 @@ func TestApplyPayloadConfigWithRoot_DisableImageGenerationChat_KeepsImageGenerat
 	}
 }
 
+func TestApplyPayloadConfigWithRoot_DisableImageGenerationPassthrough_KeepsPayloadUnchanged(t *testing.T) {
+	cfg := &config.Config{
+		SDKConfig: config.SDKConfig{DisableImageGeneration: config.DisableImageGenerationPassthrough},
+	}
+	payload := []byte(`{"tools":[{"type":"image_generation"},{"type":"function","name":"f1"}],"tool_choice":{"type":"image_generation"}}`)
+
+	// Passthrough must never inject or strip image_generation. The payload is forwarded as-is on
+	// non-images endpoints, and /v1/images/* endpoints behave like "chat" (also no removal).
+	for _, requestPath := range []string{"", "/v1/responses", "/v1/images/generations"} {
+		out := ApplyPayloadConfigWithRoot(cfg, "gpt-5.4", "openai-response", "", payload, nil, "", requestPath)
+
+		tools := gjson.GetBytes(out, "tools")
+		if !tools.Exists() || !tools.IsArray() {
+			t.Fatalf("path %q: expected tools array, got %v", requestPath, tools.Type)
+		}
+		if got := len(tools.Array()); got != 2 {
+			t.Fatalf("path %q: expected 2 tools (no removal), got %d", requestPath, got)
+		}
+		if got := tools.Array()[0].Get("type").String(); got != "image_generation" {
+			t.Fatalf("path %q: expected image_generation tool to be kept, got %q", requestPath, got)
+		}
+		if !gjson.GetBytes(out, "tool_choice").Exists() {
+			t.Fatalf("path %q: expected tool_choice to be kept", requestPath)
+		}
+	}
+}
+
 func TestApplyPayloadConfigWithRoot_DisableImageGeneration_PayloadOverrideCanRestoreImageGeneration(t *testing.T) {
 	cfg := &config.Config{
 		SDKConfig: config.SDKConfig{DisableImageGeneration: config.DisableImageGenerationAll},