feat(canvas): 生成合同 - 统一节点参数并补充模型上下文能力

web/src/components/canvas/canvas-config-node-panel.tsx

@@ -11,6 +11,7 @@ import { normalizeVideoDuration, normalizeVideoResolution } from "@/lib/video-ge
 import { defaultModelCapabilityConfig, modelCapabilityConfigFor, normalizeImageValue, normalizeVideoValue, workflowFieldChoiceValues, workflowFieldCurrentValue, workflowFieldKey, workflowFieldNumberBounds, workflowFieldRandomKey, workflowFieldSubmissionValue, workflowFieldValueError, workflowImageCapabilityConfig, workflowOutputSizeValue, workflowParameterFields, workflowVideoCapabilityConfig, workflowVideoFieldsFromJson, type WorkflowVideoFieldLike } from "@/lib/model-capabilities";
 import { defaultImageParamsForModel, modelCompatibilityError, modelRequestOptions, resolveCompatibleModel, resolveModelGenerationDefaults, type ModelRequirements } from "@/lib/model-selection";
 import { resolveCanvasWorkflowProvider } from "@/lib/canvas/canvas-workflow";
+import { canonicalGenerationMetadata } from "@/lib/canvas/generation-contract";
 import type { CanvasAudioSettingKey } from "./canvas-audio-settings-popover";
 import { useActiveTheme } from "@/stores/canvas/use-canvas-theme-store";
 import { workflowProviderPluginEnabled } from "@/lib/plugins/builtin/workflows";
@@ -66,6 +67,7 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, onConfigC
     const runtimeStatuses = usePluginStore((state) => state.runtimeStatuses);
     const theme = canvasThemes[useActiveTheme()];
     const mode = node.metadata?.generationMode === "video" || node.metadata?.generationMode === "audio" ? node.metadata.generationMode : "image";
+    node = { ...node, metadata: canonicalGenerationMetadata(node, mode) };
     const simpleMode = workspaceMode === "simple";
     const resolvedProvider = resolveCanvasWorkflowProvider(node.metadata);
     const workflowProvider = resolvedProvider;
@@ -74,16 +76,16 @@ export function CanvasConfigNodePanel({ node, isRunning, inputSummary, onConfigC
         capability: mode,
         input: inputSummary,
         videoOperation: node.metadata?.videoEditOperation,
-        videoSeconds: mode === "video" ? node.metadata?.seconds || globalConfig.videoSeconds : undefined,
+        videoSeconds: mode === "video" ? node.metadata?.seconds ?? globalConfig.videoSeconds : undefined,
         options: modelRequestOptions({
             ...globalConfig,
             size: node.metadata?.size || globalConfig.size,
             quality: node.metadata?.quality || globalConfig.quality,
             count: String(node.metadata?.count || globalConfig.count),
-            videoSeconds: node.metadata?.seconds || globalConfig.videoSeconds,
+            videoSeconds: node.metadata?.seconds ?? globalConfig.videoSeconds,
             vquality: node.metadata?.vquality || globalConfig.vquality,
-            videoGenerateAudio: node.metadata?.generateAudio || globalConfig.videoGenerateAudio,
-            videoWatermark: node.metadata?.watermark || globalConfig.videoWatermark,
+            videoGenerateAudio: node.metadata?.generateAudio ?? globalConfig.videoGenerateAudio,
+            videoWatermark: node.metadata?.watermark ?? globalConfig.videoWatermark,
             audioVoice: node.metadata?.audioVoice || globalConfig.audioVoice,
             audioFormat: node.metadata?.audioFormat || globalConfig.audioFormat,
             audioSpeed: node.metadata?.audioSpeed || globalConfig.audioSpeed,
@@ -450,6 +452,7 @@ function WorkflowSelectedLabel({ kind, label }: { kind: "app" | "workflow"; labe
 }
 
 function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: CanvasGenerationMode, requirements: ModelRequirements): AiConfig {
+    node = { ...node, metadata: canonicalGenerationMetadata(node, mode) };
     const workflowProvider = mode === "text" ? "model" : resolveCanvasWorkflowProvider(node.metadata);
     if (workflowProvider === "model") return buildModelNodeConfig(globalConfig, node, mode, requirements);
     const defaultModel = mode === "image" ? globalConfig.imageModel : mode === "video" ? globalConfig.videoModel : mode === "audio" ? globalConfig.audioModel : globalConfig.textModel;
@@ -476,7 +479,7 @@ function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: Can
     const workflowParameter = (source: string) => workflowParameters[`source:${source}`] === undefined ? "" : String(workflowParameters[`source:${source}`]);
     const normalizedImage = imageProfile ? normalizeImageValue(imageProfile, { size: workflowOutputSize || node.metadata?.size || globalConfig.size || defaultConfig.size, quality: node.metadata?.quality || workflowParameter("quality") || globalConfig.quality || defaultConfig.quality, transparentBackground: node.metadata?.transparentBackground || globalConfig.transparentBackground, count: String(node.metadata?.count || globalConfig.canvasImageCount || globalConfig.count || defaultConfig.count) }) : undefined;
     const videoProfile = mode === "video" ? capabilityProfile.video! : undefined;
-    const rawVideoSettings = { seconds: node.metadata?.seconds || workflowParameter("videoSeconds") || globalConfig.videoSeconds || defaultConfig.videoSeconds, ratio: workflowOutputSize || node.metadata?.size || globalConfig.size || defaultConfig.size, resolution: node.metadata?.vquality || workflowParameter("vquality") || globalConfig.vquality || defaultConfig.vquality };
+    const rawVideoSettings = { seconds: node.metadata?.seconds ?? workflowParameter("videoSeconds") ?? globalConfig.videoSeconds ?? defaultConfig.videoSeconds, ratio: workflowOutputSize || node.metadata?.size || globalConfig.size || defaultConfig.size, resolution: node.metadata?.vquality || workflowParameter("vquality") || globalConfig.vquality || defaultConfig.vquality };
     const normalizedVideo = videoProfile
         ? { seconds: String(rawVideoSettings.seconds), ratio: rawVideoSettings.ratio, resolution: rawVideoSettings.resolution }
         : undefined;
@@ -489,10 +492,10 @@ function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: Can
         quality: workflowParameter("quality") || normalizedImage?.quality || node.metadata?.quality || globalConfig.quality || defaultConfig.quality,
         size: normalizedImage?.size || normalizedVideo?.ratio || node.metadata?.size || globalConfig.size || defaultConfig.size,
         transparentBackground: normalizedImage?.transparentBackground || ((node.metadata?.transparentBackground || globalConfig.transparentBackground) === "true" ? "true" : "false"),
-        videoSeconds: normalizedVideo?.seconds || normalizeVideoDuration(node.metadata?.seconds || globalConfig.videoSeconds || defaultConfig.videoSeconds),
+        videoSeconds: normalizedVideo?.seconds ?? normalizeVideoDuration(node.metadata?.seconds ?? globalConfig.videoSeconds ?? defaultConfig.videoSeconds),
         vquality: String(node.metadata?.vquality || globalConfig.vquality || defaultConfig.vquality),
-        videoGenerateAudio: videoProfile?.generateAudio.supported ? node.metadata?.generateAudio || globalConfig.videoGenerateAudio || String(videoProfile.generateAudio.default) : "false",
-        videoWatermark: videoProfile?.watermark.supported ? node.metadata?.watermark || globalConfig.videoWatermark || String(videoProfile.watermark.default) : "false",
+        videoGenerateAudio: videoProfile?.generateAudio.supported ? node.metadata?.generateAudio ?? globalConfig.videoGenerateAudio ?? String(videoProfile.generateAudio.default) : "false",
+        videoWatermark: videoProfile?.watermark.supported ? node.metadata?.watermark ?? globalConfig.videoWatermark ?? String(videoProfile.watermark.default) : "false",
         audioVoice: node.metadata?.audioVoice || globalConfig.audioVoice || defaultConfig.audioVoice,
         audioFormat: node.metadata?.audioFormat || globalConfig.audioFormat || defaultConfig.audioFormat,
         audioSpeed: node.metadata?.audioSpeed || globalConfig.audioSpeed || defaultConfig.audioSpeed,
@@ -502,6 +505,7 @@ function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: Can
 }
 
 function buildModelNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: CanvasGenerationMode, requirements: ModelRequirements): AiConfig {
+    node = { ...node, metadata: canonicalGenerationMetadata(node, mode) };
     const defaultModel = mode === "image" ? globalConfig.imageModel : mode === "video" ? globalConfig.videoModel : mode === "audio" ? globalConfig.audioModel : globalConfig.textModel;
     const fallbackModel = mode === "image" ? defaultConfig.imageModel : mode === "video" ? defaultConfig.videoModel : mode === "audio" ? defaultConfig.audioModel : defaultConfig.textModel;
     const storedModel = node.metadata?.model;
@@ -544,10 +548,10 @@ function buildModelNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode
         quality: generationDefaults.quality || globalConfig.quality || defaultConfig.quality,
         size: generationDefaults.size ?? globalConfig.size ?? defaultConfig.size,
         transparentBackground: generationDefaults.transparentBackground || "false",
-        videoSeconds: generationDefaults.videoSeconds || normalizeVideoDuration(globalConfig.videoSeconds || defaultConfig.videoSeconds),
+        videoSeconds: generationDefaults.videoSeconds ?? normalizeVideoDuration(globalConfig.videoSeconds || defaultConfig.videoSeconds),
         vquality: generationDefaults.vquality ?? normalizeVideoResolution(globalConfig.vquality || defaultConfig.vquality),
-        videoGenerateAudio: videoProfile?.generateAudio.supported ? generationDefaults.videoGenerateAudio || String(videoProfile.generateAudio.default) : "false",
-        videoWatermark: videoProfile?.watermark.supported ? generationDefaults.videoWatermark || String(videoProfile.watermark.default) : "false",
+        videoGenerateAudio: videoProfile?.generateAudio.supported ? generationDefaults.videoGenerateAudio ?? String(videoProfile.generateAudio.default) : "false",
+        videoWatermark: videoProfile?.watermark.supported ? generationDefaults.videoWatermark ?? String(videoProfile.watermark.default) : "false",
         audioVoice: node.metadata?.audioVoice || globalConfig.audioVoice || defaultConfig.audioVoice,
         audioFormat: node.metadata?.audioFormat || globalConfig.audioFormat || defaultConfig.audioFormat,
         audioSpeed: node.metadata?.audioSpeed || globalConfig.audioSpeed || defaultConfig.audioSpeed,

web/src/components/canvas/canvas-node-prompt-panel.tsx

@@ -6,6 +6,7 @@ import { ArrowLeftRight, ArrowUp, AtSign, Boxes, Camera, ChevronDown, FileText,
 import { ModelPicker } from "@/components/model-picker";
 import { defaultConfig, modelOptionName, resolveModelChannel, useEffectiveConfig, type AiConfig } from "@/stores/use-config-store";
 import { resolveCanvasGenerationModel } from "@/lib/canvas/canvas-project-generation";
+import { canonicalGenerationMetadata } from "@/lib/canvas/generation-contract";
 import { clampPromptEditorModalSize, PROMPT_EDITOR_VIEWPORT_MARGIN } from "@/lib/canvas/canvas-prompt-editor-size";
 import { CreditSymbol, requestCreditCost } from "@/constant/credits";
 import { canvasThemes } from "@/lib/canvas-theme";
@@ -78,6 +79,7 @@ export function CanvasNodePromptPanel({ projectId, node, isRunning, onPromptChan
     const promptOptimizerEnabled = usePluginStore((state) => state.pluginStates[PROMPT_OPTIMIZER_PLUGIN_ID]?.effectiveEnabled ?? Boolean(state.installations.find((item) => item.manifest.id === PROMPT_OPTIMIZER_PLUGIN_ID)?.enabled));
     const simpleMode = workspaceMode === "simple";
     const mode = defaultMode(node.type);
+    node = { ...node, metadata: canonicalGenerationMetadata(node, mode) };
     const hasTextContent = node.type === CanvasNodeType.Text && Boolean(node.metadata?.content?.trim());
     const hasImageContent = node.type === CanvasNodeType.Image && Boolean(node.metadata?.content);
     const savedPrompt = node.metadata?.composerContent ?? node.metadata?.prompt ?? "";
@@ -106,17 +108,17 @@ export function CanvasNodePromptPanel({ projectId, node, isRunning, onPromptChan
             characterCount: activeReferences.filter((item) => item.kind === "character").length,
         },
         videoOperation: node.metadata?.videoEditOperation,
-        videoSeconds: mode === "video" ? node.metadata?.seconds || globalConfig.videoSeconds : undefined,
+        videoSeconds: mode === "video" ? node.metadata?.seconds ?? globalConfig.videoSeconds : undefined,
         options: modelRequestOptions({
             ...globalConfig,
             size: node.metadata?.size || globalConfig.size,
             quality: node.metadata?.quality || globalConfig.quality,
             count: String(node.metadata?.count || globalConfig.count),
             transparentBackground: node.metadata?.transparentBackground || globalConfig.transparentBackground,
-            videoSeconds: node.metadata?.seconds || globalConfig.videoSeconds,
+            videoSeconds: node.metadata?.seconds ?? globalConfig.videoSeconds,
             vquality: node.metadata?.vquality || globalConfig.vquality,
-            videoGenerateAudio: node.metadata?.generateAudio || globalConfig.videoGenerateAudio,
-            videoWatermark: node.metadata?.watermark || globalConfig.videoWatermark,
+            videoGenerateAudio: node.metadata?.generateAudio ?? globalConfig.videoGenerateAudio,
+            videoWatermark: node.metadata?.watermark ?? globalConfig.videoWatermark,
             audioVoice: node.metadata?.audioVoice || globalConfig.audioVoice,
             audioFormat: node.metadata?.audioFormat || globalConfig.audioFormat,
             audioSpeed: node.metadata?.audioSpeed || globalConfig.audioSpeed,
@@ -1034,6 +1036,7 @@ function defaultMode(type: CanvasNodeData["type"]): CanvasNodeGenerationMode {
 }
 
 export function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mode: CanvasNodeGenerationMode, requirements: ModelRequirements): AiConfig {
+    node = { ...node, metadata: canonicalGenerationMetadata(node, mode) };
     const defaultModel = mode === "image" ? globalConfig.imageModel : mode === "video" ? globalConfig.videoModel : mode === "audio" ? globalConfig.audioModel : globalConfig.textModel;
     const fallbackModel = mode === "image" ? defaultConfig.imageModel : mode === "video" ? defaultConfig.videoModel : mode === "audio" ? defaultConfig.audioModel : defaultConfig.textModel;
     const preferredModel = resolveCanvasGenerationModel(globalConfig, node.metadata?.model, mode) || resolveCanvasGenerationModel(globalConfig, defaultModel, mode) || fallbackModel;
@@ -1074,7 +1077,7 @@ export function buildNodeConfig(globalConfig: AiConfig, node: CanvasNodeData, mo
         quality: defaults.quality ?? globalConfig.quality ?? defaultConfig.quality,
         size: defaults.size ?? globalConfig.size ?? defaultConfig.size,
         transparentBackground: defaults.transparentBackground ?? "false",
-        videoSeconds: defaults.videoSeconds || normalizeVideoDuration(globalConfig.videoSeconds || defaultConfig.videoSeconds),
+        videoSeconds: defaults.videoSeconds ?? normalizeVideoDuration(globalConfig.videoSeconds ?? defaultConfig.videoSeconds),
         vquality: defaults.vquality ?? normalizeVideoResolution(globalConfig.vquality || defaultConfig.vquality),
         videoGenerateAudio: defaults.videoGenerateAudio ?? globalConfig.videoGenerateAudio ?? defaultConfig.videoGenerateAudio,
         videoWatermark: defaults.videoWatermark ?? globalConfig.videoWatermark ?? defaultConfig.videoWatermark,

web/src/components/model-capability-editor.tsx

@@ -350,7 +350,7 @@ export function ModelCapabilityEditor({ value, onChange, protocol, capability =
 }
 
 function TextCapabilityEditor({ value, onChange, protocol, disabled, section }: Pick<Props, "value" | "onChange" | "protocol" | "disabled" | "section">) {
-    const profile = value?.text || defaultModelCapabilityConfig(protocol).text!;
+    const profile = normalizeModelCapabilityConfig(value || defaultModelCapabilityConfig(protocol)).text!;
     const update = (patch: Partial<TextCapabilityConfig>) => {
         onChange?.({ version: 1, text: { ...profile, ...patch } });
     };
@@ -369,8 +369,17 @@ function TextCapabilityEditor({ value, onChange, protocol, disabled, section }:
     }
 
     if (section === "references") {
+        const contextWindowFields = (
+            <CapabilityGroup title="上下文能力" description="这是上游文本模型的能力合同，决定 Agent 本轮可保留的输入预算；不会改变运行时 checkpoint 的持久化上限。">
+                <div className="grid gap-3 sm:grid-cols-2">
+                    <NumberField label="上下文窗口 Token" value={profile.contextWindowTokens} min={4_096} max={10_000_000} disabled={Boolean(disabled)} onChange={(next) => update({ contextWindowTokens: next || 4_096 })} />
+                    <NumberField label="最大输出 Token" value={profile.maxOutputTokens} min={256} max={1_000_000} disabled={Boolean(disabled)} onChange={(next) => update({ maxOutputTokens: next || 256 })} />
+                </div>
+            </CapabilityGroup>
+        );
         return (
             <div className="admin-capability-reference-editor">
+                {contextWindowFields}
                 <div className="admin-capability-reference-grid is-three">
                     <ReferenceCard title="图片引用" description="文本模型可接收的图片范围">
                         <NumberField label="最大参考图片数" value={profile.references.maxImages} min={0} max={100} disabled={Boolean(disabled)} onChange={(next) => updateReferences({ maxImages: next || 0 })} />
@@ -394,6 +403,12 @@ function TextCapabilityEditor({ value, onChange, protocol, disabled, section }:
                 <div className="text-sm font-medium">文本理解能力</div>
                 <div className="mt-0.5 text-[var(--fs-tiny)] text-foreground/48">默认不假设支持图片或视频，只有明确配置后相关请求才会进入该模型。</div>
             </div>
+            <CapabilityGroup title="上下文能力" description="这是上游文本模型的能力合同，决定 Agent 本轮可保留的输入预算；不会改变运行时 checkpoint 的持久化上限。">
+                <div className="grid gap-3 sm:grid-cols-2">
+                    <NumberField label="上下文窗口 Token" value={profile.contextWindowTokens} min={4_096} max={10_000_000} disabled={Boolean(disabled)} onChange={(next) => update({ contextWindowTokens: next || 4_096 })} />
+                    <NumberField label="最大输出 Token" value={profile.maxOutputTokens} min={256} max={1_000_000} disabled={Boolean(disabled)} onChange={(next) => update({ maxOutputTokens: next || 256 })} />
+                </div>
+            </CapabilityGroup>
             <CapabilityGroup title="图片" description="文本模型可接收的图片参考范围">
                 <div className="grid gap-3 sm:grid-cols-2">
                     <NumberField label="最大参考图片数" value={profile.references.maxImages} min={0} max={100} disabled={Boolean(disabled)} onChange={(next) => updateReferences({ maxImages: next || 0 })} />